1170 lines
106 KiB
HTML
1170 lines
106 KiB
HTML
<!doctype html>
|
||
<html lang="en">
|
||
<head>
|
||
<meta charset="utf-8">
|
||
<title>AritmoLab — Role of AI in the Analysis of Unstructured Clinical Databases</title>
|
||
<link rel="stylesheet" href="vendor/reveal/dist/reveal.css">
|
||
<link rel="stylesheet" href="fonts.css?v=8">
|
||
<link rel="stylesheet" href="deck.css?v=41">
|
||
<link rel="stylesheet" href="slide-lens.css?v=7">
|
||
<link rel="stylesheet" href="ai-callouts.css?v=3">
|
||
<link rel="stylesheet" href="screenshot-tour.css?v=5">
|
||
</head>
|
||
<body>
|
||
<div class="reveal"><div class="slides">
|
||
|
||
|
||
<section class="arit title-slide">
|
||
<header class="head">
|
||
<img src="logo.png" alt="Policlinico San Donato">
|
||
<span class="head-org">AritmoLab · Policlinico San Donato</span>
|
||
</header>
|
||
<nav class="band">
|
||
<span class="band-section">AritmoLab · Clinical Data Platform</span>
|
||
</nav>
|
||
<div class="sbody">
|
||
<h1>Role of AI in the Analysis of Unstructured Clinical Databases</h1>
|
||
<div class="hgap"></div>
|
||
<p class="lead">How we turned a mix of structured and free-text clinical records into a research data platform — the AritmoLab platform experience.</p>
|
||
<div class="title-meta">
|
||
<span class="pill">AritmoLab</span>
|
||
<span class="what">A clinical data platform hosting cardiology data and analysis tools</span>
|
||
</div>
|
||
<div class="speakers">
|
||
<span>Dr. Marco Pancotti - MultiPhysixLab</span>
|
||
<span>Dr. Sara Paratico - I.R.C.C.S. Policlinico San Donato</span>
|
||
</div>
|
||
<p class="event">“Multidimensional Characterization of Cardiac Arrhythmias: Role of Electrocardiology in the Artificial Intelligence Era”</p>
|
||
<p class="event-where">San Donato Milanese, Milan, Italy · 2–3 October 2026</p>
|
||
</div>
|
||
<footer class="foot"><span class="g">Role of AI in the Analysis of Unstructured Clinical Databases</span><span class="g">Dr. Marco Pancotti - MultiPhysixLab</span><span class="g">Dr. Sara Paratico - Gruppo San Donato</span><span class="g">San Donato Milanese, Milan, Italy · 2–3 October 2026</span><span class="g num">01 / 13</span></footer>
|
||
<aside class="notes">
|
||
Good morning. My name is Marco Pancotti, and I led the part of the PAMP-FA project dedicated to building the AritmoLab portal at Policlinico San Donato, with support from Sara Paratico, who will co-present with me today.
|
||
The scope of the project included a data warehouse and tools for machine learning and predictive statistics to support research by the Arrhythmology Unit, directed by Professor Pappone and Professor Locati.
|
||
Today we'll focus on the use of AI to extract structured information from clinical text, followed by a brief tour of the portal and ThothII, a tool for building datamarts from natural-language requests.
|
||
</aside>
|
||
</section>
|
||
|
||
|
||
<section class="arit">
|
||
<header class="head">
|
||
<img src="logo.png" alt="">
|
||
<span class="head-org">AritmoLab · Policlinico San Donato</span>
|
||
</header>
|
||
<nav class="band">
|
||
<span class="band-section">Where we started</span>
|
||
</nav>
|
||
<div class="sbody">
|
||
<h2>Where we started - Four islands and four missing pieces</h2>
|
||
<div class="scatter">
|
||
<div class="todo-stack" style="left:30%; top:22%; width:40%">
|
||
<div class="todo" data-missing="ci" style="margin-left:0"><span class="q">?</span><span class="tx"><b>Clinical Intelligence</b><span>dashboards on the clinical history of the Unit and its patients</span></span></div>
|
||
<div class="todo" data-missing="dwh" style="margin-left:9%"><span class="q">?</span><span class="tx"><b>Datawarehouse</b><span>the single source of truth</span></span></div>
|
||
<div class="todo" data-missing="ml" style="margin-left:4%"><span class="q">?</span><span class="tx"><b>ML-ready data</b><span>cohort tables, ready for model training</span></span></div>
|
||
<div class="todo" data-missing="portal" style="margin-left:12%"><span class="q">?</span><span class="tx"><b>Professional Portal</b><span>AritmoLab — the grown-up Omics Portal</span></span></div>
|
||
</div>
|
||
<button class="isle" data-start="cardioref" style="--rot:-3deg; --size:114px; left:calc(1% + 39px); top:4%; width:230px">
|
||
<svg class="ico" viewBox="0 0 48 48" fill="none" stroke="currentColor" stroke-width="2.6" stroke-linecap="round" stroke-linejoin="round"><path d="M24 33 c-6-3.5-9-7-9-10.5 0-3 4-5.5 7-3 1 .8 1.6 1.8 2 3 .4-1.2 1-2.2 2-3 3-2.5 7 0 7 3 0 3.5-3 7-9 10.5z" fill="currentColor" stroke="none"/><path d="M22.5 21 h3 v3 h3 v3 h-3 v3 h-3 v-3 h-3 v-3 h3 z" fill="#fff" stroke="none"/></svg>
|
||
<span class="nm">Cardioref</span><span class="sc" style="width:300px">visits · operations · reports — twenty years of management on structured forms</span></button>
|
||
<button class="isle" data-start="genetic" style="--rot:3deg; --size:114px; left:calc(73% + 39px); top:4%; width:230px">
|
||
<svg class="ico" viewBox="0 0 48 48" fill="none" stroke="currentColor" stroke-width="2.6" stroke-linecap="round" stroke-linejoin="round"><path d="M18 8 C28 14 28 20 18 26 C12 30 12 35 18 40"/><path d="M30 8 C20 14 20 20 30 26 C36 30 36 35 30 40"/><path d="M17 12 H31"/><path d="M15.5 18 H32.5"/><path d="M17 24 H31"/><path d="M14.5 31 H22"/><rect x="25" y="29" width="12" height="12" rx="1.5" stroke-width="2"/><path d="M25 34 H37"/><path d="M30.5 29 V41"/></svg>
|
||
<span class="nm">Genetic data</span><span class="sc">DNA instruments — results typed into Excel by hand</span></button>
|
||
<button class="isle" data-start="omics" style="--rot:-2deg; --size:114px; left:calc(1% + 39px); top:55%; width:230px">
|
||
<svg class="ico" viewBox="0 0 48 48" fill="none" stroke="currentColor" stroke-width="2.6" stroke-linecap="round" stroke-linejoin="round"><circle cx="24" cy="24" r="20" stroke-width="2.6"/><circle cx="24" cy="24" r="4" fill="currentColor" stroke="none"/><circle cx="24" cy="10" r="3.4"/><circle cx="11.5" cy="31" r="3.4"/><circle cx="36.5" cy="31" r="3.4"/><path d="M24 20.5 V13.5"/><path d="M21 26.5 L14 29.5"/><path d="M27 26.5 L34 29.5"/></svg>
|
||
<span class="nm">Aritmolab Portal</span><span class="sc">meant to host Cardioref + genetics — still immature</span></button>
|
||
<button class="isle" data-start="ecg" style="--rot:2deg; --size:114px; left:calc(73% + 39px); top:55%; width:230px">
|
||
<svg class="ico" viewBox="0 0 48 48" fill="none" stroke="currentColor" stroke-width="2.6" stroke-linecap="round" stroke-linejoin="round"><circle cx="24" cy="24" r="20" stroke-width="2.6"/><path d="M9 24 h6 l3-9 5 16 3-7 h14" stroke-width="2.4"/><circle cx="24" cy="24" r="1.8" fill="currentColor" stroke="none"/></svg>
|
||
<span class="nm">ECG</span><span class="sc">paper strip + a CSV on request</span></button>
|
||
<div class="missing-title" style="left:calc(30% + 14px); top:calc(22% - 46px); width:40%">Missing Pieces</div>
|
||
<p class="hint" style="position:absolute; left:26%; width:48%; bottom:2px; margin:0; text-align:center">▸ click a system: how it was used — and what held it back · click a missing piece: what it would have given</p>
|
||
</div>
|
||
</div>
|
||
<footer class="foot"><span class="g">Role of AI in the Analysis of Unstructured Clinical Databases</span><span class="g">Dr. Marco Pancotti - MultiPhysixLab</span><span class="g">Dr. Sara Paratico - Gruppo San Donato</span><span class="g">San Donato Milanese, Milan, Italy · 2–3 October 2026</span><span class="g num">02 / 13</span></footer>
|
||
<aside class="notes">
|
||
The starting point was four disconnected systems:
|
||
1 - <strong>Cardioref</strong> held twenty years of electrophysiology records and supported clinical operations, but much of the information needed for research was in free text.
|
||
2 - <strong>Genetic data</strong> were manually entered into Excel, without integration with Cardioref.
|
||
3 - The <strong>Aritmolab Portal</strong> was an early prototype, not yet integrated with the source systems.
|
||
4 - <strong>ECG</strong> data were available as paper records or CSV exports on request.
|
||
|
||
Whe lacked:
|
||
5 - some longitudinal <strong>clinical analytics</strong>
|
||
6 - a shared <strong>data warehouse</strong>,
|
||
7 - a set of structured datasets for model training,
|
||
8 - and a unified <strong>clinical portal</strong>.
|
||
|
||
</aside>
|
||
</section>
|
||
|
||
|
||
<section class="arit flow-slide">
|
||
<header class="head">
|
||
<img src="logo.png" alt="">
|
||
<span class="head-org">AritmoLab · Policlinico San Donato</span>
|
||
</header>
|
||
<nav class="band">
|
||
<span class="band-section">The project</span>
|
||
</nav>
|
||
<div class="sbody">
|
||
<div class="kicker">The plan</div>
|
||
<h2>What we wanted to build</h2>
|
||
<p class="lead">One platform, two destinations: the 360° patient portal and a research-ready warehouse. Built on open-source pillars, with AI coding agents wherever they helped. Next, we’ll look at the hardest part: unstructured data.</p>
|
||
<div class="flow">
|
||
<button type="button" class="zone existed lens-target" id="plan-existed" data-lens="plan-existed" data-lens-title="What existed" data-lens-number="1" aria-label="Explore What existed" aria-expanded="false" aria-controls="slide-lens">
|
||
<span class="zlabel">What existed</span>
|
||
<span class="fcol">
|
||
<span class="fbox src"><b>Cardioref</b><span>cardiology records · procedures · letters</span></span>
|
||
<span class="fbox src"><b>Genetic data</b><span>labs · variants · nomenclature</span></span>
|
||
<span class="fbox src"><b>ECG</b><span>signals · device follow-up</span></span>
|
||
<span class="fbox src future"><b><span class="plus">+</span>Future sources</b><span>new subsystems can be connected as sources</span></span>
|
||
</span>
|
||
<template class="lens-details">
|
||
<p class="detail-intro">The starting material offered complementary views of the patient.</p>
|
||
<p><b>Clinical course · Cardioref</b>Visits, procedures and reports describe the course of care. Dates and narrative details give each event its clinical context.</p>
|
||
<p><b>Genetic findings</b>Laboratory results and variant descriptions add the genetic perspective, recorded separately from the clinical history.</p>
|
||
<p><b>Electrical activity · ECG</b>Tracings and exported signals document cardiac electrical activity, complementing the written account of the patient’s condition.</p>
|
||
<p><b>Different forms of evidence</b>Structured fields, free text and signals must retain their clinical meaning when connected. Future sources would extend this initial set.</p>
|
||
</template>
|
||
</button>
|
||
<div class="fcol ai-hit">
|
||
<button class="brain brain-btn" data-ai="ingestion" aria-label="AI contribution: the mappings" aria-pressed="false"><span class="ai-label" aria-hidden="true">AI</span><img class="artificial-brain" src="artificial-brain.svg" alt=""><span class="ai-number">6</span></button>
|
||
<div class="farrow">→</div>
|
||
</div>
|
||
<div class="zone built">
|
||
<span class="zlabel">What we built</span>
|
||
<div style="display:flex; align-items:center; gap:10px; height:100%;">
|
||
<div class="fcol" style="flex:0.8; justify-content:center"><button type="button" class="fbox lens-target" id="plan-staging" data-lens="plan-staging" data-lens-number="2" aria-label="Explore Staging" aria-expanded="false" aria-controls="slide-lens"><b>Staging</b><span>raw replica of the sources</span>
|
||
<template class="lens-details">
|
||
<p class="detail-intro">The landing area: a faithful working copy of the data brought in from each source system.</p>
|
||
<p><b>What it contains</b>Original tables, identifiers, dates and clinical text, still in the source’s format.</p>
|
||
<p><b>Why it matters</b>Subsequent processing works on this copy. The original clinical system continues its daily work, and transformations can be checked against the imported data.</p>
|
||
<p class="detail-example"><b>Clinical example</b>A Cardioref letter arrives with its original wording. “No syncope” is still text; this layer does not yet turn it into a clinical variable.</p>
|
||
</template>
|
||
</button></div>
|
||
<div class="farrow">→</div>
|
||
<div class="fcol ai" style="flex:1.25">
|
||
<button class="brain brain-btn" data-ai="integration" aria-label="AI contribution: clinical text reading" aria-pressed="false"><span class="ai-label" aria-hidden="true">AI</span><img class="artificial-brain" src="artificial-brain.svg" alt=""><span class="ai-number">7</span></button>
|
||
<button type="button" class="fbox lens-target" id="plan-integration" data-lens="plan-integration" data-lens-number="3" aria-label="Explore Integration" aria-expanded="false" aria-controls="slide-lens"><b>Integration</b><span>cleaned, normalized, deduplicated</span>
|
||
<template class="lens-details">
|
||
<p class="detail-intro">The reconciliation layer: source records become consistent, connected clinical information.</p>
|
||
<p><b>What happens here</b>Formats and terminology are standardized, duplicates reconciled, and records linked through patient and event identifiers.</p>
|
||
<p><b>Where AI contributes</b>It extracts conditions, procedures and drug-challenge outcomes from clinical text. Negation and context matter; extraction quality needs validation.</p>
|
||
<p class="detail-example"><b>Clinical example</b>“No syncope” is not a positive finding. A family history of Brugada must remain distinct from the patient’s own diagnosis.</p>
|
||
</template>
|
||
</button>
|
||
<div class="fcap">AI reads the clinical text: pathologies, procedures, drug-challenge outcomes</div>
|
||
</div>
|
||
<div class="farrow">→</div>
|
||
<div class="fcol" style="flex:1.05; justify-content:center"><button type="button" class="fbox lens-target" id="plan-star-schema" data-lens="plan-star-schema" data-lens-number="4" aria-label="Explore Data warehouse" aria-expanded="false" aria-controls="slide-lens"><b>Data warehouse</b><span>star schema · organized for analysis</span>
|
||
<template class="lens-details">
|
||
<p class="detail-intro">The shared analytical database: integrated clinical data reorganized for research across patients and over time.</p>
|
||
<p><b>How a star schema works</b>“Facts” represent events or measurements, such as a procedure or test. “Dimensions” describe their context, such as the patient, date and procedure type.</p>
|
||
<p><b>Why it matters</b>Researchers can filter, group and compare records using common definitions, without reconstructing every relationship from the original hospital tables.</p>
|
||
<p class="detail-example"><b>Clinical question it can support</b>How many patients underwent a given procedure each year, and how does that distribution vary by age group?</p>
|
||
</template>
|
||
</button></div>
|
||
<div class="farrow">→</div>
|
||
<div class="fcol ai" style="flex:1.25">
|
||
<button class="brain brain-btn" data-ai="datamarts" aria-label="AI contribution: datamart generation" aria-pressed="false"><span class="ai-label" aria-hidden="true">AI</span><img class="artificial-brain" src="artificial-brain.svg" alt=""><span class="ai-number">8</span></button>
|
||
<button type="button" class="fbox lens-target" id="plan-datamarts" data-lens="plan-datamarts" data-lens-number="5" aria-label="Explore Datamarts" aria-expanded="false" aria-controls="slide-lens"><b>Datamarts</b><span>research-ready marts</span>
|
||
<template class="lens-details">
|
||
<p class="detail-intro">Focused datasets derived from the warehouse for a specific research question, study or dashboard.</p>
|
||
<p><b>What they define</b>The cohort, time window, variables and level of detail: for example, one row per patient or one row per procedure.</p>
|
||
<p><b>How ThothII helps</b>A plain-English question becomes a proposed SQL query through a guided workflow with human review. The resulting dataset supports analysis and portal dashboards.</p>
|
||
<p class="detail-example"><b>Illustrative study dataset</b>Patients who underwent a drug-challenge test, with test date, result and selected clinical characteristics. Its cohort and variable definitions must be agreed before interpreting results.</p>
|
||
</template>
|
||
</button>
|
||
<div class="fcap">AI builds them on demand from plain-English questions (ThothII)</div>
|
||
</div>
|
||
<div class="farrow">→</div>
|
||
<div class="fcol" style="flex:1">
|
||
<div class="fbox end"><b>AritmoLab Data Warehouse</b><span>queried for research</span></div>
|
||
<div class="fbox end"><b>AritmoLab Portal</b><span>management & exploration</span></div>
|
||
</div>
|
||
</div>
|
||
</div>
|
||
</div>
|
||
<div class="pillars">
|
||
<span class="agents">AI coding agents — Anthropic · OpenAI — used wherever they helped, 360° in the code</span>
|
||
<span class="pillar"><b>Airflow</b> · orchestration</span>
|
||
<span class="pillar"><b>Superset</b> · dashboards</span>
|
||
<span class="pillar"><b>Django</b> · portal scaffolding</span>
|
||
<span class="pillar"><b>Authentik</b> · auth — GSD LDAP</span>
|
||
</div>
|
||
</div>
|
||
<footer class="foot"><span class="g">Role of AI in the Analysis of Unstructured Clinical Databases</span><span class="g">Dr. Marco Pancotti - MultiPhysixLab</span><span class="g">Dr. Sara Paratico - Gruppo San Donato</span><span class="g">San Donato Milanese, Milan, Italy · 2–3 October 2026</span><span class="g num">03 / 13</span></footer>
|
||
<aside class="notes">
|
||
Here is the whole project on one slide.
|
||
|
||
On the left, what already existed, three hospital data sources:
|
||
(1) <strong>What existed</strong>: <strong>Cardioref</strong>, our electrophysiology records dataset, the <strong>genetic data</strong> coming from the labs, the <strong>ECG signals</strong>, and, in the future, other sources.
|
||
|
||
On the right, what we built in AritmoLab:
|
||
(2) a staging copy of all relevant Cardioref data, (3) an integration layer where the data is cleaned and normalized, (4) a star-schema warehouse, where the data are reorganized as facts and dimensions, and (5) research datamarts, generated from the data warehouse and presented through dashboards embedded in the portal.
|
||
|
||
Everywhere you see the brain symbol, that's where AI works for us.
|
||
6 - The <strong>mappings</strong> that bring the sources in were themselves drafted by AI, then revised by humans
|
||
7 - At <strong>integration</strong>, AI reads the clinical text.
|
||
8 - At the end, AI builds the <strong>datamarts</strong> on demand from plain-English questions.
|
||
That was the plan. And to build it, we used AI everywhere it helped — drafting configs, reading clinical text, and at the very end, producing the datamarts themselves.
|
||
|
||
Next, we’ll look at the hardest part: unstructured data. My colleague, Sara Paratico,do will show you exactly how the text reading works. But first, the first step of the climb: the mappings.
|
||
</aside>
|
||
</section>
|
||
|
||
|
||
<section class="arit center-v problem">
|
||
<header class="head">
|
||
<img src="logo.png" alt="">
|
||
<span class="head-org">AritmoLab · Policlinico San Donato</span>
|
||
</header>
|
||
<nav class="band">
|
||
<span class="band-section">The problem</span>
|
||
</nav>
|
||
<div class="sbody">
|
||
<div class="kicker">From free text to a research platform</div>
|
||
<h2>The clinical truth lives in unstructured columns</h2>
|
||
<div class="quote">
|
||
<div class="tr-grid">
|
||
<p class="it">“Il paziente riferisce sincope ricorrente; ECG basale con sopraslivellamento ST in V1–V3; test provocativo con flecainide positivo per pattern Brugada.”</p>
|
||
<p class="en"><span class="en-tag">EN</span>“The patient reports recurrent syncope; baseline ECG with ST-segment elevation in V1–V3; flecainide provocation test positive for a Brugada pattern.”</p>
|
||
</div>
|
||
</div>
|
||
<div class="pipeline" id="pipeline">
|
||
<div class="stage"><b>Sources</b><span>Cardioref · Genetic data<br>Omics Portal · ECG</span><span class="brain-ai"><svg viewBox="0 0 32 32" fill="none" stroke="currentColor" stroke-width="2.4" stroke-linecap="round" stroke-linejoin="round"><path d="M16 4 C11 2 5.5 4 6 8.5 C2.5 10 2 14.5 4.5 17 C2.5 20 4 24.5 8 25 C9 28 14 29.5 16 27"/><path d="M16 4 C21 2 26.5 4 26 8.5 C29.5 10 30 14.5 27.5 17 C29.5 20 28 24.5 24 25 C23 28 18 29.5 16 27"/><path d="M16 4.5 V26.5"/><path d="M16 10 h4.5"/><circle cx="23" cy="10" r="1.7" fill="currentColor" stroke="none"/><path d="M16 15.5 h-4.5"/><circle cx="8.5" cy="15.5" r="1.7" fill="currentColor" stroke="none"/><path d="M16 21 h4.5"/><circle cx="23" cy="21" r="1.7" fill="currentColor" stroke="none"/></svg></span></div>
|
||
<div class="arrow">→</div>
|
||
<div class="stage"><b>Staging</b><span>raw replica</span><span class="brain-ai"><svg viewBox="0 0 32 32" fill="none" stroke="currentColor" stroke-width="2.4" stroke-linecap="round" stroke-linejoin="round"><path d="M16 4 C11 2 5.5 4 6 8.5 C2.5 10 2 14.5 4.5 17 C2.5 20 4 24.5 8 25 C9 28 14 29.5 16 27"/><path d="M16 4 C21 2 26.5 4 26 8.5 C29.5 10 30 14.5 27.5 17 C29.5 20 28 24.5 24 25 C23 28 18 29.5 16 27"/><path d="M16 4.5 V26.5"/><path d="M16 10 h4.5"/><circle cx="23" cy="10" r="1.7" fill="currentColor" stroke="none"/><path d="M16 15.5 h-4.5"/><circle cx="8.5" cy="15.5" r="1.7" fill="currentColor" stroke="none"/><path d="M16 21 h4.5"/><circle cx="23" cy="21" r="1.7" fill="currentColor" stroke="none"/></svg></span></div>
|
||
<div class="arrow">→</div>
|
||
<div class="stage ai"><b>Integration</b><span>3NF · text mining</span><span class="brain-ai"><svg viewBox="0 0 32 32" fill="none" stroke="currentColor" stroke-width="2.4" stroke-linecap="round" stroke-linejoin="round"><path d="M16 4 C11 2 5.5 4 6 8.5 C2.5 10 2 14.5 4.5 17 C2.5 20 4 24.5 8 25 C9 28 14 29.5 16 27"/><path d="M16 4 C21 2 26.5 4 26 8.5 C29.5 10 30 14.5 27.5 17 C29.5 20 28 24.5 24 25 C23 28 18 29.5 16 27"/><path d="M16 4.5 V26.5"/><path d="M16 10 h4.5"/><circle cx="23" cy="10" r="1.7" fill="currentColor" stroke="none"/><path d="M16 15.5 h-4.5"/><circle cx="8.5" cy="15.5" r="1.7" fill="currentColor" stroke="none"/><path d="M16 21 h4.5"/><circle cx="23" cy="21" r="1.7" fill="currentColor" stroke="none"/></svg></span></div>
|
||
<div class="arrow">→</div>
|
||
<div class="stage"><b>Data Warehouse</b><span>star schema</span></div>
|
||
<div class="arrow">→</div>
|
||
<div class="stage hot" id="marts"><b>Marts</b><span>dbt · Superset</span></div>
|
||
</div>
|
||
<div class="midrow">
|
||
<div class="concepts">
|
||
<div class="lbl">four challenges to face · the answers come next</div>
|
||
<div class="citem"><span class="n">01</span><div><b>Volume without structure</b><span class="d">twenty years of records locked in free-text columns</span></div></div>
|
||
<div class="citem"><span class="n">02</span><div><b>Ambiguous language</b><span class="d">negation · family history · bilingual shorthand</span></div></div>
|
||
<div class="citem"><span class="n">03</span><div><b>No shared ontology</b><span class="d">every system names conditions its own way</span></div></div>
|
||
<div class="citem"><span class="n">04</span><div><b>Proving reliability</b><span class="d">extraction must be demonstrably correct</span></div></div>
|
||
</div>
|
||
<div class="railwrap" id="railwrap">
|
||
<div class="conn-drop" id="conn-drop"></div>
|
||
<div class="conn-elbow" id="conn-elbow"></div>
|
||
<div class="rail" id="rail">
|
||
<div class="lbl">outputs of the datamarts</div>
|
||
<div class="row"><svg class="oicon" viewBox="0 0 34 34" fill="none" stroke="currentColor" stroke-width="2.2" stroke-linecap="round"><path d="M4 29 h26"/><rect x="7" y="17" width="5" height="12" rx="1"/><rect x="15" y="9" width="5" height="20" rx="1"/><rect x="23" y="13" width="5" height="16" rx="1"/></svg><b>Dashboards</b><span class="d">clinical review on Superset</span></div>
|
||
<div class="row"><svg class="oicon" viewBox="0 0 34 34" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round"><circle cx="7" cy="7" r="2.6"/><circle cx="7" cy="27" r="2.6"/><circle cx="17" cy="17" r="3.4"/><circle cx="27" cy="7" r="2.6"/><circle cx="27" cy="27" r="2.6"/><path d="M9.3 8.8 L14.6 14.6"/><path d="M9.3 25.2 L14.6 19.4"/><path d="M19.4 14.6 L24.7 8.8"/><path d="M19.4 19.4 L24.7 25.2"/></svg><b>Machine Learning</b><span class="d">cohort tables for model training</span></div>
|
||
<div class="row"><svg class="oicon" viewBox="0 0 34 34" fill="none" stroke="currentColor" stroke-width="2.2" stroke-linecap="round"><path d="M3 27 L11 19 L17 22 L23 13"/><path d="M23 13 L30 6" stroke-dasharray="3 3.4"/><circle cx="30" cy="6" r="2.2" fill="currentColor" stroke="none"/><path d="M3 31 h28" opacity=".35"/></svg><b>Predictive Statistics</b><span class="d">outcomes & risk analysis</span></div>
|
||
</div>
|
||
</div>
|
||
</div>
|
||
<div class="thoth-wrap" id="thothwrap">
|
||
<div class="t-conn" id="t-conn"></div>
|
||
<div class="t-head" id="t-head"></div>
|
||
<div class="thoth-col">
|
||
<div class="thoth-box"><span class="brain-ai"><svg viewBox="0 0 32 32" fill="none" stroke="currentColor" stroke-width="2.4" stroke-linecap="round" stroke-linejoin="round"><path d="M16 4 C11 2 5.5 4 6 8.5 C2.5 10 2 14.5 4.5 17 C2.5 20 4 24.5 8 25 C9 28 14 29.5 16 27"/><path d="M16 4 C21 2 26.5 4 26 8.5 C29.5 10 30 14.5 27.5 17 C29.5 20 28 24.5 24 25 C23 28 18 29.5 16 27"/><path d="M16 4.5 V26.5"/><path d="M16 10 h4.5"/><circle cx="23" cy="10" r="1.7" fill="currentColor" stroke="none"/><path d="M16 15.5 h-4.5"/><circle cx="8.5" cy="15.5" r="1.7" fill="currentColor" stroke="none"/><path d="M16 21 h4.5"/><circle cx="23" cy="21" r="1.7" fill="currentColor" stroke="none"/></svg></span>ThothII — natural-language → SQL over the DWH, builds the final datamarts</div>
|
||
</div>
|
||
</div>
|
||
</div>
|
||
<footer class="foot"><span class="g">Role of AI in the Analysis of Unstructured Clinical Databases</span><span class="g">Dr. Marco Pancotti - MultiPhysixLab</span><span class="g">Dr. Sara Paratico - Gruppo San Donato</span><span class="g">San Donato Milanese, Milan, Italy · 2–3 October 2026</span><span class="g num">04 / 13</span></footer>
|
||
<aside class="notes">
|
||
Here is one real sentence from a discharge letter — in Italian, as the clinicians wrote it. In this single sentence there is a diagnosis (syncope), an ECG finding (ST elevation in V1–V3), and a drug-challenge result (flecainide positive for Brugada pattern). For a human cardiologist this is readable in two seconds. For a database, this is just a blob of text in a column.
|
||
The structured tables — demographics, procedures, dates — only tell half the story. The rest is locked inside these free-text fields, in every hospital system we have.
|
||
</aside>
|
||
</section>
|
||
|
||
|
||
<section class="arit center-v stats">
|
||
<header class="head">
|
||
<img src="logo.png" alt="">
|
||
<span class="head-org">AritmoLab · Policlinico San Donato</span>
|
||
</header>
|
||
<nav class="band">
|
||
<span class="band-section">Text analysis</span>
|
||
</nav>
|
||
<div class="sbody">
|
||
<div class="kicker">What the text mining produced</div>
|
||
<h2>Deterministic AI, audited numbers</h2>
|
||
<div class="strip">
|
||
<div class="sstat"><span class="n">58,438</span><span class="l">letters read</span></div>
|
||
<div class="sstat"><span class="n">73,389</span><span class="l">pathologies extracted</span></div>
|
||
<div class="sstat"><span class="n">10,908</span><span class="l">tests parsed</span></div>
|
||
<div class="sstat"><span class="n">2,307</span><span class="l">Brugada patients</span></div>
|
||
<div class="sstat"><span class="n">≈ 240</span><span class="l">letters / month</span></div>
|
||
</div>
|
||
<div class="cycle" id="cycle">
|
||
<svg class="cycle-svg" viewBox="0 0 1184 270">
|
||
<defs><marker id="arrw" viewBox="0 0 10 10" refX="8" refY="5" markerWidth="6.5" markerHeight="6.5" orient="auto-start-reverse"><path d="M0 0L10 5L0 10z" fill="#cb333b"/></marker></defs>
|
||
<g id="cycle-arcs" fill="none" stroke="#cb333b" stroke-width="2">
|
||
<path marker-end="url(#arrw)"/>
|
||
<path marker-end="url(#arrw)"/>
|
||
<path marker-end="url(#arrw)"/>
|
||
<path marker-end="url(#arrw)"/>
|
||
</g>
|
||
</svg>
|
||
<div class="vcomp" style="left:592px; top:50px"><svg class="cicon" viewBox="0 0 34 34" fill="none" stroke="currentColor" stroke-width="2.2" stroke-linecap="round" stroke-linejoin="round"><rect x="7" y="4" width="20" height="26" rx="2.5"/><path d="M12 11h10"/><path d="M12 16h10"/><path d="M12 21h6"/></svg><b>Letter reader</b><span class="d">every discharge letter, Italian & English</span></div>
|
||
<div class="vcomp" style="left:992px; top:135px"><svg class="cicon" viewBox="0 0 34 34" fill="none" stroke="currentColor" stroke-width="2.2" stroke-linecap="round" stroke-linejoin="round"><circle cx="14.5" cy="14.5" r="8"/><path d="M20.5 20.5 L29 29"/><path d="M11 14.5 h7"/><path d="M14.5 11 v7"/></svg><b>Clinical matcher</b><span class="d">recognises diagnoses, procedures, test outcomes</span></div>
|
||
<div class="vcomp" style="left:592px; top:220px"><svg class="cicon" viewBox="0 0 34 34" fill="none" stroke="currentColor" stroke-width="2.2" stroke-linecap="round" stroke-linejoin="round"><path d="M17 3.5 L28 8 v8.5 c0 7-4.7 11.6-11 14 C10.7 28.1 6 23.5 6 16.5 V8 Z"/><path d="M12.5 16.5 l3.2 3.2 L23.5 12"/></svg><b>Context guard</b><span class="d">negations & family history kept apart</span></div>
|
||
<div class="vcomp" style="left:192px; top:135px"><svg class="cicon" viewBox="0 0 34 34" fill="none" stroke="currentColor" stroke-width="2.2" stroke-linecap="round" stroke-linejoin="round"><circle cx="17" cy="7" r="3"/><circle cx="7" cy="26.5" r="3"/><circle cx="27" cy="26.5" r="3"/><path d="M17 10 v7"/><path d="M17 17 L8 23.5"/><path d="M17 17 L26 23.5"/></svg><b>Ontology sorter</b><span class="d">every finding into its clinical category</span></div>
|
||
</div>
|
||
<div class="note-row">
|
||
<div class="nitem"><b>Traceable.</b> Every finding carries the version of the rules that produced it (<code>pattern_version</code>): any number can be rebuilt years later.</div>
|
||
<div class="nitem"><b>Audited.</b> A new rule set goes live only after it proves ≥95% accuracy on 100 records reviewed by hand.</div>
|
||
</div>
|
||
</div>
|
||
<footer class="foot"><span class="g">Role of AI in the Analysis of Unstructured Clinical Databases</span><span class="g">Dr. Marco Pancotti - MultiPhysixLab</span><span class="g">Dr. Sara Paratico - Gruppo San Donato</span><span class="g">San Donato Milanese, Milan, Italy · 2–3 October 2026</span><span class="g num">05 / 13</span></footer>
|
||
<aside class="notes">
|
||
[DRAFT — Sara] I'll take you inside these numbers. We read 58,438 discharge letters — every single one, every night. From them the text miner extracted 73,389 pathology records, classified into our clinical ontology. It parsed 10,908 drug-challenge tests, distinguishing the therapy from the actual test. And at the end of the chain: 2,307 patients with a confirmed Brugada pattern — a cohort nobody could have built by hand.
|
||
This is not a one-off migration: the pipeline runs every night, and today it processes on average two hundred and forty letters a month. Four simple pieces do the work: a letter reader for the Italian and English text, a clinical matcher that recognises diagnoses, procedures and test outcomes, a context guard that keeps negations and family history apart, and the ontology sorter that files every finding into its clinical category.
|
||
The key word here is deterministic: no black box. Every number can be traced back to the rule that produced it.
|
||
</aside>
|
||
</section>
|
||
|
||
|
||
<section class="arit nlp">
|
||
<header class="head">
|
||
<img src="logo.png" alt="">
|
||
<span class="head-org">AritmoLab · Policlinico San Donato</span>
|
||
</header>
|
||
<nav class="band">
|
||
<span class="band-section">Text analysis</span>
|
||
</nav>
|
||
<div class="sbody">
|
||
<div class="kicker">Clinical NLP</div>
|
||
<h2>Reading clinical text is not keyword matching</h2>
|
||
<div class="hgap"></div>
|
||
<div class="split">
|
||
<div class="left">
|
||
<ul class="points">
|
||
<li><b>Negation</b> — “fibrillazione atriale <i>esclusa</i>”, “<i>non</i> FA” <span class="g">(“AF ruled out”, “no AF”)</span>: the rule scans a window around every match before deciding</li>
|
||
<li><b>Family ≠ patient</b> — “padre con FA” <span class="g">(“father with AF”)</span> is attributed to the family member, not to the patient</li>
|
||
<li><b>Two languages</b> — every pattern matches IT and EN forms: fibrillazione atriale / atrial fibrillation</li>
|
||
<li><b>Abbreviations & noise</b> — FA, f.a., “TA 140/90” (blood pressure, not a diagnosis)</li>
|
||
<li><b>One field, many statements</b> — the text is split into clauses before matching</li>
|
||
</ul>
|
||
</div>
|
||
<div class="right">
|
||
<div class="code">
|
||
<div class="code-head"><span>what the rules see</span><span>examples from the corpus</span></div>
|
||
<pre>“<i>padre con</i> fibrillazione atriale”
|
||
<span class="en">“father with atrial fibrillation”</span>
|
||
→ patologia: FA · attribuzione: <b>familiare</b>
|
||
<span class="en">pathology: AF · attribution: family</span>
|
||
|
||
“fibrillazione atriale <i>esclusa</i>”
|
||
<span class="en">“atrial fibrillation ruled out”</span>
|
||
→ patologia: FA · <b>negato</b>
|
||
<span class="en">pathology: AF · negated</span>
|
||
|
||
“test provocativo con flecainide
|
||
positivo per pattern Brugada”
|
||
<span class="en">“flecainide provocation test, positive for Brugada pattern”</span>
|
||
→ test: flecainide · esito: <b>POSITIVO</b>
|
||
<span class="en">test: flecainide · outcome: POSITIVE</span></pre>
|
||
</div>
|
||
</div>
|
||
</div>
|
||
</div>
|
||
<footer class="foot"><span class="g">Role of AI in the Analysis of Unstructured Clinical Databases</span><span class="g">Dr. Marco Pancotti - MultiPhysixLab</span><span class="g">Dr. Sara Paratico - Gruppo San Donato</span><span class="g">San Donato Milanese, Milan, Italy · 2–3 October 2026</span><span class="g num">06 / 13</span></footer>
|
||
<aside class="notes">
|
||
[DRAFT — Sara] Why can't we just search for "FA"? Because clinical text lies to naive search. "Fibrillazione atriale esclusa" contains the words of a diagnosis but negates it — so every rule scans a window around the match, looking for negation cues. "Padre con FA" is real atrial fibrillation — but in the father, not the patient: we record it as family history. Letters mix Italian and English, abbreviations collide ("TA" is blood pressure, not a therapy), and one field can contain five different statements — so we split the text into clauses first.
|
||
Each of these problems has a specific, versioned solution. Sara to expand with real corpus examples.
|
||
</aside>
|
||
</section>
|
||
|
||
|
||
<section class="arit center-v">
|
||
<header class="head">
|
||
<img src="logo.png" alt="">
|
||
<span class="head-org">AritmoLab · Policlinico San Donato</span>
|
||
</header>
|
||
<nav class="band">
|
||
<span class="band-section">Text analysis</span>
|
||
</nav>
|
||
<div class="sbody">
|
||
<div class="kicker">The clinical ontology</div>
|
||
<h2>Two tiers, one clinical order</h2>
|
||
<div class="onto-split">
|
||
<div class="onto-left">
|
||
<div class="tier-h">TIER 1 · arrhythmic<span class="cnt">11</span></div>
|
||
<div class="cats">
|
||
<span class="cat fill"><b>AF</b><span class="it">FA</span></span><span class="cat fill"><b>Brugada</b></span><span class="cat fill"><b>Flutter</b></span><span class="cat fill"><b>AT</b><span class="it">TA</span></span><span class="cat fill"><b>AVNRT</b><span class="it">TRN</span></span><span class="cat fill"><b>PSVT</b><span class="it">WPW / TPSV</span></span><span class="cat fill"><b>VT</b><span class="it">TV</span></span><span class="cat fill"><b>VF</b><span class="it">FV</span></span><span class="cat fill"><b>Ventricular ectopy</b><span class="it">Extrasistolia V.</span></span><span class="cat fill"><b>Syncope</b><span class="it">Sincope</span></span><span class="cat fill"><b>Long QT</b><span class="it">QT lungo</span></span>
|
||
</div>
|
||
<div class="tier-h">TIER 2 · structural<span class="cnt">5</span></div>
|
||
<div class="cats">
|
||
<span class="cat"><b>AV block</b><span class="it">Blocco AV</span></span><span class="cat"><b>Bundle branch block</b><span class="it">Blocco di branca</span></span><span class="cat"><b>Cardiomyopathy</b><span class="it">Cardiomiopatia</span></span><span class="cat"><b>Heart failure</b><span class="it">Scompenso</span></span><span class="cat"><b>Valvular disease</b><span class="it">Valvulopatia</span></span>
|
||
</div>
|
||
<div class="onto-points">
|
||
<div class="oitem"><span class="n">01</span><div><b>Order encodes clinical precedence</b> <span class="d">— “Brugada” is matched before “TV”, so “substrato per TV” <span class="g">(“substrate for VT”)</span> can't mask a Brugada pattern</span></div></div>
|
||
<div class="oitem"><span class="n">02</span><div><b>Synonyms live inline</b> <span class="d">— FA / f.a. / fib. atriale / atrial fibrillation → one canonical label</span></div></div>
|
||
<div class="oitem"><span class="n">03</span><div><b>Versioned like software</b> <span class="d">— semver for the pattern library, <code>pattern_version</code> on every extracted row</span></div></div>
|
||
</div>
|
||
</div>
|
||
<div class="out-panel">
|
||
<div class="out-h">The output: text becomes recorded data</div>
|
||
<div class="out-item"><span class="n">01</span><div><b>Quantitative data on the records</b><span class="d">every procedure and implant the text analysis reads is written back as structured, quantitative values on the record that describes it</span></div></div>
|
||
<div class="out-item"><span class="n">02</span><div><b>Straight into the DWH flow</b><span class="d">these records enter the data-warehouse generation like any other source</span></div></div>
|
||
<div class="out-item"><span class="n">03</span><div><b>As if typed at the visit</b><span class="d">the data lands exactly as if clinicians had keyed it in themselves during the visits</span></div></div>
|
||
</div>
|
||
</div>
|
||
</div>
|
||
<footer class="foot"><span class="g">Role of AI in the Analysis of Unstructured Clinical Databases</span><span class="g">Dr. Marco Pancotti - MultiPhysixLab</span><span class="g">Dr. Sara Paratico - Gruppo San Donato</span><span class="g">San Donato Milanese, Milan, Italy · 2–3 October 2026</span><span class="g num">07 / 13</span></footer>
|
||
<aside class="notes">
|
||
[DRAFT — Sara] Extraction needs a target vocabulary — that's the ontology. Tier 1 holds the eleven arrhythmological categories that matter most for our research; Tier 2 holds five structural conditions. The order of the rules is itself clinical knowledge: Brugada patterns are tested before TV, because "substrato per TV" often appears in Brugada reports and would otherwise mask the diagnosis.
|
||
The ontology is versioned like software: a semantic version for the pattern library, stamped on every extracted row. When we add a synonym or fix a rule, the change is traceable — and the data can be rebuilt.
|
||
And this is the output of the whole work: reading a letter writes structured, quantitative values back onto the records that describe the procedures and implants it contains — and those records enter the warehouse generation exactly as if the clinicians had typed them during the visits. Text becomes data, indistinguishable from bedside data entry.
|
||
Sara: review the category list and the Italian labels before final. Labels are now English-first (audience is mostly foreign) — please confirm the EN terms, especially PSVT as the umbrella for WPW/TPSV and AT/AVNRT for TA/TRN.
|
||
</aside>
|
||
</section>
|
||
|
||
|
||
<section class="arit center-v validation">
|
||
<header class="head">
|
||
<img src="logo.png" alt="">
|
||
<span class="head-org">AritmoLab · Policlinico San Donato</span>
|
||
</header>
|
||
<nav class="band">
|
||
<span class="band-section">Text analysis</span>
|
||
</nav>
|
||
<div class="sbody">
|
||
<!-- Verified against aritmolab/chirone-etl on Gitea, commit 2ab00106188f298c1a0c1e2c48b3a4dbc66c45b5.
|
||
Exact sources and interpretation limits: ../slides/08-validation-sources.md.
|
||
SC-003/004 are targets; ~893 refers to false-positive new-letter records, not distinct patients. -->
|
||
<div class="kicker">Validation</div>
|
||
<h2>Trusting the text is a process, not a promise</h2>
|
||
<div class="onto-split">
|
||
<div class="onto-left">
|
||
<div class="onto-points">
|
||
<button type="button" class="oitem lens-target" id="validation-quality" data-lens="validation-quality" data-lens-number="1" data-lens-placement="center" data-lens-caption="Validation · 01 / 04" aria-label="Explore Quality gates" aria-expanded="false" aria-controls="slide-lens"><span class="n">01</span><span><b>Quality gates</b> <span class="d">— targets: ≥95% procedure accuracy on 100 manual reviews; ≥85% pathology coverage</span></span>
|
||
<template class="lens-details">
|
||
<p class="detail-intro">We separate correct classification from how often the rules find a condition.</p>
|
||
<p><b>Procedure classification</b>The specification sets a target of at least 95% agreement with a manual review of 100 randomly selected records: did the system assign the right procedure type?</p>
|
||
<p><b>Pathology coverage</b>The 85% target asks whether at least one recognized arrhythmia is extracted from a record, assuming most procedures concern known arrhythmias. It does not measure whether every diagnosis is correct or every disease is found.</p>
|
||
<p class="detail-example"><b>How to interpret the numbers</b>These are acceptance targets. Clinical review is still needed to detect incorrect labels and missed findings; coverage alone cannot establish clinical accuracy.</p>
|
||
</template>
|
||
</button>
|
||
<button type="button" class="oitem lens-target" id="validation-review" data-lens="validation-review" data-lens-number="2" data-lens-placement="center" data-lens-caption="Validation · 02 / 04" aria-label="Explore Clinical criteria" aria-expanded="false" aria-controls="slide-lens"><span class="n">02</span><span><b>Clinical criteria</b> <span class="d">— explicit criteria determine which patients belong in the analysis; a disease name alone is not enough</span></span>
|
||
<template class="lens-details">
|
||
<p class="detail-intro">Finding a disease name is the first step. Inclusion in an analysis depends on what the text means and the criteria for that cohort.</p>
|
||
<p><b>Read the context</b>“Father with Brugada” concerns a relative; “test negative for Brugada” reports a negative result; “suspected Brugada” expresses uncertainty. None of these phrases alone establishes a diagnosis in the patient.</p>
|
||
<p><b>Make the selection explicit</b>Queries apply the cohort criteria to prepare the dataset. Superset displays the result. The diagnosis view first excludes negated findings and findings attributed to relatives.</p>
|
||
<p class="detail-example"><b>A rule used in this project</b>For Brugada and long QT syndrome, that view also requires a positive provocative test or an ablation for the condition, at patient level. Other pathologies use only the first filter. These are project inclusion rules, not a universal diagnostic standard.</p>
|
||
</template>
|
||
</button>
|
||
<button type="button" class="oitem lens-target" id="validation-feedback" data-lens="validation-feedback" data-lens-number="3" data-lens-placement="center" data-lens-caption="Validation · 03 / 04" aria-label="Explore Feedback loop" aria-expanded="false" aria-controls="slide-lens"><span class="n">03</span><span><b>Feedback loop</b> <span class="d">— v1.3.1 fixed missed negations behind ~893 false-positive Brugada records</span></span>
|
||
<template class="lens-details">
|
||
<p class="detail-intro">The 11 June 2026 audit identified about 893 false-positive Brugada records in the newer letter format.</p>
|
||
<p><b>The error and the fix</b>“Test alla flecainide negativo per sindrome di Brugada” was treated as an affirmed finding: the rules recognized “negato”, but missed “negativo”. Version 1.3.1 added “negativo”, “negativa” and “negativi” before and after the condition.</p>
|
||
<p><b>Make the correction repeatable</b>Regression tests require that this sentence still produces a Brugada finding, now marked as negated. A separate test checks that an affirmed Brugada diagnosis remains positive. Reprocessing applies the revised rules to historical letters.</p>
|
||
<p class="detail-example"><b>What changed</b>The finding is reclassified, not erased. The ~893 figure counts affected records, not necessarily distinct patients; the cohort filters can now exclude those negated findings.</p>
|
||
</template>
|
||
</button>
|
||
<button type="button" class="oitem lens-target" id="validation-context" data-lens="validation-context" data-lens-number="4" data-lens-placement="center" data-lens-caption="Validation · 04 / 04" aria-label="Explore Nothing is silently dropped" aria-expanded="false" aria-controls="slide-lens"><span class="n">04</span><span><b>Nothing is silently dropped</b> <span class="d">— negated and family-attributed findings are stored too, filtered only at the mart layer</span></span>
|
||
<template class="lens-details">
|
||
<p class="detail-intro">A recognized finding can be kept even when it is negated or refers to someone else.</p>
|
||
<p><b>Store context alongside the finding</b>The extractor looks up to 60 characters before and after the match, stopping at a full stop, semicolon or line break. It records whether the finding is negated and whether it concerns the patient or a relative.</p>
|
||
<p><b>Filter when building the research dataset</b>Those flags travel through integration and the warehouse to the mentions dataset. The confirmed-diagnosis view excludes negated and family findings, while the underlying dataset remains available for other analyses.</p>
|
||
<p class="detail-example"><b>Preservation has a defined scope</b>Repeated matches for the same pathology are consolidated, preferring an affirmed patient finding. “Suspected” or “to exclude” is not automatically treated as a negation. These limits remain explicit for clinical review.</p>
|
||
</template>
|
||
</button>
|
||
</div>
|
||
<div class="roi roi-quiet"><b class="k">Quality gate</b>Manual comparison checks procedure accuracy. Pathology coverage and clinical correctness are assessed separately.</div>
|
||
</div>
|
||
<button type="button" class="out-panel lens-target" id="validation-advantages" data-lens="validation-advantages" data-lens-number="5" data-lens-title="The advantages" data-lens-placement="center" data-lens-caption="Validation · Advantages" aria-label="Explore the advantages: speed, precision, determinism" aria-expanded="false" aria-controls="slide-lens">
|
||
<span class="out-h">The advantage: speed, precision, determinism</span>
|
||
<span class="out-item"><span class="n">01</span><span><b>Machine speed</b><span class="d">twenty years of letters are read by rules, not by hand; when a pattern improves, the whole archive can be reprocessed</span></span></span>
|
||
<span class="out-item"><span class="n">02</span><span><b>Explicit quality targets</b><span class="d">≥95% procedure accuracy on a manual sample; ≥85% pathology coverage, with clinical review of the resulting cohort</span></span></span>
|
||
<span class="out-item"><span class="n">03</span><span><b>Deterministic by design</b><span class="d">the same letter and rule version yield the same extraction: no LLM sampling, no randomness, results traceable to versioned rules</span></span></span>
|
||
<template class="lens-details">
|
||
<p class="detail-intro">Speed, precision and determinism each address a different part of making clinical text usable for research.</p>
|
||
<p><b>01 · Speed: apply a correction across the archive</b>Once defined, the extraction rules process letters automatically. An improved rule can be applied again to historical text, without manually relabelling every record.</p>
|
||
<p><b>02 · Precision: make errors visible and correctable</b>Explicit quality targets, clinical context and cohort filters make the results inspectable. The Brugada correction shows how an audit can identify a recurring error, and regression tests can protect the fix. Quality still needs measurement and clinical review.</p>
|
||
<p><b>03 · Determinism: reproduce the extraction</b>With the same input, rule version and configuration, extraction produces the same findings. There is no language-model sampling at this step. Versioning explains why a result changes when the rules change.</p>
|
||
</template>
|
||
</button>
|
||
</div>
|
||
</div>
|
||
<footer class="foot"><span class="g">Role of AI in the Analysis of Unstructured Clinical Databases</span><span class="g">Dr. Marco Pancotti - MultiPhysixLab</span><span class="g">Dr. Sara Paratico - Gruppo San Donato</span><span class="g">San Donato Milanese, Milan, Italy · 2–3 October 2026</span><span class="g num">08 / 13</span></footer>
|
||
<aside class="notes">
|
||
<p><button type="button" data-note-lens="validation-quality">1. Quality gates.</button> We set a 95% procedure-accuracy target, against 100 manual reviews. The 85% pathology target measures coverage, not diagnostic accuracy.</p>
|
||
<p><button type="button" data-note-lens="validation-review">2. Clinical criteria.</button> Finding a disease name in a letter is only the first step. We apply explicit clinical criteria to decide which patients belong in the analysis.</p>
|
||
<p><button type="button" data-note-lens="validation-feedback">3. Feedback loop.</button> Version 1.3.1 corrected missed negations in roughly 900 Brugada records, with regression tests protecting the fix.</p>
|
||
<p><button type="button" data-note-lens="validation-context">4. Nothing is silently dropped.</button> Negated and family findings retain their context. Research datasets filter them explicitly, so preserving a finding does not mean counting it as the patient’s diagnosis.</p>
|
||
<p><button type="button" data-note-lens="validation-advantages">5. The advantages.</button> We can reprocess the archive quickly, inspect the rules behind each result, and reproduce the same extraction with the same rules.</p>
|
||
</aside>
|
||
</section>
|
||
|
||
<!-- ============ 09 · ARITMOLAB TODAY ============ -->
|
||
<section class="arit">
|
||
<header class="head">
|
||
<img src="logo.png" alt="">
|
||
<span class="head-org">AritmoLab · Policlinico San Donato</span>
|
||
</header>
|
||
<nav class="band">
|
||
<span class="band-section">AritmoLab today</span>
|
||
</nav>
|
||
<div class="sbody">
|
||
<div class="kicker">The portal, today</div>
|
||
<h2>AritmoLab — a quick tour</h2>
|
||
<div class="screenshot-tour" data-screenshot-auto-open aria-label="AritmoLab tour, screenshots 1 to 7">
|
||
<button type="button" data-screenshot="home" data-screenshot-number="1" data-screenshot-title="Home page" aria-haspopup="dialog"><img src="screenshots/1-HomePage.png" alt="" loading="lazy"><span><b>1</b> Home page</span></button>
|
||
<button type="button" data-screenshot="patients" data-screenshot-number="2" data-screenshot-title="Patient list" aria-haspopup="dialog"><img src="screenshots/2-PatientList.png" alt="" loading="lazy"><span><b>2</b> Patient list</span></button>
|
||
<button type="button" data-screenshot="profile" data-screenshot-number="3" data-screenshot-title="Patient profile" aria-haspopup="dialog"><img src="screenshots/3-PatientGeneralData.png" alt="" loading="lazy"><span><b>3</b> Patient profile</span></button>
|
||
<button type="button" data-screenshot="history" data-screenshot-number="4" data-screenshot-title="Clinical history" aria-haspopup="dialog"><img src="screenshots/4-PatientDetail.png" alt="" loading="lazy"><span><b>4</b> Clinical history</span></button>
|
||
<button type="button" data-screenshot="procedure" data-screenshot-number="5" data-screenshot-title="Procedure details" aria-haspopup="dialog"><img src="screenshots/5-PatientProcedureDetails.png" alt="" loading="lazy"><span><b>5</b> Procedure details</span></button>
|
||
<button type="button" data-screenshot="dashboards" data-screenshot-number="6" data-screenshot-title="Dashboard catalogue" aria-haspopup="dialog"><img src="screenshots/6-DashboardsList.png" alt="" loading="lazy"><span><b>6</b> Dashboard catalogue</span></button>
|
||
<button type="button" data-screenshot="brugada" data-screenshot-number="7" data-screenshot-title="Brugada dashboard" aria-haspopup="dialog"><img src="screenshots/7-Brugada-dashboard.png" alt="" loading="lazy"><span><b>7</b> Brugada dashboard</span></button>
|
||
</div>
|
||
<p class="hint">Select a screen to enlarge · Follow the tour from 1 to 7</p>
|
||
</div>
|
||
<footer class="foot"><span class="g">Role of AI in the Analysis of Unstructured Clinical Databases</span><span class="g">Dr. Marco Pancotti - MultiPhysixLab</span><span class="g">Dr. Sara Paratico - Gruppo San Donato</span><span class="g">San Donato Milanese, Milan, Italy · 2–3 October 2026</span><span class="g num">09 / 13</span></footer>
|
||
<aside class="notes">
|
||
<p><button type="button" data-note-screenshot="home">1. Home page.</button> The home page summarises around 57,000 patients, by age, sex and geographical origin. It is the starting point for exploring the archive.</p>
|
||
<p><button type="button" data-note-screenshot="patients">2. Patient list.</button> Search filters help us find a patient or study participant, open their record, or export the results.</p>
|
||
<p><button type="button" data-note-screenshot="profile">3. Patient profile.</button> The profile brings demographic and clinical fields together, with access to procedures, devices, diagnostic examinations and genetics.</p>
|
||
<p><button type="button" data-note-screenshot="history">4. Clinical history.</button> A dated timeline brings together clinical notes, discharge letters and procedures, retaining the source of each event.</p>
|
||
<p><button type="button" data-note-screenshot="procedure">5. Procedure details.</button> Here, an ablation record shows the treated arrhythmias, procedural details, recorded complications and conclusions.</p>
|
||
<p><button type="button" data-note-screenshot="dashboards">6. Dashboard catalogue.</button> We then move from individual records to dashboards covering departmental activity, procedures, devices and genetics.</p>
|
||
<p><button type="button" data-note-screenshot="brugada">7. Brugada dashboard.</button> The funnel separates text mentions, filtered mentions, confirmed cases and ablation outcomes. Other charts describe sex, age, annual diagnoses and the timing of pre- and post-assessments.</p>
|
||
</aside>
|
||
</section>
|
||
<!-- ============ 10 · CRISIS ============ -->
|
||
<section class="arit crisis">
|
||
<header class="head">
|
||
<img src="logo.png" alt="">
|
||
<span class="head-org">AritmoLab · Policlinico San Donato</span>
|
||
</header>
|
||
<nav class="band">
|
||
<span class="band-section">The crisis</span>
|
||
</nav>
|
||
<div class="sbody">
|
||
<div class="kicker">All good? Not yet.</div>
|
||
<h2>The warehouse speaks SQL. Research needs more.</h2>
|
||
<div class="hgap"></div>
|
||
<div class="crisis-points" aria-label="Four barriers between the warehouse and research">
|
||
<button type="button" class="oitem lens-target" id="crisis-language" data-lens="crisis-language" data-lens-number="1" data-lens-title="From clinical question to SQL" data-lens-placement="center" data-lens-caption="The research gap · 01 / 04" aria-expanded="false" aria-controls="slide-lens"><span class="n">01</span><span><b>Clinical questions need translation</b><span class="d">Clinicians define the question; a query must express it across linked tables.</span></span>
|
||
<template class="lens-details">
|
||
<p class="detail-intro">Knowing what to ask is different from knowing how the database stores the answer.</p>
|
||
<p><b>The clinical question</b>“How many patients with confirmed Brugada underwent an ablation?” requires agreed definitions of the cohort and procedure.</p>
|
||
<p><b>The engineering task</b>SQL must connect diagnoses, patients and procedures, apply the time window and count each patient once, even when several records describe the same person.</p>
|
||
<p class="detail-example"><b>The bridge</b>Clinical expertise defines the meaning. A reviewed query turns that meaning into an explicit, checkable selection.</p>
|
||
</template>
|
||
</button>
|
||
<button type="button" class="oitem lens-target" id="crisis-intelligence" data-lens="crisis-intelligence" data-lens-number="2" data-lens-title="Health Intelligence: describe what happened" data-lens-placement="center" data-lens-caption="The research gap · 02 / 04" aria-expanded="false" aria-controls="slide-lens"><span class="n">02</span><span><b>Health Intelligence needs shared definitions</b><span class="d">Volumes, diagnoses and outcomes need consistent groups, periods and denominators.</span></span>
|
||
<template class="lens-details">
|
||
<p class="detail-intro">A dashboard needs agreed indicators, not simply a chart drawn over raw records.</p>
|
||
<p><b>Define what is counted</b>Patients, admissions and procedures answer different questions. For an outcome percentage, specify which patients are eligible and the observation period.</p>
|
||
<p><b>Prepare comparable summaries</b>Aggregate by year, procedure or patient group using the same definitions. Keep missing information visible so that changes in documentation are not mistaken for changes in care.</p>
|
||
<p class="detail-example"><b>Example</b>Annual ablation volumes describe activity. An outcome percentage also needs a defined denominator and follow-up window.</p>
|
||
</template>
|
||
</button>
|
||
<button type="button" class="oitem lens-target" id="crisis-prediction" data-lens="crisis-prediction" data-lens-number="3" data-lens-title="Predictive research: build the study table" data-lens-placement="center" data-lens-caption="The research gap · 03 / 04" aria-expanded="false" aria-controls="slide-lens"><span class="n">03</span><span><b>Predictive research needs a study dataset</b><span class="d">A defined cohort, consistent variables and outcomes measured over an agreed period.</span></span>
|
||
<template class="lens-details">
|
||
<p class="detail-intro">Linked clinical records must become a table shaped around the study question.</p>
|
||
<p><b>Define one row</b>Choose the unit of analysis: for example, one patient or one procedure. Place the selected characteristics in columns, with consistent units and explicit handling of missing values.</p>
|
||
<p><b>Respect the timeline</b>Define when prediction would occur and when the outcome is assessed. Predictor variables must contain only information available at that prediction time.</p>
|
||
<p class="detail-example"><b>Example</b>To study outcomes after ablation, separate pre-procedure characteristics from later observations. This prevents future information from leaking into the prediction.</p>
|
||
</template>
|
||
</button>
|
||
<button type="button" class="oitem lens-target" id="crisis-datamarts" data-lens="crisis-datamarts" data-lens-number="4" data-lens-title="Datamarts: the preparation bottleneck" data-lens-placement="center" data-lens-caption="The research gap · 04 / 04" aria-expanded="false" aria-controls="slide-lens"><span class="n">04</span><span><b>Hand-made datamarts are the bottleneck</b><span class="d">Each question requires selection, joins, checks and a reproducible dataset.</span></span>
|
||
<template class="lens-details">
|
||
<p class="detail-intro">A datamart is a focused dataset prepared for a particular analysis.</p>
|
||
<p><b>More than writing SQL</b>Someone must agree the cohort, connect the sources, resolve duplicate records and check missing values and patient counts. A query can run successfully and still answer the wrong question.</p>
|
||
<p><b>Make the work repeatable</b>Keep the selection rules, query and checks together, so the dataset can be rebuilt when the data or study definition changes.</p>
|
||
<p class="detail-example"><b>Where assistance helps</b>AI can draft the query. Clinicians and engineers still review its meaning and results before using the dataset.</p>
|
||
</template>
|
||
</button>
|
||
</div>
|
||
<div class="roi"><b class="k">The gap</b>Twenty years of data, one warehouse — and no fast road from a research question to an answer.</div>
|
||
</div>
|
||
<footer class="foot"><span class="g">Role of AI in the Analysis of Unstructured Clinical Databases</span><span class="g">Dr. Marco Pancotti - MultiPhysixLab</span><span class="g">Dr. Sara Paratico - Gruppo San Donato</span><span class="g">San Donato Milanese, Milan, Italy · 2–3 October 2026</span><span class="g num">10 / 13</span></footer>
|
||
<aside class="notes">
|
||
<p><strong>0. What is SQL?</strong><br>SQL means Structured Query Language. It tells a database what to select, connect and count.</p>
|
||
<p><button type="button" data-note-lens="crisis-language">1. Clinical questions.</button><br>“How many patients with confirmed Brugada underwent an ablation?” We must define confirmation and count each patient once.</p>
|
||
<p><button type="button" data-note-lens="crisis-intelligence">2. Health Intelligence.</button><br>Dashboards need agreed definitions, time periods and denominators to make comparisons meaningful.</p>
|
||
<p><button type="button" data-note-lens="crisis-prediction">3. Predictive research.</button><br>Study tables separate characteristics known before prediction from outcomes observed afterwards.</p>
|
||
<p><button type="button" data-note-lens="crisis-datamarts">4. Datamarts.</button><br>A datamart is an analysis dataset with explicit selection rules and repeatable checks.</p>
|
||
<p><strong>5. The gap.</strong><br>The gap is between clinical meaning and database instructions: storing data does not automatically make a question answerable.</p>
|
||
</aside>
|
||
</section>
|
||
|
||
<!-- ============ 11 · THOTHII TOUR ============ -->
|
||
<section class="arit">
|
||
<header class="head">
|
||
<img src="logo.png" alt="">
|
||
<span class="head-org">AritmoLab · Policlinico San Donato</span>
|
||
</header>
|
||
<nav class="band">
|
||
<span class="band-section">A tour of ThothII</span>
|
||
</nav>
|
||
<div class="sbody">
|
||
<div class="kicker">ThothII, step by step</div>
|
||
<h2>From question to datamart — the app</h2>
|
||
<div class="screenshot-tour screenshot-tour-compact" data-screenshot-auto-open aria-label="ThothII tour, screenshots 1 to 12">
|
||
<button type="button" data-screenshot="thothii-start" data-screenshot-number="1" data-screenshot-title="Starting point" aria-haspopup="dialog">
|
||
<img src="screenshots/thothii/01-StartingPoint.png?v=4" alt="" loading="lazy"><span><b>1</b> Starting point</span>
|
||
</button>
|
||
<button type="button" data-screenshot="thothii-disambiguation-1" data-screenshot-number="2" data-screenshot-title="Disambiguation · 1" aria-haspopup="dialog">
|
||
<img src="screenshots/thothii/02-Disambiguation01.png?v=4" alt="" loading="lazy"><span><b>2</b> Disambiguation · 1</span>
|
||
</button>
|
||
<button type="button" data-screenshot="thothii-disambiguation-final" data-screenshot-number="3" data-screenshot-title="Question clarified" aria-haspopup="dialog">
|
||
<img src="screenshots/thothii/03-DisambiguationFinal.png?v=4" alt="" loading="lazy"><span><b>3</b> Question clarified</span>
|
||
</button>
|
||
<button type="button" data-screenshot="thothii-schema-1" data-screenshot-number="4" data-screenshot-title="Schema linking · 1" aria-haspopup="dialog">
|
||
<img src="screenshots/thothii/04-SchemaLinking01.png?v=4" alt="" loading="lazy"><span><b>4</b> Schema linking · 1</span>
|
||
</button>
|
||
<button type="button" data-screenshot="thothii-schema-final" data-screenshot-number="5" data-screenshot-title="Schema confirmed" aria-haspopup="dialog">
|
||
<img src="screenshots/thothii/05-CloseSchemaLinking.png?v=4" alt="" loading="lazy"><span><b>5</b> Schema confirmed</span>
|
||
</button>
|
||
<button type="button" data-screenshot="thothii-cte-plan" data-screenshot-number="6" data-screenshot-title="CTE planning" aria-haspopup="dialog">
|
||
<img src="screenshots/thothii/06-CTEPlanning.png?v=4" alt="" loading="lazy"><span><b>6</b> CTE planning</span>
|
||
</button>
|
||
<button type="button" data-screenshot="thothii-cte-1" data-screenshot-number="7" data-screenshot-title="CTE · 1" aria-haspopup="dialog">
|
||
<img src="screenshots/thothii/07-CTE01.png?v=4" alt="" loading="lazy"><span><b>7</b> CTE · 1</span>
|
||
</button>
|
||
<button type="button" data-screenshot="thothii-sql" data-screenshot-number="8" data-screenshot-title="Final SQL" aria-haspopup="dialog">
|
||
<img src="screenshots/thothii/08-FinalSQL.png?v=5" alt="" loading="lazy"><span><b>8</b> Final SQL</span>
|
||
</button>
|
||
<button type="button" data-screenshot="thothii-datamart" data-screenshot-number="9" data-screenshot-title="Datamart" aria-haspopup="dialog">
|
||
<img src="screenshots/thothii/09-DatamartProduction.png?v=5" alt="" loading="lazy"><span><b>9</b> Datamart</span>
|
||
</button>
|
||
<button type="button" data-screenshot="thothii-memory" data-screenshot-number="10" data-screenshot-title="Saving memory" aria-haspopup="dialog">
|
||
<img src="screenshots/thothii/10-MemorySaving.png?v=5" alt="" loading="lazy"><span><b>10</b> Saving memory</span>
|
||
</button>
|
||
<button type="button" data-screenshot="thothii-finish" data-screenshot-number="11" data-screenshot-title="Final step" aria-haspopup="dialog">
|
||
<img src="screenshots/thothii/11-FinalStep.png?v=5" alt="" loading="lazy"><span><b>11</b> Final step</span>
|
||
</button>
|
||
<button type="button" data-screenshot="thothii-return" data-screenshot-number="12" data-screenshot-title="Back to start" aria-haspopup="dialog">
|
||
<img src="screenshots/thothii/12-BackToStartingPoint.png?v=5" alt="" loading="lazy"><span><b>12</b> Back to start</span>
|
||
</button>
|
||
</div>
|
||
<p class="hint">Select a screen to enlarge · Follow the tour from 1 to 12</p>
|
||
</div>
|
||
<footer class="foot"><span class="g">Role of AI in the Analysis of Unstructured Clinical Databases</span><span class="g">Dr. Marco Pancotti - MultiPhysixLab</span><span class="g">Dr. Sara Paratico - Gruppo San Donato</span><span class="g">San Donato Milanese, Milan, Italy · 2–3 October 2026</span><span class="g num">11 / 13</span></footer>
|
||
<aside class="notes">
|
||
<div data-popup-notes="thothii-start"><h3>From a research question to SQL</h3><div data-popup-notes-body><p>The task has changed. We are no longer extracting structured, coded values from the text within clinical records. We are now starting from a natural-language request to retrieve the records that contain the relevant values.</p><p>Research questions are often difficult to express clearly and unambiguously, and their clinical concepts do not map directly to the underlying database structure. As a result, writing the right SQL query can take hours, with repeated testing and refinement before it accurately reflects the intended research question.</p></div></div>
|
||
<div data-popup-notes="thothii-disambiguation-1"><h3>Clarifying the question with the researcher</h3><div data-popup-notes-body>
|
||
<p><strong>1. Workflow phase.</strong> At the top, the phase indicator shows where we are in the process. Here, phase one is highlighted: clarifying the research question.</p>
|
||
<p><strong>2. AI working time.</strong> The timer measures how long the AI has been working, excluding the time the human reviewer spends considering the options and making decisions.</p>
|
||
<p><strong>3. Model activity.</strong> On the left, we can follow the AI's running explanation of its analysis: the interpretations it considers, the evidence it consults, and the rationale for its proposals.</p>
|
||
<p><strong>4. Proposed interpretations.</strong> In the centre, the AI asks the researcher to choose between alternative interpretations. Here, the question is how to identify an ablation for atrial fibrillation in the database. This choice determines which patients enter the study.</p>
|
||
<p><strong>5. Reviewer actions.</strong> The reviewer can select a proposed option, or use <strong>Go back</strong> to revisit the previous step, <strong>Exit</strong> to stop the current workflow, or <strong>Other — specify</strong> to describe a different interpretation in free text.</p>
|
||
</div></div>
|
||
<div data-popup-notes="thothii-disambiguation-final"><h3>From clarification to an agreed study question</h3><div data-popup-notes-body>
|
||
<p><strong>Bringing the decisions together.</strong> The AI combines the original request with the researcher's answers into an explicit study definition. It records what we mean by each clinical concept, which patients to include, the time window, and the result we want.</p>
|
||
<p><strong>The clarified question.</strong> In this example: how many distinct patients had their first recorded ablation for atrial fibrillation between 2020 and 2023? We identify the procedure from the treated pathology, and find each patient's first AF ablation across the entire available database history before applying the date filter.</p>
|
||
<p><strong>Making the interpretation actionable.</strong> The summary links these choices to the relevant database fields and selection rules, and displays the checks completed during clarification. These agreed criteria guide the subsequent schema linking and SQL construction.</p>
|
||
<p><strong>Human confirmation.</strong> The researcher reviews the consolidated definition. <strong>Save and proceed</strong> confirms it and closes phase one; <strong>Reject</strong> sends it back for revision.</p>
|
||
</div></div>
|
||
<div data-popup-notes="thothii-schema-1"><h3>Finding the data needed to answer the question</h3><div data-popup-notes-body>
|
||
<p><strong>Linking clinical meaning to data.</strong> We are now in phase four, schema linking: identifying where the information needed to answer the agreed question is stored. Tables group related records; columns hold specific details, such as a patient identifier, a procedure date, or the condition treated.</p>
|
||
<p><strong>Proposed tables.</strong> The AI presents candidate tables and explains why each is relevant. Here, one table provides the ablation episodes and their dates; another identifies the pathology treated, allowing us to distinguish atrial fibrillation ablations from other procedures.</p>
|
||
<p><strong>Selecting the columns.</strong> The column controls let the reviewer inspect and adjust the fields selected for the query. We need the information required to identify patients, recognise the relevant procedures, and apply the agreed time criteria.</p>
|
||
<p><strong>Review before proceeding.</strong> The researcher and a technical reviewer can check this mapping together, adjust the selection, and confirm it. This establishes which data will support the SQL query; the relationships between the selected tables are reviewed next.</p>
|
||
</div></div>
|
||
<div data-popup-notes="thothii-schema-final"><h3>Schema linking complete: from clinical meaning to database structure</h3><div data-popup-notes-body>
|
||
<p><strong>What schema linking means.</strong> Schema linking connects the concepts in the clarified research question to the database: which tables contain the information, which columns represent each concept, and how records from different tables must be connected.</p>
|
||
<p><strong>Our clinical example.</strong> Here, we connect the treated pathology to the corresponding ablation episode, associate that episode with its date and patient, and specify the rules for identifying each patient's first recorded AF ablation and applying the 2020–2023 window. The intended result is a count of distinct patients.</p>
|
||
<p><strong>What completion means.</strong> These choices are now consolidated into a documented mapping, with the relationships, selection rules, and validation checks shown for review. We have an explicit specification of where the answer will come from and how the relevant data fit together.</p>
|
||
<p><strong>The final approval.</strong> By selecting <strong>Save and proceed</strong>, the reviewer approves this mapping for the next stage: planning and building the SQL query. Query execution and verification of the resulting patient count still follow.</p>
|
||
</div></div>
|
||
<div data-popup-notes="thothii-cte-plan"><h3>CTEs: building the query step by step</h3><div data-popup-notes-body><p><strong>CTE stands for Common Table Expression.</strong> It is a named intermediate result within an SQL query.</p><p>CTEs break a complex research question into smaller, logical steps—for example, identifying AF ablations, finding the first procedure for each patient, and selecting the study population.</p><p>This makes the query easier to understand, check, and modify, helping us verify that each step reflects the intended clinical criteria.</p></div></div>
|
||
<div data-popup-notes="thothii-cte-1"><h3>Reviewing a CTE: purpose, SQL, and results</h3><div data-popup-notes-body>
|
||
<p><strong>Purpose and position.</strong> Each CTE is presented as a reviewable step. At the top, we see its name and position in the sequence: here, the first of three. A short explanation describes its clinical purpose and the reasoning behind the selection rules.</p>
|
||
<p><strong>The SQL implementation.</strong> The code shows how that purpose is translated into database operations. In this example, it connects the treated pathology to the ablation episode and selects AF ablations, returning the patient identifier, episode identifier, and procedure date.</p>
|
||
<p><strong>Tests and a data preview.</strong> The screen reports the test status and execution time, followed by a preview of the returned records. The ten rows shown are a limited preview, not the total study population. Successful execution still requires a check that the results make clinical sense.</p>
|
||
<p><strong>Human review.</strong> The reviewer can compare the explanation, code, and sample data before choosing <strong>Save and proceed</strong> or <strong>Reject</strong>. This makes each intermediate step inspectable before it contributes to the final query.</p>
|
||
</div></div>
|
||
<div data-popup-notes="thothii-sql"><h3>Titolo8</h3><div data-popup-notes-body><p>testo8</p></div></div>
|
||
<div data-popup-notes="thothii-datamart"><h3>Reusing the query: daily datamarts or research on demand</h3><div data-popup-notes-body>
|
||
<p>Once the query generation and review workflow is complete, the validated SQL query can be reused beyond the current session.</p>
|
||
<p><strong>Daily datamart generation.</strong> The query can be integrated into the broader ETL pipeline and scheduled to run every day. This allows the datamart to be rebuilt or refreshed systematically as new source data become available, using the same agreed selection rules.</p>
|
||
<p><strong>Reuse by researchers.</strong> Alternatively, the query can simply be saved as an SQL file and made available to researchers, who can inspect it, run it when needed, or adapt it for a subsequent study.</p>
|
||
<p>In both cases, the workflow produces a reusable query that captures the reviewed interpretation of the research question.</p>
|
||
</div></div>
|
||
<div data-popup-notes="thothii-memory"><h3>Saving clinical clarifications for future queries</h3><div data-popup-notes-body>
|
||
<p><strong>What we retain.</strong> Clarifying a research question can produce knowledge that is useful beyond the current study: for example, an agreed interpretation of a clinical term or how a procedure is represented in the local data.</p>
|
||
<p><strong>The researcher chooses.</strong> At the end of the workflow, ThothII presents candidate clarifications for reuse. The reviewer selects which ones to save as shared knowledge for the workspace.</p>
|
||
<p><strong>How this helps next time.</strong> When a related question is asked, ThothII can retrieve these clarifications and propose them to the reviewer. The reviewer checks whether they apply to the new question before using them. This helps avoid repeating the same clarification work while keeping each study's interpretation under human control.</p>
|
||
</div></div>
|
||
<div data-popup-notes="thothii-finish"><h3>A saved session documents how the result was reached</h3><div data-popup-notes-body>
|
||
<p>At the end of the workflow, ThothII finalizes a session that brings together the generated SQL statement and the documented process that led to it.</p>
|
||
<p>The session preserves the research question, its agreed interpretation, the mapping to the database, the intermediate query steps, and the recorded validation results.</p>
|
||
<p>Human involvement is captured through the recorded clarifications and review decisions: what the reviewer selected, approved, or rejected along the way. These records document how the interaction between the system and the researcher shaped the final query.</p>
|
||
<p>Researchers can revisit the session to inspect the SQL and understand the choices behind it. This provides a traceable record of how the original question became the final result.</p>
|
||
<p><strong>Time and cost for this example.</strong> The AI processing took less than five minutes, excluding the time spent on human review. Using DeepSeek V4 Flash, the estimated model cost for this run was about five cents.</p>
|
||
</div></div>
|
||
<div data-popup-notes="thothii-return"><h3>Expert supervision and deployment options</h3><div data-popup-notes-body>
|
||
<p><strong>Supervision requires knowledge of the data.</strong> The workflow must be supervised by someone who understands the meaning and content of the database tables and fields. AI can make plausible guesses about what a field represents or how records should be connected, and these guesses can become hallucinations. The reviewer's role is to challenge those assumptions and check them against the actual data and its clinical meaning.</p>
|
||
<p><strong>The researcher defines the dataset.</strong> Human judgement is essential to decide which information the study needs, which records meet the required quality criteria, and which time windows and temporal rules should apply. The responsible researcher must confirm these choices so that the resulting dataset addresses the research question. Successful SQL execution alone does not establish that the dataset is suitable for the study.</p>
|
||
<p><strong>Server or authorised workstation.</strong> At Policlinico San Donato, ThothII runs on a server. The application can also be installed on the PC of a qualified staff member who is authorised to access the database remotely. The application runs on that workstation, while the database server executes the SQL through the authorised connection.</p>
|
||
<p><strong>A local deployment option.</strong> Open models such as Qwen3.8-27B or Gemma 4 26B A4B are candidates for running the AI component on premises. Their suitability for this workflow should be assessed on representative research questions, including the quality of the generated SQL and the reliability of the review steps.</p>
|
||
<p><strong>Hardware within reach.</strong> With quantized models, a compact server equipped with a suitable GPU is a practical deployment option. A GPU budget of a few thousand euros is a planning target; the required memory and final cost depend on the model, context length, and number of concurrent sessions.</p>
|
||
<p><strong>Where the query runs.</strong> SQL execution takes place on the internal database server through controlled application code. The language model helps construct and review the query; the database engine executes it.</p>
|
||
<p><strong>Keeping model processing local.</strong> Hosting the session model on premises allows its prompts and responses to remain within the organisation. With an external model, the information included in prompts and tool results must be checked and filtered to prevent sensitive data from leaving the environment.</p>
|
||
<p><small>Model references: <a href="https://huggingface.co/Qwen/Qwen3.8-27B" target="_blank" rel="noopener noreferrer">Qwen3.8-27B model card</a>; <a href="https://ai.google.dev/gemma/docs/core" target="_blank" rel="noopener noreferrer">Gemma 4 models and memory requirements</a>.</small></p>
|
||
</div></div>
|
||
</aside>
|
||
</section>
|
||
<!-- ============ 12 · HAPPY ENDING ============ -->
|
||
<section class="arit">
|
||
<header class="head">
|
||
<img src="logo.png" alt="">
|
||
<span class="head-org">AritmoLab · Policlinico San Donato</span>
|
||
</header>
|
||
<nav class="band">
|
||
<span class="band-section">The happy ending</span>
|
||
</nav>
|
||
<div class="sbody">
|
||
<div class="kicker">From datamart to discovery</div>
|
||
<h2>One datamart — many questions answered</h2>
|
||
<div class="charts">
|
||
<div class="chart"><b>Multivariate analysis</b><span>[SVG chart — next step]</span></div>
|
||
<div class="chart"><b>Machine learning</b><span>[SVG chart — next step]</span></div>
|
||
</div>
|
||
<p class="hint">[Illustrative data, modeled on published arrhythmology predictors — Brugada focus]</p>
|
||
</div>
|
||
<footer class="foot"><span class="g">Role of AI in the Analysis of Unstructured Clinical Databases</span><span class="g">Dr. Marco Pancotti - MultiPhysixLab</span><span class="g">Dr. Sara Paratico - Gruppo San Donato</span><span class="g">San Donato Milanese, Milan, Italy · 2–3 October 2026</span><span class="g num">12 / 13</span></footer>
|
||
<aside class="notes">[DRAFT — next step: two SVG charts, illustrative data modeled on real literature] And this is the happy ending. From one datamart, the Unit can run a multivariate analysis — which factors truly drive arrhythmic risk in Brugada patients — and train machine-learning models on the same table. What used to take weeks of manual data preparation now takes minutes. The AI did not replace the researcher: it gave the researcher back their time.</aside>
|
||
</section>
|
||
|
||
<section class="arit title-slide">
|
||
<header class="head">
|
||
<img src="logo.png" alt="">
|
||
<span class="head-org">AritmoLab · Policlinico San Donato</span>
|
||
</header>
|
||
<nav class="band">
|
||
<span class="band-section">Thank you</span>
|
||
</nav>
|
||
<div class="sbody">
|
||
<div class="kicker">Questions?</div>
|
||
<h1>Thank you</h1>
|
||
<div class="hgap"></div>
|
||
<p class="lead">Want the deep dives? The technical walkthroughs of the text miner, the mapping agents and ThothII are coming as videos on our Substack.</p>
|
||
<div class="speakers" style="margin-top:32px">
|
||
<span>Dr. Marco Pancotti - MultiPhysixLab</span>
|
||
<span>Dr. Sara Paratico - I.R.C.C.S. Policlinico San Donato</span>
|
||
</div>
|
||
</div>
|
||
<footer class="foot"><span class="g">Role of AI in the Analysis of Unstructured Clinical Databases</span><span class="g">Dr. Marco Pancotti - MultiPhysixLab</span><span class="g">Dr. Sara Paratico - Gruppo San Donato</span><span class="g">San Donato Milanese, Milan, Italy · 2–3 October 2026</span><span class="g num">13 / 13</span></footer>
|
||
<aside class="notes">
|
||
[DRAFT] Thank you for your attention. If you want to go deeper — the text miner, the mapping agents, ThothII — we are publishing the technical walkthroughs as videos on our Substack; the link is on the final version of this deck. And now, happy to take your questions.
|
||
</aside>
|
||
</section>
|
||
|
||
</div></div>
|
||
|
||
<!-- AI contribution popup -->
|
||
<div class="ai-pop" id="aiPop">
|
||
<svg class="ai-line"><line x1="0" y1="0" x2="0" y2="0"></line><circle cx="0" cy="0" r="7"></circle></svg>
|
||
<div class="ai-pop-card" role="dialog" aria-modal="true">
|
||
<button class="ai-pop-x" aria-label="Close">✕</button>
|
||
<div class="ai-pop-kicker" id="aiPopKicker">AI contribution</div>
|
||
<div class="ai-pop-title"><span class="brain brain-btn" id="aiPopBrain"><img class="artificial-brain" src="artificial-brain.svg" alt=""></span><b id="aiPopTitle"></b></div>
|
||
<div id="aiPopBody"></div>
|
||
</div>
|
||
</div>
|
||
|
||
<script src="vendor/reveal/dist/reveal.js"></script>
|
||
<script src="vendor/reveal/plugin/notes/notes.js?v=3"></script>
|
||
<script src="slide-lens.js?v=4"></script>
|
||
<script src="screenshot-tour.js?v=4"></script>
|
||
<script>
|
||
Reveal.initialize({
|
||
width: 1280, height: 720, margin: 0,
|
||
hash: true, controls: false, progress: false,
|
||
transition: 'fade', overview: true,
|
||
plugins: [RevealNotes],
|
||
});
|
||
|
||
// jump dropdown — one select per slide, in the red band
|
||
const JUMP_TITLES = ['Title', 'Where we started', 'What we wanted to build',
|
||
'The problem', 'What the text miner reads', 'Hard problems of clinical NLP', 'The clinical ontology',
|
||
'Trust the text', 'AritmoLab today', 'All good? Not yet', 'A tour of ThothII',
|
||
'From datamarts to models', 'Thank you'];
|
||
document.querySelectorAll('.arit .band').forEach((band) => {
|
||
const sel = document.createElement('select');
|
||
sel.className = 'jump';
|
||
sel.setAttribute('aria-label', 'Jump to slide');
|
||
sel.innerHTML = JUMP_TITLES.map((t, i) => `<option value="${i}">${String(i + 1).padStart(2, '0')} · ${t}</option>`).join('');
|
||
sel.addEventListener('change', () => { Reveal.slide(parseInt(sel.value, 10)); sel.blur(); });
|
||
band.appendChild(sel);
|
||
});
|
||
Reveal.on('slidechanged', () => {
|
||
const idx = Reveal.getIndices().h;
|
||
document.querySelectorAll('.arit .band select.jump').forEach((s) => { s.selectedIndex = idx; s.blur(); });
|
||
});
|
||
|
||
// slide "The problem": rail drops from Marts center, ThothII arrow rises to Marts.
|
||
// Reveal scales the canvas — all rect offsets are divided by the current scale.
|
||
function alignProblem() {
|
||
const pipe = document.getElementById('pipeline');
|
||
if (!pipe || !pipe.closest('section').classList.contains('present')) return;
|
||
const s = Reveal.getScale() || 1;
|
||
const pr = pipe.getBoundingClientRect();
|
||
const marts = document.getElementById('marts').getBoundingClientRect();
|
||
const rail = document.getElementById('rail').getBoundingClientRect();
|
||
const rwrap = document.getElementById('railwrap').getBoundingClientRect();
|
||
const drop = document.getElementById('conn-drop');
|
||
const elbow = document.getElementById('conn-elbow');
|
||
const cx = (marts.left + marts.width / 2 - rwrap.left) / s;
|
||
const railLeft = (rail.left - rwrap.left) / s;
|
||
const railTop = (rail.top - rwrap.top) / s;
|
||
const mBottom = (marts.bottom - rwrap.top) / s;
|
||
drop.style.left = (cx - 1) + 'px';
|
||
drop.style.top = mBottom + 'px';
|
||
drop.style.height = Math.max(0, railTop - mBottom) + 'px';
|
||
elbow.style.left = railLeft + 'px';
|
||
elbow.style.width = Math.max(0, cx - railLeft) + 'px';
|
||
elbow.style.top = (railTop - 1) + 'px';
|
||
const twrap = document.getElementById('thothwrap');
|
||
const box = twrap.querySelector('.thoth-box');
|
||
const tr = twrap.getBoundingClientRect();
|
||
const br = box.getBoundingClientRect();
|
||
const conn = document.getElementById('t-conn');
|
||
const head = document.getElementById('t-head');
|
||
const x = tr.width / s - 10;
|
||
const top = (pr.bottom - tr.top) / s + 2;
|
||
const boxTop = (br.top - tr.top) / s;
|
||
conn.style.left = x + 'px';
|
||
conn.style.top = top + 'px';
|
||
conn.style.height = Math.max(0, boxTop - top) + 'px';
|
||
head.style.left = x + 'px';
|
||
head.style.top = (top - 7) + 'px';
|
||
}
|
||
addEventListener('load', alignProblem);
|
||
addEventListener('resize', alignProblem);
|
||
Reveal.on('slidechanged', alignProblem);
|
||
Reveal.on('ready', alignProblem);
|
||
if (document.fonts && document.fonts.ready) document.fonts.ready.then(alignProblem);
|
||
|
||
// slide "What the text miner reads": curved arrows between the four components,
|
||
// drawn from the measured edges of each block so every arrow starts and lands precisely.
|
||
function alignCycle() {
|
||
const cyc = document.getElementById('cycle');
|
||
if (!cyc || !cyc.closest('section').classList.contains('present')) return;
|
||
const s = Reveal.getScale() || 1;
|
||
const cr = cyc.getBoundingClientRect();
|
||
const boxes = [...cyc.querySelectorAll('.vcomp')].map(c => {
|
||
const r = c.getBoundingClientRect();
|
||
const l = (r.left - cr.left) / s, t = (r.top - cr.top) / s, w = r.width / s, h = r.height / s;
|
||
return { l, t, w, h, cx: l + w / 2, cy: t + h / 2 };
|
||
});
|
||
function edgePoint(from, to, box) {
|
||
const dx = to.cx - from.cx, dy = to.cy - from.cy;
|
||
let t = 1;
|
||
if (dx > 0) t = Math.min(t, ((box.l + box.w) - from.cx) / dx);
|
||
if (dx < 0) t = Math.min(t, (box.l - from.cx) / dx);
|
||
if (dy > 0) t = Math.min(t, ((box.t + box.h) - from.cy) / dy);
|
||
if (dy < 0) t = Math.min(t, (box.t - from.cy) / dy);
|
||
return { x: from.cx + dx * t, y: from.cy + dy * t };
|
||
}
|
||
const paths = cyc.querySelectorAll('#cycle-arcs path');
|
||
const seq = [0, 1, 2, 3];
|
||
for (let k = 0; k < 4; k++) {
|
||
const a = boxes[seq[k]], b = boxes[seq[(k + 1) % 4]];
|
||
const start = edgePoint(a, b, a), end = edgePoint(b, a, b);
|
||
const mid = { x: (start.x + end.x) / 2, y: (start.y + end.y) / 2 };
|
||
let ox = mid.x - 592, oy = mid.y - 135;
|
||
const len = Math.hypot(ox, oy) || 1;
|
||
const cpx = mid.x + (ox / len) * 46, cpy = mid.y + (oy / len) * 46;
|
||
paths[k].setAttribute('d',
|
||
`M ${start.x.toFixed(1)} ${start.y.toFixed(1)} Q ${cpx.toFixed(1)} ${cpy.toFixed(1)} ${end.x.toFixed(1)} ${end.y.toFixed(1)}`);
|
||
}
|
||
}
|
||
addEventListener('load', alignCycle);
|
||
addEventListener('resize', alignCycle);
|
||
Reveal.on('slidechanged', alignCycle);
|
||
Reveal.on('ready', alignCycle);
|
||
if (document.fonts && document.fonts.ready) document.fonts.ready.then(alignCycle);
|
||
|
||
const AI_CONTRIBS = {
|
||
ingestion: {
|
||
title: 'The mappings',
|
||
body: `
|
||
<p class="lead">The mappings that turn three raw sources into clean data are <b>crafted by AI</b> — <b>revised by humans</b>.</p>
|
||
<ul class="ai-pop-list">
|
||
<li><b>AI drafts</b> — LLM agents write the mappings from prompts and schema samples</li>
|
||
<li><b>What is a mapping?</b> — a plain instruction sheet that tells the system, field by field, where each piece of data comes from and where it has to land</li>
|
||
<li><b>At scale</b> — <span class="n">19,764</span> lines of instructions, all machine-validated</li>
|
||
<li><b>Humans revise</b> — every line is checked, then recorded in <b>git</b>, the archive that keeps every version of the instructions and can undo any change</li>
|
||
</ul>
|
||
<div class="roi"><b class="k">Positive effects</b>Weeks of hand-writing became days of reviewing.</div>`,
|
||
},
|
||
integration: {
|
||
title: 'Reading the clinical text',
|
||
body: `
|
||
<p class="lead">A deterministic, bilingual (IT/EN) pattern library reads the free text inside the nightly pipeline.</p>
|
||
<ul class="ai-pop-list">
|
||
<li><b>Pathologies</b> — <span class="n">73,389</span> extracted from <span class="n">58,438</span> discharge letters, into a two-tier clinical ontology</li>
|
||
<li><b>Procedures ↔ pathologies</b> — linked at clause level</li>
|
||
<li><b>Drug-challenge tests</b> — <span class="n">10,908</span> parsed (flecainide, ajmaline, adrenaline, isoprenaline)</li>
|
||
<li><b>No black box</b> — semver rules, <code>pattern_version</code> on every row, <span class="n">≥95%</span> accuracy gate on 100 manual reviews</li>
|
||
</ul>
|
||
<div class="roi"><b class="k">Positive effects</b>58,438 letters of free text became an analysable research asset — automatically, every night.</div>`,
|
||
},
|
||
datamarts: {
|
||
title: 'Datamarts from plain English',
|
||
body: `
|
||
<p class="lead">Researchers ask in plain English; the AI writes the SQL over the star schema.</p>
|
||
<ul class="ai-pop-list">
|
||
<li><b>Ask</b> — “how many Brugada patients had an effective ablation?”</li>
|
||
<li><b>Generate</b> — AI builds the SQL and assembles a curated datamart</li>
|
||
<li><b>Serve</b> — results flow to Superset for statistics and machine learning</li>
|
||
<li><b>In the loop</b> — the researcher reviews every proposed query before it runs</li>
|
||
</ul>
|
||
<div class="roi"><b class="k">Positive effects</b>A new research datamart in minutes instead of weeks of hand-written SQL — with the researcher approving every query.</div>`,
|
||
},
|
||
};
|
||
const pop = document.getElementById('aiPop');
|
||
const card = pop.querySelector('.ai-pop-card');
|
||
const line = pop.querySelector('.ai-line line');
|
||
const dot = pop.querySelector('.ai-line circle');
|
||
let activePopup = null;
|
||
const closePop = () => {
|
||
pop.classList.remove('open');
|
||
document.querySelectorAll('.brain-btn[aria-pressed="true"]').forEach(button => button.setAttribute('aria-pressed', 'false'));
|
||
activePopup = null;
|
||
SlideLens.close();
|
||
ScreenshotTour.close();
|
||
window.dispatchEvent(new Event('presentationchange'));
|
||
};
|
||
Reveal.on('slidechanged', closePop);
|
||
const placeNear = (brainEl) => {
|
||
const b = brainEl.getBoundingClientRect();
|
||
const vw = window.innerWidth, vh = window.innerHeight;
|
||
card.style.width = '';
|
||
card.style.maxHeight = '';
|
||
if (pop.classList.contains('is-ai')) {
|
||
const margin = 24, gap = Math.min(150, Math.max(80, vw * .12));
|
||
const roomRight = vw - margin - b.right - gap;
|
||
const roomLeft = b.left - gap - margin;
|
||
const onRight = roomRight >= roomLeft;
|
||
const by = b.top + b.height / 2;
|
||
let left, top, sx, sy, ex, ey;
|
||
if (Math.max(roomRight, roomLeft) >= 260) {
|
||
card.style.width = `${Math.min(540, Math.max(roomRight, roomLeft))}px`;
|
||
const cw = card.offsetWidth, ch = card.offsetHeight;
|
||
left = onRight ? b.right + gap : b.left - gap - cw;
|
||
top = Math.max(margin, Math.min(vh - margin - ch, by - ch / 2));
|
||
sx = onRight ? b.right + 8 : b.left - 8; sy = by;
|
||
ex = onRight ? left : left + cw;
|
||
ey = Math.max(top + 24, Math.min(by, top + ch - 24));
|
||
} else {
|
||
// Narrow windows: use a vertical connector instead of covering the icon.
|
||
const below = vh - b.bottom >= b.top;
|
||
const verticalGap = 54;
|
||
card.style.maxHeight = `${Math.max(80, (below ? vh - b.bottom : b.top) - verticalGap - margin)}px`;
|
||
left = Math.max(margin, Math.min(vw - margin - card.offsetWidth, b.left + b.width / 2 - card.offsetWidth / 2));
|
||
top = below ? b.bottom + verticalGap : b.top - verticalGap - card.offsetHeight;
|
||
sx = b.left + b.width / 2; sy = below ? b.bottom + 8 : b.top - 8;
|
||
ex = Math.max(left + 24, Math.min(sx, left + card.offsetWidth - 24));
|
||
ey = below ? top : top + card.offsetHeight;
|
||
}
|
||
card.style.left = `${left}px`; card.style.top = `${top}px`;
|
||
line.setAttribute('x1', sx); line.setAttribute('y1', sy);
|
||
line.setAttribute('x2', ex); line.setAttribute('y2', ey);
|
||
dot.setAttribute('cx', sx); dot.setAttribute('cy', sy);
|
||
return;
|
||
}
|
||
const cw = card.offsetWidth, ch = card.offsetHeight;
|
||
const bx = b.left + b.width / 2, by = b.top + b.height / 2;
|
||
const gap = 52;
|
||
let left;
|
||
if (bx + gap + cw < vw - 12) left = bx + gap;
|
||
else if (bx - gap - cw > 12) left = bx - gap - cw;
|
||
else left = Math.max(12, Math.min(vw - cw - 12, bx - cw / 2));
|
||
const top = Math.max(12, Math.min(vh - ch - 12, by - ch / 2));
|
||
card.style.left = Math.round(left) + 'px';
|
||
card.style.top = Math.round(top) + 'px';
|
||
const ly = Math.max(top + 24, Math.min(by, top + ch - 24));
|
||
const ax = left < bx ? left + cw : left;
|
||
line.setAttribute('x1', bx); line.setAttribute('y1', by);
|
||
line.setAttribute('x2', ax); line.setAttribute('y2', ly);
|
||
dot.setAttribute('cx', bx); dot.setAttribute('cy', by);
|
||
};
|
||
const MISSING_CONTRIBS = {
|
||
ci: { kicker: 'The missing piece · 20 years ago', title: 'Clinical Intelligence — the history, made visible',
|
||
body: `<p class="lead">Twenty years of clinical history, with no way to show it.</p>
|
||
<ul class="ai-pop-list"><li><b>What it is</b> — dashboards on the Unit's activity and its patients: volumes, incidence, outcomes</li>
|
||
<li><b>Who needed it</b> — management, audit, clinical research</li>
|
||
<li><b>Why it never existed</b> — the data was in prose; building dashboards on text was impractical</li></ul>
|
||
<div class="roi"><b class="k">The gap</b>No fast answer to “how are we doing?” — every question meant a manual query.</div>` },
|
||
dwh: { kicker: 'The missing piece · 20 years ago', title: 'Datawarehouse — one analysis-ready home',
|
||
body: `<p class="lead">One home for every record the Unit produces: text, structured fields, genetics, ECG.</p>
|
||
<ul class="ai-pop-list"><li><b>What it is</b> — a single schema joining all the sources, designed for analysis</li>
|
||
<li><b>Why it never existed</b> — each system kept its own silo; joining them meant hand-stitching exports</li></ul>
|
||
<div class="roi"><b class="k">The gap</b>Every research question started with manual exports — and ended in fragile spreadsheets.</div>` },
|
||
ml: { kicker: 'The missing piece · 20 years ago', title: 'ML-ready data — cohorts for research',
|
||
body: `<p class="lead">Tables shaped for statistics and model training: clean, wide, reproducible.</p>
|
||
<ul class="ai-pop-list"><li><b>What it is</b> — one row per patient, features beside outcomes, rebuildable at will</li>
|
||
<li><b>Why it never existed</b> — building it by hand from free text and separate silos took weeks</li></ul>
|
||
<div class="roi"><b class="k">The gap</b>Predictive research on the Unit's data stayed out of reach.</div>` },
|
||
portal: { kicker: 'The missing piece · 20 years ago', title: 'Professional Portal — the 360° view',
|
||
body: `<p class="lead">The 360° view of the patient: Cardioref and genetics finally in one place.</p>
|
||
<ul class="ai-pop-list"><li><b>What it is</b> — the grown-up Omics Portal: management, care and research on the same platform</li>
|
||
<li><b>Why it never existed</b> — the idea was there; the integrated data to feed it was not</li></ul>
|
||
<div class="roi"><b class="k">The gap</b>A single door for patients and clinicians — promised, never delivered.</div>` },
|
||
};
|
||
const START_CONTRIBS = {
|
||
cardioref: {
|
||
kicker: 'The starting point · 20 years ago',
|
||
title: 'Cardioref — the daily workhorse',
|
||
body: `
|
||
<p class="lead">For twenty years, the database running the Arrhythmology Unit.</p>
|
||
<ul class="ai-pop-list">
|
||
<li><b>Used for</b> — patient records, the Unit's daily activity, planning and logging of everything that was done</li>
|
||
<li><b>Good at it</b> — as a management database, it did (and still does) its job</li>
|
||
<li><b>The catch</b> — most clinical information lives in free text, outside the structured fields</li>
|
||
</ul>
|
||
<div class="roi"><b class="k">Problem</b>Almost unusable for research and data extraction: the knowledge is there — locked in prose.</div>`,
|
||
},
|
||
genetic: {
|
||
kicker: 'The starting point · 20 years ago',
|
||
title: 'Genetic data — an island',
|
||
body: `
|
||
<p class="lead">A dedicated system collecting the genetic results of our patients.</p>
|
||
<ul class="ai-pop-list">
|
||
<li><b>Used for</b> — storing genetic tests, variants and lab reports</li>
|
||
<li><b>The catch</b> — its only output is Excel sheets, with no integration into Cardioref</li>
|
||
</ul>
|
||
<div class="roi"><b class="k">Problem</b>Genotypes on one side, clinical records on the other — no way to join them.</div>`,
|
||
},
|
||
omics: {
|
||
kicker: 'The starting point · 20 years ago',
|
||
title: 'Aritmolab Portal — born too early',
|
||
body: `
|
||
<p class="lead">A 360° patient-management system, still in its infancy.</p>
|
||
<ul class="ai-pop-list">
|
||
<li><b>Used for</b> — the first patient-management features beyond the hospital electronic health record</li>
|
||
<li><b>The catch</b> — embryonic functionality, and no integration with Cardioref</li>
|
||
</ul>
|
||
<div class="roi"><b class="k">Problem</b>A promising idea without a bridge: it could not see the clinical history stored elsewhere.</div>`,
|
||
},
|
||
ecg: {
|
||
kicker: 'The starting point · 20 years ago',
|
||
title: 'ECG subsystems — one per device',
|
||
body: `
|
||
<p class="lead">A constellation of ECG systems, each with its own software and its own silo.</p>
|
||
<ul class="ai-pop-list">
|
||
<li><b>Used for</b> — acquiring and storing ECG signals</li>
|
||
<li><b>The catch</b> — no machine-to-machine communication: the only export is a CSV file, on request</li>
|
||
</ul>
|
||
<div class="roi"><b class="k">Problem</b>Every ECG meant a manual download — automation was impossible by design.</div>`,
|
||
},
|
||
};
|
||
const openPop = (key, brainEl, dict) => {
|
||
const c = (dict || AI_CONTRIBS)[key];
|
||
if (!c) return;
|
||
SlideLens.close();
|
||
ScreenshotTour.close();
|
||
const isAI = (dict || AI_CONTRIBS) === AI_CONTRIBS;
|
||
pop.classList.toggle('is-ai', isAI);
|
||
document.querySelectorAll('.brain-btn[aria-pressed="true"]').forEach(button => button.setAttribute('aria-pressed', 'false'));
|
||
if (isAI && brainEl) brainEl.setAttribute('aria-pressed', 'true');
|
||
document.getElementById('aiPopBrain').style.display = isAI ? 'flex' : 'none';
|
||
document.getElementById('aiPopKicker').textContent = c.kicker || 'AI contribution';
|
||
document.getElementById('aiPopTitle').textContent = c.title;
|
||
document.getElementById('aiPopBody').innerHTML = c.body;
|
||
pop.classList.add('open');
|
||
if (brainEl) placeNear(brainEl);
|
||
activePopup = { kind: isAI ? 'ai' : dict === START_CONTRIBS ? 'start' : 'missing', key };
|
||
window.dispatchEvent(new Event('presentationchange'));
|
||
};
|
||
document.addEventListener('click', (e) => {
|
||
const screenshot = e.target.closest('[data-screenshot]');
|
||
if (screenshot) {
|
||
closePop();
|
||
ScreenshotTour.open(screenshot);
|
||
return;
|
||
}
|
||
const lens = e.target.closest('[data-lens]');
|
||
if (lens) {
|
||
const wasOpen = SlideLens.active?.key === lens.dataset.lens;
|
||
closePop();
|
||
if (!wasOpen) SlideLens.open(lens);
|
||
return;
|
||
}
|
||
if (e.target.closest('.slide-lens-close')) { closePop(); return; }
|
||
const b = e.target.closest('.brain-btn[data-ai]');
|
||
if (b) { openPop(b.dataset.ai, b, AI_CONTRIBS); return; }
|
||
const s = e.target.closest('.isle[data-start]');
|
||
if (s) { openPop(s.dataset.start, s, START_CONTRIBS); return; }
|
||
const m = e.target.closest('.todo[data-missing]');
|
||
if (m) { openPop(m.dataset.missing, m, MISSING_CONTRIBS); return; }
|
||
if (e.target.closest('.ai-pop-x') || e.target === pop) closePop();
|
||
if (SlideLens.active && !e.target.closest('.slide-lens')) closePop();
|
||
});
|
||
// Escape closes the popup first (capture), without toggling Reveal's overview
|
||
document.addEventListener('keydown', (e) => {
|
||
if (e.key === 'Escape' && (pop.classList.contains('open') || SlideLens.active || ScreenshotTour.active)) {
|
||
closePop(); e.preventDefault(); e.stopPropagation();
|
||
}
|
||
const block = e.target.closest('button[data-lens], button[data-screenshot]');
|
||
if (block && (e.key === 'Enter' || e.key === ' ')) {
|
||
e.preventDefault(); e.stopPropagation();
|
||
if (!e.repeat) block.click();
|
||
return;
|
||
}
|
||
if (e.repeat || e.altKey || e.ctrlKey || e.metaKey || e.target.closest('input, select, textarea, [contenteditable="true"]')) return;
|
||
const lens = [...(Reveal.getCurrentSlide()?.querySelectorAll('[data-lens-number], [data-screenshot-number]') || [])]
|
||
.find(button => (button.dataset.lensNumber || button.dataset.screenshotNumber) === e.key);
|
||
if (lens) { lens.click(); e.preventDefault(); e.stopPropagation(); }
|
||
}, true);
|
||
|
||
// Small shared interface for the presenter console and its synchronized audience view.
|
||
const popupSelector = '[data-lens], [data-screenshot], .brain-btn[data-ai], .isle[data-start], .todo[data-missing]';
|
||
const popupIdentity = el => {
|
||
const kind = ['lens', 'screenshot', 'ai', 'start', 'missing'].find(k => el.dataset[k]);
|
||
return { kind, key: el.dataset[kind] };
|
||
};
|
||
const popupDictionary = kind => ({ ai: AI_CONTRIBS, start: START_CONTRIBS, missing: MISSING_CONTRIBS })[kind];
|
||
window.PresentationControls = {
|
||
snapshot: () => ({ index: Reveal.getIndices().h, popup: ScreenshotTour.active || SlideLens.active || activePopup }),
|
||
slides: () => Reveal.getSlides().map(slide => ({
|
||
title: slide.querySelector('h1, h2')?.textContent.trim() || '',
|
||
notes: slide.querySelector('aside.notes')?.innerHTML || '',
|
||
popupNotes: [...slide.querySelectorAll('aside.notes [data-popup-notes]')].map(note => ({
|
||
key: note.dataset.popupNotes,
|
||
title: note.querySelector('h3').textContent,
|
||
html: note.querySelector('[data-popup-notes-body]').innerHTML,
|
||
})),
|
||
popups: [...slide.querySelectorAll(popupSelector)].map(el => {
|
||
const id = popupIdentity(el);
|
||
if (id.kind === 'screenshot') return { ...id, title: el.dataset.screenshotTitle };
|
||
const source = id.kind === 'lens' ? document.getElementById(id.key) : null;
|
||
return { ...id, title: source ? source.dataset.lensTitle || source.querySelector('b').textContent : popupDictionary(id.kind)[id.key].title };
|
||
}),
|
||
})),
|
||
openPopup: id => {
|
||
const el = [...Reveal.getCurrentSlide().querySelectorAll(popupSelector)]
|
||
.find(el => { const candidate = popupIdentity(el); return candidate.kind === id.kind && candidate.key === id.key; });
|
||
if (el && id.kind === 'screenshot') {
|
||
if (ScreenshotTour.active?.key !== id.key) { closePop(); ScreenshotTour.open(el); }
|
||
} else if (el && id.kind === 'lens') {
|
||
if (SlideLens.active?.key !== id.key) { closePop(); SlideLens.open(el); }
|
||
else SlideLens.layout();
|
||
} else if (el) openPop(id.key, el, popupDictionary(id.kind));
|
||
},
|
||
closePopup: closePop,
|
||
};
|
||
// Open the first tour stop only on slide entry, never on hover or focus.
|
||
// The thumbnail stays unobstructed; synchronized state can still close or replace the popup.
|
||
const openInitialScreenshot = () => {
|
||
if (new URLSearchParams(location.search).get('view') === 'thumbnail') return;
|
||
const first = Reveal.getCurrentSlide()?.querySelector('[data-screenshot-auto-open] [data-screenshot-number="1"]');
|
||
if (first) window.PresentationControls.openPopup(popupIdentity(first));
|
||
};
|
||
Reveal.on('slidechanged', openInitialScreenshot);
|
||
if (Reveal.isReady()) openInitialScreenshot();
|
||
else Reveal.on('ready', openInitialScreenshot);
|
||
addEventListener('resize', () => {
|
||
if (activePopup) window.PresentationControls.openPopup(activePopup);
|
||
});
|
||
</script>
|
||
<script src="presenter-bridge.js"></script>
|
||
</body>
|
||
</html>
|