# Outline v4 — Role of AI in the Analysis of Unstructured Clinical Databases

Updated 2026-09-18 against the actual deck, rather than the earlier proposed structure.

Deck: [`../deck/index.html`](../deck/index.html) — Reveal.js 5.1.0, **14 slides**, 16:9
(1280 × 720), with local fonts and speaker notes on every slide.
Review status: [`../prototype/deck-overview.html`](../prototype/deck-overview.html).
The titles below are the actual slide headings; some overview labels are shorter or older.

Presenters: **MP** = Dr. Marco Pancotti · **SP** = Dr.ssa Sara Paratico.
The notes explicitly identify Sara on slides 05–08. The final handoff around slide 04
and the remaining speaker allocation should be confirmed during rehearsal.
Per-slide timings and total running time need a new rehearsal; the previous 15-slide
timing estimate no longer describes this deck.

## Actual slide sequence

| # | Actual slide title | Content / format | Slide review status | Speaker notes |
|---|---|---|---|---|
| 01 | Role of AI in the Analysis of Unstructured Clinical Databases | AritmoLab introduction, clinical data platform and presenters | FINAL | Present; no DRAFT marker |
| 02 | Four islands and four missing pieces | Interactive map: Cardioref, Genetic data, Omics Portal and ECG subsystems; Clinical Intelligence, Datawarehouse, ML-ready data and Professional Portal as missing pieces | FINAL | Present; no DRAFT marker |
| 03 | What we wanted to build | Architecture and data flow toward the patient portal and research warehouse; clickable AI contribution cards | FINAL | Present; no DRAFT marker |
| 04 | The clinical truth lives in unstructured columns | Italian clinical sentence with English translation; why free text needs structured interpretation | FINAL | DRAFT |
| 05 | Deterministic AI, audited numbers | Text-mining outputs: 58,438 letters, 73,389 pathologies, 10,908 tests and 2,307 Brugada patients; extraction pipeline | FINAL | DRAFT — Sara |
| 06 | Reading clinical text is not keyword matching | Clinical NLP challenges, including negation and distinguishing family history from the patient's conditions | FINAL | DRAFT — Sara |
| 07 | Two tiers, one clinical order | Clinical ontology: 11 arrhythmic and 5 structural categories; text becomes recorded data | FINAL | DRAFT — Sara |
| 08 | Trusting the text is a process, not a promise | Five centered detail lenses: four validation controls and the three advantages | REFINED | ~120 words; five hover cues |
| 09 | AritmoLab — a quick tour | Portal tour; currently six screenshot placeholders | DRAFT | DRAFT |
| 10 | The warehouse speaks SQL. Research needs more. | Gap between the warehouse and clinical/research use: SQL access, aggregates and research cohorts | DRAFT | DRAFT |
| 11 | Ask the warehouse in plain English | ThothII: natural-language question, SQL proposal, human review and a research-ready datamart | DRAFT | DRAFT |
| 12 | From question to datamart — the app | ThothII walkthrough; currently six screenshot placeholders | DRAFT | DRAFT |
| 13 | One datamart — many questions answered | Multivariate analysis and machine learning; two chart placeholders with an illustrative-data disclosure | DRAFT | DRAFT — charts still to be created |
| 14 | Thank you | Questions and invitation to future technical walkthroughs on Substack | DRAFT | DRAFT |

`FINAL` and `REFINED` reproduce the overview's recorded review status. They do not imply
that speaker notes or clinical claims have received a separate final validation.
Figures above describe the content currently displayed in the deck.

## Current presentation behavior

- Slide 02 has four clickable source systems and four clickable missing pieces,
  each opening an explanatory popup.
- Slide 03 has clickable brain icons that open AI contribution cards with a connector
  to the selected icon.
- Slide 08 has four centered lens popups. In the presenter console, hovering over
  a numbered control or its corresponding heading in the notes opens the detail
  in both preview and audience view. Escape closes it. Its quality wording separates
  the original procedure-accuracy target from pathology coverage. The ~893
  false-positive records and v1.3.1 negation fix are confirmed in the ETL sources;
  see [source notes](08-validation-sources.md) for precise scope and limitations.
- Its advantages panel highlights on hover and opens a fifth centered lens on
  click, explaining speed, precision and determinism. The fifth speaker-note cue
  and presenter control open the same popup.
- Slide changes use a fade transition. Popups fade in and their cards move/scale into
  place; clickable source icons and brain buttons have hover effects.
- A dropdown on each slide jumps to another slide. Reveal.js provides keyboard
  navigation and the overview.
- The `S` key opens the Reveal.js speaker view when served locally. Speaker scripts
  live in each slide's `<aside class="notes">` in `deck/index.html`.
- Source Sans 3 and Source Serif 4 are bundled under `deck/fonts/`; the deck does not
  depend on Google Fonts.

## Remaining work

- Review slide 08 for final approval and slides 09–14 individually.
- Replace the six placeholders on slide 09 with real anonymized or demo portal
  screenshots. The in-slide authoring note still requests 6–7 screenshots; settle
  that count when inserting the images.
- Replace the six placeholders on slide 12 with screenshots of the ThothII workflow.
  Its authoring note likewise still requests 6–7 screenshots.
- Create the two charts on slide 13, identify the supporting literature, and preserve
  the disclosure if the data remain illustrative rather than measured results.
- Validate the speaker scripts on slides 04–14, including Sara's clinical section
  on slides 05–08, and confirm the handoff between presenters.
- Confirm the “twenty years” wording, clinical figures, and quality-gate claims before
  presenting them as final facts.
- Confirm the closing Substack reference and whether the promised videos will be
  available by the talk.
- Rehearse the actual 14-slide sequence and set timings, including popup explanations
  and any live demo chosen in addition to the screenshot walkthroughs.

The previous standalone slides on AI co-engineering, genetic reconciliation, lessons
learned and a live-demo divider are not separate slides in the current sequence.
Their old slide numbers, review tasks and timings should not be reused.

## Local preview

From the repository root:

```bash
python3 -m http.server 8000 --bind 127.0.0.1 --directory presentation
```

- Deck: <http://localhost:8000/deck/index.html>
- Review overview: <http://localhost:8000/prototype/deck-overview.html>
- Presenter console: <http://localhost:8000/presenter/> — notes, interactive preview,
  popup controls and a separate audience window for an extended HDMI display.
  See [`../presenter/README.md`](../presenter/README.md) for operation and scope.

Keep `presentation/` as the server root: the overview uses absolute `/deck/` URLs.
Stop the server with Ctrl+C.