PRISM
Status: concept — the front door of the PRISM documentation.
PRISM (Predictive Recommendations for Improved Screening in Medicine) turns anonymized insurance-claims histories into early-screening suggestions. This page is the map of the documentation; the recommended first read is What is PRISM?.
PRISM in one breath
Each patient's history is a six-column markdown table of standardized billing codes; a model's only job is to continue the table with the next plausible rows. A screening suggestion "fires" when that continuation contains a diagnostic TEST code. An ensemble of independently trained models — one per disjoint data pool — each continues the same patient and votes; the signal is how many independently agree, never any single model. PRISM is constructive-only (it can only suggest a test, never deny care), is paid only when a suggestion leads to a documented early detection, and operates as a Public Benefit Corporation.
The evidence backbone of this documentation is the 2026 synthetic proof of concept: a deliberately artificial experiment that steered a five-model ensemble to surface a fabricated condition's screening test earlier than its training data did. It succeeded, and its results are quoted alongside an equally plain account of what they do not prove. Production-scale material appears throughout, always marked as vision, never as built.
How to read the status lines
Every article opens with a status line that says how much weight to put on it.
| tag | meaning |
|---|---|
concept | durable design principle of PRISM |
demonstrated | validated end-to-end in the 2026 synthetic prototype |
vision | production design intent; not yet built |
history | record of past work; superseded, kept for provenance |
Tags may be compound ("concept, demonstrated"). For audience-specific paths through the articles, see the reading guide.
Section map
| section | what it covers | where to start |
|---|---|---|
| Start | orientation: the one-page overview, reading paths, glossary, and the organization | What is PRISM? |
| Problem | the screening gap, and why only insurers see the whole patient journey | The screening gap |
| Data | the six-column timeline format, its standardized codesets, and anonymity by architecture | The patient timeline format |
| Method | sequence completion, pools and consensus, the two training rounds, evaluation without a hold-out | Sequence completion |
| Prototype | the 2026 synthetic proof of concept end-to-end, with results and worked examples | Prototype overview |
| Safeguards | the constructive-only architecture and the clinical decision support posture | Constructive-only |
| Business | results-based compensation, zero integration, and the Public Benefit Corporation | Results-based compensation |
| Vision | production scale, continuous screening and retraining, open collaboration | Production scale |
| History | the 2024 seed, the 2025 first prototype, and the full timeline | Timeline |
Where to begin
If you read one article, read What is PRISM? — the whole system on a page, from problem to method to prototype to what comes next. If you know your role — physician, insurer, engineer, researcher — the reading guide lays out a path for each audience. And if a term looks unfamiliar, the glossary defines every one and links to the article that owns it.