PRISM — Polysubstance Risk & Interaction Surveillance for Mental-health prescribing

Phase 3 of the PRISM platform.

Clinical decision support checks drug–drug interactions. It largely ignores drug–substance and substance–substance risk — which is where overdose mortality actually comes from (opioid + benzodiazepine, opioid + alcohol).

PRISM is a reasoning layer that closes that gap and grounds an LLM in retrieved evidence rather than recall. It ships as an MCP server with 47 tools and a web platform, both driven by the same validated cascade.

Not a medical device. PRISM surfaces evidence with its provenance. It does not give clinical advice, and the validation harness asserts that it never crosses that line.

---

The 7-level evidence cascade

Level	Source	Offline?
1	Curated interaction knowledge base	yes
2	OpenFDA drug labeling	live
3	CYP450 / transporter kinetics	yes
4	Pharmacodynamic stacking (e.g. CNS depression)	yes
5	Drug-class combinations	yes
6	FAERS disproportionality signals (Phase 2)	yes
7	External interaction sources	live

Severity is the maximum across levels; confidence scales with the number of independent levels that fire (1 → low, 4+ → very_high). The model cannot invent an interaction — it can only report what a level returned, tagged with that level's source.

On top of the per-pair checks sits compound-risk synthesis: triple CNS-depressant stacks, naloxone candidacy, cocaethylene formation, Beers criteria, teratogenicity — plus automatic screening triggers (AUDIT, DAST-10, CUDIT-R, GAD-7, PHQ-9).

Proxy matching means street terms resolve to pharmacology: heroin → opioid class, street benzos → benzodiazepine class.

Population prior

Phase 1's NSDUH model is imported as a national baseline, so the system can answer "how risky is this patient's profile relative to the population?" rather than only "do these two drugs interact?" — 1,188 risk profiles, 60 annual trend rows, and the logistic coefficients for real-time scoring.

Validation

14 clinical scenarios run end-to-end through the live server and PostgreSQL:

• 2 cases flagged life_threatening; 0 instances of overstepping into advice
• Case 14 regression (benzodiazepine + opioid) returns life_threatening across levels 1, 4, 5, 6, 7 with fda_black_box=True — it returned nothing before the cascade was built, and it is the reason the cascade exists
• Audit rows persist to cascade_results, compound_risk_alerts, and screening_queue

The public repo ships a synthetic 14-case set that reproduces identical severities for all 14 cases. See scenarios/README.md.

Layout

src/
  server.py         FastMCP server, 47 registered tools
  cascade.py        the 7-level engine (DB-optional, unit-testable)
  cascade_tools.py  DB-backed cascade tools
  extra_tools.py    remaining registry tools
  live_apis.py      OpenFDA (L2) + RxNorm (L1) hooks
  api/app.py        FastAPI bridge for the web UI
  db/
    schema.sql              24 base tables
    migrations_path3.sql    +5 cascade tables (29 total)
    seed.py                 curated KB (+ private case loader)
    load_scenarios.py       scenario loader (synthetic or private)
    import_path3_data.py    imports Phase 1 + Phase 2 exports
web/                Next.js 14 app (8 pages, Tailwind, Recharts)
scenarios/          synthetic validation cases
validate_scenarios.py

Setup

python -m venv venv && ./venv/bin/pip install -r requirements.txt
createdb prism_db
psql prism_db -f src/db/schema.sql
psql prism_db -f src/db/migrations_path3.sql

python -m src.db.seed --reference-only      # curated KB, no patient data
python -m src.db.load_scenarios --replace   # 14 synthetic cases
python src/db/import_path3_data.py          # Phase 1 + Phase 2 exports
python validate_scenarios.py

import_path3_data.py reads the sibling NSDUH/ and FAERS/ checkouts by default; override with NSDUH_RESULTS and FAERS_RESULTS environment variables.

Running the server

python -m src.server           # stdio  (Claude Desktop, Claude Code)
python -m src.server --http    # HTTP   on :8000

Register with an MCP host:

{ "mcpServers": {
    "prism": {
      "command": "/path/PRISM/venv/bin/python",
      "args": ["-m", "src.server"],
      "cwd": "/path/PRISM",
      "env": { "DATABASE_URL": "postgresql://localhost/prism_db" }
}}}

License

MIT (code). Clinical scenario source material is not distributed. Not a medical device; not reviewed by any regulatory authority; must not be used to direct patient care.