Published: 2 February 2025

RESEARCH PAPER COMING SOON

In this project, we designed a RAG framework tethered to a database summarising drug’s interactions to create a system capable of alerting clinicians. Globally, the World Health Organization (WHO) estimates that over 3 million deaths annually are due to unsafe care, and unsafe care is a leading cause of death. In the United States, adverse drug events (ADEs) are estimated to cause around 250,000 deaths each year and are the third leading cause of death.

Building a RAG framework for drug–drug interaction alerts

Designing a retrieval-augmented system that reads prescriptions, checks trusted interaction data, and surfaces clear alerts for polymedicated patients.

Clinical note: decision support only; not a substitute for clinical judgement or local policy.

Why it matters

Polypharmacy is common, and risk depends on combinations plus context (age, renal/hepatic function, pregnancy, comorbidities). Clinicians need concise, explainable alerts that appear in their workflow without adding to alert fatigue.

What it does

Ingests prescriptions (structured meds and key free text) and patient context from the EHR.
Retrieves evidence from a curated interaction corpus (severity, mechanism, management, citations), layered with local formulary policy.
Generates alerts: a short summary, mechanism and one actionable recommendation, all tied to sources.
Delivers UX as an inline banner with a “details” drawer; everything is audit-logged.

How it works (brief)

Normalise drug names to a canonical vocabulary (for example RxNorm/ATC); keep route, strength and form.
Enrich with patient factors and lightweight flags (for example “QT-risk present”).
Hybrid retrieval: lexical match for exact substances plus biomedical embeddings for class/mechanism language; pre-filter by route, dose and patient modifiers; score by match strength, source quality and recency.
Constrained generation: the model sees only retrieved passages and a compact patient summary; it must cite every claim and escalate uncertainty when evidence is thin.

Safety and governance

Confidence thresholds and contradiction checks; deterministic rules for absolute contraindications.
Clinical safety case; versioned corpus with reviews and scheduled refreshes.
Minimum necessary PHI; encryption; inference within the clinical boundary; no model training on patient data.
Live metrics: precision/recall against a labelled set, severity accuracy, citation correctness, override rates and equity across groups.

What success looks like

Grounded, context-aware alerts that clinicians can trust, fewer false positives, and a maintainable pathway to keep evidence and models up to date — piloted in silent mode alongside existing tools, then rolled out with measured impact.