Drug Interaction Misses & Contraindication Blindness

Issue: Model Recommends Drug Combination Without Flagging Known Dangerous Interactions

Frequency: Common

Symptoms

Model recommends two drugs known to interact dangerously
Interaction database not consulted or outdated
New drug interactions discovered post-deployment
Patient takes dangerous drug combo; adverse event

Root Cause Drug interactions are discrete rules (Drug A + Drug B → risk level X). Models sometimes learn statistical patterns but miss logical rules. Interaction database must be comprehensive and up-to-date. Missing one rare but dangerous interaction is a patient safety issue.

Example

Scenario: Medication recommender system
Patient: On warfarin (blood thinner)
Model recommends: Ibuprofen (analgesic) for pain
Known interaction: Warfarin + NSAID → Increased bleeding risk
Model: Doesn't flag (interaction not in training data or database outdated)
Patient: Takes both; develops GI bleeding
Impact: Life-threatening adverse event; liability

Key Statistics

Known interactions: 30,000+ documented
Coverage of interactions in training: Often <50%
New interactions discovered: 100-200 per year

Mitigations

Prevention

Authoritative drug interaction database with schema enforcement: Implement a canonical drug interaction knowledge base (DrugBank, Micromedex, FDA-maintained) with required quarterly updates. Before recommendation generation, apply constraint satisfaction validation: for each recommended drug combination, query interaction matrix to verify no flags exist. Fail-safe: if database lookup fails, default to “cannot recommend” rather than silent pass. Root cause mitigation: Prevents reliance on statistical patterns by enforcing hard rules from verified sources.
Multi-layer interaction detection with fallback verification: Deploy layered checking: (1) Direct database lookup (exact drug codes), (2) Semantic similarity fallback for brand-name variants, (3) Drug class-level contraindication rules (NSAIDs + anticoagulants). Use PharmGKB for genetic-interaction rules when patient pharmacogenomics available. Root cause: Catches interactions missed by single-point lookup.
Pharmacovigilance pipeline integration: Integrate post-market adverse event feeds (FDA MedWatch, EudraVigilance) with monthly batch processing to update interaction database. Flag new drug interactions discovered in past 6 months as “alert-only” recommendations requiring explicit pharmacist override. Root cause: Captures newly discovered interactions before statistical patterns emerge.

Detection & Response

Drug-drug interaction audit logging: For every medication recommendation, log: (a) recommended drug, (b) current medications, (c) interaction database lookup result, (d) whether interaction was flagged, (e) recommendation decision. Implement real-time alerting on “missed interactions”: when patient reports adverse event, retroactively check if known interaction existed at recommendation time. Target: 100% detection rate for interactions in database.
Pharmacovigilance signal detection: Monitor post-recommendation adverse events by drug combination. Track “adverse event rate per combination” metric. Alert when combination shows >2 adverse events in 30-day window (indicates potential unknown interaction). Implement collaborative filtering across patient population to detect emerging patterns not yet in formal database.

Architecture Patterns

Interaction Rule Engine: Centralized rule database (tables: drug-codes, interaction-rules, severity-levels, alternatives). Before recommendation, query engine returns interaction severity and suggested alternatives. Engine backed by PharmGKB and FDA sources. Updates run on fixed schedule (weekly for patches, monthly for major).
Pharmacovigilance Feedback Loop: MedWatch events feed → signal detection (clustering on drug combinations) → escalation to pharmacy committee → database update → re-scoring of past recommendations in audit log.
Constraint Satisfaction Layer: Hard-coded rules for absolute contraindications (e.g., “any combination with warfarin + NSAID → BLOCK”). Allows only whitelisted combinations or those with explicit pharmacist review.

Key Metrics

Metric	Target	Alert Threshold	Measurement Method
Interaction Detection Rate	100%	<99%	# of known interactions correctly flagged / total recommendations with potential interactions
False Negative Rate	<0.1%	>0.5%	# of missed interactions / total interaction opportunities
Database Freshness	<7 days	>14 days	Time since last update to interaction database
Post-Recommendation Adverse Event Rate	<0.01%	>0.05%	# of adverse events within 7 days of recommendation / total recommendations
Pharmacist Override Rate	<1%	>5%	# of interactions requiring pharmacist approval / total flagged interactions

Alerts & Escalation

Alert	Condition	Severity	Response
Known Interaction Not Flagged	Drug combination present in database but not flagged in recommendation (audit discovers post-event)	CRITICAL	Immediate investigation; review recommendation logic; potential patient harm assessment; escalate to clinical leadership
Adverse Event Post-Recommendation	Patient reports adverse event within 7 days of drug recommendation; audit finds known interaction existed	CRITICAL	Halt similar recommendations pending review; notify prescriber; initiate pharmacovigilance case
Emerging Signal Detected	>2 adverse events for same drug combination in 30-day window not previously flagged as dangerous	HIGH	Escalate to pharmacy committee; flag combination as “signal under investigation”; require pharmacist review for future recommendations