Contraindication Omission

Issue: Agent Recommends or Approves a Medication Without Checking Patient-Specific Contraindications Beyond Drug-Drug Interactions

Frequency: Common

Symptoms

• Drug-drug interaction checks pass, but the model misses condition-based contraindications (e.g., beta-blockers in decompensated asthma, NSAIDs in renal failure)
• Allergy history present in the chart but not cross-checked against the recommended drug class
• Contraindications tied to lab values (e.g., QT-prolonging drugs with a known prolonged QTc) are not flagged because the agent only checks the medication list, not the full chart
• Pregnancy/lactation contraindications missed when the model’s interaction check is scoped only to other medications

Root Cause Most automated interaction-checking focuses narrowly on pairwise drug-drug interactions because that is the most structured and widely available dataset (RxNorm/DDI databases). Condition-based, lab-based, and demographic-based contraindications require integrating across the medication list, problem list, lab results, and patient demographics simultaneously — a broader reasoning task that narrow DDI-checking pipelines do not perform, and that general-purpose LLM agents may skip unless explicitly prompted to check each category.

Example

Scenario: Patient with chart-documented severe asthma and chronic renal insufficiency
New prescription: Non-selective beta-blocker for hypertension + standing NSAID for pain
Drug-drug interaction check: Passes (no interaction between the two new drugs)
Missed contraindications: Beta-blocker contraindicated in severe asthma; NSAID contraindicated in renal insufficiency
Impact: Risk of bronchospasm and further renal decline; both preventable with chart-aware contraindication checking

Key Statistics

• Condition-based contraindications are missed at a meaningfully higher rate than drug-drug interactions in systems that scope checking to medication-pair lookups only
• A substantial share of preventable adverse drug events in hospitalized patients stem from contraindications relative to a pre-existing condition, not drug-drug interactions
• Integrating problem-list and lab-value data into contraindication checking has been shown to materially increase detection of clinically significant contraindications versus medication-list-only checks

Mitigation Strategies

Prevention

1. Mandatory multi-source contraindication matrix validation: Implement a structured pre-prescription gate that requires explicit checking across five dimensions before recommendation: (a) medication-interaction matrix (DDI), (b) condition-contraindication lookup (patient problems + drug class), (c) lab-value range validation (renal function, QTc, LFTs), (d) allergy/intolerance database cross-check, (e) demographic rules (pregnancy, pediatric, geriatric). Fail-safe: if any dimension incomplete, return “cannot recommend - missing data” rather than proceed. Root cause mitigation: Prevents narrow DDI-only checking by enforcing breadth of constraint checking.

2. Typed contraindication database with inheritance rules: Build contraindication knowledge base indexed by: drug-class (not just single drug), condition (hierarchical: asthma → reactive-airway-disease), lab-ranges (eGFR <30, QTc >450ms), and demographic flags. Use RxNorm drug hierarchy to catch class-level contraindications. Root cause: Catches contraindications that would be missed by single-drug-code lookup.

3. Chart-aware recommendation workflow with decision audit trail: Modify recommendation generation to output structured justification: “Checked: [medication-list: OK], [problem-list: [asthma flagged]], [labs: [eGFR 25]], [allergies: OK], [demographics: OK]. Gate check result: [BLOCKED - beta-blocker contraindicated in asthma]”. Require this audit trail visible to prescriber before approval. Root cause: Explicit enumeration prevents omission of key data sources.

Detection & Response

1. Contraindication detection stratified by source: Instrument prescription recommendations to track: (a) % requiring DDI-only check (medication-list lookup), (b) % requiring condition-based check (problem-list), (c) % requiring lab-based check (vital-signs/labs). Alert when condition-based or lab-based checks are systematically skipped. Target: 100% of recommendations include all 5 dimensions checked.

2. Adverse event root-cause attribution: Post-adverse-event analysis: for each adverse event, retroactively check if known contraindication existed at recommendation time. Classify events by contraindication type (DDI, condition-based, lab-based). Alert when >2 events in 30 days share same missed contraindication type.

Architecture Patterns

1. Multi-Layer Contraindication Engine: Centralized service with five rule layers: (1) PharmGKB DDI database, (2) Mayo-Clinic condition-contraindication taxonomy, (3) Lab-value rule engine (eGFR thresholds, QTc intervals), (4) Allergy database (SNOMED CT), (5) Demographic rules. Each layer returns [OK / WARN / BLOCK]. Recommendation blocked if any layer returns BLOCK.

2. Chart-Aware Recommendation Gate: Recommendation engine passes full chart context (medications, problems, labs, allergies, demographics) to contraindication service. Service returns structured decision with audit trail. Prescriber sees decision + rationale before approval.

3. Lab-Triggered Rules Engine: Automated rule generation from clinical guidelines: “IF eGFR < 30 THEN flag NSAIDs, ACE-inhibitors, certain antibiotics”. Rules updated quarterly from current guidelines (UpToDate, KDIGO, etc.).

Key Metrics

Alerts & Escalation

References

• Multi-model assurance analysis showing large language models are highly vulnerable to adversarial hallucination attacks during clinical decision support
• A Comprehensive Survey on the Trustworthiness of Large Language Models in Healthcare