Comorbidity Neglect in Treatment Recommendations

Issue: Model Recommends Treatment Optimal for Primary Condition But Dangerous for Comorbidities Patient Has

Frequency: Common

Symptoms

• Treatment recommended for Disease A is contraindicated in Disease B
• Model optimizes for primary condition; ignores secondary conditions
• Patient has multiple conditions; model sees only one
• Treatment plan unsafe when full medical history considered

Root Cause Models often trained on single-disease datasets or optimize for single target. Don’t learn interactions between diseases and treatments. Medical records may be siloed; full history not available to model. Multi-morbidity is common but under-represented in training data.

Example

Scenario: Treatment recommendation for diabetes
Patient: Type 2 diabetes + chronic kidney disease (CKD)
Model recommends: SGLT2 inhibitor (good for diabetes)
Reality: SGLT2i also helps CKD! This is actually beneficial.
Counter-example: Model recommends ACE inhibitor
Patient also has: Hyperkalemia (elevated potassium)
ACE inhibitor contraindicated (raises potassium)
Impact: Patient gets worse instead of better

Key Statistics

• Patients with comorbidities: >50% in typical elderly population
• Treatment contraindications for comorbidities: 20-40% of recommendations affected
• Adverse event rate due to comorbidity oversight: 5-15%

Mitigation Strategies

1. Multi-Disease Modeling: Train on multi-condition patients; optimize jointly
2. Full History Review: Ensure all diagnoses available to model
3. Contraindication Matrix: Explicit lookup of treatment vs. all patient conditions
4. Clinician Override: Flag when treatment has contraindication; require confirmation

Metrics

• Sensitivity to comorbidities (does model consider them?)
• Contraindication detection rate
• Treatment safety by comorbidity count

Alerts

• Treatment contraindicated by comorbidity → Warn

References

• Multimorbidity in Medical AI
• Treatment Optimization with Comorbidities