A Serverless Pharmacogenomic Risk Dashboard: Translating Ensemble Models and Model‐Based Scenario Rules to Clinical Decision Support

ABSTRACT

The “last mile” problem in healthcare AI—translating high‐performance models into accessible, privacy‐preserving point‐of‐care tools—remains unsolved for pharmacogenomic (PGx) risk assessment. No existing platform integrates opioid and polypharmacy risk scoring, model‐based scenario analysis, and CPIC‐based PGx patient cards within a single privacy‐first serverless architecture. We designed, developed, and implemented the PGx Risk Dashboard: a serverless system combining an S3‐hosted static frontend with AWS Lambda containerized backends. Fifty‐six bin‐level ensembles (2 cohorts × 7 age bands × 4 density bins), each comprising CatBoost/XGBoost/XGBoost‐RF base learners, are packaged within an AWS Lambda container, with aggregate fallback artifacts used for sparse cells. Partial‐input imputation using training‐set medians handles real‐world data sparsity; the PGx Patient Card executes stateless CPIC lookups for 573 deterministic gene‐drug logic‐verification cases without storing any PII. Lambda cold‐start latency was mean 2100 ms (SD 250 ms); warm inference latency was mean 6 ms (SD 1 ms), meeting the sub‐100 ms target. Prediction stability was maintained under sparse inputs (≤ 5 features provided): mean |Δ p̂ | = 0.10 vs. full‐feature baseline. The PGx Risk Dashboard demonstrates technical feasibility for low‐latency, privacy‐first PGx risk decision support without EHR integration. We maintain the association‐versus‐causation distinction to support safer clinical use, avoid overstating treatment‐effect evidence from observational outputs, and preserve a clear path to prospective causal‐effect validation as next steps.