A Serverless Pharmacogenomic Risk Dashboard: Translating Ensemble Models and Model-Based Scenario Rules to Clinical...

Area of Science:

• Computational Biology
• Bioinformatics
• Health Informatics

Background:

• The
• last mile
• challenge persists in translating healthcare AI models, particularly pharmacogenomic (PGx) risk assessment, into practical point-of-care tools.
• Existing platforms lack integration of opioid and polypharmacy risk scoring, scenario analysis, and patient-specific PGx information within a privacy-preserving architecture.

Purpose of the Study:

• To design, develop, and implement a novel serverless architecture, the PGx Risk Dashboard, for privacy-first pharmacogenomic risk assessment.
• To integrate opioid and polypharmacy risk scoring, model-based scenario analysis, and Clinical Pharmacogenetics Implementation Consortium (CPIC)-based patient cards.

Main Methods:

• Developed a serverless system using an S3-hosted static frontend and AWS Lambda containerized backends.
• Ensembled 56 bin-level models (CatBoost, XGBoost, XGBoost-RF) within AWS Lambda containers, utilizing aggregate fallback artifacts for sparse data.
• Implemented partial-input imputation and stateless CPIC lookups for gene-drug logic verification, ensuring no Personally Identifiable Information (PII) is stored.

Main Results:

• Achieved low-latency inference: mean cold-start 2100ms (SD 250ms), mean warm inference 6ms (SD 1ms), meeting the sub-100ms target.
• Demonstrated prediction stability under sparse input conditions (mean |Δp̂| = 0.10 vs. full-feature baseline).
• Successfully integrated PGx risk scoring, scenario analysis, and CPIC-based patient cards without requiring Electronic Health Record (EHR) integration.

Conclusions:

• The PGx Risk Dashboard proves the technical feasibility of low-latency, privacy-first PGx risk decision support.
• The system facilitates safer clinical use by maintaining the association-versus-causation distinction.
• This work establishes a foundation for prospective causal-effect validation of PGx interventions.