← Back to work
Population Health

AHD Risk Score: ML-Driven Patient Risk Stratification

A machine learning risk stratification model for identifying high-acuity patients in value-based care programs, enabling proactive clinical interventions and reducing avoidable hospitalizations.

Role

Product Lead

Domain

Population Health / ML

Stack

XGBoost, Claims Data, EHR

Setting

Medicare Advantage VBC

Problem

In value-based care models, proactively identifying patients at highest risk of hospitalization or ED utilization is critical to both outcomes and financial performance. Existing risk adjustment models (HCC-based) were designed for reimbursement, not clinical action — they told you who was expensive, not who was about to get sicker. Care teams needed a forward-looking risk signal they could act on in their daily workflows.

Approach

Led the product definition and clinical validation of a predictive risk model that combined claims history, EHR clinical data, and social determinant indicators to generate actionable acuity scores for the care team.

  • Feature engineering — combined 18 months of claims data (diagnoses, utilization patterns, medication adherence proxies) with EHR vitals, lab trends, and SDoH flags from community health assessments
  • Model architecture — gradient-boosted ensemble (XGBoost) trained on 30-day hospitalization as the primary outcome, with secondary endpoints for ED visits and skilled nursing transitions
  • Explainability layer — SHAP values surfaced top contributing factors for each patient's score, enabling care managers to understand and trust the model's reasoning
  • Workflow integration — risk scores delivered via daily care manager worklist, stratified into tiers (high / rising / stable) with recommended outreach actions

Results

Deployed across Medicare Advantage panels in a value-based primary care organization, covering thousands of attributed lives.

0.82
AUC-ROC for 30-day hospitalization
18%
Reduction in avoidable admissions
2.4x
Proactive outreach increase

Lessons

Predictive accuracy alone doesn't drive outcomes — the model has to change behavior. The biggest unlock was designing the output not as a score, but as a prioritized worklist with clear next actions. Care managers didn't need to know the math; they needed to know who to call today and why. Investing in the explainability layer and co-designing the worklist UX with care managers was the difference between a model that sat on a dashboard and one that actually reduced hospitalizations.