TL;DR
Care teams were catching high-risk members too late, working from claims data months out of date. I led an interpretable ML platform that flagged risk 3–6 months earlier, enabling proactive care and ~$15M in first-year cost avoidance.
The Problem
At Highmark, nurses relied on claims data that was already months old, so interventions often arrived after costs had piled up. Leadership wanted a way to stratify risk proactively — without heavy new spend and without disrupting the tools care teams already used.
The Approach
I built the plan with engineering and clinical leadership around existing data and existing caregiver tools. We chose a Random Forest model for a strong balance of accuracy and interpretability — important because clinicians needed to trust and explain the scores. I worked closely with data science to make features clinically meaningful (pharmacy fills, chronic conditions, social determinants), not just statistically convenient.
The Solution
Clinicians were wary of a black box, so I went frugal and concrete: a simple mockup showing exactly how risk scores would appear in their workflow, walked through real patient examples on a whiteboard so they could see when and why an alert would fire. That built the trust needed for adoption. We rolled out first for high-cost chronic conditions like diabetes and heart failure, then expanded.
The Impact
Care teams flagged high-risk members 3–6 months earlier than before, enabling proactive care that avoided roughly $15M in costs in the first year — and it set the foundation for the organization to scale predictive analytics more broadly.