Medicare Risk Adjustment

Medicare’s risk adjustment system is used to estimate beneficiaries’ expected healthcare costs based on factors such as age, sex, and medical diagnoses. These predictions help determine how much Medicare pays private Medicare Advantage insurers to cover each patient. The primary goal of risk adjustment is to ensure that insurers receive appropriate funding for patients with different healthcare needs, while reducing incentives to avoid sicker, more expensive patients.

However, the current risk adjustment model is only about 12% accurate.

My main task at the Healthcare Transformation Institute at Penn is to improve the accuracy and equity of the risk adjustment algorithm used in Medicare Advantage programs. The goals are to incentivize care for medically vulnerable individuals and create parity between Medical Advantage and traditional Medicare programs.

In this role, I work with Drs. Ezekiel Emanuel, Daniel Shenfeld, and Ravi Parikh. In our paper, A Machine Learning Model to Improve Risk Adjustment Accuracy in Medicare, we show how we:

  • Developed a machine learning Medicare risk-adjustment model that outperformed CMS-HCC (v28) by 3x on log-cost fit using the same age, sex, and diagnosis-code (ICD-10) inputs

  • Built scalable Python pipelines to train and evaluate cost-prediction models on 4M+ Medicare beneficiaries, using embeddings, clustering, and XGBoost

  • Quantified model gains across key subpopulations and simulated downstream payment effects, estimating $750M–$3.25B in potential annual Medicare savings from reduced favorable selection

 

Speaking at a panel on the use of AI in data science. I was the School of Medicine representative among the panelists. The event was hosted by the University of Pennsylvania’s Data Driven Discovery Initiative (DDDI).