A new machine-learning–driven risk-adjustment model could significantly reshape how Medicare Advantage (MA) plans are paid—and potentially save Medicare billions in overpayments each year.
In a March study published in Health Services Research, researchers from the University of Pennsylvania found that a model called Franklin is three times more accurate than the Centers for Medicare & Medicaid Services’ (CMS) current Hierarchical Condition Category (HCC) model at predicting one-year Medicare costs, despite using the same underlying data inputs: age, sex, and diagnosis codes.
The findings are significant because risk adjustment underpins MA payment, shaping plan revenue, benefit design, and incentives around coding and enrollment. Persistent inaccuracies in CMS’ current model have fueled concerns and increased scrutiny from lawmakers and regulators about favorable selection and systematic overpayments.
By improving predictive accuracy, especially for lower-cost beneficiaries, the Franklin model reduces the financial returns to aggressive coding and favorable selection. Simulation results suggest that if applied to MA risk adjustment, Franklin could save Medicare between $750 million and $3.25 billion annually, depending on the degree of favorable selection in the market.
CMS’ current model vs Franklin
CMS uses the HCC model to risk-adjust payments for more than 65 million beneficiaries across Medicare Advantage, accountable care organizations, and Affordable Care Act marketplaces. However, researchers note that the model struggles to accurately predict costs for beneficiaries with few or no qualifying diagnosis codes.
Nearly 47 percent of Medicare beneficiaries lack HCCs, meaning their payments are adjusted based only on age and sex. For this large group, the mismatch between predicted and actual costs results in systematic overpayments of roughly $1,200 per person per year, creating strong incentives for favorable selection in MA.
Franklin is designed as a plug-and-play replacement for the current HCC model. Rather than introducing new data sources, it uses machine-learning techniques to analyze the same Medicare claims data CMS already collects. Researchers built and trained the model using a 20 percent sample of 2018 traditional Medicare claims and later validated its performance using 2022–23 data.
Unlike HCC, which maps a limited set of diagnoses into fixed categories, Franklin analyzes beneficiaries’ full diagnostic profiles, drawing on nearly all ICD-10 diagnosis codes. Key design features include:
· Broader diagnostic capture: While HCC includes fewer than 15 percent of diagnosis codes, Franklin incorporates nearly all available codes.
· Nonlinear modeling: The model allows for nonlinear relationships and negative weights, making it more resistant to targeted “gaming” through selective coding.
· Clusters instead of hierarchies: Diagnoses are grouped dynamically based on real-world code co-occurrence rather than fixed category hierarchies.
Together, these design choices shift risk adjustment away from narrow, highly transparent payment-boosting codes and toward a broader, less gameable representation of patient health.
In head-to-head testing using the same Medicare claims data, Franklin consistently outperformed HCC across metrics and populations. The model was especially effective at identifying the lowest-cost 20 percent of beneficiaries, where overpayments are most concentrated. It also reduced large over- and under-predictions by roughly half compared with HCC.
Accuracy improvements were seen across multiple subgroups, including women, people with disabilities, rural residents, and Black and Hispanic beneficiaries, which suggests the model could also mitigate disparities embedded in current payment methods.
Implications for MA payment
Simulation studies indicate that applying Franklin to MA risk adjustment could reduce favorable selection–related overpayments by $23 to $99 per member per year, with the exact savings depending on how aggressively plans engage in selection behavior. At today’s MA enrollment levels, that translates into $750 million to $3.25 billion in annual Medicare savings.
In an article for Penn LDI, researchers caution that Franklin is not entirely immune to gaming and represents just one step toward addressing MA overpayments. The research team is also exploring the software and analytic infrastructure needed to support CMS adoption. However, because Franklin relies on existing data and is designed as a drop-in replacement for HCC, it offers CMS a near-term opportunity to modernize risk adjustment without redesigning the entire MA payment system and potentially reshaping incentives and payment accuracy well before a full methodological overhaul.