Enhancing Medicaid Social Risk Identification: Comparing Linear and Non-Linear Machine Learning Models with Expert Heuristics.

Abstract

Conventional risk identification methods often fail to capture the complex, non-linear patterns of social risks, reducing the effectiveness of operationalizing Social Determinants of Health within Medicaid Managed Care Organizations. The utility of machine learning models is investigated in this paper to overcome the performance limitations of the current expert heuristic scoring systems in practice. We develop a suite of linear and non-linear machine learning models based on a real-world Medicaid dataset containing six SDOH domains to predict individuals at risk of unmet social needs. Results showed that the non-linear Random Forest model with an AUC of 0.72 significantly outperformed both linear models (AUC 0.58) and MPC expert heuristics. The gain in discriminatory performance stems from the non-linear model's ability to capture higher-order interactions between SDOH domains. Our work advocates for the integration of non-linear ML into Medicaid workflows, offering a more robust, data-driven, and scalable path to accurately target high-risk members for proactive outreach, ultimately improving patient outcomes and reducing avoidable healthcare expenditures.