(118) Beyond Rule-Based Risk Stratification: Redefining Risk for Better Care Coordination in Managed Care Through Artificial Intelligence

Background: : Proactively identifying evolving risk and high-risk members is essential to improving outcomes and costs in managed care. Traditional rule-based models, built on static thresholds, often miss the full complexity of member health profiles. In contrast, Artificial Intelligence (AI) offers a dynamic, data-driven approach that captures evolving risk patterns across multiple variables with greater predictive accuracy.

Objective: : To evaluate how an AI algorithm outperforms traditional rule-based approaches in member risk stratification and how variables such as comorbidities, SDoH, demographics, lifestyle, utilization patterns, HEDIS care gaps, and lab results improve precision and granularity of risk differentiation.

Methods: : We conducted a retrospective analysis using de-identified medical, pharmacy claims, and clinical data from a large payer population. Natural Language Processing (NLP) was applied to longitudinal health data — including comorbidities, medications, HEDIS measures, and SDoH — to create unique health risk profiles. This captured the evolving, nonlinear nature of individual risk over time. Profiles were fused with structured data like lab results and utilization, then regressed against real-world cost metrics. This integrated approach enabled dynamic, personalized risk scoring and cohesive stratification into actionable risk categories.

Results: : To benchmark our in-house AI model, we compared its performance against the CMS-HCC risk adjustment model using the same member population. Risk strata—high, medium, and low—were derived based on predicted risk scores from both models. The AI model produced distinct strata with average per-member costs of $31,078 (high-risk), $2,680 (medium-risk), and $439 (low-risk), versus $22,219, $4,525, and $1,953 under CMS-HCC, showing AI better segregates members. Strata quality was quantified using the Davies-Bouldin Index (DBI), where lower values indicate better-defined strata. Across all risk tiers, the AI model outperformed the rule-based approach (DBI for high: AI = 0.47 vs. HCC = 0.55; medium: AI = 0.55 vs. HCC = 0.58; low: AI = 0.52 vs. HCC = 0.61), indicating more compact and well-separated member groupings by AI.

Conclusions: : AI significantly improves member risk stratification over traditional models by recognizing complex variable interactions and dynamic risk signals. Incorporating AI-driven stratification into managed care workflows can enhance early intervention strategies, optimize resource allocation, and advance value-based care.

Sponsorship:: Datalink Software LLC