Do First Generics Reach Medicaid Pharmacy Markets Equally After Approval? An ML.NET Framework for Measuring Diffusion and Access Gaps

Regulatory approval of a first generic drug is often treated as a meaningful access milestone because it permits generic entry for a previously brand-only product. However, approval does not guarantee that utilization appears immediately or broadly across payer markets. This study develops and evaluates an end-to-end, reproducible SQL Server and ML.NET workflow for measuring and forecasting post-approval diffusion of first generic drugs across state-level Medicaid pharmacy markets. FDA first-generic approval records from 2017–2024 were linked with Medicaid State Drug Utilization Data from 2017–2025. The SQL Server pipeline stages FDA approval records and annual Medicaid utilization files, deduplicates launch records, maps generic and brand names to Medicaid utilization records, constructs quarterly launch-level diffusion metrics, exports model-ready datasets, and produces final result tables. Two complementary distribution measures are defined: Numeric Distribution, the percentage of active state-level Medicaid molecule markets where generic utilization is observed, and Weighted Distribution, the share of molecule-level prescription volume represented by those generic-observed states. Their difference, Access Gap, identifies high-volume-first diffusion, where generic utilization appears in large Medicaid markets before becoming geographically broad. After deduplication, the final analysis dataset contained 278 included first-generic launch records, 3,202 quarterly diffusion observations across quarters 1–13 after approval, 1,064 first-four-quarter observations, and 265 complete first-year launch cases. Among complete Q1–Q4 cases, 202 launches were classified as Fast adopters, 6 as Medium adopters, and 57 as Slow adopters based on fourth-quarter Weighted Distribution. A year-based ML.NET multiclass model trained on 2017–2022 launch observations and tested on 2023–2024 observations achieved 90.78% micro accuracy, 65.08% macro accuracy, and log loss of 0.4734. The model separated Fast and Slow adoption patterns well, with F1-scores of 92.49% and 96.00%, respectively, but did not identify the rare Medium class. The results show that early Medicaid diffusion signals can support first-year adoption monitoring, especially for distinguishing broad versus weak diffusion, while intermediate adoption remains difficult to model because of class imbalance and threshold sensitivity. The main contribution is a transparent, reproducible public-data framework for moving from regulatory approval records to interpretable diffusion metrics and early adoption forecasting.