Shared-Benchmark Regime Decomposition for Nonstationary Clinical Decisions

Structured clinical decision sequences, such as tabular EHRs and irregularly sampled physiological signals, are partially observed and often nonstationary: similar measured states can lead to different actions at different times. We introduce Shared-Benchmark Regime Decomposition (SBRD), a latent-variable framework that separates a shared benchmark value function from switching constraint-regime deviations. A continuous-time latent dynamics layer constructs belief states and benchmark action values from structured health records, while a sparse regime layer recovers persistent, interpretable wedges in observed decisions. On ICU sepsis treatment (MIMIC-IV) and chronic CKD-MBD management, SBRD improves held-out action prediction over stationary and interpretable behavioral baselines and yields clinically meaningful regime profiles. The recovered wedges support benchmark-relative constraint-relaxation diagnostics that reveal where treatment complexity or protocol-related pressures account for recoverable benchmark value.