Navigator Synthetic Demo: Reproducible Demonstrations of STREAM-SMR and Offline Reinforcement Learning for ICU Registry Analytics

A self-contained, dependency-light reproduction of two core analytic methods used in the JIPAD Navigator, demonstrated entirely on synthetic ICU-registry data so that the algorithms can be inspected and reproduced without access to any real patient data. STREAM-SMR — a scalar Kalman filter on the log standardized mortality ratio (SMR) for streaming quality surveillance. On synthetic data it recovers per-facility latent quality (corr = 0.84 with ground truth, RMSE = 0.21) and flags 3 outlier facilities, tracking an injected change-point faster than a cumulative baseline. Offline RL demonstration — behaviour-constrained Fitted-Q-Iteration on a synthetic ICU MDP, presented with an honest account of the conservatism-vs-performance trade-off and the limits of off-policy evaluation (logged behaviour 53.5% survival; constrained policy 70.8%; unconstrained 85.0%; reference severity-matched escalation 75.6%). Both methods run end-to-end from a single script (run_all.py) using only NumPy / SciPy / scikit-learn / matplotlib, producing two figures and a results summary deterministically (seed 20260628). The production Navigator uses PyTorch and d3rlpy; this demo reimplements the same ideas with a minimal dependency footprint for friction-free reproducibility. Code is MIT-licensed; documentation is CC-BY-4.0.