Clinical deterioration unfolds as coupled, partially observed multivariate physiological trajectories rather than a single diagnostic label. We introduce PGP-Clinical-TimeKAN, a trajectory‑first framework for joint probabilistic forecasting of multivariate physiology. The core combines missingness‑aware temporal encoders, a soft organ‑system prior, patient‑specific relational graphs, nonlinear Kolmogorov‑Arnold messages, and a low‑rank multivariate Student‑t head. Evaluation uses a frozen MIMIC‑IV derived cohort with 6,882 patients and 54,694 windows, leveraging 24‑hour histories to predict six‑hour futures. Across five seeds and 13 model variants, PGP-Clinical-TimeKAN attains the second‑lowest normalized MAE (0.37727 ± 0.00029) and the lowest RMSE (0.52656 ± 0.00034). Compared with deterministic TimeKAN, MAE improves by 0.52%. For probabilistic forecasting, the marginal NLL is 0.66380 and CRPS 0.27301. Empirical coverage reaches 0.533, 0.831, and 0.958 for nominal 50%, 80%, and 95% intervals. Removing relational structure yields the largest ablation loss. Increasing covariance rank boosts joint likelihood but has little impact on point‑wise accuracy. A trajectory‑derived risk score remains weaker than a dedicated GRU‑D classifier (AUROC 0.603 vs 0.650), limiting the clinical claim. Thus, joint trajectory forecasting offers an inspectable intermediate task, yet accurate physiology forecasts alone do not guarantee a calibrated event detector.
Review