PGP-Clinical-TimeKAN Forecasts Clinical Trajectories with Probabilistic Modeling

Summary

The paper introduces PGP-Clinical-TimeKAN, a trajectory-first AI framework for jointly forecasting partially observed, correlated clinical physiology rather than predicting a single diagnostic label. It combines missingness-aware temporal encoders, a soft organ-system prior, patient-specific relations, nonlinear Kolmogorov-Arnold messages, and a low-rank multivariate Student-t output head. On 24-hour patient histories and six-hour forecasts from a frozen MIMIC-IV-derived cohort of 6,882 patients and 54,694 windows, the method was evaluated against 13 models across five random seeds. It achieved the second-lowest normalized MAE, 0.37727 +/- 0.00029, and the lowest RMSE, 0.52656 +/- 0.00034, reducing MAE by 0.52% compared with deterministic TimeKAN. Its probabilistic forecasts obtained a marginal NLL of 0.66380 and a CRPS of 0.27301, with empirical coverage of 0.533, 0.831, and 0.958 for nominal 50%, 80%, and 95% intervals. The largest ablation loss came from removing relational structure, while increasing covariance rank improved joint likelihood but barely changed point accuracy. A risk score derived from the predicted trajectories remained weaker than a dedicated GRU-D classifier, with AUROC of 0.603 versus 0.650. The authors therefore present joint trajectory forecasting as an inspectable intermediate task, while cautioning that accurate physiology forecasts do not by themselves establish a calibrated clinical event detector.