PGP-Clinical-TimeKAN
A clinical AI framework designed for joint probabilistic forecasting of multivariate physiological indicators. It integrates modules for missing-value-aware encoding, organ-system priors, patient-specific relationship modeling, and KAN-based message passing.
Timeline (last 90 days)
The framework replaces the traditional MLP with a KAN network for graph message passing to enhance nonlinear fitting capability.
On probabilistic prediction metrics, the framework achieves a marginal negative log-likelihood (NLL) of 0.66380 and a continuous ranked probability score (CRPS) of 0.27301.
PGP-Clinical-TimeKAN replaces the traditional MLP with a KAN network for graph message passing to enhance nonlinear fitting capability.
Low-rank parameterization decomposes the full covariance matrix into the form Σ = WW^T + D, reducing the parameter count from O(d²) to O(dr).
The framework's prediction intervals achieve empirical coverage rates of 0.533, 0.831, and 0.958 at nominal confidence levels of 50%, 80%, and 95%, respectively.
PGP-Clinical-TimeKAN's marginal negative log-likelihood (NLL) is 0.66380, and its Continuous Ranked Probability Score (CRPS) is 0.27301.
PGP-Clinical-TimeKAN's empirical coverage rates at nominal 50%, 80%, and 95% confidence levels are 0.533, 0.831, and 0.958, respectively.
Ablation experiments found that removing relational structure modeling leads to the greatest performance loss.
PGP-Clinical-TimeKAN is a framework for joint probabilistic forecasting of multivariate physiological indicators, adopting a trajectory-first design philosophy.
The study reveals that accurate physiological indicator prediction does not automatically equate to a well-calibrated event detector, and there exists a semantic gap between physiological trajectory prediction and clinical event detection.
6 more timeline events
All Facts (16)
The framework replaces the traditional MLP with a KAN network for graph message passing to enhance nonlinear fitting capability.
50%UnverifiedOn probabilistic prediction metrics, the framework achieves a marginal negative log-likelihood (NLL) of 0.66380 and a continuous ranked probability score (CRPS) of 0.27301.
50%UnverifiedPGP-Clinical-TimeKAN replaces the traditional MLP with a KAN network for graph message passing to enhance nonlinear fitting capability.
50%UnverifiedLow-rank parameterization decomposes the full covariance matrix into the form Σ = WW^T + D, reducing the parameter count from O(d²) to O(dr).
50%UnverifiedThe framework's prediction intervals achieve empirical coverage rates of 0.533, 0.831, and 0.958 at nominal confidence levels of 50%, 80%, and 95%, respectively.
50%UnverifiedPGP-Clinical-TimeKAN's marginal negative log-likelihood (NLL) is 0.66380, and its Continuous Ranked Probability Score (CRPS) is 0.27301.
50%UnverifiedPGP-Clinical-TimeKAN's empirical coverage rates at nominal 50%, 80%, and 95% confidence levels are 0.533, 0.831, and 0.958, respectively.
50%UnverifiedAblation experiments found that removing relational structure modeling leads to the greatest performance loss.
50%UnverifiedPGP-Clinical-TimeKAN is a framework for joint probabilistic forecasting of multivariate physiological indicators, adopting a trajectory-first design philosophy.
50%UnverifiedThe study reveals that accurate physiological indicator prediction does not automatically equate to a well-calibrated event detector, and there exists a semantic gap between physiological trajectory prediction and clinical event detection.
50%UnverifiedThe Student-t distribution was chosen over the Gaussian distribution because its heavier tails better capture outliers and extreme physiological events in clinical data.
50%UnverifiedRisk scores derived from predicted trajectories performed worse than a dedicated GRU-D classifier on clinical event detection tasks (AUROC 0.603 vs. 0.650).
50%UnverifiedPGP-Clinical-TimeKAN integrates five technical modules: a missing-value-aware temporal encoder, soft organ system priors, patient-specific relationship modeling, nonlinear Kolmogorov-Arnold message passing, and a low-rank multivariate Student-t distribution.
50%UnverifiedThe dataset used in this study contains 6,882 patients and 54,694 time windows, with the experimental setup using 24 hours of historical data to predict physiological indicator changes over the next 6 hours.
50%UnverifiedPGP-Clinical-TimeKAN achieved a normalized MAE of 0.37727 ± 0.00029, ranking second, and an RMSE of 0.52656 ± 0.00034, reaching the best level among all models.
50%UnverifiedCompared to the deterministic TimeKAN model, PGP-Clinical-TimeKAN reduced MAE by 0.52%.
50%