Physics-Informed Neural Networks for Clinical Time-Series Forecasting with Clinical Utility

Abstract

Clinical time-series forecasting is essential for anticipating patient deterioration and supporting proactive interventions, yet traditional machine learning approaches often operate as black boxes, limiting trust and adoption in critical care. To address this challenge, we implement a Physics-Informed Neural Network (PINN) that integrates physiological constraints, such as fluid balance and pharmacokinetics, directly into the learning process, ensuring predictions remain consistent with known medical principles. The study focuses on forecasting laboratory and vital sign trends while simultaneously predicting deterioration risk (ICU stay or in-hospital death) from a dataset containing demographics, preoperative labs, and intraoperative measurements. By embedding physical laws into the loss function, the model enhances interpretability, generalization, and plausibility, offering a bridge between data-driven predictions and physiological reasoning. Results show strong regression performance (RMSE = 11.75, MAE = 8.46) and competitive classification metrics (Accuracy = 0.84, AUC=0.82), though recall highlights the challenge of capturing all high-risk cases. These findings demonstrate that PINNs not only improve predictive robustness but also provide clinically meaningful insights, marking a step toward trustworthy, physically grounded AI for healthcare.

Ähnliche Arbeiten