TIME: A Taylor-Inspired Mixed-Effects Model for IDH Prediction

Intradialytic hypotension (IDH) is a critical complication in hemodialysis that increases morbidity and treatment risks, yet existing machine learning methods inadequately address session-level physiological state heterogeneity and fail to balance linear and nonlinear feature interactions, limiting predictive accuracy and interpretability.

We developed TIME (Taylor-Inspired Mixed-Effects Model), which explicitly separates linear interactions, higher-order nonlinear dependencies, and residual terms inspired by Taylor series expansion, while incorporating a physiological state embedding layer to capture session-level physiological state heterogeneity revealed through hierarchical clustering on autoencoder-derived latent representations of hematologic indicators selected using SHAP analysis and clinical knowledge. Using data from 532 hemodialysis patients (18,309 records), we compared TIME against 19 state-of-the-art baseline models, including blood pressure-stratified subcohort and out-of-distribution analyses.

Hierarchical clustering identified two distinct hematological states with different IDH risks (57.93% vs. 55.75%, $p=6.60\times 10^{-7}$). TIME achieved 0.6929 accuracy, 0.7435 F1-score, and 0.3663 MCC on the full cohort, the highest AUC across all three subcohorts (0.7010, 0.6795, and 0.7201), and strong out-of-distribution performance (accuracy 0.7053, F1 0.7377, MCC 0.4009). Integrating TIME's architecture into baseline models improved MCC by 2.67 percentage points on average.

TIME effectively predicts IDH by modeling session-level physiological state heterogeneity and balancing feature interactions, with robust generalizability across diverse populations.

TIME advances precision medicine in dialysis care by enabling early risk identification and personalized treatment strategies, while revealing broader immune, nutritional, and fluid-related markers of IDH risk.