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

Area of Science:

• Nephrology
• Biomedical Engineering
• Machine Learning

Background:

• Intradialytic hypotension (IDH) is a significant complication during hemodialysis, associated with increased patient morbidity and treatment risks.
• Current machine learning models struggle with the heterogeneity of patient physiological states during dialysis sessions and balancing linear/nonlinear feature interactions, limiting prediction accuracy and interpretability.

Purpose of the Study:

• To develop an advanced machine learning model, TIME (Taylor-Inspired Mixed-Effects Model), for accurate and interpretable prediction of intradialytic hypotension (IDH).
• To address limitations in existing models by explicitly handling session-level physiological state heterogeneity and complex feature interactions.

Main Methods:

• Developed TIME, a novel model separating linear, nonlinear, and residual terms using Taylor series expansion.
• Incorporated a physiological state embedding layer using hierarchical clustering on autoencoder-derived hematologic indicators (selected via SHAP and clinical knowledge).
• Validated TIME on 18,309 records from 532 hemodialysis patients, comparing it against 19 baseline models with subcohort and out-of-distribution analyses.

Main Results:

• Hierarchical clustering revealed two distinct hematological states with significantly different IDH risks.
• TIME achieved superior performance with 0.6929 accuracy, 0.7435 F1-score, and 0.3663 MCC on the full cohort.
• TIME demonstrated strong performance across subcohorts (AUC up to 0.7201) and out-of-distribution testing (accuracy 0.7053, F1 0.7377).
• Integrating TIME's architecture improved baseline models' MCC by an average of 2.67 percentage points.

Conclusions:

• TIME effectively predicts IDH by modeling physiological heterogeneity and balancing feature interactions, showing robust generalizability.
• TIME advances precision medicine in dialysis by enabling early risk identification and personalized treatment strategies.
• The model highlights broader immune, nutritional, and fluid-related markers associated with IDH risk.