Real- Time Intradialytic Hypotension Prediction Model Using an Improved Equilibrium Optimizer for Feature Selection

Background and Objective: Intradialytic hypotension (IDH) is a severe complication of hemodialysis associated with cardiovascular health and mortality. Predicting IDH using conventional statistical models can be challenging due to its complex etiology. In this study, we proposed an improved binary Equilibrium Optimizer (BiEO) algorithm for feature selection to attain the optimal combination of input features. Methods: The paper proposed an improved binary version of the Equilibrium Optimizer algorithm, aimed at enhancing its optimization performance regarding identification of patient-specific model parameters. The utility of a novel time factor defi-nition and a fitness-based selection rule improved the exploration and exploitation ability. We also developed an IDH prediction infrastructure based on XGBoost, a practical machine learning algorithm. This system provided real-time predictions of IDH occurrence possibility within the next 60 minutes based on systolic blood pressure measurements. To provide model interpretability, SHAP analysis was conducted. Furthermore, we compared our results with those of two previous works to validate feasibility. Results: (1) The optimal feature combination attained by BiEO was consistent with clinical knowledge, indicating that blood pressure and its variation had the most significant impact on the model's prediction. (2) Our XGBoost model demonstrated a slight improvement in performance on our self-collected dataset compared to other two referenced works, as measured by several performance indices. (3) The number of input features decreased by nearly 40 % without sacrificing model performance. Conclusions: Accurate IDH prediction with reduced model dimension is now feasible with feature selection using BiEO. This development brings benefits for real-time IDH prediction in clinical application, preventing intradialytic hypotension, and enhancing quality of treatments.