A dynamic stacking machine learning framework and prototype software module for real-time hemodialysis hypotension monitoring

Intradialytic hypotension (IDH) remains a critical complication in hemodialysis (HD). While machine learning offers predictive potential, a significant gap remains between theoretical algorithms and clinical deployment. This research develops a dynamic stacking machine learning framework and an embedded software module for real-time IDH monitoring.

Data comprising 2,228,650 HD sessions were gathered from a cohort of 1075 HD patients. Thirteen critical feature parameters were dynamically extracted to capture physiological changes during dialysis. We constructed a stacking ensemble model by amalgamating multiple base learners, linking the model's findings to established physiological science. To translate this algorithm into practice, we engineered an prototype software module designed to interface seamlessly with HD Information Systems. This module integrates real-time data ingestion, dynamic feature processing, and visual prediction.

Physiological analysis identified HD duration and predialysis mean arterial pressure as pivotal risk indicators, consistent with clinical mechanisms. The binary effect of diabetes further highlighted the model's ability to capture nonlinear interactions. The proposed stacking model achieved a precision of 0.89 (95% confidence interval [CI]: 0.87-0.91) and AUC of 0.94 (95% CI: 0.92-0.95). Crucially, the prototype simulation of the module demonstrated successful offline tracking of physiological trends and provided risk stratification, converting complex algorithmic outputs into actionable clinical alerts.

The developed stacking machine learning framework, encapsulated within an embedded software module, offers an effective solution for predicting IDH events during HD treatment. It can enhance monitoring efficiency and facilitate timely prevention strategies, although external validation is necessary before considering any real-world clinical implementation.