Proactive Prediction of Alarms in Hemodialysis Machines: A Comparative Evaluation of AI Models for Clinical Prognostic Support

Background/Objectives: Hemodialysis is a treatment for patients with kidney disease, yet it remains risk-prone due to potential technical or physiological instabilities. Dialysis machines rely on reactive safety mechanisms that trigger alarms only when predefined thresholds are reached. This often alerts clinical staff only after a complication has already occurred, contributing to potential delays in interventions. This study aims to develop a proactive prognostic framework capable of anticipating system instabilities before they transform into clinical adverse events. Methods: This work proposes a predictive approach for the proactive detection of patient- or process-related alarm conditions using ten machine learning models and one deep learning architecture. A dataset consisting of over two million records sampled at 0.5 s intervals from 95 dialysis sessions was assembled by merging machine logs of sensor readings and the binary machine status (alarm/normal). Results: Among the models tested, Random Forest emerged as the most robust predictive tool, achieving an F1 score higher than 0.85 for predicting alarms up to one minute in advance, with precision (~0.83), recall (~0.87), specificity (~0.99), AUROC (~0.99), and AUPRC (~0.94) values indicating strong performance, and stable performance across validation and test sets. A progressive decrease in performance was observed for longer prediction time windows. Conclusions: These findings demonstrate the feasibility of predicting dialysis-machine-generated alarms in advance using machine learning and deep learning models. Providing advance warning could give clinicians the opportunity to anticipate impending alarm conditions and, where appropriate, implement timely corrective actions, potentially improving treatment management and the operational reliability of renal replacement therapies.