AI-Driven Early Warning and Prediction System for Intraoperative Hypotension Using Graph Neural Networks and Time-Series Biosignals

Abstract

Intraoperative hypotension (IOH) is a frequent and consequential perioperative complication that often develops silently and is associated with myocardial injury, acute kidney injury, and increased mortality, making early prediction essential for timely intervention. This research addresses the prediction of IOH five minutes before onset by leveraging physiological waveforms, signal-derived relationships, and clinical context. While most existing models treat biosignal channels as independent inputs, ignoring their physiological coupling, our work proposes a hybrid fusion approach combining temporal, relational, and clinical representations. We designed an Adaptive Graph Neural Network branch to model relational structure between signal-derived features, and a bidirectional LSTM branch to capture their temporal evolution. Both branches were fused with a clinical and statistical feature encoder within a unified LSTM-GNN architecture, aiming to assess the added value of combining relational and temporal learning over single-paradigm baselines. Trained and evaluated on the VitalDB dataset under leak-free, case-level splitting, the proposed model achieved an AUROC of 93.9% and AUPRC of 87% , with 97.9% specificity and 72.4 % sensitivity on an independent test set. These results confirm the value of integrating inter-signal relational modeling into multimodal IOH prediction, and demonstrate its feasibility for real-world perioperative monitoring through the HypoPredict prototype. Keywords: Intraoperative hypotension, Early prediction, Physiological waveforms, Graph Neural Network, LSTM, Multimodal fusion.