Predicting intraoperative hypotension using scalogram images and dual deep learning models

Intraoperative hypotension (IOH) is associated with an increased risk of postoperative complications, including myocardial injury, renal damage, and mortality. While AI-based models such as the Hypotension Prediction Index (HPI) have been commercialized, concerns remain regarding overestimated performance due to selection bias and limited interpretability. This study aims to develop a robust, multi-modal deep learning model that improves early prediction of IOH using Invasive Blood Pressure (IBP) waveforms, with an emphasis on class-specific performance and fair evaluation protocols. We collected a large dataset of intraoperative arterial waveform recordings and annotated hypotension events. We designed a transformer-based architecture—integrating Vision Transformer (ViT) for scalogram images and TabNet for feature-engineered ABP data—to autonomously learn dynamic arterial patterns while maintaining interpretability. The proposed multi-modal model achieved competitive performance, particularly outperforming baselines in minority-class recall, which is crucial for imbalanced medical tasks. At a 5-minute prediction horizon, the model demonstrated an AUROC of 0.94 and an AUPRC of 0.78. To address the issue of selection bias, we ”Redefined classes based on clinical occurrence; while prior studies often required the entire observation window to remain strictly below or above certain thresholds, we defined a hypotensive event as any window containing at least one mean arterial pressure (MAP) measurement of ≤ 65 mmHg. Furthermore, to ensure rigorous performance evaluation and prevent data leakage, we implemented patient-level cross-validation. This approach advances the field by ensuring fairness in model evaluation and providing actionable insights in the intraoperative environment