Open-Source Neural Network vs. Simple Hemodynamic Rules for Intraoperative Hypotension Prediction

Importance. Intraoperative hypotension (IOH) is associated with postoperative morbidity and mortality. Machine learning algorithms such as the Hypotension Prediction Index are deployed clinically but may not outperform simple mean arterial pressure (MAP) thresholds, and the choice of IOH definition may shape performance as much as the choice of model. Objective. To compare an open-source convolutional neural network (CNN) trained on MAP time series against two rule-based comparators (a MAP threshold rule and a three-feature trend-based logistic regression) across four IOH definitions using deployment-relevant metrics. Design, Setting, and Participants. Diagnostic study using the publicly available VitalDB database; 500 noncardiac surgery cases yielded 209,213 prediction windows. Main Outcomes and Measures. Discrimination, calibration, net benefit, alert burden, and lead time across four definitions (MAP <65 mm Hg for ≥1 or ≥5 minutes, <55 mm Hg for ≥1 minute, and ≥20% decrease from baseline). Results. For the three absolute definitions, all models performed comparably (AUROC 0.71–0.75, overlapping 95% confidence intervals [CIs]), with similar net benefit, alert burden (44–48 alerts/h), and lead time (~7.5 minutes). For the relative-decrease definition, the trend model incorporating baseline (AUROC 0.78; 95% CI, 0.74–0.82) outperformed the CNN (0.54; 95% CI, 0.51–0.58) and the MAP threshold (0.49; 95% CI, 0.44–0.53), with nonoverlapping CIs. Conclusions and Relevance. A neural network offered no advantage over a one-line MAP rule for absolute-threshold hypotension. For a percentage-drop definition, only the model incorporating patient baseline performed well. Aligning model and metric with the clinically relevant definition mattered more than model complexity.