Machine Learning for Intraoperative Hypotension Prediction: A Systematic Review of Model Performance, Validation, and Clinical Impact

This is the authors' abstract. We don't add key points for this paper.

Intraoperative hypotension (IOH) is common and is associated with postoperative myocardial and kidney injury and mortality. Machine-learning (ML) systems can forecast impending hypotension, but high discrimination does not necessarily imply reliable calibration, transportability, or clinical benefit. To systematically evaluate ML models for IOH prediction, with emphasis on data sources, model architecture, prediction horizon, discrimination, calibration, external validation, clinical implementation, and patient-centered outcomes. MEDLINE, Embase, Web of Science, the Cochrane Library, IEEE Xplore, and Scopus were searched from inception to 25 June 2025. Adult operating-room studies developing, validating, or prospectively evaluating ML models for future IOH or immediate post-induction hypotension were eligible. Only primary peer-reviewed journal articles or peer-reviewed conference proceedings were included in the primary synthesis. Prediction-model studies were appraised with PROBAST, randomized trials with RoB 2, and non-randomized impact studies with ROBINS-I domains. Owing to heterogeneity in IOH definitions, labeling, prediction horizons, and outcome metrics, synthesis was narrative. The source search identified 1,248 records. After deduplication and screening, 94 full texts were assessed. Eligibility re-audit removed secondary reviews, non-peer-reviewed preprints, a post-cutoff publication, and a purely preoperative case-level model, leaving 37 primary peer-reviewed studies for qualitative synthesis. Most waveform-based models targeted MAP <65 mmHg within 5-15 minutes and commonly reported AUROC values above 0.85, but calibration and external validation were uncommon. HPI-guided management reduced IOH burden in several randomized trials, although at least one large pilot trial found no benefit and a 2025 spine-surgery trial did not significantly reduce the primary TWA endpoint. Across trials, evidence for reduced postoperative organ injury or mortality remained insufficient. ML can forecast short-horizon IOH with high discrimination, especially from continuous arterial waveforms, but current evidence is limited by single-center development, non-independent datasets, selective sampling, sparse calibration assessment, and uncertain incremental value over simple MAP-based strategies. Future work should prioritize external and prospective validation, calibration, alarm burden, decision-curve analysis, treatment-aware modeling, and adequately powered trials of patient-centered outcomes.