Deep learning algorithms for predicting intraoperative hypotension: A systematic review and meta-analysis.

BACKGROUND

Intraoperative hypotension (IOH) is associated with myocardial injury, acute kidney injury, perioperative stroke, and 30-day mortality. Deep learning (DL) refers to a family of machine-learning methods that learn multiple levels of data representation through nonlinear transformations. In perioperative monitoring, DL models can learn patterns from time-varying physiological signals and combine them with clinical variables to estimate impending IOH risk. DL algorithms applied to continuous physiological waveforms therefore represent a rapidly expanding paradigm for early IOH prediction, but their pooled diagnostic accuracy has not been synthesised within a unified diagnostic-accuracy framework.

OBJECTIVE

To quantify the pooled diagnostic accuracy of DL-based IOH prediction models and to identify factors that modify their performance.

METHODS

PubMed, Embase, Web of Science, and the Cochrane Library were searched through March 2026. Methodological quality was appraised with PROBAST+AI and certainty of evidence with GRADE. A bivariate random-effects model generated pooled sensitivity, specificity, and SROC-AUC; heterogeneity was quantified by τ2(Se), τ2(Sp), and ρ. Subgroup analyses (prediction window, DL architecture, modality, IOH reference, analysis unit) and random-effects meta-regression were performed.

RESULTS

Twelve studies were included; nine contributed 22 validation datasets. Pooled sensitivity was 0.78 (95% CI 0.73-0.81), specificity 0.88 (0.82-0.92), and SROC-AUC 0.87 (0.83-0.90); diagnostic odds ratio 24.7 (16.1-37.9), positive likelihood ratio 6.31, negative likelihood ratio 0.26. Heterogeneity was τ2(Se) = 0.25, τ2(Sp) = 1.04, ρ = -0.28. The 5-min window achieved the highest performance (sensitivity 0.81; specificity 0.91). Meta-regression identified DL architecture as the only significant moderator of specificity (P = 0.02), with hybrid CNN-RNN exceeding pure CNN (β = 1.77, 95% CI 0.45-3.09). Deeks' test was non-significant (P = 0.06). GRADE certainty was Low.

CONCLUSIONS

DL models for IOH prediction achieve moderate diagnostic accuracy, with hybrid CNN-RNN architectures and 5-min prediction windows showing the most favourable performance. Universal absence of formal calibration assessment, scarce external validation, and geographic concentration constrain immediate clinical translation.

SYSTEMATIC REVIEW REGISTRATION: PROSPERO CRD420261377604 Reporting guideline: PRISMA-DTA.