JMIR Publications

ABSTRACT

Background:

Intraoperative hypotension (IOH) is associated with myocardial injury, acute kidney injury, perioperative stroke, and 30-day mortality, yet conventional blood pressure monitoring remains reactive rather than anticipatory. Deep learning (DL) algorithms applied to continuous physiological waveforms represent a rapidly expanding paradigm for early IOH prediction, but the comparative performance of distinct DL architectures and the influence of prediction-window length, input data modality, IOH reference standard, and analysis unit on diagnostic accuracy have not been systematically synthesised.

Objective:

To quantify the pooled diagnostic accuracy of DL-based IOH prediction models and to identify methodological and clinical 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 the PROBAST+AI tool and overall certainty of evidence with the GRADE framework. A bivariate random-effects model generated pooled sensitivity, specificity, and the area under the summary receiver operating characteristic (SROC) curve, with heterogeneity quantified by τ²(Se), τ²(Sp), and the inter-study correlation ρ. Threshold effect was tested with Spearman’s correlation, publication bias with Deeks’ test, and clinical utility with Fagan’s nomogram. Prespecified subgroup analyses (prediction window, DL architecture, input modality, IOH reference standard, analysis unit) and Bayesian random-effects meta-regression explored heterogeneity sources.

Results:

Twelve studies were included; nine contributed 22 validation datasets to the quantitative synthesis. The 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); the diagnostic odds ratio was 24.7 (16.1–37.9), positive likelihood ratio 6.31, and negative likelihood ratio 0.26. Heterogeneity was τ²(Se) = 0.25, τ²(Sp) = 1.04, and ρ = −0.28; no significant threshold effect was detected (Spearman ρ = 0.29, P = 0.20). The 5-minute window achieved the highest performance (sensitivity 0.81, 95% CI 0.77–0.85; specificity 0.91, 0.84–0.95). 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); no covariate significantly moderated sensitivity. Deeks’ test showed no statistically significant publication bias (P = 0.06). At a 10% pre-test probability, post-test probabilities were 41% (positive) and 3% (negative). GRADE certainty was Low.

Conclusions:

Deep learning models for IOH prediction achieve moderate diagnostic accuracy, with hybrid CNN-RNN architectures and 5-minute prediction windows showing the most favourable performance. The universal absence of formal calibration assessment, scarce external validation, and geographic concentration of the evidence base constrain immediate clinical translation. Prospective multinational validation with mandatory calibration reporting and patient-level evaluation is required before DL-based IOH alerts can be safely integrated into perioperative decision support. Clinical Trial: PROSPERO CRD420261377604.