Artificial Intelligence–Assisted Electrocardiogram Interpretation Among Surgeons: Diagnostic Augmentation and Workflow Integration in Perioperative Risk Assessment

Abstract

Objective To determine whether artificial intelligence–assisted electrocardiogram interpretation improves surgeons' identification of occult cardiovascular pathology and high-risk perioperative indicators, and to characterise the usability, trust and workflow factors governing its integration into surgical practice.

Design Retrospective observational cohort study with an embedded prospective clinician reader study. Participants interpreted a standardised set of ECGs before and after AI-generated probability outputs, separated by a two-week washout period.

Setting Preoperative surgical assessment and perioperative decision-making.

Participants A cohort of 37,060 adult patients undergoing major non-cardiac surgery was analysed to assess the diagnostic efficacy of convolutional neural network models. The reader component comprised attending surgeons and surgical residents who evaluated a standardised set of 100 preoperative 12-lead ECG traces, with and without AI-generated probability outputs, to identify occult cardiovascular pathology.

Main outcome measures Diagnostic performance measured through sensitivity, specificity and area under the receiver operating characteristic curve; usability assessed with the System Usability Scale; subjective clinician confidence and ease of telemetry interpretation assessed with 4-point Likert-scale items; and changes in clinician-directed therapeutic adjustments following model-assisted interpretation.

Results Sixty-eight per cent of participating surgeons integrated AI probability scores into their clinical workflow, with a notable preference for these tools among residents compared with senior faculty. Surgeons supported by AI interpretation showed a marked increase in the identification of high-risk perioperative indicators and significantly outperformed non-assisted assessments in predictive accuracy, with improved sensitivity in recognising subtle arrhythmic patterns and a reduction in diagnostic oversights among complex, high-risk surgical candidates. Adoption was nonetheless hindered by the need for manual data entry and by concerns regarding the transparency of algorithmic outputs, which impeded trust in non-intuitive clinical recommendations. Conversely, seamless mobile integration and high-speed data access proved essential in overcoming these operational hurdles, significantly enhancing practical deployment.

Conclusions AI-assisted ECG interpretation augmented surgeon decision-making, with the greatest gains among less experienced readers and in complex, high-risk cases. Translation into routine practice depends less on discrimination than on eliminating manual data entry, providing interpretable outputs, and embedding the tool in existing mobile and electronic workflows. Prospective evaluation of patient-centred outcomes is now required.