Evaluation of nursing performance during coronary artery bypass grafting surgeries using machine learning analysis of surgical videos: a multicenter retrospective study.

backgroundCoronary Artery Bypass Grafting (CABG) is a high-risk, complex cardiac procedure that relies heavily on effective nursing collaboration to ensure surgical safety and improve patient outcomes. The advancement of machine learning (ML) and computer vision technologies, surgical video analysis has emerged as a promising tool for objective, quantitative assessment of clinical performance compared to traditional methods.

objectiveTo develop and validate an ML-based model using surgical video analysis to objectively evaluate nursing performance during CABG, identify inefficient operational patterns, and generate personalized improvement recommendations for nurses.

methodsA total of 1000 surgical video clips of CABG procedures were collected from three tertiary hospitals. Each clip was annotated by two senior cardiac nursing experts and one cardiovascular surgeon to label key nursing collaborative behaviors, including timeliness of instrument delivery, compliance with sterile procedures, and accuracy of intraoperative assistance. A Three-Dimensional Convolutional Neural Network (3D CNN) was constructed to analyze sequential nursing actions from the video clips, establishing a quantitative evaluation model for nursing operation quality. Supervised contrastive learning (SupCon) was applied to learn discriminative features between high-performance and low-performance nursing behaviors.

resultsThe annotated dataset achieved high inter-annotator agreement (Cohen’s kappa = 0.82; Krippendorff’s alpha = 0.81), indicating reliable labeling quality. Annotation batch analysis showed no significant decline in agreement over time (kappa range: 0.80–0.84), ruling out fatigue effects. The 3D CNN-based nursing performance evaluation model outperformed baseline models, achieving an accuracy of 91.7% (95% confidence interval [CI]: 89.8%–93.2%, verified via 1000 bootstrap iterations), precision of 90.3% (95% CI: 88.1%–92.2%), recall of 92.1% (95% CI: 90.0%–93.9%), F1-score of 91.2% (95% CI: 89.2%–92.9%), and AUC of 0.95 (95% CI: 0.93–0.97). Four inefficient operational patterns were identified via contrastive learning: delayed instrument delivery (37.2%), non-compliant sterile operations (28.5%), redundant auxiliary steps (21.3%), and inaccurate instrument handover (13.0%). Personalized recommendations based on these patterns achieved high clinical applicability (satisfaction rate 87.5% among senior cardiac nurses).

conclusionOur ML-based surgical video analysis model provides an objective, accurate, and efficient tool for nursing performance evaluation in CABG. Prospective validation linking predicted scores to clinical outcomes and external validation across diverse settings are needed before widespread clinical adoption.