Aim: Surgeon’s intraoperative decisions significantly impact patient outcomes. In the reconciliation cycle, interoperative decisions are guided by probabilistic reasoning, which is informed by the evolving intraoperative features. This paper aims to compare the utility of a traditional logistic regression (LR) model for critical view of safety (CVS) achievement to Bayesian network (BN) maps using intraoperative features. It hypothesizes that BN mapping better integrates with surgeon heuristics. Methods: Using prospectively gathered intraoperative data, BN maps were developed and tested to determine their ability to predict critical view of safety achievement. Performance was compared to traditional logistic regression models to consider their utility in practice. Results: In total, 4,663 patients were identified. Of these patients, 2,837 (61%) presented acutely and 3,122 (67%) were female. CVS was achieved in 4,131 (92%) of patients. In total, four BN were developed. Optimal performance was seen in model 2 with an AUC of 0.879 (0.798-0.960) (P < 0.001). Selecting a cut-off of 0.6 gave an optimized sensitivity of 99% and a specificity of 45% for CVS achievement. In comparison to this, for the combined acute LR model, ROC curve analysis gave an AUC of 0.829 (0.787-0.872 ) (P < 0.001). A cut-off of 75% probability resulted in a sensitivity of 95% and a specificity of 38% for CVS achievement. Conclusion: The present study illustrates how BN modeling can map to surgeon decision making to facilitate reasoning in complex environments. Further work is needed to facilitate data capture and implementation. Despite this, they represent a promising avenue for intuitive decision support tools.
Read full abstract