Creation of an effective colorectal anastomotic leak early detection tool using an artificial neural network.

PURPOSE: Anastomotic leaks greatly increase both morbidity and mortality amongst colorectal patients. Earlier detection of leaks leads to improved patient outcomes; however, diagnosis often proves difficult due to heterogeneous presentation and varied differential diagnosis. The purpose of the study was to create an artificial neural network (ANN) capable of accurately identifying patients at risk of developing a post-operative colorectal anastomotic leak.
METHODS: A genetic ANN was trained and validated on a retrospective patient cohort. Two comparative groups were identified: those with anastomotic leaks confirmed at re-operation with a control group of patients with a post-operative delayed recovery, but in whom leak was excluded and no re-operation required.
RESULTS: Seventy-six patients were identified: 20 confirmed leaks and 56 controls. No significant difference in the baseline features between leak and control groups in terms of age (leaks 65.9 years [SD 9.29] controls 58.3 years [SD 17.0)], P = 0.054). Utilising backwards variable selection, ANN maintained 19 input variables. Internal validation of the ANN produced a sensitivity of 85.0 %, specificity of 82.1 %, and AUC of 0.89 for correct identification of clinical anastomotic leaks. Of the 20 confirmed leaks, the model correctly identified 17 and misclassified 10 control patients in the clinical leak category. External validation on 12 consecutive pilot prospective patients produced a specificity of 83.3 %.
CONCLUSIONS: ANNs can be created to accurately detect clinical anastomotic leaks in the early post-operative period using routinely available clinical data. Further prospective ANN testing is required to confirm generalisability. ANNs may provide useful real-world tools for improving patient safety and outcomes.