Systematic review and meta-analysis of the role of machine learning in predicting postoperative complications following colorectal surgery: how far has machine learning come?

Abstract

Background: To systematically evaluate the clinical utility of machine learning in predicting postoperative outcomes following colorectal surgery.

Methods: A systematic literature search was conducted using PubMed, MEDLINE, Embase, and Google Scholar. Clinical studies investigating the role of machine learning models in predicting postoperative complications following colorectal surgery were included. Outcome measure was area under the curve for the model under investigation. The area under the curve and standard error were pooled using a random effects model to estimate the overall effect size. Statistical analyses were performed using the MedCalc (version 23) software, and the results presented as forest plots.

Results: Eighteen eligible articles were included. These reported outcomes on postoperative complications, namely anastomotic leak, mortality, prolonged length of hospitalization, re-admission rates, risk of bleeding, paralytic ileus occurrence, and surgical site infection. Pooled area under the curve for anastomotic leak was 0.813 [standard error: 0.031, 95% confidence interval (0.753-0.873)]; mortality 0.867 [standard error: 0.015, 95% confidence interval (0.838-0.896)]; prolonged length of stay 0.810 [standard error: 0.042, 95% confidence interval (0.728-0.892)]; and surgical site infection 0.802 [standard error: 0.031, 95% confidence interval (0.742-0.862)], respectively.

Conclusion: Machine learning methods and techniques are displaying promising clinical utility and applicability in accurately predicting the risk of developing complications following colorectal surgery. Future well-designed, adequately powered, multi-center studies are needed to investigate the usefulness and generalizability of these novel approaches in optimizing peri-operative surgical care.

Keywords: colorectal surgery; machine learning; postoperative complications; predictive modeling; systematic review and meta-analysis.

Figure 1.

PRISMA flow chart.

Figure 2.

Risk of bias (ROB) and applicability assessment of the included studies.

Figure 3.

Forest plots of the AUC for (a) anastomotic leak; (b) surgical site infection; (c) mortality; and (d) prolonged length of hospital stay. ROC, receiver operator curve.