Abstract 101P
Background
Traditional nomograms for identifying patients at high risk of colorectal anastomotic leakage lack accuracy. Given the complex, multifactorial nature of leakage, machine learning (ML) offers a promising predictive alternative. We developed and internally validated a ML model for its prediction. Aim: To externally validate a ML model's performance for predicting colorectal anastomotic leakage intraoperatively using intraoperatively available variables.
Methods
Data was used from a large prospective study in which modifiable riskfactors were determined in a prospective fashion. This multicenter, prospective study involved adult patients undergoing elective colorectal resection. It evaluates the model's predictive accuracy presented metrics such as the area under the curve (AUC), sensitivity, and specificity.
Results
The ML model underwent training and validation utilizing data from 2,483 patients enrolled in the LekCheck study, across 14 participating hospitals, with 7.6% experiencing colorectal anastomotic leakage (CAL). Preliminary internal validation showed promising AUC values for gradient boosting (0.90), random forest (0.89), and artificial neural networks (0.84), indicating high predictive capability.
Conclusions
This study shows promising construction and internal validation of a ML based prediction model for colorectal anastomotic leakage. We have started a prospective external validation, possibly transforming colorectal surgical practice by providing an accurate prediction of the possibility of the occurrence of a life-threatening complication.
Clinical trial identification
Editorial acknowledgement
Legal entity responsible for the study
The authors.
Funding
Medtronic.
Disclosure
All authors have declared no conflicts of interest.