Postoperative bleeding risk prediction for patients undergoing colorectal surgery

Area of Science:

• Medical Informatics
• Surgical Outcomes Research
• Machine Learning in Healthcare

Background:

• Postoperative bleeding remains a significant concern with limited consensus on risk factors.
• Electronic medical record data offers a rich source for identifying bleeding predictors.
• Predicting high-risk patients is crucial for proactive management and resource optimization.

Purpose of the Study:

• To leverage machine learning and electronic medical record data for novel postoperative bleeding risk factor identification.
• To develop predictive models for identifying patients at high risk of postoperative bleeding.
• To enhance the understanding of factors influencing postoperative bleeding after colorectal surgery.

Main Methods:

• Retrospective analysis of 13,399 colorectal surgery patients (1998-2015).
• Extraction and evaluation of 299 potential predictors from electronic medical records.
• Comparison of logistic regression and gradient boosting machine models using AUC-ROC and AUC-PR metrics.

Main Results:

• Postoperative bleeding occurred in 12.5% of patients.
• Gradient boosting machine achieved a superior AUC-ROC of 0.822 compared to logistic regression (0.735).
• Novel predictors identified by gradient boosting machine included nutrition, weakness, patient mobility, and activity level.

Conclusions:

• Functional capacity measures (nutrition, mobility, activity) are significant novel predictors of postoperative bleeding.
• Machine learning models effectively predict postoperative bleeding risk.
• Improved risk assessment allows for better resource allocation in managing postoperative bleeding.