Machine Learning Models Using Post-operative CRP Trends to Predict Colorectal Anastomotic Leak: A Pilot Study.

Area of Science:

• Colorectal surgery outcomes
• Postoperative complication prediction
• Biomarker analysis in surgical patients

Background:

• Anastomotic leak (AL) is a critical complication following colorectal resection, leading to significant patient morbidity.
• Current reliance on static C-reactive protein (CRP) thresholds for AL prediction has limitations.
• Emerging research explores the predictive value of CRP trajectory (changes over time) for AL.

Purpose of the Study:

• To evaluate machine learning (ML) models for predicting AL using postoperative CRP thresholds and trajectory data.
• To compare the performance of ML models against conventional univariate CRP metrics.
• To assess the accuracy of dynamic CRP analysis in early postoperative risk stratification.

Main Methods:

• Retrospective analysis of elective large bowel resections (2020-2025).
• Development and validation of logistic regression, Lasso, Random Forest (RF), and Extreme Gradient Boosting (XGB) models.
• Utilized CRP levels from postoperative days 1-3 (absolute values, percentage change, net difference) with five-fold cross-validation.

Main Results:

• The XGB model integrating absolute CRP values and percentage change achieved the highest predictive performance (AUC 0.91).
• ML models incorporating both CRP thresholds and trajectory data outperformed models using only absolute CRP values.
• XGB and RF models demonstrated superior accuracy and balanced sensitivity/specificity compared to Lasso regression.

Conclusions:

• Machine learning models leveraging dynamic CRP data (thresholds and trajectory) significantly enhance anastomotic leak prediction accuracy.
• Dynamic, data-driven analysis of CRP offers improved early postoperative risk stratification for colorectal surgery patients.
• Further multi-center validation is warranted to confirm the generalizability of these ML-based predictive models.