A Novel Deep Learning Model for Predicting Colorectal Anastomotic Leakage: A Pioneer Multicenter Transatlantic Study.
Miguel Mascarenhas1,2,3,4,5, Francisco Mendes1,2,3,5, Filipa Fonseca6
1Precision Medicine Unit, Department of Gastroenterology, Unidade Local de Saúde São João, 4200-319 Porto, Portugal.
Journal of Clinical Medicine
|August 14, 2025
Summary
This study developed an AI model using intraoperative video to predict colorectal anastomotic leaks (CAL). The model achieved 99.6% accuracy, offering real-time risk assessment to potentially prevent complications.
Area of Science:
- Surgical Innovation
- Artificial Intelligence in Medicine
- Medical Imaging Analysis
Background:
- Colorectal anastomotic leak (CAL) is a severe complication impacting patient outcomes.
- Current CAL risk assessment lacks real-time predictive capabilities.
- Need for advanced tools to improve surgical safety in colorectal procedures.
Purpose of the Study:
- To develop an artificial intelligence (AI) model for predicting colorectal anastomotic leak (CAL).
- To utilize intraoperative laparoscopic video recordings for real-time CAL risk assessment.
- To enhance surgical decision-making and patient safety during colorectal surgery.
Main Methods:
- A convolutional neural network (CNN) was trained on annotated frames from colorectal surgery videos.
- Data from three international centers included 5356 frames from 26 patients (2007 with CAL, 3349 normal).
- Four CNN architectures were evaluated, with performance measured by accuracy, sensitivity, specificity, and AUROC.
Main Results:
- The best-performing AI model achieved 99.6% accuracy and 99.6% AUROC.
- High sensitivity (99.2%) and specificity (100.0%) demonstrated reliable CAL prediction.
- Heatmaps provided visual explanations, identifying critical regions influencing the model's predictions.
Conclusions:
- This is the first AI model for CAL prediction using intraoperative video analysis.
- The model's high accuracy suggests a potential paradigm shift in surgical decision-making.
- Further validation with larger datasets could enable real-time interventions to prevent CAL.
