A. Gurau1, M. Brock1, J. Ha1 1Johns Hopkins University School Of Medicine, Thoracic Surgery, Baltimore, MD, USA
Introduction:
Anastomotic leak is a significant complication after esophagectomy. This study aimed to develop and validate a machine-learning model to predict anastomotic leaks after esophagectomy using a large multivariate clinical dataset.
Methods:
IRB exemption was obtained. The NSQIP Esophagectomy Targeted participant user data file (PUF) for 2016-2019 was utilized, comprising 4,350 patients. The dataset was split into training (80%, n=3,480) and validation (20%, n=870) subsets. A distributed random forest model was fit to the training data. Variables included demographic characteristics, preoperative workup and treatment-related factors, and in-hospital parameters. Model performance was assessed on both training and validation subsets using accuracy, sensitivity, specificity, and area under the ROC curve.
Results:
The mean patient age was 64 years (standard deviation 10.4). The distributed random forest model predicted esophageal anastomotic leaks with the following performance: Training subset: accuracy 88.48%, specificity 90.25%, sensitivity 77%, area under the receiver operating characteristic curve (AUC-ROC) 0.93. Validation subset: accuracy 87.2%, specificity 92.2%, sensitivity 49.5%, AUC-ROC 0.77. At 95% specificity, sensitivity was 33%, demonstrating a significant trade-off between specificity and sensitivity.
Conclusion:
The random forest model demonstrated good accuracy and specificity in predicting esophageal anastomotic leaks, suggesting the capability to model complex multivariate clinical data. However, the model's decreased sensitivity and AUC during validation indicate limitations in reliably ruling out leaks. Future improvements may include identifying additional predictive variables and training with a larger dataset. With further refinement, this model shows potential for clinical risk stratification.