Development and Validation of Machine Learning Models to Predict Readmission After Colorectal Surgery

Area of Science:

• Surgical Outcomes Research
• Health Informatics
• Machine Learning in Medicine

Background:

• Colorectal surgery readmissions are frequent, leading to patient complications and increased healthcare costs.
• Traditional logistic regression models have shown limited accuracy in predicting these readmissions.
• Machine learning offers potential for improved prediction by identifying complex patterns in patient data.

Purpose of the Study:

• To develop a more accurate predictive model for 30-day readmission after colorectal surgery using machine learning techniques.
• To compare the performance of machine learning models against traditional logistic regression.

Main Methods:

• Utilized the National Quality Improvement Program (NSQIP) database (2012-2019) for patient data.
• Developed and compared three machine learning models: random forest (RF), gradient boosting (XGB), and neural network (NN).
• Evaluated model performance using the area under the receiver operating characteristic curve (AUROC) and compared against a logistic regression (LR) model.

Main Results:

• The dataset comprised 213,827 patients, with a 10.8% readmission rate.
• Neural network (NN) achieved the highest AUROC (0.751), outperforming logistic regression (LR) (0.684).
• Random forest (RF) and XGBoost (XGB) also showed improved performance (AUROCs 0.749 and 0.745, respectively). Key predictors included ileus, length of stay, surgical site infection, and ostomy placement.

Conclusions:

• Machine learning models significantly outperform traditional methods for predicting colorectal surgery readmissions.
• The developed machine learning model demonstrates potential for improving patient outcomes by enabling targeted interventions.
• External validation of this model could facilitate its clinical implementation to reduce readmission rates.