Machine learning models for predicting surgical intervention in colorectal cancer

Aim

We aimed to develop and validate a machine learning (ML) model to predict surgical intervention in colorectal cancer (CRC) patients in the state of São Paulo, Brazil, using clinical and sociodemographic data as predictors.

Methods

We conducted a longitudinal analysis using data from the Fundação Oncocentro de São Paulo (FOSP) database, which included CRC cases diagnosed between 2000 and 2023. We defined the primary outcome as surgical intervention and analyzed 29 predictor variables, including clinical, demographic, and socioeconomic factors. We evaluated six ML algorithms (Random Forest, Gradient Boosting, LightGBM, CatBoost, Logistic Regression, and Decision Trees). Data was divided into training (70 %) and test (30 %) sets and preprocessing steps were applied, including normalization, one-hot encoding, and addressing class imbalance. We assessed model performance using AUC-ROC, accuracy, precision, recall, F1-score, and specificity. SHAP was used to interpret variable importance.

Results

The dataset comprised 72,038 participants, 17,852 in the group that did not undergo surgery and 54,186 in the group that did. The Random Forest model achieved the highest performance, with an AUC of 0.94, accuracy of 0.82, and F1-score of 0.87. Key predictors included treatment-related factors (e.g., time between diagnosis and treatment), tumor stage, age, and socioeconomic indicators (e.g., municipal human development index). Geographic accessibility, such as travel time to healthcare facilities, also significantly influenced predictions.

Conclusion

This study demonstrates the potential of ML models, particularly Random Forest, to predict surgical necessity in CRC patients by integrating clinical and sociodemographic data.