Machine learning based on body composition radiomics for predicting early recurrence in colorectal cancer: a multicenter study.

BACKGROUND

Early recurrence (ER) in colorectal cancer (CRC) leads to dismal outcomes. Current pTNM staging fails to capture the host's systemic pathophysiological status. We developed an interpretable machine learning (ML) model based on preoperative CT body composition radiomics to predict ER in CRC.

METHODS

This multicenter study enrolled 917 patients who underwent radical resection across three independent institutions, and the cohort was partitioned into a training set (n = 548) and two external test sets (n = 263 and n = 106). We extracted 1,896 radiomic features from four body composition compartments: skeletal muscle (SM), subcutaneous adipose tissue (SAT), intermuscular adipose tissue (IMAT), and visceral adipose tissue (VAT) at the L3 level on CT. Following feature selection using LASSO and Boruta algorithms, eight ML algorithms were evaluated. The optimal classifier was integrated with clinical risk factors. SHAP was utilized for model interpretability.

RESULTS

An 11-feature radiomics signature was identified. The Random Forest model demonstrated optimal generalization, yielding AUCs of 0.807, 0.776, and 0.750 in the training and two test sets, respectively. SHAP analysis revealed that IMAT (46.1%) and SM (42.9%) features were primary drivers, and increased SM textural uniformity may reflect adverse muscle quality and possible myosteatosis-related tissue alterations. The individualized radiomics risk score effectively stratified patients, demonstrating significantly divergent recurrence-free and overall survival across all cohorts (p < 0.05). Decision curve analysis confirmed superior net clinical benefit over pTNM staging.

CONCLUSION

This interpretable ML model may improve ER risk stratification in CRC and provide a quantitative tool for individualized postoperative surveillance.