An interpretable machine learning model based on MRI radiomics and GAME score for predicting early recurrence after...

To develop and validate machine learning models based on preoperative magnetic resonance imaging(MRI) and baseline clinical characteristics for predicting early recurrence(ER) in patients with colorectal liver metastases(CRLM) treated with thermal ablation(TA).

Patients with CRLM who underwent TA between January 2016 and December 2021 at two hospitals in China were allocated. Clinical and MRI data were used to develop and validate the clinical model, radiomics machine learning (R-ML) model, and combined clinical-radiomics model to predict ER after TA. The prognostic performance of the genetic and morphological evaluation (GAME) score and the Fong score was also compared (Supplementary Material). The best-performing algorithm among eight machine learning methods was selected to establish the R-ML model. Model performance was assessed through receiver operating characteristic (ROC) curve analysis, calibration plots, decision curve analysis (DCA), and survival analyses.

A total of 187 consecutive patients were enrolled (114 for the training cohort, 48 for the testing cohort, and 25 for the external test cohort). The GAME score showed better prognostic performance than the Fong score (Supplementary Material). The largest diameter of liver metastases (OR: 5.760, 95% CI: 2.130-16.700; P < 0.001) and the GAME group (OR: 0.093, 95% CI: 0.007-0.985; P = 0.040) were independent risk factors for ER. The XGBoost-based R-ML model performed best across cohorts. In external validation, the combined model (AUC = 0.772, P = 0.015) demonstrated superior predictive capacity to both the clinical (AUC = 0.647, P = 0.380) and R-ML models (AUC = 0.743, P = 0.056).

The combined model incorporating preoperative MRI-derived radiomics features and clinical parameters serves as a valuable tool for predicting ER risk in patients with CRLM undergoing TA therapy.