Prognostic prediction model for rectal cancer based on CMS subtype indicators and SHAP-based interpretable analysis

To develop an interpretable radiomics model for predicting 3-year progression-free survival (PFS) in rectal cancer (RC) by integrating consensus molecular subtype (CMS) indicators and Shapley additive explanations (SHAP) analysis with ultra-high b-value diffusion-weighted imaging (DWI).

This retrospective study included 153 RC patients randomly assigned to training (n = 107) and validation (n = 46) cohorts. All patients underwent 3.0T DWI with b-values of 0, 1500, 2000, and 3000 s/mm². Radiomic features were extracted from stretched exponential model parametric maps. Radscore1 (Conventional radiomic features ) and Radscore2 (CMS surrogate-related features ) were separately calculated with least absolute shrinkage and selection operator Cox regression (LASSO-Cox). Five prognostic models were constructed with clinical factors and Radscore in evaluating 3-year PFS. Model performance was evaluated using time-dependent receiver operating characteristic (ROC) curves, calibration curves, and net reclassification improvement (NRI) analysis.

In the validation cohort, radscore2 (area under the curve (AUC) = 0.787) outperformed radscore1 (AUC = 0.751) in predicting 3-year PFS. Model 5 (clinical + radscore2) achieved the highest AUC (0.823) with good calibration. NRI analysis showed significant improvement for model 5 compared with the clinical model (NRI = 0.540, p = 0.001). SHAP analysis identified α_logarithm_ngtdm_Contrast (0.463) and α_gradient_firstorder_TotalEnergy (0.445) as the most important features.

The integration of CMS-surrogate guided feature selection and SHAP analysis could construct an interpretable radiomics model with good performance in predicting 3-year PFS in rectal cancer.