Optimizing machine learning-based multimodal radiomics for predicting IDH status in gliomas: A SHAP-based multicenter study

Abstract

Background

Predicting isocitrate dehydrogenase (IDH) status is crucial in glioma management. Conventional MRI (cMRI) has limitations, but the clinical translation of radiomics and machine learning (ML) is often limited by single-center datasets and poor model interpretability.

Purpose

To develop and validate an interpretable, multicenter ML model integrating cMRI with functional sequences (DWI and PWI) for predicting IDH status in gliomas.

Material and Methods

This retrospective study included 180 patients from four institutions (150 training, 30 external test). Radiomics features were extracted from cMRI (T1WI, T2WI, FLAIR, T1CE), DWI, and DSC-PWI (CBV maps). After feature selection, multiparametric MRI-based fusion radiomics models were built and compared using three ML algorithms across four segmentation strategies. The optimal model was explained using SHapley Additive exPlanation (SHAP).

Results

The full-modality model (cMRI + DWI + PWI) with 3Dmodified segmentation achieved the best performance, with area under the curve of 0.840 (training) and 0.810 (external test). Incorporating functional sequences significantly improved prediction over cMRI alone. SHAP analysis identified key predictive features and provided individualized visual explanations for model decisions.

Conclusion

The developed ML-SHAP model, integrating conventional and functional MRI, reliably predicts IDH status and demonstrates generalizability across multiple centers. This interpretable tool shows potential for supporting preoperative molecular diagnosis in glioma.