Ensemble learning-based radiomics model for predicting radiation-induced temporal lobe injury in nasopharyngeal...

To develop and validate a multimodal ensemble machine learning model integrating multi-sequence magnetic resonance imaging (MRI) radiomics, clinical characteristics, and hematological biomarkers for early prediction of Radiation-induced temporal lobe injury (RTLI) in Nasopharyngeal carcinoma (NPC) patients before radiotherapy.

A total of 161 NPC patients treated with intensity-modulated radiation therapy (IMRT) were retrospectively analyzed and randomly assigned to training (n = 113) and validation (n = 48) sets in a 7:3 ratio. Radiomic features were extracted from pretreatment T1WI, CE-T1WI, T2WI, and DWI, with features showing ICC > 0.75 retained. After SMOTE balancing, Elastic Net (EN) was used for feature selection to generate EN-scores, and Random Forest (RF) produced RF-scores. These, together with two SVM-based scores obtained from demographic and hematological biomarkers, were combined into an ensemble ERSS (EN-RF-SVM-SVM) model. Model performance was evaluated using ROC analysis, calibration, and decision curve analysis.

The ERSS model demonstrated superior predictive performance compared with single-sequence, multi-sequence MRI integration models and LR model. The AUCs of the ERSS model were 0.957 in the training set and 0.968 in the validation set. Calibration curves showed excellent agreement between predicted and observed outcomes. DCA indicated that the ERSS model provided the highest net clinical benefit across a wide range of threshold probabilities compared with other models.

The ERSS multimodal ensemble learning model provides a highly accurate and clinically meaningful tool for early prediction of RTLI in NPC patients. By integrating multi-sequence MRI radiomics, hematological biomarkers, and clinical factors, the ERSS model enables individualized risk assessment and may assist in optimizing radiotherapy planning and follow-up strategies.