Purpose:
To develop and evaluate an attention-based multimodal model for improving the prediction of overall survival in patients with glioblastoma by integrating clinical and imaging features.

Material and methods:
Data from 358 glioblastoma patients were obtained from The Cancer Imaging Archive. The dataset included clinical data and three magnetic resonance imaging (MRI) sequences. Three categories of features – clinical, radio­mics, and deep features – were used for survival prediction. Radiomics features were extracted using PyRadiomics, while deep features were extracted from MRI images using a Swin UNEt TRansformers. An attention-based multi­modal predictive model was designed to integrate these three feature sets. The DeepHit framework was used for survival modeling. Model performance was evaluated using the concordance index (C-index) and the Brier score. In addition, accuracy, precision, recall, area under the curve (AUC), and F1-score were calculated to predict mortality risk at different time intervals.

Results:
The proposed model achieved a C-index of 0.91 and a Brier score of 0.14. For identifying mortality risk within 6 months, the model achieved an accuracy of 0.93, precision of 0.87, recall of 0.92, AUC of 0.96, and an F1-score of 0.89. For mortality risk within 12 months, the corresponding values were 0.91, 0.87, 0.85, 0.92, and 0.88, respectively. For mortality risk within 24 months, the model achieved an accuracy of 0.75, precision of 0.71, recall of 0.70, AUC of 0.94, and an F1-score of 0.80.

Conclusions:
The proposed model, leveraging a combined radiomics-based approach, demonstrated the capability to predict overall survival in patients with glioblastoma. However, to ensure generalizability, further validation using multicenter datasets is required.