AI model integrating dosimetric and radiomic features for predicting cognitive impairment after brain radiotherapy.
Authors
Affiliations (3)
Affiliations (3)
- Beijing Technology and Business University, Key Laboratory of Geriatric Nutrition and Health, Ministry of Education, Beijing, 100048, China. Electronic address: [email protected].
- Department of Respiratory and Critical Care Medicine, The First Affiliated Hospital of Nanjing Medical University, Nanjing, 210029, China. Electronic address: [email protected].
- Beijing Technology and Business University, Key Laboratory of Geriatric Nutrition and Health, Ministry of Education, Beijing, 100048, China. Electronic address: [email protected].
Abstract
Cognitive impairment is one of the most debilitating late effects of brain radiotherapy, affecting memory, attention, and executive function in a substantial proportion of survivors. Accurate prospective identification of patients at high risk would enable adaptive dose constraints, prophylactic neuroprotective interventions, and personalised cognitive rehabilitation planning.To develop and validate a multimodal artificial intelligence model that integrates dosimetric parameters, MRI-derived radiomic features, and clinical variables for predicting radiation-induced cognitive impairment at six months after completion of brain radiotherapy.A retrospective cohort of 412 patients who received brain radiotherapy between 2016 and 2023 was assembled from two tertiary neuro-oncology centres. Cognitive impairment was defined as a decline of at least 1.5 standard deviations from baseline on at least two domains of a standardised neuropsychological battery at six months. Dosimetric features were extracted from treatment planning systems for seven organs at risk. Radiomic features (n = 1316) were extracted from pre-treatment T1-weighted and T2-FLAIR MRI using PyRadiomics. Feature selection combined LASSO regression with recursive feature elimination. A gradient-boosted ensemble model with attention-based feature fusion was trained on 289 patients and independently validated on 123 patients. The proposed AI Fusion model achieved an AUC of 0.871 (95% CI 0.821-0.921) on the independent validation set, significantly outperforming dosimetric-only (AUC 0.714), radiomic-only (AUC 0.748), and clinical-only (AUC 0.682) models. Sensitivity was 81.4% and specificity was 79.6% at the optimal threshold. SHAP analysis identified mean hippocampal dose (D_mean), wavelet-LLH first-order energy from hippocampal region, GLCM entropy, and V30Gy as the four most predictive features. Decision curve analysis demonstrated net benefit superiority of the fusion model across threshold probabilities from 15% to 75%. Integrating dosimetric and radiomic features into a gradient-boosted AI model substantially improves prediction of cognitive impairment after brain radiotherapy. The model provides a clinically actionable risk stratification tool that could guide individualised treatment planning and post-treatment surveillance.