Identifying pre-treatment risk factors for cancer-related cognitive decline in patients with breast cancer

PURPOSE: To explore pre-treatment risk factors for cognitive decline in patients with breast cancer using a machine learning approach applied to a comprehensive multimodal clinical, biological, and neuroimaging dataset. METHODS: Sixty-seven women with early breast cancer were assessed at diagnosis (T0), 8 months (T1), and 17 months after diagnosis (T2). Cognitive decline was defined using the Reliable Change Index. Patients were classified as showing no decline or decline at either follow-up. Five feature sets were evaluated: (1) patient characteristics, treatment, and psychosocial measures; (2) inflammatory and neural health markers, (3) structural brain volumes, (4) resting-state functional MRI connectivity, and (5) diffusion MRI measures. Random forest models were first trained on feature set 1, then sequentially combined with sets 2-5 to explore their additive predictive value. Each model underwent standardized preprocessing, recursive feature selection of the top 6 predictors, and tuning before random forest classification. A final composite model was constructed by pooling the six top predictors from each feature set to assess potential complementary multimodal information. Feature contributions were examined using SHAP values. RESULTS: Of 67 patients, 33 (49%) experienced cognitive decline following treatment. Models achieved prediction accuracies of 76%, improving up to 81% when MRI measures and/or serum markers were included. Key baseline predictors of cognitive decline included more aggressive subtypes, planned systemic therapy, perceived stress, and limited cognitive and brain reserve. CONCLUSION: Machine learning explored potential pre-treatment risk factors for cancer-related cognitive decline in patients with breast cancer. These findings highlight potential risk factors that could support risk-stratification.