Explainable machine learning to predict muscle loss during radiotherapy for oral cavity cancer.
Authors
Affiliations (8)
Affiliations (8)
- Department of Radiation Oncology, Changhua Christian Hospital, Changhua, Taiwan.
- Department of Radiation Oncology, MacKay Memorial Hospital, Taipei, Taiwan; Department of Death Care Service, MacKay Junior College of Medicine, Nursing, and Management, Taipei, Taiwan.
- Institute of Biomedical Informatics, National Yang Ming Chiao Tung University, Taipei, Taiwan.
- Department of Radiology, MacKay Memorial Hospital, Taipei, Taiwan.
- Department of Otolaryngology and Head Neck Surgery, MacKay Memorial Hospital, Taipei, Taiwan.
- Department of Radiation Oncology, MacKay Memorial Hospital, Taipei, Taiwan.
- Institute of Biomedical Informatics, National Yang Ming Chiao Tung University, Taipei, Taiwan; Ph.D. Program of Interdisciplinary Medicine, National Yang Ming Chiao Tung University, Taipei, Taiwan. Electronic address: [email protected].
- Department of Radiation Oncology, MacKay Memorial Hospital, Taipei, Taiwan; School of Medicine, College of Medicine, MacKay Medical University, New Taipei City, Taiwan. Electronic address: [email protected].
Abstract
Muscle loss during adjuvant radiotherapy is associated with poor survival outcomes in patients with oral cavity cancer (OCC). This study aimed to develop and validate an explainable machine learning model to predict muscle loss. This study included 1,024 patients with OCC (derivation cohort, 636 patients; external validation cohort, 388 patients) who underwent surgery and adjuvant radiotherapy between 2010 and 2021. Muscle mass was measured using computed tomography at the C3 vertebral level before and after radiotherapy, with "muscle loss" defined as a decrease of ≥4.2%. Random Forest (RF), eXtreme Gradient Boosting (XGBoost), and Categorical Boosting (CatBoost) models were trained using clinical and dosimetric data to predict muscle loss. Model performance was evaluated using the area under the curve (AUC). The SHapley Additive exPlanations method was applied for interpretation. Muscle loss occurred in 166 (26.1%) and 98 (25.3%) patients in the derivation and external validation cohorts, respectively. The RF model outperformed XGBoost and CatBoost in the external validation cohort (AUC: 0.913, 0.892, and 0.904, respectively). Top predictors included pre-radiotherapy Mini-Nutritional Assessment scores, mean radiation doses to the superior/middle pharyngeal constrictor muscle and supraglottic larynx, and chemotherapy. A nonlinear dose-toxicity relationship was observed between the mean dose to the swallowing structures and muscle loss. The model provided patient-level interpretations, identifying specific contributors for individual cases. An explainable model could predict muscle loss and identify specific risk factors. This approach may enable clinicians to tailor interventions and radiotherapy planning to mitigate muscle loss.