KA-TMoE: A Deep Learning Method Using Time-Series CT Radiomics to Predict Post-Radiotherapy Rib Fractures in NSCLC Patients.

Abstract

Background and purposeRib fracture is a recognized clinical complication in medically inoperable patients with non-small cell lung cancer (NSCLC) undergoing stereotactic body radiotherapy (SBRT), leading to diminished quality of life and delayed recovery. There remains an unmet need for reliable tools to predict rib fracture risk to support individualized prognosis. This study aimed to develop and validate a deep learning model for predicting post-SBRT rib fractures using time-series CT radiomics.Material and methodsThis retrospective study collected CT scans from three timepoints in 67 NSCLC patients, comprising over 1600 individual ribs. We proposed a novel Knowledge-aware Temporal Mixture of Experts (KA-TMoE) model that integrates longitudinal CT radiomics with radiomic grouping knowledge to estimate fracture risk at the rib level. Model performance and interpretability were evaluated.ResultsThe KA-TMoE model demonstrated favorable predictive performance in the validation cohort, achieving an area under the receiver operating characteristic curve (AUC) of 0.792. Exploratory DeLong testing was generally consistent with the observed performance differences between KA-TMoE and the ablation variants, suggesting that both longitudinal information and radiomics-grouping knowledge contributed to model performance. Mann-Whitney U tests demonstrated significant differences in model output distributions across cohorts. Time-to-event analysis showed that the model-predicted high-risk group had a higher risk of fracture than the low-risk group (hazard ratio = 10.82; p ConclusionKA-TMoE showed potential as a preliminary rib-level risk-stratification framework for predicting post-SBRT rib fractures in NSCLC patients. It may support earlier personalized risk stratification, closer surveillance, and timely supportive evaluation.