Time-series deep learning radiomics for predicting post-radiotherapy rib fractures in non-small cell lung cancer

Background and purpose: Rib 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. This study aimed to develop and validate a deep learning model for predicting post-radiotherapy rib fracture using time-series CT radiomics. Material and methods: This study retrospectively collected CT scans from 67 NSCLC patients, comprising 1,605 individual ribs as separate instances. We proposed a novel Knowledge-aware Temporal Mixture of Experts (KA-TMoE) model that integrates radiomics from sequential CT scans to estimate fracture risk for each rib. Model performance was evaluated using area under the curve (AUC), sensitivity, specificity, and F1 score. Model interpretability was achieved using SHapley Additive exPlanations analysis, which attributed predictive value to each input feature. Results: The KA-TMoE model demonstrated strong predictive performance, achieving favorable AUC in the validation cohort (0.792). The DeLong test confirmed statistically significant improvements over ablation variants, underscoring the importance of integrating temporal data and domain knowledge. High sensitivity (0.85) and specificity (0.78) reflected a well-balanced trade-off, surpassing alternative approaches. Whitney U tests further supported its robustness, which showed significant differences in output distributions across cohorts. Among the top 20 most influential features, half originated from three-month postoperative radiomics, emphasizing the critical role of temporal information. Conclusion: The KA-TMoE model provides a robust, accurate framework for predicting rib fractures after SBRT in NSCLC patients. Its predictive power enables personalized risk assessment, better patient management, and optimized clinical prognosis.