A deep learning model for accurately predicting cancer-specific survival in patients with primary bone sarcoma of the extremity: a population-based study

• Published in: Clinical & translational oncology 2024-03, Vol.26 (3), p.709-719
• Main Authors: Cheng, Debin, Liu, Dong, Li, Xian, Mi, Zhenzhou, Zhang, Zhao, Tao, Weidong, Dang, Jingyi, Zhu, Dongze, Fu, Jun, Fan, Hongbin
• Format: Article
• Language: English
• Subjects: Bone Neoplasms - secondary; Deep Learning; Extremities - pathology; Humans; Medicine; Medicine & Public Health; Nomograms; Oncology; Osteosarcoma - pathology; Osteosarcoma - therapy; Prognosis; Research Article; Sarcoma - pathology
• ISSN: 1699-3055; 1699-3055
• DOI: 10.1007/s12094-023-03291-6

Purpose Primary bone and joint sarcomas of the long bone are relatively rare neoplasms with poor prognosis. An efficient clinical tool that can accurately predict patient prognosis is not available. The current study aimed to use deep learning algorithms to develop a prediction model for the prognosis of patients with long bone sarcoma. Methods Data of patients with long bone sarcoma in the extremities was collected from the Surveillance, Epidemiology, and End Results Program database from 2004 to 2014. Univariate and multivariate analyses were performed to select possible prediction features. DeepSurv, a deep learning model, was constructed for predicting cancer-specific survival rates. In addition, the classical cox proportional hazards model was established for comparison. The predictive accuracy of our models was assessed using the C-index, Integrated Brier Score, receiver operating characteristic curve, and calibration curve. Results Age, tumor extension, histological grade, tumor size, surgery, and distant metastasis were associated with cancer-specific survival in patients with long bone sarcoma. According to loss function values, our models converged successfully and effectively learned the survival data of the training cohort. Based on the C-index, area under the curve, calibration curve, and Integrated Brier Score, the deep learning model was more accurate and flexible in predicting survival rates than the cox proportional hazards model. Conclusion A deep learning model for predicting the survival probability of patients with long bone sarcoma was constructed and validated. It is more accurate and flexible in predicting prognosis than the classical CoxPH model.