Deep-Learning-Based Model for the Prediction of Cancer-Specific Survival in Patients with Spinal Chordoma.

Objective: Spinal chordomas are locally aggressive and frequently recurrent tumors with a poor prognosis. Previous studies focused on a Cox regression model to predict the survival of patients with spinal chordoma. We aimed to develop a more effective model based on deep learning for prognosis prediction in spinal chordoma.

Methods: Patients with spinal chordoma were gathered from the SEER database. Cox regression analysis was conducted to compare the influence of different clinical characteristics on cancer-specific survival. Two deep learning models, namely, DeepSurv and NMTLR, were developed, alongside 2 classic models, for the purpose of comparison. Performance of these models was evaluated by concordance index, Integrated Brier Score, receiver operating characteristic curves, Kaplan-Meier curves, and calibration curves.

Results: A total of 258 spinal chordoma patients were included in the current study. The median follow-up time was 94 ± 52 months. Variables used for prognosis prediction consisted of age, primary site, tumor size, histologic grade, extension of surgery, tumor invasion, and metastasis. Comparing with conventional models, each deep learning model showed superior predictive performance, the C-index on the test cohort is 0.830 for DeepSurv and 0.804 for NMTLR, respectively. The DeepSurv model represented the best performance, with area under the curve of 0.843 in predicting 5-year survival and 0.880 in predicting 10-year survival.

Conclusions: We successfully constructed a deep learning model to predict the CSS of spinal chordoma patients and proved that it was more accurate and practical than conventional prediction model. Our deep learning model has the potential to guide clinicians in better care planning and decision-making.