Development and Validation of Deep Learning Models for Ovarian Clear Cell Carcinoma Survival

Abstract

AbstractBackground Ovarian clear cell carcinoma (OCCC) is a rare and distinct histologic subtype of epithelial ovarian carcinomas. Few studies have explored the use of deep learning models for predicting survival in OCCC. Our study aims to compare the performance of deep learning models with a multivariate Cox regression model in predicting survival for OCCC patients. Methods In this population-based cohort study, we extracted 926 patients diagnosed with OCCC between 2010 and 2017 from the Surveillance, epidemiology, and end results (SEER) database. Three algorithms, including DeepSurv and neural multi-task logistic regression (NMTLR) based on neural networks, and RSF based on ensemble learning, were chosen for training. Additionally, a Cox proportional hazard model was constructed for comparison purposes. The algorithm was externally validated on an independent test cohort, comprising 134 OCCC patients diagnosed between January 2005 and July 2021 in Qilu Hospital of Shandong University. The model's performance was assessed using the C-index and IBS (Integrated Brier Score), while the accuracy of predicting 1-, 3-, and 5-year survival was evaluated using ROC and AUC. Furthermore, a user-friendly interface was developed to facilitate the use of deep learning models for predicting survival. Results The deep learning model has demonstrated promising results in predicting overall survival (OS) for OCCC patients, outperforming the Cox proportional hazard model. DeepSurv consistently exhibited superior prediction performance compared to the Cox proportional hazard model in both the SEER training set (C-index: 0.781 vs. 0.724) and the independent China test set (C-index: 0.836 vs. 0.829). Additionally, the DeepSurv model displayed significantly higher AUC values for 3-year and 5-year OS in the China cohort when compared to the Cox proportional hazard model (AUC for 3-year OS: 0.844 vs. 0.836; AUC for 5-year OS: 0.821 vs. 0.817). Moreover, we developed a user-friendly graphical interface that allows for visualization of the deep learning model. Conclusions This study appears that deep learning models hold more promising than traditional linear regression models in predicting OS in OCCC patients. However, it is important to note that further large-scale, real-world studies are required to validate and substantiate this model.