Development and validation of a deep transfer learning-based multivariable survival model to predict overall survival in lung cancer.

ABSTRACT

Background: Numerous deep learning-based survival models are being developed for various diseases, but those that incorporate both deep learning and transfer learning are scarce. Deep learning-based models may not perform optimally in real-world populations due to variations in variables and characteristics. Transfer learning, on the other hand, enables a model developed for one domain to be adapted for a related domain. Our objective was to integrate deep learning and transfer learning to create a multivariable survival model for lung cancer. Methods: We collected data from 601,480 lung cancer patients in the Surveillance, Epidemiology, and End Results (SEER) database and 4,512 lung cancer patients in the First Affiliated Hospital of Guangzhou Medical University (GYFY) database. The primary model was trained with the SEER database, internally validated with a dataset from SEER, and externally validated through transfer learning with the GYFY database. The performance of the model was compared with a traditional Cox model by C-indexes. We also explored the model's performance in the setting of missing data and generated the artificial intelligence (AI) certainty of the prediction. Results: The C-indexes in the training dataset (SEER full sample) with DeepSurv and Cox model were 0.792 (0.791-0.792) and 0.714 (0.713-0.715), respectively. The values were 0.727 (0.704-0.750) and 0.692 (0.666-0.718) after applying the trained model in the test dataset (GYFY). The AI certainty of the DeepSurv model output was from 0.98 to 1. For transfer learning through fine-tuning, the results showed that the test set could achieve a higher C-index (20% vs. 30% fine-tuning data) with more fine-tuning dataset. Besides, the DeepSurv model was more accurate than the traditional Cox model in predicting with missing data, after random data loss of 5%, 10%, 15%, 20%, and median fill-in missing values. Conclusions: The model outperformed the traditional Cox model, was robust with missing data and provided the AI certainty of prediction. It can be used for patient self-evaluation and risk stratification in clinical trials. Researchers can fine-tune the pre-trained model and integrate their own database to explore other prognostic factors.