Performance vs. Privacy: Evaluating the Performance of Predicting Second Primary Cancer in Lung Cancer Survivors with Privacy-preserving Approaches

Abstract

Deep learning has been widely used in the medical field to support medical decision making. Simultaneously, with the rise of data privacy protection, accessing clinical records across different institutions has become a possible challenge. Several approaches, such as federated and transfer learning, have been proposed to train models without accessing all the records from each institution, but the performance of these privacy-preserved models may not be as good as centralized approaches, which aggregate all records to build a centralized model. To explore the potential of privacy-preserving second primary cancer (SPC) prediction of lung cancer survivors using real-world data, we evaluated the performance of federated learning, transfer learning, learning with a single institution, and traditional centralized learning. We trained machine learning models using data from four hospitals and compared the model performances of learning from a single institution, centralized learning, federated learning, and transfer learning. The results show that federated learning outperformed other learning strategies in three of the four sites (AUROC from 0.733 to 0.777). However, only Site 6 showed that federated learning significantly outperformed all the other learning strategies (P < 0.05). In summary, federated learning can develop a unified model for the multiple institutions while maintaining data security.