Federated-learning-based prognosis assessment model for acute pulmonary thromboembolism

Abstract Background Acute pulmonary thromboembolism (PTE) is a common cardiovascular disease and recognizing low prognosis risk patients with PTE accurately is significant for clinical treatment. This study evaluated the value of federated learning (FL) technology in PTE prognosis risk assessment while ensuring the security of clinical data. Methods A retrospective dataset consisted of PTE patients from 12 hospitals were collected, and 19 physical indicators of patients were included to train the FL-based prognosis assessment model to predict the 30-day death event. Firstly, multiple machine learning methods based on FL were compared to choose the superior model. And then performance of models trained on the independent (IID) and non-independent identical distributed(Non-IID) datasets was calculated and they were tested further on Real-world data. Besides, the optimal model was compared with pulmonary embolism severity index (PESI), simplified PESI (sPESI), Peking Union Medical College Hospital (PUMCH). Results The area under the receiver operating characteristic curve (AUC) of logistic regression(0.842) outperformed convolutional neural network (0.819) and multi layer perceptron (0.784). Under IID, AUC of model trained using FL(Fed) on the training, validation and test sets was 0.852 ± 0.002, 0.867 ± 0.012 and 0.829 ± 0.004. Under Real-world, AUC of Fed was 0.855 ± 0.005, 0.882 ± 0.003 and 0.835 ± 0.005. Under IID and Real-world, AUC of Fed surpassed centralization model(NonFed) (0.847 ± 0.001, 0.841 ± 0.001 and 0.811 ± 0.001). Under Non-IID, although AUC of Fed (0.846 ± 0.047) outperformed NonFed (0.841 ± 0.001) on validation set, it (0.821 ± 0.016 and 0.799 ± 0.031) slightly lagged behind NonFed (0.847 ± 0.001 and 0.811 ± 0.001) on the training and test sets. In practice, AUC of Fed (0.853, 0.884 and 0.842) outshone PESI (0.812, 0.789 and 0.791), sPESI (0.817, 0.770 and 0.786) and PUMCH(0.848, 0.814 and 0.832) on the training, validation and test sets. Additionally, Fed (0.842) exhibited higher AUC values across test sets compared to those trained directly on the clients (0.758, 0.801, 0.783, 0.741, 0.788). Conclusions In this study, the FL based machine learning model demonstrated commendable efficacy on PTE prognostic risk prediction, rendering it well-suited for deployment in hospitals.