Surgical Aggregation: A Federated Learning Framework for Harmonizing Distributed Datasets with Diverse Tasks

Many large-scale chest x-ray datasets have been curated for the detection of abnormalities using deep learning, with the potential to provide substantial benefits across many clinical applications. However, these datasets focus on detecting a subset of disease labels that could be present, thus limiting their clinical utility. Furthermore, the distributed nature of these datasets, along with data sharing regulations, makes it difficult to share and create a complete representation of disease labels. To that end, we propose surgical aggregation, a federated learning framework for aggregating and harmonizing knowledge from distributed datasets with different disease labels into a 'global' deep learning model. We utilized surgical aggregation to harmonize the NIH (14 labels) and CheXpert (13 labels) datasets into a global model with the ability to predict all 20 unique disease labels and compared it to the performance of 'baseline' models trained individually on both datasets. We observed that the global model resulted in excellent performance across held-out test sets from both datasets with an average AUROC of 0.75 and 0.74 respectively when compared to the baseline average AUROC of 0.81 and 0.71. On the MIMIC external test set, we observed that the global model had better generalizability with average AUROC of 0.80, compared to the average AUROC of 0.74 and 0.76 respectively for the baseline models. Our results show that surgical aggregation has the potential to develop clinically useful deep learning models by aggregating knowledge from distributed datasets with diverse tasks -- a step forward towards bridging the gap from bench to bedside.