Aggregate and transfer knowledge of functional connectivity of brain for detecting autism spectrum disorder for multi-site research

Abstract

Background

The optimization of deep learning models significantly relies on the availability of extensive, high-quality data, a requirement often challenging in multi-institutional healthcare research due to data privacy issues. Federated learning, a novel paradigm enabling model training on distributed datasets without the necessity for data sharing, emerges as a viable solution to this challenge.

Methods

This study explores the application of federated learning towards the detection of autism spectrum disorder by harnessing the functional connectivity attributes of resting state-fMRI data sourced from the ABIDE repository, encompassing contributions from 20 data sets of 17 distinct medical institutions. The performance of models trained through federated learning was compared against two benchmark models - one assimilating data from all participating institutions and the other utilizing data from a single institution. Our methodology involved the implementation of two federated learning models, namely equivalent averaging and weighted averaging, with the latter reflecting the extent of participation from the respective institutions.

Results

Experimental results of baseline models indicated that 20 detection models, each trained using an institution’s own data, yielded a mean accuracy of 0.507 (mean precision of 0.493 and mean AUC of 0.511), while the detection model trained on the integrated dataset achieved an accuracy of 0.68 (precision of 0.72 and AUC of 0.68). The federated learning models employing equivalent averaging and weighted averaging achieved mean accuracies of 0.662 (precision of 0.654 and AUC of 0.664) and 0.647 (precision of 0.649 and AUC of 0.652), respectively. While local models trained on each site's own data achieved only 74.5% accuracy compared to models trained on integrated data, both equivalent averaging federated learning and weighted averaging federated learning approached over 95% of the performance of models trained on integrated data (97.4% and 95.1%, respectively). However, there was no clear superiority between the two federated learning models in the evaluation performances. For most performance metrics, the federated models outperformed the single local models, and in some instances, federated models outperformed the integrated models.

Conclusions

This study employed federated learning for the detection of autism spectrum disorder using functional connectivity features extracted from fMRI, aiming to elucidate the potential of federated learning in multicenter clinical trials. By integrating and propagating knowledge modeled with data from each participating institution, federated learning proved to be effective in enhancing model performance in scenarios where large-scale data collection is challenging due to prevalent privacy concerns.