JMIR Publications

ABSTRACT

Background:

Mood disorder has emerged as a serious concern for public health, especially, bipolar disorder has a less favorable prognosis than depression. Although prompt recognition of depression conversion to bipolar disorder is needed, early prediction is challenging due to overlapping symptoms. Recently, there are attempts to develop the prediction model using federated learning. Federated learning in medical fields is a method for training multi-institutional machine learning models without patient-level data sharing.

Objective:

This study aimed to develop and validate a federated, differentially private multi-institutional bipolar transition prediction model.

Methods:

This retrospective study enrolled patients diagnosed with a first depressive episode at five tertiary hospitals in South Korea. We developed model for predicting bipolar transition using data from four institutions, 17,631 patients. Also, we used data from 4,541 patients for external validation from one institution. We created standardized pipelines to extract large-scale clinical features from four institutions without any code modification. Further, we performed feature selection in a federated environment for computational efficiency and applied differential privacy to gradient updates. The study performed feature selection in a federated environment and applied differential privacy to gradient updates. We compared the federated and four local models developed with each hospital data on internal and external validation datasets.

Results:

In the internal dataset, 279 out of 17,631 patients showed bipolar disorder transition. In the external dataset, 60 out of 4,541 patients showed bipolar disorder transition. The average performance of the federated model in the internal test (area under the curve [AUC]: 0.726) and external validation (AUC: 0.719) datasets was higher than other locally developed models (AUC: 0.642-0.707 and AUC: 0.642-0.699, respectively).

Conclusions:

This study developed and validated a differentially private federated model using distributed multi-institutional psychiatric data with standardized pipelines in a real-world environment. The federated model performed better than models using local data only.