Evaluation of Federated Learning Using Standardized EHR Data in Japan.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Surgical Outcomes Research

Background:

• Multi-institutional data sharing is crucial for developing robust predictive models in healthcare.
• Patient privacy concerns often hinder the sharing of sensitive electronic health record (EHR) data.
• Predicting prolonged air leaks (PAL) after video-assisted thoracoscopic surgery (VATS) is clinically significant.

Purpose of the Study:

• To apply federated learning (FL) for developing a predictive model for PAL after VATS.
• To evaluate the feasibility of FL in multi-institutional data sharing while preserving patient privacy.
• To assess the performance of an FL-based predictive model compared to centralized approaches.

Main Methods:

• Utilized standardized EHR data from two Japanese hospitals.
• Implemented federated learning (FL) to train a predictive model by exchanging only model parameters.
• Ensured patient privacy by not sharing underlying patient data.

Main Results:

• Achieved high discriminatory accuracy for predicting PAL using FL.
• Demonstrated that FL can enhance predictive model accuracy in healthcare settings.
• Observed that FL models may be influenced by data volume from participating institutions.

Conclusions:

• Federated learning (FL) offers a viable solution for privacy-preserving multi-institutional data sharing in healthcare.
• FL can facilitate the development of accurate predictive models for surgical outcomes like PAL.
• Further validation is needed to address potential biases in FL models related to data volume.