Privacy-preserving federated learning models for healthcare data

Abstract

This study investigates privacy-preserving Federated Learning (FL) models for analyzing multi-institutional healthcare data without sharing sensitive patient information. The research addresses the growing demand for collaborative model development while ensuring compliance with privacy regulations. Data were collected from five tertiary care hospitals in the United States, comprising 48,500 anonymized patient records related to cardiovascular disease. A federated deep learning framework was implemented, incorporating homomorphic encryption and differential privacy to secure model updates. Performance was evaluated using accuracy, precision, recall, F1-score, and area under the ROC curve (AUC). Statistical analyses, including paired t-tests, ANOVA, chi-square tests, and Pearson correlation, were employed to assess differences in model performance and associations between clinical features and outcomes. Results showed that privacy-preserving techniques caused a small but statistically significant reduction in accuracy (-1.6%) and recall (-2.2%), while maintaining strong predictive capability (AUC ? 0.94). Key predictors included blood pressure, age, and cholesterol levels. These findings demonstrate that privacy-preserving FL offers a practical balance between data protection and predictive accuracy, supporting its potential for secure healthcare analytics.