Privacy-preserving Model Training for Disease Prediction Using Federated Learning with Differential Privacy

Area of Science:

• Health Informatics
• Machine Learning in Healthcare
• Computational Biology

Background:

• Machine learning (ML) requires extensive data for robust model development in health sciences.
• Sharing sensitive patient data across institutions for ML is restricted by privacy concerns.
• Federated learning (FL) offers a solution for decentralized model training.

Purpose of the Study:

• To develop and evaluate a federated learning algorithm incorporating differential privacy.
• To ensure data privacy during the training of ML models on distributed datasets.
• To assess the feasibility of applying this framework in clinical settings.

Main Methods:

• Implemented a federated learning algorithm with differential privacy.
• Trained a neural network model on gene expression data for breast cancer prediction.
• Compared the performance of the private federated model against a non-private single-site model.

Main Results:

• The federated learning model with differential privacy achieved comparable accuracy and precision to a non-private model.
• The algorithm effectively preserved privacy while training on distributed data.
• Demonstrated successful application in predicting breast cancer status.

Conclusions:

• The proposed federated learning algorithm with differential privacy is effective for privacy-preserving ML in healthcare.
• This framework allows clinical data scientists to build private models on federated datasets.
• The approach facilitates collaborative research without compromising patient confidentiality.