Balancing privacy and performance in healthcare: A federated learning framework for sensitive data

Area of Science:

• Artificial Intelligence
• Machine Learning in Healthcare
• Data Privacy

Background:

• Healthcare data is highly sensitive and requires robust privacy measures.
• Federated learning (FL) enables collaborative model training without centralizing data.
• Balancing privacy, model performance, and efficiency in FL for healthcare is challenging.

Purpose of the Study:

• To design and evaluate a privacy-preserving federated learning (PPFL) framework for sensitive healthcare data.
• To balance robust privacy, model performance, and computational efficiency.
• To promote user trust in machine learning applications within healthcare.

Main Methods:

• Integrated differentially private stochastic gradient descent (DPSGD) into a federated learning (FL) pipeline.
• Evaluated the framework on the Stroke Prediction Dataset.
• Measured model utility (accuracy, F1), privacy (ε), resource usage, and trust features against baselines.

Main Results:

• Achieved 93% accuracy in stroke risk prediction with a final privacy budget of ε 0.69.
• Demonstrated minimal computational overhead.
• Outperformed existing methods in privacy-utility trade-off and provided real-time privacy feedback.

Conclusions:

• The PPFL framework enables effective and trustworthy privacy-preserving machine learning in healthcare.
• The framework is suitable for resource-constrained settings.
• Future work includes extending model architectures, regulatory alignment, and user trust assessment.