Enhanced Diabetes Detection via a Privacy-Preserving Federated Learning Framework

Background:The integration of artificial intelligence (AI) in healthcare depends on striking a balance between patient privacy and clinical utility.The standard methods often compromise one for the other, preventing the development of trustworthy healthcare AI.Objective: This paper aims to resolve the privacy-utility trade-off by developing an enhanced federated learning framework with adaptive differential privacy (DP) mechanisms that are optimized for clinical data.Methods: We implement and compare several different methods, from the most centralized deep learning to various federated configurations with formal DP guarantees.Our improved framework involves adaptive noise scheduling and quality-weighted federated averaging on top of a federated neural network framework.We validate on two major diabetes screening datasets: Diabetes Health Indicators (BRFSS 2015) and National Health and Nutrition Examination Survey (NHANES 2015-2016), including comprehensive clinical measurements.Results: This paper presents a favourable balance between privacy protection and clinical utility for both datasets.It offers strong formal differential privacy guarantees and good diagnostic performance, achieving high ranking accuracy with clinical risk prioritization.The model demonstrates generalization robustness by capturing clinically meaningful risk factors aligned with established medical guidelines, confirming that the applied privacy-preserving mechanisms do not compromise clinical relevance.Conclusion: Our framework meaningfully advances the privacy-utility trade-off healthcare AI, by offering tunable formal privacy guarantees while ensuring strong clinical performance.The approach is highly generalizable across diverse data collection methodologies and maintains clinically relevant feature representations, thus allowing safe adoption in sensitive medical domains.