Differential privacy for medical deep learning: methods, tradeoffs, and deployment implications.

Area of Science:

• Medical Deep Learning
• Privacy-Preserving Technologies
• Health Informatics

Background:

• Differential privacy (DP) is crucial for safeguarding sensitive patient data in medical deep learning (DL).
• Implementing DP effectively without sacrificing clinical utility or equity presents significant challenges.
• Existing research lacks a comprehensive synthesis of DP applications in medical DL.

Purpose of the Study:

• To conduct a scoping review of differential privacy applications in medical deep learning.
• To analyze DP's impact on clinical utility and equity in both centralized and federated learning settings.
• To identify current gaps and future directions for privacy-preserving medical DL.

Main Methods:

• A systematic search strategy was employed to identify relevant studies up to March 2025.
• Included studies spanned various medical data modalities and deep learning tasks.
• Data extraction focused on DP mechanisms, privacy budgets, performance metrics, and fairness evaluations.

Main Results:

• DP, particularly DP-SGD, can preserve clinically acceptable performance at moderate privacy levels (ϵ ≈ 10), especially in medical imaging.
• Strict privacy settings (ϵ ≈ 1) often result in significant accuracy degradation, more pronounced in smaller or heterogeneous datasets.
• Fairness evaluations are infrequent, with some studies indicating DP may exacerbate subgroup performance disparities.

Conclusions:

• DP is a viable tool for medical DL, but careful parameter tuning is needed to balance privacy and utility.
• There is a critical need for standardized fairness auditing and reporting in privacy-preserving medical DL.
• Future research should prioritize developing equitable and clinically robust DP methods for diverse medical datasets.