Learning to Diagnose Privately: DP-Powered LLMs for Radiology Report Classification​

Scientific Achievement

  • DP-LoRA-enabled private multi-abnormality radiology classification with only moderate performance loss.​
  • Differential privacy reduced semantic memorization while preserving clinically useful classification performance.​

Significance and Impact​

  • Enables privacy-conscious clinical LLM adaptation without sharing sensitive patient training data.​
  • Compact private models can support hospital deployment with lower memory and GPU requirements.​

Technical Approach​

  • LoRA limited trainable parameters, reducing noisy updates required for differential privacy.​
  • DP-SGD clipped gradients and added Gaussian noise for formal privacy guarantees.​
  • Weak supervision generated multi-abnormality labels from specialized radiology language models.​
  • Experiments varied privacy budgets and LoRA ranks to quantify privacy-utility tradeoffs.​

Publication(s) for this work: Bhattacharjee, P., Tian, F., Rubin, G. D., Lo, J. Y., Merchant, N., Hanson, H. A., Gounley, J., & Tandon, R. (2026). Learning to Diagnose Privately: DP-Powered LLMs for Radiology Report Classification. IEEE Access, 14, 53501-53512. https://doi.org/10.1109/ACCESS.2026.3679277


Media Contacts