Why patient data cannot be easily forgotten?

Ruolin Su, Xiao Liu, Sotirios A. Tsaftaris

Abstract

Rights provisioned within data protection regulations permit patients to request that knowledge about their information be eliminated by data holders. With the advent of AI learned on data, one can imagine that such rights can extend to requests for forgetting knowledge of patient's data within AI models. However, forgetting patients' imaging data from AI models is still an under-explored problem. In this paper, we study the influence of patient data on model performance and formulate two hypotheses for a patient's data: either they are common and similar to other patients or form edge cases, i.e., unique and rare cases. We show that it is not possible to easily forget patient data. We propose a targeted forgetting approach to perform patient-wise forgetting. Extensive experiments on the benchmark Automated Cardiac Diagnosis Challenge dataset showcase the improved performance of the proposed targeted forgetting approach as opposed to a state-of-the-art method.

Introduction

As artificial intelligence becomes deeply integrated into healthcare, data privacy regulations like the General Data Protection Regulation (GDPR) grant patients the "right to be forgotten." In the context of deep learning, this mandate goes beyond deleting raw files; it requires ensuring that a trained model no longer retains or can be exploited to reveal sensitive information about a specific patient.

While the gold standard for forgetting is retraining a model from scratch without the target data, this is often computationally prohibitive for large-scale or decentralized systems. "Machine Unlearning" or "Scrubbing" methods aim to update model weights post-training to remove specific influences. However, most existing methods were developed for general computer vision datasets where data is relatively uniform. This paper explores why these methods fail in the heterogeneous landscape of medical imaging.

The Challenge: Common Clusters vs. Edge Cases

The authors propose a framework where patient data falls into two categories: Common Clusters (patients with features similar to many others) and Edge Cases (patients with unique anatomical or pathological features).

Comparison of unlearning strategies and the two-hypothesis framework for patient data distribution.

In standard datasets like CIFAR-10, most samples belong to common clusters, making unlearning relatively straightforward due to data redundancy. In contrast, medical datasets like the Automated Cardiac Diagnosis Challenge (ACDC) are highly "long-tailed." The authors found that over 60% of patients in the ACDC dataset are edge cases. When a patient is an edge case, their unique features are encoded in specific model weights. Standard scrubbing methods, which add noise to weights deemed "uninformative" to the remaining data, often inadvertently damage these unique weights, causing the model's performance on other rare cases to degrade significantly.

Methodology: Targeted Forgetting

To address the limitations of standard scrubbing, the researchers introduced Targeted Forgetting. This approach utilizes the Fisher Information Matrix (FIM) to surgically identify and modify only the weights that are most informative for the specific patient being forgotten ($D_f$).

  1. Scrubbing (Baseline): Identifies weights not highly informative to the retained data ($D_r$) and adds noise to them.
  2. Targeted Forgetting (Proposed): Directly identifies weights highly informative to the forgotten data ($D_f$) and adds noise specifically to them.

By focusing on the information used by the target patient rather than what is ignored by the majority, Targeted Forgetting minimizes collateral damage to the model's general diagnostic capabilities.

Experimental Evaluation

The method was validated using the ACDC dataset, which contains cardiac MRI data from 100 patients, preprocessed into over 14,000 2D images. This dataset includes various pathologies, such as Hypertrophic Cardiomyopathy (HCM), which contribute to the high frequency of edge cases.

Example cardiac MRI slice from the ACDC dataset representing a Hypertrophic Cardiomyopathy (HCM) case.

Results and Discussion

The experiments revealed several key insights:

  • Generalization Performance: On edge-case patients, Targeted Forgetting maintained significantly lower test errors (approx. 0.29–0.31) compared to standard Scrubbing (approx. 0.74) when high levels of forgetting were required.
  • Noise Efficiency: Targeted Forgetting achieved "complete forgetting" (matching the performance of a model never trained on that patient) using nearly ten times less noise than the scrubbing method. For one specific edge case (Patient 94), Targeted Forgetting required an average noise value of 3.00E-06, while Scrubbing required 2.33E-05.
  • The "Hardness" of Forgetting: The study demonstrated a direct correlation between a patient's uniqueness and the difficulty of forgetting them. The more an edge case contributes to the model's specialized knowledge, the harder it is to remove that knowledge without affecting overall accuracy.

Conclusion

Forgetting patient data in medical AI is a complex task due to the high prevalence of unique edge cases. While standard unlearning techniques fail by over-scrubbing informative weights, the proposed Targeted Forgetting approach offers a more precise solution. Future work will focus on automatically detecting patient hypotheses (common vs. edge case) and validating the method across diverse imaging modalities and pathologies to ensure clinical utility is maintained alongside privacy.

Summary updated

Abstract

Rights provisioned within data protection regulations, permit patients to request that knowledge about their information be eliminated by data holders. With the advent of AI learned on data, one can imagine that such rights can extent to requests for forgetting knowledge of patient's data within AI models. However, forgetting patients' imaging data from AI models, is still an under-explored problem. In this paper, we study the influence of patient data on model performance and formulate two hypotheses for a patient's data: either they are common and similar to other patients or form edge cases, i.e. unique and rare cases. We show that it is not possible to easily forget patient data. We propose a targeted forgetting approach to perform patient-wise forgetting. Extensive experiments on the benchmark Automated Cardiac Diagnosis Challenge dataset showcase the improved performance of the proposed targeted forgetting approach as opposed to a state-of-the-art method.
BibTeX
@inproceedings{su2022why,
  author = {Ruolin Su and Xiao Liu and Sotirios A. Tsaftaris},
  title = {{Why patient data cannot be easily forgotten?}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2022},
  url = {https://openreview.net/forum?id=Q9Emu3ksQ6},
}
Figure