TL;DR
Data protection regulations like the "right to be forgotten" pose a major challenge for medical AI, where completely retraining a model to purge a patient's data is often infeasible. This paper reveals that over 60% of patients in medical datasets act as unique "edge cases," causing existing machine unlearning methods to severely degrade model accuracy. To solve this, the authors propose targeted forgetting, an approach that selectively perturbs only the weights most informative to the specific patient being removed.
What Problem Does This Paper Address?
When a patient requests the deletion of their medical records, regulations (such as GDPR) imply their data should also be removed from trained AI models. Simply deleting records from a database does not prevent deep neural networks from memorizing patient traits, leaving them vulnerable to model inversion and membership inference attacks.
Retraining a model from scratch without the patient’s data is the ideal "golden standard," but it is computationally prohibitive and often impossible in decentralized setups like federated learning. While machine unlearning (or "scrubbing") methods exist for general computer vision tasks, standard techniques fail when applied to medical imaging on a patient-by-patient basis.
Methods
The paper contrasts two strategies for post-hoc weight modification:
- Standard Scrubbing (Baseline): Following Golatkar et al. (2020), this method computes the Fisher Information Matrix (FIM) over the retaining dataset ($D_r$) to identify which weights matter most to the remaining data. It then adds noise to weights deemed uninformative to $D_r$ to scrub forgotten data ($D_f$).
- Targeted Forgetting (Proposed): Instead of calculating the FIM on the massive retaining dataset, the authors calculate the FIM directly on the forgetting set ($D_f$). By inverting the focus, the algorithm injects calibrated Gaussian noise specifically into the model parameters that are highly informative about the patient to be forgotten, leaving unaffected weights intact: $$\mathcal{S}_T(\mathbf{w}) = \mathbf{w} + \left(\lambda_T \sigma_{hT}^2\right)^{\frac{1}{4}} F_{D_f}(\mathbf{w})^{\frac{1}{4}}$$ In practice, the authors store only the diagonal of the FIM and perturb a small subset (e.g., the top 1%) of the most informative weights.
Datasets
- Automated Cardiac Diagnosis Challenge (ACDC): A clinical benchmark of 4D cardiac MRI scans from 100 patients (split into 90 training and 10 testing patients), preprocessed into 224×224 2D slices (14,724 training images, 1,464 test images) spanning five diagnostic categories.
- CIFAR-10: 60,000 natural images across 10 classes, used as a comparative baseline to demonstrate dataset bias differences between general vision and medical domains.
Key Insights and Ideas
- A Tale of Two Hypotheses: Medical patient data generally falls into one of two categories:
- Common Clusters: Data that share visual representations with many other patients.
- Edge Cases: Unique, rare cases residing in the "long tail" of the data distribution.
- Medical Data is Dominated by Edge Cases: When retraining models from scratch without a candidate subset on CIFAR-10, golden standard error on the omitted data stays between 10% and 25%. In contrast, removing a single patient in ACDC yields golden standard errors ranging from 0% to 100%. Over 60% of ACDC patients have a golden standard error exceeding 50%, demonstrating that medical datasets consist heavily of distinct edge cases.
- Why Scrubbing Fails on Medical Data: Because edge-case patient features are unique, their weights appear uninformative relative to the aggregate retaining set ($D_r$). Scrubbing mistakenly injects noise across weights associated with all edge cases, tanking generalization across the remaining dataset and test set.
- The Common Cluster Dilemma: If a patient belongs to a common cluster, forgetting that patient completely is nearly impossible without hurting the model's ability to classify the rest of the cluster.
Key Results
The authors evaluated performance using Error ($1 - \text{Accuracy}$) on retaining ($D_r$), forgetting ($D_f$), and test ($D_{\text{test}}$) sets:
- Matching Golden Standard (Edge Cases):
- For Patient 94 (an extreme edge case, golden error on $D_f = 1.00$, $D_{\text{test}} = 0.24$):
- Scrubbing: Matches $D_f$ error ($1.00$), but retaining error balloons to $0.46$ and test error degrades sharply to $0.56$.
- Targeted Forgetting: Matches $D_f$ error ($1.00$), keeps retaining error at $0.23$, and limits test error to $0.30$.
- Without Golden Standard Calibration (High Noise):
- When pushing noise to force full unlearning ($1.00$ error on $D_f$) without knowing the golden target:
- For Patient 94, scrubbing yields a catastrophic test error of $0.75$, whereas targeted forgetting maintains a test error of $0.31$.
- Scrubbing required significantly larger noise magnitudes (e.g., $2.33 \times 10^{-5}$ vs. $3.00 \times 10^{-6}$ for targeted forgetting) to achieve unlearning, needlessly disrupting unrelated representations.
Limitations and Future Work
- Common Cluster Degradation: Neither method can cleanly forget a patient belonging to a dense, common cluster without degrading model accuracy for other patients in that same cluster.
- Data Access Assumption: The method requires temporary access to the specific patient's data ($D_f$) at unlearning time to calculate the local Fisher Information Matrix.
- Evaluation Scope: Experiments were conducted exclusively on cardiac MRI; empirical testing on additional imaging modalities (e.g., CT, histopathology) remains future work.
- Automated Hypothesis Classification: Future work must develop automated diagnostic tools to determine whether a requested patient is an edge case or cluster case prior to applying unlearning protocols.
Summary updated