Decades-old anonymized medical data may cause AI misdiagnoses, study finds

The authors evaluated four large medical datasets – MIMIC‑CXR, MIMIC‑IV‑ED, MIMIC‑ECG and the HEEDB collection – covering from tens of thousands up to more than 1.8 million patients. In some cases, predicted probabilities shifted by more than 70 percentage points, and the effect persisted for over 25 years in the HEEDB ECG data. Tests showed lower sensitivity for new conditions, while accuracy improved when the patient’s health had not changed. Applying differential privacy during training reduced the bias, and patient‑level differential privacy almost eliminated it, though it cannot be used on already anonymised data.

The researchers note that as model capacity grows under scaling laws, the absolute number of individuals affected by memorisation bias could rise sharply, raising safety and regulatory concerns for medical AI deployments that rely on nationally sourced data.