Anonymized Medical Data May Cause AI Misdiagnoses Decades Later

A joint German-UK study found AI models trained on anonymized medical records can memorize patients and carry their past conditions into current predictions. In one case, a model gave only a 15% heart attack probability versus 73% for a model without historical data.

• Effect persists for decades after a record enters the dataset
• Predicted probability shifts reached 70 percentage points
• Tested on 4 datasets: ECG, chest X-ray and EHR
• German breast cancer screening uses a model on 1.2 million mammograms