Researchers identify hidden data leakage flaw in brain-scan survival models

Researchers found that a common mistake in building AI models for predicting patient survival from brain scans (EEG recordings) makes the models look better than they actually are — they're accidentally using information from test data during training, which inflates their accuracy. They built a new model structure that prevents this leak and keeps predictions honest even when tested on completely new patients.

This matters because EEG-based survival prediction for comatose patients after cardiac arrest is already being used in hospitals — and if the models are fooling themselves about how accurate they are, doctors are making life-or-death decisions based on inflated performance numbers. The paper shows a systematic way to catch and fix this hidden flaw, which could apply to other medical AI systems that use long sensor recordings segmented into pieces.