BridgeMIL lifts EEG diagnosis 4.28 points

BridgeMIL reached 76.57% mean accuracy across three EEG disease datasets and five model backbones, 4.28 percentage points above the strongest baseline. The preprint reports first place in 14 of 15 dataset-backbone settings after separating window-level representation learning from patient-level supervision instead of assigning every EEG segment the patient’s diagnosis. For teams confronting the validation burden behind consumer AI health products, the transferable lesson is to align labels with the unit of prediction; retrospective benchmark accuracy, however, is not clinical validation.