BridgeMIL Decouples EEG Instance Learning for Better Disease Diagnosis

Traditional EEG-based disease diagnosis often involves segmenting long recordings into short instances and then assigning the subject's overall diagnosis label to each instance. This approach assumes all instances provide equally reliable diagnostic evidence, which is often not true. While Multiple Instance Learning (MIL) avoids this by treating each subject as a "bag" of instances, the scarcity of subjects compared to instances in EEG datasets can limit the quality of learned representations in end-to-end MIL systems. To overcome these limitations, a new two-stage framework called BridgeMIL has been developed. In its first stage, BridgeMIL pretrains an encoder without relying on inherited instance labels. Instead, it aligns temporally nearby windows and independently sampled within-subject sub-bags, using variance and covariance regularization to prevent collapse and reduce redundancy. The second stage transfers this pretrained encoder to an attention-based MIL aggregator, applying supervision only to the subject-level predictions. Feature retention mechanisms are used to limit representation drift. Experiments across three EEG disease datasets and five backbones show BridgeMIL achieving significantly higher mean accuracy, outperforming strong baselines. This highlights the importance of aligning supervision with the subject-level objective while effectively learning from abundant EEG instances without mislabeling them.