Abstract:

We report the rank-1 system for NeuroMM-2026 Track-1, binary EEG spike detection scored by AUPRC. The system combines domain-aware preprocessing, five time–frequency representations, ImageNet-pretrained vision backbones adapted from prior competition solutions, and a regularized Nelder–Mead ensemble, reaching an inductive score of 0.9711. A conventional pseudo-labelling warm-up improves this to 0.9778, after which our candidate-specialist self-training procedure adds high-confidence released candidates at full weight and reaches 0.9846. Because this final regime is transductive and in-sample by design, we evaluate it with a held-out-fold replay, which suggests partial transfer while not excluding candidate-specific memorization. The main lesson is that, on this saturated EEG benchmark, broad model diversity is insufficient; the decisive lever is many confident pseudo-labelled candidates, full commitment, and candidate-excluded real-label validation. Code: https://github.com/GG3BBE0/PushingPastSaturation-NeuroMM-2026-Track1.