Late Fusion of Time-Domain and Time-Frequency Sleep EEG for Future Cognitive Impairment Prediction

Nahyun Lee1, Kim TaeHwan2, TaeYoung Kang2, DongJae Lee3, DONGGUN KIM4, Jeongbin You5, younghyuk kim4, Seungsang Oh4, Wuming Gong6, Il-Youp Kwak2
1Chung Ang University, 2Chung-Ang University, 3¹Department of Anesthesiology and Pain Medicine, Korea University Guro/Anam Hospital, Seoul, Republic of Korea ²Department of Physiology, Korea University College of Medicine, Seoul, Republic of Korea, 4korea university, 5Koera University, 6University of Minnesota


Abstract

Clinical polysomnography may contain markers of cognitive decline years before diagnosis, while heterogeneous montages and acquisition conditions complicate cross-site prediction. Team CAU_KU developed a compact late-fusion system for predicting cognitive impairment one to six years after polysomnography. Using 6,600 multicenter recordings, our pipeline standardizes six EEG derivations from variable PSG configurations and samples quality-controlled windows across the night. Time-domain and time-frequency branches are trained separately, calibrated at the subject level, and fused by probability averaging. The final system achieves an age-conditioned AUROC of 0.746 on the official Challenge validation set. These results support the value of complementary sleep EEG representations within routine polysomnography for cross-site prediction of future cognitive impairment.