MF-PNet: A Lightweight Multi-Frequency Progressive Neural Network for Depression Classification Deployed on Wearable EEG Sensors
Abstract
Low-channel wearable EEG offers a compact acquisition configuration for depression recognition, but its limited spatial coverage and the computational constraints of edge devices present challenges for representation learning and embedded inference. This paper presents MF-PNet, a multi-frequency progressive neural network designed for three-channel EEG. Five frequency-specific branches process the canonical EEG bands in parallel, while a full-band branch progressively aggregates their representations through layer-wise lateral transfer and squeeze-and-excitation-based channel recalibration. Temporal and channel attention are further incorporated to emphasize informative EEG segments and frontal channels. Subject-level leave-one-subject-out cross-validation was conducted on a balanced in-house cohort of 146 participants. Across five random seeds, MF-PNet achieved an accuracy and F1-score of <inline-formula> <tex-math notation="LaTeX">${98}.{8}\%\pm {1}.{0}\%$ </tex-math></inline-formula> under auditory stimulation. Within-dataset evaluation on the independent Figshare and MODMA cohorts yielded accuracies of <inline-formula> <tex-math notation="LaTeX">${92}.{1}\%\pm {2}.{0}\%$ </tex-math></inline-formula> and <inline-formula> <tex-math notation="LaTeX">${66}.{4}\%\pm {6}.{5}\%$ </tex-math></inline-formula>, respectively. For embedded implementation, MF-PNet was distilled into a MobileNetV3-based student containing 75.51 K parameters. The student achieved an accuracy of <inline-formula> <tex-math notation="LaTeX">${99}.{3}\%\pm {1}.{2}\%$ </tex-math></inline-formula>, and its INT8 representation retained an accuracy of <inline-formula> <tex-math notation="LaTeX">${98}.{5}\%\pm {0}.{9}\%$ </tex-math></inline-formula>. On the STM32U575CGT6 MCU, the implementation required 217.15 KB of ROM and 215.78 KB of RAM, with an inference latency of 46.00 ms per 2-s EEG window and a power consumption of 43.2 mW. The resulting framework integrates progressive cross-band EEG representation learning, teacher–student compression, INT8 quantization, and MCU execution into a unified pipeline for three-channel wearable EEG depression recognition.