AOR：面向 EEG 通道缺失的自适应正交回归抑郁特征选择方法

Objective. Depression induces severe mental and physical impairments, highlighting the clinical need for rapid and reliable detection. Electroencephalography (EEG) directly reflects neural electrical activity and serves as a promising tool for depression investigation. However, the inherent non-stationarity, redundancy, and noise susceptibility of EEG signals, together with the common issue of incomplete channel data due to electrode loss or severe noise during acquisition, compromise the reliability of diagnostic models. Existing EEG feature selection methods lack adaptability to such incomplete data, leading to distorted feature evaluation and degraded performance. Approach. To address this limitation, an adaptive orthogonal regression (AOR) method is proposed for EEG feature selection for depression detection under incomplete channel conditions. The method integrates three core innovations: embedding missing channel information into the orthogonal regression formulation to mitigate the impact of incomplete channels, incorporating adaptive channel-weighted learning to automatically quantify the relative importance of different channels, and applying global redundancy minimization to eliminate inter-feature redundancy while enhancing discriminability. The optimization problem is solved via Generalized Power Iteration and Augmented Lagrangian Multiplier methods. Main results. AOR was validated on the MODMA and PRED-d003 datasets across 3-channel, 64-channel, and 128-channel configurations under single-modal and multi-modal scenarios. Experimental results demonstrate that AOR consistently outperforms state-of-the-art methods across various missing ratios and channel configurations. The average cross-subject depression recognition accuracy reaches 92.25% and 95.12% across different missing ratios on the two datasets. Additionally, feature weight analysis reveals that EEG kurtosis consistently exhibits high discriminative power across different configurations. Significance. The proposed AOR method provides a robust and efficient solution for depression feature selection under incomplete EEG channel conditions, offering reliable diagnostic performance in real-world clinical scenarios.