Portable electroencephalography in early detection of depression: Progress and future directions

Table 3 Portable electroencephalography in depression detection

Feature type	Models	Key indicators	Main results	Ref.
Frequency-domain (PSD)	Support vector machine, random forest, CNN, RNN; feature combination	Decreased frontal α power; increased β power; abnormal θ/δ power	Frontal EEG + PSD: 90.7% accuracy; combined with nonlinear features improves stability	[34,38,39,43]
Time-domain	LSTM, feature fusion (time + frequency), classifiers	Amplitude, peak, mean value	Used for signal evaluation; combination improves accuracy	[34,41,44,45]
Time-frequency (STFT/WT)	STFT, WT + nonlinear features, classifiers	α/β power; WT adapts to time scales	WT extracts nonlinear features; high accuracy; significant power asymmetry	[38,40,42,45]
Nonlinear features	Nonlinear features + classifiers	SampEn, ApEn, FD, HFD, LZC	Lower complexity in depression; LZC and HFD achieve approximately 90% accuracy	[40,43]
Functional connectivity	Coherence, PLV + classifiers	Decreased frontal coherence, PLV; reduced frequency synchronization	Reduced interhemispheric connectivity; linked to emotion regulation	[34,38]

PSD: Power spectral density; CNN: Convolutional neural network; RNN: Recurrent neural network; EEG: Electroencephalography; LSTM: Long short-term memory; STFT: Short-time Fourier transform; WT: Wavelet transform; SampEn: Sample entropy; ApEn: Approximate entropy; FD: Fractal dimension; HFD: Higuchi fractal dimension; LZC: Lempel-Ziv complexity; PLV: Phase locking value.