نهمین کنگره نقشه برداری مغز

نمایش مقاله

Title:

Decoding Anxiety-Spectrum Disorders with Neurophysiological Biomarkers and Machine Learning

Subject:

Artificial Intelligence and Brain Mapping

Presenter Author:
Name:	Vahid Asayesh

Author:
Name 1:	Vahid Asayesh - Corresponding - Presenter
Affiliation 1:	NPCindex Research Company, Tabriz, Iran, Asayesh Psychiatric Clinic, Tabriz, Iran

Name 2:	Majid Torabi
Affiliation 2:	NPCindex Research Company, Tabriz, Iran, Asayesh Psychiatric Clinic, Tabriz, Iran

Name 3:	Mehdi Dehghani
Affiliation 3:	NPCindex Research Company, Tabriz, Iran

Name 4:	Sepideh Akhtari-Khosroshahi
Affiliation 4:	NPCindex Reasearch Company, Tabriz, Iran

Name 5:	Maedeh Akhtari-Khosroshahi
Affiliation 5:	NPCindex Reasearch Company, Tabriz, Iran

Name 6:
Affiliation 6:

Name 7:
Affiliation 7:

Name 8:
Affiliation 8:

Name 9:
Affiliation 9:

Name 10:
Affiliation 10:

Abstract

Original Research

introduction :

Anxiety-spectrum disorders like generalized anxiety disorder (GAD), panic disorder (PD), and social phobia/social anxiety disorder (SAD) are common, disabling, and frequently overlapping, complicating differential diagnosis and delaying effective care [1–4]. Electroencephalography (EEG) offers a noninvasive physiological complement to symptom scales, and quantitative EEG (QEEG) may provide objective biomarkers for clinical decisions [5]. Prior QEEG work reports group differences in spectral power and connectivity, yet linear metrics alone may miss relevant dynamics [6, 7]. Nonlinear measures like Higuchi fractal dimension (HFD), Lempel–Ziv complexity (LZC), and correlation dimension (CD) capture signal complexity and can reveal distinctions unseen in frequency analyses [8, 9]. We integrate spectral (absolute, relative, cross-band ratios) and nonlinear (HFD, LZC, CD) EEG features from resting-state 19-channel recordings to test whether GAD, PD, and SAD are reliably differentiable [5, 8, 9].

Methods:

Sixty patients were enrolled (GAD n=20, PD n=20, SAD n=20), diagnosed per DSM-5 [1]. Resting-state eyes-closed (EC) EEG was recorded for 5 min with a 19-channel Mitsar device (250 Hz; 10–20 montage; average reference). Data were down-sampled to 125 Hz, filtered (0.1–50 Hz band-pass, 50 Hz notch), and cleaned via ICA, wavelet ICA, and FASTER [2–4]. Extracted features comprised absolute, relative, and cross-band spectral power and nonlinear HFD, LZC, and CD [5–7]. Omnibus Kruskal–Wallis tests with false-discovery-rate (FDR) control (α=0.05) assessed group effects; significant results prompted pairwise Mann–Whitney tests with Cliff’s δ [8–10]. Features significant in all three contrasts were retained, z-scored, and ranked by minimum-Redundancy/Maximum-Relevance (mRMR) [11]. Classification proceeded incrementally by adding ranked features; performance peaked at 23 features.

Results:

Both spectral and nonlinear features significantly differentiated groups. Kruskal–Wallis with FDR identified robust effects over frontal and central electrodes, particularly for absolute and relative delta, theta, and beta power, consistent with fronto-central involvement in anxiety regulation [12, 13]. Cross-band ratios (notably delta–beta and theta–gamma) were also discriminative, reflecting arousal and cortical excitability differences [14]. Nonlinear indices (HFD, LZC) further improved separation, supporting evidence that reduced signal complexity marks psychiatric states [15, 16]. Multiple feature–channel pairs remained significant across all pairwise contrasts, indicating robustness. After mRMR ranking, classification with the optimal 23-feature subset yielded: support vector machine (SVM) 92.36% accuracy (92.30% sensitivity, 96.10% specificity), multilayer perceptron (MLP) 93.30% accuracy (93.31% sensitivity, 94.64% specificity), and residual fully-connected network (ResFCN) 94.95% accuracy (94.96% sensitivity, 97.47% specificity). Accuracy declined when adding features beyond 23, indicating redundancy and overfitting.

Conclusion:

Integrating spectral and nonlinear QEEG features enabled reliable differentiation of GAD, PD, and SAD despite clinical overlap. The most consistent effects localized to frontal–central regions (delta/theta/beta), while cross-band ratios indexed arousal-related differences [12–14]. Nonlinear metrics (HFD, LZC) contributed complementary information, aligning with the view that reduced neural complexity characterizes psychiatric dysfunction [15, 16]. A compact 23-feature panel maximized performance across models, with ResFCN outperforming SVM and MLP. These findings support QEEG-based biomarker panels as objective aids for anxiety-spectrum differential diagnosis; larger, independent cohorts and external validation remain essential for clinical translation.

Keywords:

Electroencephalography, Machine Learning, Non-linear Neuromarkers, Anxiety Disorders, Classification.