Scholarlink

Background <p>At present, the clinical diagnosis of obsessive-compulsive disorder (OCD) is primarily based on interviews with patients by experienced psychiatrists, which is inherently limited by its subjectivity. Therefore, there is a pressing clinical need to identify objective biomarkers of OCD. Although the combination of EEG and data-driven method based on machine learning offers promise for identifying objective neurophysiological biomarkers, current researches either rely on only small sample sizes, lack interpretability, use only one single dataset which cannot validate model generalizability, or utilize only one type of EEG feature which cannot evaluate the relative utility of different features as potential biomarkers of OCD under a unified data-driven framework.</p> Method <p>This study employed two independent datasets (Dataset 1: OCD = 35, healthy controls = 37; Dataset 2: OCD = 21, healthy controls = 21). Eight EEG features were extracted and were sent to six machine learning (ML) classifiers to classify OCD from healthy controls. After optimizing the hyperparameters and the most relevant feature set on Dataset 1, each ML model was retrained on Dataset 1 and then were tested on an independent external test set (Dataset 2) to assess their generalizability. Feature contributions were interpreted using SHapley Additive Explanations (SHAP) analysis.</p> Result <p>Among the eight EEG feature sets, the phase-locking value (PLV) features achieved the highest classification performance across all machine learning models. The LightGBM classifier outperformed others ML classifiers, reaching an accuracy of 86.6% on Dataset 1 and 83.3% on Dataset 2 (independent external test). The data-driven method based on machine learning selected the most important 16 PLV features and SHAP-based feature importance analysis identified alpha band PLV (F4-P3), delta band PLV (P3-O1, F3-O1), and theta band PLV (C3-T4) as the most influential contributors to model predictions.</p> Conclusion <p>This study developed an explainable machine learning framework based on PLV functional connectivity feature to enable accurate and generalizable OCD classification. These results highlight the potential of PLV especially long-range PLV connectivity between frontal and parietal/occipital lobe as a reliable EEG biomarker for objective and clinically applicable OCD diagnosis.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Exploring potential resting-state EEG biomarkers of obsessive-compulsive disorder based on explainable machine learning analysis of independent training and test samples

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月15日

查看全文
  • Zongya Zhao,
  • Junming Wang,
  • Yanxiang Niu,
  • Mengyue Qiu,
  • Mingjie Zhu,
  • Mingcai Li,
  • Huicong Ren,
  • Zhixian Gao,
  • Chang Wang,
  • Wu Ren,
  • Xuezhi Zhou,
  • Mingchao Qi,
  • Zhaohui Zhang,
  • Zihao Xu,
  • Yi Yu

摘要

Background

At present, the clinical diagnosis of obsessive-compulsive disorder (OCD) is primarily based on interviews with patients by experienced psychiatrists, which is inherently limited by its subjectivity. Therefore, there is a pressing clinical need to identify objective biomarkers of OCD. Although the combination of EEG and data-driven method based on machine learning offers promise for identifying objective neurophysiological biomarkers, current researches either rely on only small sample sizes, lack interpretability, use only one single dataset which cannot validate model generalizability, or utilize only one type of EEG feature which cannot evaluate the relative utility of different features as potential biomarkers of OCD under a unified data-driven framework.

Method

This study employed two independent datasets (Dataset 1: OCD = 35, healthy controls = 37; Dataset 2: OCD = 21, healthy controls = 21). Eight EEG features were extracted and were sent to six machine learning (ML) classifiers to classify OCD from healthy controls. After optimizing the hyperparameters and the most relevant feature set on Dataset 1, each ML model was retrained on Dataset 1 and then were tested on an independent external test set (Dataset 2) to assess their generalizability. Feature contributions were interpreted using SHapley Additive Explanations (SHAP) analysis.

Result

Among the eight EEG feature sets, the phase-locking value (PLV) features achieved the highest classification performance across all machine learning models. The LightGBM classifier outperformed others ML classifiers, reaching an accuracy of 86.6% on Dataset 1 and 83.3% on Dataset 2 (independent external test). The data-driven method based on machine learning selected the most important 16 PLV features and SHAP-based feature importance analysis identified alpha band PLV (F4-P3), delta band PLV (P3-O1, F3-O1), and theta band PLV (C3-T4) as the most influential contributors to model predictions.

Conclusion

This study developed an explainable machine learning framework based on PLV functional connectivity feature to enable accurate and generalizable OCD classification. These results highlight the potential of PLV especially long-range PLV connectivity between frontal and parietal/occipital lobe as a reliable EEG biomarker for objective and clinically applicable OCD diagnosis.