Scholarlink

<p>Attention Deficit Hyperactivity Disorder (ADHD) is a neurodevelopmental disorder characterized by impairments in attention, impulsivity, and executive function. Current diagnostic approaches rely primarily on subjective behavioral assessments, highlighting the need for objective and interpretable neurophysiological biomarkers. This study proposes a Multi-Objective Rank Scored–SHapley Additive exPlanations (MRS-SHAP) framework for EEG-based ADHD analysis, integrating explainability, temporal stability, and performance-aware feature selection within a unified pipeline. A publicly available dataset comprising 121 children (61 ADHD, 60 controls; age 7–12 years), recorded using a 19-channel EEG system, was utilized. A total of 998 features were extracted across spectral, nonlinear, and connectivity domains and dataset does not include subtype-specific ADHD labels; therefore, the analysis focuses on general ADHD classification. The proposed approach combines SHAP-based feature importance, Dynamic Time Warping (DTW)-based temporal stability, and cross-validated performance weighting to derive a compact, robust, and explainable feature subset. All preprocessing, resampling (SMOTE-ENN), feature selection, and model training steps were implemented within a strictly nested, leakage-free cross-validation framework to ensure unbiased evaluation. Under segment-level evaluation, the CatBoost classifier achieved high performance (Accuracy = 98.40%, F1-score = 98.40%), representing an upper-bound estimate. However, subject-level validation using GroupKFold yielded more realistic performance (Accuracy ≈ 79.5%, AUC ≈ 0.86), while evaluation on a held-out test set (~ 20% of subjects) produced a more conservative estimate (Accuracy ≈ 65.7%, AUC ≈ 0.73). This performance gap highlights the impact of inter-subject variability and emphasizes the importance of strict subject-level validation in EEG-based modeling. Comparative analysis demonstrates that the proposed feature selection approach outperforms conventional methods such as PCA, ANOVA, and Recursive Feature Elimination (RFE). SHAP-based interpretability identifies beta-band activity, EEG microstate dynamics, and fronto-central connectivity as key discriminative features, consistent with established neurophysiological evidence of ADHD-related network dysfunction. Ablation and sensitivity analyses further confirm the complementary contributions of feature importance and temporal stability, with minimal dependence on synthetic resampling. These findings indicate that integrating explainability and temporal stability enhances the robustness and interpretability of EEG-based feature selection. However, the results should be interpreted as methodological validation rather than clinical evidence, as the study is limited to a single dataset and may be influenced by confounding factors such as medication exposure. Future work will focus on external validation across independent cohorts and clinically controlled settings to assess generalizability and translational relevance.</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

MRS-SHAP: Multi-objective explainable EEG feature selection framework for ADHD detection

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月11日

查看全文
  • Mohammed Atheef G A,
  • Shyam Pranav G,
  • Omkar S Powar

摘要

Attention Deficit Hyperactivity Disorder (ADHD) is a neurodevelopmental disorder characterized by impairments in attention, impulsivity, and executive function. Current diagnostic approaches rely primarily on subjective behavioral assessments, highlighting the need for objective and interpretable neurophysiological biomarkers. This study proposes a Multi-Objective Rank Scored–SHapley Additive exPlanations (MRS-SHAP) framework for EEG-based ADHD analysis, integrating explainability, temporal stability, and performance-aware feature selection within a unified pipeline. A publicly available dataset comprising 121 children (61 ADHD, 60 controls; age 7–12 years), recorded using a 19-channel EEG system, was utilized. A total of 998 features were extracted across spectral, nonlinear, and connectivity domains and dataset does not include subtype-specific ADHD labels; therefore, the analysis focuses on general ADHD classification. The proposed approach combines SHAP-based feature importance, Dynamic Time Warping (DTW)-based temporal stability, and cross-validated performance weighting to derive a compact, robust, and explainable feature subset. All preprocessing, resampling (SMOTE-ENN), feature selection, and model training steps were implemented within a strictly nested, leakage-free cross-validation framework to ensure unbiased evaluation. Under segment-level evaluation, the CatBoost classifier achieved high performance (Accuracy = 98.40%, F1-score = 98.40%), representing an upper-bound estimate. However, subject-level validation using GroupKFold yielded more realistic performance (Accuracy ≈ 79.5%, AUC ≈ 0.86), while evaluation on a held-out test set (~ 20% of subjects) produced a more conservative estimate (Accuracy ≈ 65.7%, AUC ≈ 0.73). This performance gap highlights the impact of inter-subject variability and emphasizes the importance of strict subject-level validation in EEG-based modeling. Comparative analysis demonstrates that the proposed feature selection approach outperforms conventional methods such as PCA, ANOVA, and Recursive Feature Elimination (RFE). SHAP-based interpretability identifies beta-band activity, EEG microstate dynamics, and fronto-central connectivity as key discriminative features, consistent with established neurophysiological evidence of ADHD-related network dysfunction. Ablation and sensitivity analyses further confirm the complementary contributions of feature importance and temporal stability, with minimal dependence on synthetic resampling. These findings indicate that integrating explainability and temporal stability enhances the robustness and interpretability of EEG-based feature selection. However, the results should be interpreted as methodological validation rather than clinical evidence, as the study is limited to a single dataset and may be influenced by confounding factors such as medication exposure. Future work will focus on external validation across independent cohorts and clinically controlled settings to assess generalizability and translational relevance.