Machine Learning-Driven Analysis of Temporal Pupil Dynamics for Interpretable ADHD Diagnosis (Preprint)

Published: Feb. 3, 2025

Attention-deficit/hyperactivity disorder (ADHD) is a prevalent neurodevelopmental characterized by persistent patterns of inattention and hyperactivity. Current diagnostic methods rely heavily on subjective measures, such as clinical interviews behavior rating scales, which are prone to bias variability. Objective biomarkers, essential for reliable standardized diagnosis, remain elusive. Pupillometry, measures dynamic pupil responses associated with cognitive attentional processes, offers promising avenue objective ADHD diagnostics. However, existing studies often overlook clinically relevant features fail prioritize model interpretability, hindering their potential implementation. This study aims develop interpretable machine learning models utilizing temporal dynamics classify control groups, aiming improved accuracy explainability. utilized already published pupillometry data from 49 participants, including 21 controls 28 ADHD-diagnosed children, 17 assessed both on-medication off-medication. Data were collected during visuospatial working memory task designed evaluate processes. The preprocessed remove noise analysis. Pupil behavioral first identified based literature review conducted the population. final set was determined through statistical analyses using mixed block-wise ANOVA assess significance. Binary classification developed differentiate participants. evaluated progressively, starting derived only dynamics, then incorporating performance metrics, finally reaction time metrics. Performance area under receiver operating characteristic curve. ensured interpretability selection statistically significant features, supported review, that contribute meaningfully task. Key included median size (blocks 1 3), dilation contraction rates 4 8), time, each exhibiting distinct Visualizations heatmaps feature importance charts highlighted relevance, providing transparency in models' decision-making demonstrated robust these features. Models trained exclusively achieved best 86.7% an AUROC score 0.884. Incorporating metrics performance, achieving 88.9% 0.931. integration resulted highest 90%, 0.93, sensitivity 100%, specificity 80.8%. highlights leveraging biomarker ADHD. By focusing proposed offer practical trustworthy approach advancing development tools use.

Language: Английский

Dual temporal pathway model of emotion processing based on dynamic network reconfiguration analysis of EEG signals

Acta Psychologica, Journal Year: 2025, Volume and Issue: 255, P. 104912 - 104912

Published: March 14, 2025

Language: Английский