Attention deficit hyperactivity disorder assessment through objective measures: POV glasses and machine learning approach

TL;DR

This study uses POV cameras and machine learning to objectively assess ADHD in children by analyzing their body movements during a structured interaction.

Contribution

The study introduces a novel method for ADHD assessment using movement-based features and machine learning to reduce subjective bias in diagnosis.

Findings

ADHD children showed significantly higher global activity index compared to controls.

AdaBoost classifier achieved 81.82% accuracy in distinguishing ADHD from control groups.

Movement patterns in specific body regions correlated with parent-reported hyperactivity scores.

Abstract

The diagnosis of Attention-Deficit/Hyperactivity Disorder (ADHD) largely relies on clinical interviews and parent/teacher-report rating scales, which are vulnerable to subjective bias. Therefore, there is an increasing need for objective measures to complement existing assessment approaches. The aim of this study was to objectively quantify children’s body movement during a controlled, semi structured interaction, to examine differences between children with and without ADHD, and to evaluate the cross-sectional discriminative capacity of these movement-based features using machine learning methods.

This study employed a cross-sectional, observational case–control design including 37 children diagnosed with ADHD and 29 typically developing children aged 7–11 years. Psychiatric diagnoses were established using the DSM-5–based K-SADS PL interview. Video recordings were obtained during a standardized 5-minute instructional interaction using a researcher-worn point-of-view (POV) camera. Body movement measures of the head, upper limbs, and lower limbs were extracted from the video recordings using MediaPipe Pose. Movement data were statistically compared between groups, followed by classification analyses using machine learning algorithms.

The global activity index was significantly higher in the ADHD group compared to the control group (p = 0.003). Regional analyses revealed significant group differences in shoulder, elbow, ankle, foot, and head movements. A significant positive correlation was found between the global activity index and parent-reported hyperactivity scores (r = 0.28, p = 0.025). In the machine learning analyses, the AdaBoost classifier demonstrated the highest performance, achieving an accuracy of 81.82% and a ROC–AUC value of 0.85.

This study demonstrates that video-based movement analyses obtained during controlled, semi-structured interactions may capture motor activity patterns associated with ADHD. The findings are expected to contribute to the development of digital behavioral markers that may complement existing clinical assessment approaches in the context of ADHD evaluation.

Figures