# Leveraging point-of-view camera and MediaPipe for objective hyperactivity assessment in preschool ADHD

## TL;DR

This study explores using camera recordings and movement analysis to objectively assess hyperactivity in preschool children with ADHD.

## Contribution

It introduces a low-burden, vision-driven method using POV cameras and MediaPipe for quantifying hyperactivity in preschool settings.

## Key findings

- Hyperactive children showed significantly greater movement in distal limb segments compared to non-hyperactive peers.
- A machine learning model achieved 84.31% accuracy in classifying hyperactivity risk based on movement features.
- Distal limb movements were identified as the most informative features for classification.

## Abstract

Attention-Deficit/Hyperactivity Disorder (ADHD) often emerges in early childhood, with hyperactivity and impulsivity constituting the most prominent symptoms during the preschool period. Current assessment approaches rely largely on clinical interviews and behavior rating scales, which are susceptible to subjectivity and contextual bias. Objective, ecologically valid, and low-burden methods for quantifying hyperactivity in preschool settings remain limited.

This observational, cross-sectional study investigated whether movement-based features extracted from teacher-worn point-of-view (POV) video recordings could differentiate preschool children at risk for ADHD-related hyperactivity from non-hyperactive peers. Fifty-one preschool children (48–60 months) participated in a standardized, three-minute storytelling interaction conducted in a familiar classroom environment. Video recordings were processed using MediaPipe pose estimation to derive region-specific movement indices across multiple body segments. Group differences were examined using statistical analyses. In addition, supervised machine learning models were applied to evaluate classification performance based on movement features.

Children in the hyperactivity-risk group exhibited significantly greater movement across several body regions, particularly distal upper- and lower-limb segments, compared to non-hyperactive peers. Machine learning analyses indicated promising classification performance, with the support vector machine achieving an accuracy of 84.31%, sensitivity of 80.0%, specificity of 87.10%, and an area under the receiver operating characteristic curve (AUC) of 0.83. Permutation-based feature importance analyses highlighted distal limb movements as the most informative features for classification.

These findings suggest that POV-based, vision-driven assessment provides a promising, objective, and ecologically valid approach for quantifying hyperactivity-related motor behavior in preschool children. While not intended as a standalone diagnostic tool, this low-burden framework may serve as a valuable complement to existing screening practices and support early identification efforts in educational settings.