Journal DOI 10.68956/journal.human-performance-futures

Creative Commons source reading · Machine learning for performance

The Development of a Real-Time Artificial Intelligence System Using a Convolutional Neural Network for the Assessment of Infant Head Control

Akihiro Hosono; Katsura Ito; Mari Ichikawa

Originally published in JMA J · 2026-05-29

From the article

Selected source text. JournalOS removes markup and shortens excerpts where indicated. The full article provides the complete context.

Abstract

Introduction:To develop a real-time artificial intelligence (AI) system using a convolutional neural network for the assessment of infant head control (good or poor) during the traction response.Methods:Infants aged 3-6 months who received community-based group health checkups and whose guardians provided informed consent were enrolled in this study. Videos were recorded during the pull-to-sit test conducted by a pediatrician. The collected videos were divided into training and test datasets. Training data were extracted as image frames and labeled according to the pediatrician’s judgment. After supervised machine learning, the AI system was evaluated using the test dataset, and the sensitivity, specificity, and detection rate were calculated.Results:Altogether, 184 videos were collected and divided into the training (n = 110) and…

Results

From a total of 128 infants, 184 videos were collected and divided into 110 videos for the training dataset and 74 videos for the test dataset. Discrepancies were observed between the two reviewers in two videos (overall agreement, 99.2%), both of which were included in the test dataset. In these discrepant cases, the videos were conservatively classified as screening-positive (i.e., head lag-negative) based on clinical judgment, prioritizing patient safety. Given that one infant underwent a retest, the total number of videos included two retest videos recorded from both sides. The training dataset consisted of 103 videos showing infants with good head control and seven videos showing infants with poor head control, with 6,576 images extracted for supervised machine learning. The testing dataset with 74 videos included 64 videos…

Discussion

In this study, we developed and evaluated a real-time AI system based on a convolutional neural network to assess infant head control during the pull-to-sit test. The results indicated that the proposed AI system achieved high specificity in identifying good head control (absence of head lag), whereas sensitivity for poor head control remained limited, suggesting a supportive rather than a standalone role in infant health assessments. A key challenge in the evaluation of infant head control is the lack of fully standardized assessment criteria and the inherent subjectivity of clinical judgment. To address this issue, head control in the present study was evaluated by pediatricians using criteria consistent with commonly used clinical and neurological assessment scales, including the Premie-Neuro Examination, Head Control Scale,…

Editorial reading guide

  • What data created and tested the model, and was evaluation independent?
  • What is the cost of a false positive, false negative or out-of-distribution decision?