Exploring within-session separability between video-annotated self-injurious behavior and accelerometer-derived features in individuals with severe intellectual disabilities
Published in Frontiers in Digital Health, 2026
Summary
This brief research report explored whether video-annotated self-injurious behavior (SIB) events could be distinguished from non-SIB activity using wrist-worn accelerometer features within the same observation session. Two individuals with autism spectrum disorder and severe intellectual disabilities participated. Triaxial acceleration recordings were synchronized with video annotations, and movement-related features were compared between annotated SIB events and non-SIB periods.
The two cases showed contrasting patterns. In Case A, several acceleration-derived features distinguished the seven annotated SIB events relatively clearly from non-SIB activity. For the representative 0.25-second rolling range of vector magnitude, an exploratory threshold retained six of seven events while producing one false-positive candidate segment. In Case B, the six annotated SIB events overlapped substantially with high-acceleration non-SIB movements; capturing all six events required a highly permissive threshold that also produced five false-positive candidate segments.
Rather than developing or validating an automated classifier, this study characterized within-person, within-window separability as a preliminary constraint on accelerometer-based monitoring. Threshold results were exploratory and should not be interpreted as validated detection performance; false-positive candidate segments were not independently confirmed by video re-review.
From an HCI and digital health perspective, the study illustrates a video-anchored measurement workflow for assessing when wearable sensing can provide distinguishable event-level information in a particular observation context. It offers a foundation for person-specific calibration and human-in-the-loop monitoring systems, while highlighting the need to evaluate sensor limitations, annotation reliability, and variation across activities and settings.
Research line
- Measuring: operationalizing event-level separability between annotated target behavior and non-target activity within an observation window
- Wearable Sensing: comparing wrist-worn triaxial acceleration features with video-annotated events
- Human Activity Recognition: characterizing event-capture and false-positive candidate tradeoffs before classifier development
- Digital Health: evaluating the feasibility and limits of person-specific monitoring for self-injurious behavior
- Human-in-the-Loop Systems: using video-anchored calibration and candidate-segment review as potential steps toward semi-automated monitoring
Links
Recommended citation
Matsuda, S., Nakatani, K., Harada, A., Suzuki, K., & Inoue, M. (2026). Exploring within-session separability between video-annotated self-injurious behavior and accelerometer-derived features in individuals with severe intellectual disabilities. Frontiers in Digital Health, 8, Article 1909340. https://doi.org/10.3389/fdgth.2026.1909340
Recommended citation: Matsuda, S., Nakatani, K., Harada, A., Suzuki, K., & Inoue, M. (2026). Exploring within-session separability between video-annotated self-injurious behavior and accelerometer-derived features in individuals with severe intellectual disabilities. *Frontiers in Digital Health, 8*, Article 1909340. https://doi.org/10.3389/fdgth.2026.1909340
Download Paper
