Peer-reviewed article
ASC-emotion: A privacy-aware dataset for analysing emotional dysregulation and engagement in children with Autism.
Zakia Batool Turabee, David J Brown, Mufti Mahmud, Andreas Oikonomou, Muhammad Arifur Rahman et al. ·
Summary
Written by AI, not yet AICS's view · claude-haiku-4-5-20251001This is a technical paper describing the development of a machine learning dataset for detecting emotional dysregulation in children with autism. The researchers created a privacy-protected video dataset and tested three AI models to identify affective states, achieving 96% accuracy. The work addresses a gap in available datasets for autism research but does not evaluate an intervention or measure child outcomes.
Key points
- A new dataset was created specifically for detecting emotional dysregulation and engagement in children with autism
- Privacy-preserving measures were implemented to protect vulnerable participants in video data collection
- Three machine learning models (BiLSTM, GraphConv, PointCNN+LSTM) were tested and achieved 96% accuracy in detecting affective states
- The dataset and models enable early detection of meltdowns to support timely intervention
- The work was funded by an Erasmus+ project focused on AI applications in autism support
Limitations
This is a dataset development paper, not an intervention evaluation. It does not report on actual implementation of the detection system with children, effectiveness of any interventions, or real-world impact on child well-being.
Abstract
Predicting emotional dysregulation events in children with Autism is essential for timely mitigation of triggering events and prevention of further escalation of the situation. However, there is a scarcity of accessible and standarised datasets for use in AI-based research associated with challenging behaviours in children with ASC. To address this gap, we have curated a novel privacy-preserved dataset as part of an Erasmus+ funded project (AI-TOP-2020-1-UK01-KA201-079,167). The dataset was validated using three machine learning architectures which gave an accuracy of 96% in detecting the affective states related to learning in children with Autism. The key contributions of this study are: 1. Development of an Autism meltdown dataset exemplifies methodological rigor, advances an urgent clinical challenge through early detection and intervention, and enables broad impact by promoting reproducibility, benchmarking and translational health outcomes. 2. Implementation of privacy preserving measures addresses ethical concerns regarding the use of video data with this vulnerable population as part of the machine learning pipeline. 3. High levels of accuracy are demonstrated via empirical validation of the dataset through three machine learning models (BiLSTM, Graphical Neural Network - EdgeConv, PointCNN+LSTM) for detecting affective states related to learning and physiological arousal in children with Autism.