WLA4ND: a Wearable Dataset of Learning Activities for Young Adults with Neurodiversity to Provide Support in Education
Hui Zheng, Pattiya Mahapasuthanon, Yujing Chen, Huzefa Rangwala, Anya S Evmenova, Vivian Genaro Motti · 2021 · Proceedings of the 23rd International ACM SIGACCESS Conference on Computers and Accessibility (ASSETS '21) · doi:10.1145/3441852.3471220
Summary
This paper introduces WLA4ND (Wearable Learning Activities for Neurodiversity), the first wearable sensor dataset of learning activities collected from young adults with neurodiversity. While existing wearable sensor datasets focus on fitness, daily living, and locomotion activities from neurotypical people, no prior dataset included learning-related activity data from neurodiverse individuals — a gap that could introduce bias into AI-based assistive technologies. Eight young adults with neurodiversity (including ASD, ADHD, Down Syndrome, and Intellectual and Developmental Disability) participated in 32 learning sessions. Participants were alumni of Mason LIFE, an inclusive postsecondary education program at George Mason University. Data was collected using a Fossil Gen 5 smartwatch running a custom Android Wear OS application that recorded accelerometer, gyroscope, gravity, linear acceleration, and heart rate data at 100Hz. The four learning activities — reading with follow-up Q&A (easy and medium levels), typing from slides, and prompt writing — were selected through brainstorming sessions with three special education investigators. Three coders annotated the data into six ground-truth labels (Read, Write, Write Q&A, Type, Off-task, Rest) by watching video recordings of sessions, achieving a Fleiss' Kappa inter-rater reliability of 98.79%. The dataset totals 14,211 data points across eight participants and is publicly available on GitHub alongside classification code.
Key findings
Five state-of-the-art activity recognition models were evaluated on WLA4ND, demonstrating that existing Human Activity Recognition (HAR) technologies can effectively classify learning activities from neurodiverse populations when trained on representative data. For user-dependent evaluation, the Convolutional Recurrent Neural Network (CRNN) achieved the highest balanced accuracy of 92.2% with an F1-score of 83.7%. For user-independent evaluation (leave-one-user-out), the Federated Multi-Task Hierarchical Attention Model (FATHOM) performed best with 91.8% balanced accuracy and 91.8% F1-score. Deep learning models (CRNN, FATHOM, MLP) consistently outperformed conventional Logistic Regression by 16-26% in balanced accuracy. The main classification challenge was distinguishing Write Q&A from Read activities, as both involve reading on paper. Feature analysis via FATHOM's attention mechanism revealed both variations and commonalities across participants with different neurodiverse conditions — accelerometer features were consistently important across participants, while heart rate and gravity sensor contributions varied. The heart rate sensor proved useful for distinguishing Rest from Off-task activities, which are otherwise similar in hand movement patterns.
Relevance
This research addresses a fundamental gap in inclusive AI: the absence of representative datasets from neurodiverse populations. Without such datasets, machine learning models trained solely on neurotypical data risk producing biased and ineffective assistive technologies for the approximately 17% of children in the US who present neurodiverse conditions. The practical implications are significant — smartwatch-based activity recognition could enable less intrusive monitoring and intervention in inclusive education settings, replacing or supplementing the constant presence of human support staff. For example, detecting Off-task behaviour could trigger gentle smartwatch reminders, while tracking sustained focus could deliver positive reinforcement — all without the stigmatization that verbal interventions from caregivers can cause. The dataset being openly available on GitHub enables other researchers to develop and evaluate activity recognition models for neurodiverse users. For accessibility practitioners, this work highlights the importance of collecting data from the actual target populations of assistive technologies rather than assuming that models trained on neurotypical data will generalize.
Tags: neurodiversity · wearable technology · machine learning · activity recognition · inclusive education · datasets · smartwatch · ADHD · autism · intellectual disability · Down syndrome