Abstract
Children with Autism Spectrum Disorder (ASD) often experience a significant "gap" between their outward behavior and internal affective states. This study investigates whether Electrodermal Activity (EDA) can reveal the socio-cognitive load during social interactions with an Embodied Conversational Agent (ECA). By analyzing physiological signals during the pauses before a child speaks, researchers predicted response latency with up to 70% accuracy, uncovering hidden layers of social stress that are not always visible to observers.
Introduction
For children on the autism spectrum, traditional observational methods (video or audio) often fail to capture internal stress or intense cognitive effort. A child might appear calm while experiencing a high "socio-cognitive load." This research explores Electrodermal Activity (EDA)—a measure of sympathetic nervous system activity via sweat glands—as a window into these internal states. Specifically, the study examines verbal response latency (the time taken to respond to a prompt) as a behavioral marker of this internal processing, asking whether the body's physiology differs when a child struggles to find the right words.
Methodology
The study utilized the "Rachel ECA Interaction Corpus," involving 9 verbally fluent children with ASD (aged 6-12) interacting with a virtual character named "Rachel" in a "Wizard of Oz" setup.
Latency and Interval Definition
Researchers calculated the time between the end of the ECA's turn and the start of the child's response. Responses were categorized as "short" or "long" based on the 70th percentile of durations for each individual child.
Feature Extraction and Classification
48 physiological features were extracted from the EDA signal during a 5–10 second "effect interval" following the ECA's prompt. These included time-domain statistical moments, extrema (peaks and valleys), and frequency-domain power spectral density. A K-Nearest Neighbor (KNN) classifier was then used to distinguish between short and long latency intervals based solely on these physiological features.
Key Results and Insights
The system achieved classification accuracies ranging from 50.3% to 70.3%, highlighting the high degree of individuality in ASD social processing.
- Heterogeneity in Arousal: The most striking finding was the variability between subjects. For some children, long latencies correlated with higher arousal, suggesting stressful "thinking time." Conversely, others showed higher arousal during short latencies, indicating that rapid social interaction itself was more stressful than the pauses used for cognitive processing.
- Temporal Window: The most discriminative physiological information typically appeared 5 to 8 seconds after a prompt. If measured too early, the body hasn't reacted; too late, and the signal fades.
- Feature Importance: The baseline arousal level (the first three moments of the EDA signal) was the most consistently useful feature across all subjects.
Conclusion and Future Work
This proof-of-concept study demonstrates that EDA provides a "complementary view" of a child's state that is often invisible to audio-visual observation. While limited by a small sample size, the results suggest that physiology can serve as a reliable indicator of socio-cognitive load. Future work aims to integrate multiple modalities—such as acoustic, lexical, and gestural features—and explore "physiological synchrony" between children and parents. Ultimately, these insights could lead to real-time, dynamically personalized virtual interfaces that adapt to a child's internal stress levels.
Summary updated