Stress Detection of Autistic Adults during Simulated Job Interviews Using a Novel Physiological Dataset and Machine Learning
Miroslava Migovich, Deeksha Adiani, Michael Breen, Amy Swanson, Timothy J. Vogus, Nilanjan Sarkar · 2024 · ACM Transactions on Accessible Computing · doi:10.1145/3639709
Summary
This research addresses the significant employment barriers faced by autistic adults, with job interviews identified as a primary obstacle. The researchers developed a physiological stress detection system integrated with CIRVR (Career Interview Readiness in Virtual Reality), a simulated interview platform designed with input from autistic job seekers and career counselors. Fifteen autistic young adults (ages 17-27) participated in simulated data entry job interviews while wearing an Empatica E4 wristband that collected physiological signals: photoplethysmography (heart rate, blood volume pulse, interbeat interval), electrodermal activity (skin conductance level and response), temperature, and 3-axis accelerometer data. A Registered Behavior Technician manually labeled stress levels on a 1-10 scale every 15 seconds while reviewing video recordings, creating ground truth labels for three stress classes (low, medium, high). The study represents the first publicly available dataset of physiological stress data specifically from autistic individuals during simulated interviews, addressing a significant gap in affective computing research. The interview structure included greeting, previous work experience, technical questions, education, personal questions, and an opportunity for the interviewee to ask questions—with built-in interruptions (phone calls, door knocks) designed to simulate realistic stressors.
Key findings
Five supervised machine learning algorithms were evaluated for stress detection. For individual models (80/20 train/test split), Elastic Net Regression achieved the highest accuracy at 84.8% with an F1 score of 0.85. For group models using leave-one-out cross validation, Support Vector Regression performed best with 75.4% accuracy and F1 of 0.64—still useful for deployment scenarios where individual calibration sessions are impractical. The stress visualization analysis revealed that open-ended questions without single correct answers produced the highest stress and greatest variation among participants. The "Questions" section—when interviewees were asked if they had any questions about the job—had the highest average stress (1.00 on a 0-2 scale). The Personal section, which included questions about handling difficult situations with coworkers, followed closely (0.99) and showed the widest variation in stress responses. The data visualization tools—including line graphs with question markers, sunburst diagrams showing stress by interview section, and comparative charts—enabled identification of specific questions that triggered stress spikes for individual participants, providing actionable insights for both interviewees and interviewers.
Relevance
This work has direct implications for improving employment outcomes for autistic adults, who face unemployment rates far exceeding the general population despite often possessing skills required for available positions. The physiological approach to stress detection is particularly valuable for autistic individuals, for whom traditional self-report measures may be less reliable due to differences in interoception and emotional recognition. For accessibility practitioners and employers, the findings highlight specific interview adaptations that could reduce stress: providing structured rather than open-ended questions, avoiding or preparing candidates for interruptions, and giving advance notice about the "do you have any questions" portion. The visualization tools offer a model for how employers could review and adapt their interview processes to be more neurodiversity-inclusive. The dataset itself—made available by the researchers—enables future work on affective computing systems designed specifically for autistic users, addressing a critical gap since most existing physiological stress detection research has focused on neurotypical populations. Future applications could include real-time stress feedback during interview training or adaptive interview systems that adjust difficulty based on detected stress levels.
Tags: autism · employment · job interview · physiological sensing · affective computing · machine learning · stress detection · wearable sensors · virtual reality