Source-linked AI summary
Stress Detection Using Wearable Physiological and Sociometric Sensors
Oscar Martinez Mozos, Virginia Sandulescu, Sally Andrews, David Ellis, Nicola Bellotto, Radu Dobrescu, Jose Manuel Ferrandez
TL;DR
Stress can negatively affect health and remains difficult to monitor automatically in everyday social settings. This paper combines wearable physiological and sociometric sensors with machine-learning classifiers to distinguish stressful from neutral situations during the TSST. The combined approach discriminated stress with high confidence and was evaluated for real-time detection, modality-specific usefulness, and feature discriminativeness.
Problem
Stress is a significant social problem, and automatic early detection could improve quality of life and reveal mechanisms underlying negative behaviours and cognitions.
Method
The study combines physiological wearable measurements with sociometric activity and voice measurements, then applies classifiers to TSST data.
Results
The combined sensor measurements accurately discriminate stressful from neutral TSST situations and support real-time stress-state decisions.
Takeaways & Limitations
Combining physiological and sociometric sensors provides a useful tool for real-time stress detection in controlled social situations.
Takeaways & Limitations
The combined sensors may be uncomfortable for long-term daily use, with the physiological sensor more suitable for controlled environments.
Abstract
from arXiv · showhide
Stress remains a significant social problem for individuals in modern societies. This paper presents a machine learning approach for the automatic detection of stress of people in a social situation by combining two sensor systems that capture physiological and social responses. We compare the performance using different classifiers including support vector machine, AdaBoost, and k-nearest neighbor. Our experimental results show that by combining the measurements from both sensor systems, we could accurately discriminate between stressful and neutral situations during a controlled Trier social stress test (TSST). Moreover, this paper assesses the discriminative ability of each sensor modality individually and considers their suitability for real-time stress detection. Finally, we present an study of the most discriminative features for stress detection.
1. Introduction
Stress harms health and daily functioning, while shortages of caregivers motivate wearable technologies for automatic monitoring. This paper combines physiological and social activity measurements to detect stress during social situations.
- Stress can impair attention, memory, and decision-making, while prolonged high stress correlates with anxiety, depression, and premature ageing.
- 13.3 million working days are lost annually to stress-related illnesses, and stress costs UK enterprises around 8.4 million pounds.
- Declining caregiver availability motivates technologies that automatically monitor physical and mental health during daily activities and social interaction.
- The proposed system combines physiological signals and sociometric activity measurements, using classifiers to distinguish stressful from neutral TSST situations.
- The study uses personalized binary classifiers because stress responses vary substantially between individuals.
2. Related Work
Prior stress-detection systems use physiological, activity, or voice signals, but often target restricted tasks or settings. This work combines wearable physiological and social measurements for personalized stress detection in social situations.
- Reported physiological systems achieve accuracies from 82.8% to 96.6%, but some omit classifier details or are difficult to deploy as everyday wearables.
- Earlier systems commonly rely on physiological measurements alone, whereas this study adds sociometric activity and social-context information.
- The TSST provides neutral and stressful conditions in a social setting rather than through a computer-based task.
- The proposed wearable system permits freer movement than desktop-based workplace systems while incorporating voice with physiological and activity measurements.
- Unlike a prior five-participant study, this work combines physiological and activity signals and evaluates 18 participants.
- Sociometric badges capture unconscious social signals that may complement incomplete or inaccurate verbal or written reports of cognitive states.
3. Sensor Modalities
The system uses two wearable modalities: a wrist-worn physiological sensor and a conference-badge sociometric sensor. Their signals provide physiological, movement, speech, and voice features for stress analysis.
- The physiological wristband records electrodermal activity, photoplethysmogram, and heart-rate variability through finger electrodes.
- Figure 1 contrasts the wristband form factor of the physiological sensor with the conference-badge form factor of the sociometric sensor.
- Physiological signals are acquired at 1000 Hz, down-sampled to 10 Hz, filtered, and processed for photoplethysmogram templates and recorded features.
- The sociometric badge includes a microphone, accelerometer, Bluetooth transmitter, and infrared transmitter, although this study uses only microphone and accelerometer measurements.
- Sociometric features represent body movement, posture, speaking, silence, speech volume, volume consistency, frequency peaks, and voice pitch.
4. Experimental Protocol
The protocol uses the Trier Social Stress Test to contrast neutral activities with socially stressful speaking and cognitive tasks, while measuring participants’ anxiety before and after the session.
- TSST design: The TSST combines public speaking and cognitive tasks to induce stress in a controlled social setting.Neutral tasks are treated as non-stressful, while speaking and cognitive tasks are treated as stressful conditions.
- Session sequence: Each session begins with a neutral task, followed by presentation preparation, public speaking, a cognitive task, and a final neutral task.The neutral tasks use predefined questions; the speaking task is a mock job interview, and the cognitive task requires counting backward by 13 from 1022.
- Anxiety assessment: Participants completed the State Trait Anxiety Inventory before and after the TSST to measure state anxiety and also report trait anxiety.The questionnaire contains separate 20-item state and trait sections rated on a four-point scale.
- Protocol scope: The experiment used 18 participants who completed the TSST individually under a protocol with approximately specified task durations.Task times were tentative and varied by a few seconds between participants.
5. Data Collection
Data were collected from 18 participants wearing synchronized physiological and sociometric sensors, producing participant-specific labeled time-series datasets for stress classification.
- Participants and recording: Eighteen students aged 18 to 39 completed individual TSST sessions while wearing both sensor systems.The sample included males and females, and the study received ethics approval.
- Sensor integration: Physiological and sociometric signals were recorded simultaneously, synchronized by timestamp, and resampled to a common 10 Hz frequency.The physiological signals included EDA and PPG; the sociometric badge supplied activity signals.
- Dataset construction: Each participant’s synchronized samples were stored in a separate dataset and labeled stressed or neutral according to the corresponding TSST task.Neutral activities were assumed to represent non-stressed states.
- Extracted features: Collected features included PPG templates, EDA, heart-rate variability, and body-movement rate aligned with the ongoing TSST task.Example feature signals were shown for participants P6 and P16.
6. Classification and Evaluation
The study compares personalized binary classifiers for stress detection using participant-specific sensor datasets, evaluates them with standard classification metrics, and examines task-aligned signals and feature discrimination.
- Classifiers: The study compares support vector machines and AdaBoost for binary discrimination between stressed and non-stressed examples.The SVM experiments include radial-basis-function and linear kernels, while AdaBoost combines one-dimensional weak classifiers.
- Signal interpretation: The task timeline in the signal examples distinguishes neutral, preparation, public-speaking, cognitive, and final-neutral periods.The figure presents example signals for participants P6 and P16 alongside the task performed at each time.
- Personalized classification: Personalized classifiers were trained separately for each participant using synchronized physiological and activity signals.Each participant’s dataset was divided into training and testing subsets using stratified random selection.
- Evaluation setup: Each classifier uses 75% of a participant’s data for training and 25% for testing with matched class distributions.The split is performed independently for each participant-specific dataset.
- Evaluation metrics: Accuracy, precision, and recall are used to evaluate the personalized classifiers, with higher values indicating better results.Stress examples are treated as the positive class for these metrics.
7. Experimental Results
The experiments evaluate personalized stress classifiers using combined physiological and sociometric signals, compare sensor modalities, and examine discriminative features during the TSST. Combined sensing generally performs best, while AdaBoost supports accurate personalized detection with a small feature subset.
- Experimental setup: Each participant completed one TSST while wearing both physiological and sociometric sensors, enabling analysis of synchronized stress-related signals.
- Classifier comparison: AdaBoost and RBF-kernel SVM achieved similar combined-sensor results and outperformed the linear SVM, with AdaBoost providing slightly higher accuracy and precision.The authors identify both AdaBoost and RBF-based SVM as suitable classifiers, while noting AdaBoost requires no individually trained parameters.
- Classifier comparison: Accuracy, precision, and recall were similar across participants for the personalized AdaBoost classifiers, supporting consistent stress detection across individuals.
- Classifier comparison: The AdaBoost confusion matrix showed high values on its main diagonal, indicating many test examples were assigned their correct stress or non-stress labels.
- Classifier comparison: During complete TSST sessions, the classifier usually predicted the correct state, although false alarms occurred around transitions between relaxed and stressed situations.The authors suggest these transitions may require time and identify them as a topic for future study.
- Sensor modality comparison: The combined sensor modality produced better classification results than either single modality, while the sociometric sensor performed better than the physiological sensor alone.The authors relate the sociometric result to continuous speaking and permitted movement during the TSST.
- Feature analysis: AdaBoost ranked recurring discriminative features including eda, posact, hz3f, bmact, and amp3f, with respective appearance rates of 16%, 15%, 14%, 13%, and 13%.
- Feature analysis: Using only the five most discriminative features still produced high classification results, indicating that many device-derived features could be omitted from personalized classifiers.
8. Conclusion
The paper combines wearable physiological and sociometric sensors for real-time stress detection, validating the approach in controlled TSST sessions. It also identifies practical boundaries for everyday use, including sensor comfort and the limited generalizability of controlled stress testing.
- Conclusion: The combined wearable system provides accurate stress-state decisions at any time and is presented as useful for real-time stress detection.The authors suggest this capability could support future feedback and early warning systems.
- Conclusion: Controlled TSST sessions provided the validation setting, but they do not completely represent general everyday activities.The authors describe controlled validation as a first step toward more reliable results.
- Conclusion: Long-term comfort limits the combined solution for day-to-day stress detection, especially because the wireless physiological sensor may be uncomfortable.The sociometric sensor can be worn as a badge for longer periods, while the physiological sensor may suit controlled environments better.
- Conclusion: The sociometric badge alone achieved high classification rates, although lower than the combined physiological-and-sociometric solution, and may suit everyday stress monitoring.Its ease of wear and ability to record social interactions support further study in daily activities and social settings.
- Conclusion: Sensors similar to those in the sociometric badge are available in intelligent bracelets, smartphones, and smartwatches for future longer-term stress studies.The authors propose examining these devices alongside self-report beyond the psychological laboratory.