Scientists have taught an artificial intelligence to read stress directly from the body, and the results suggest that a simple wrist-worn sensor paired with deep learning can spot exam anxiety with remarkable accuracy. A research team led by Büşra Kocaçınar of Istanbul Kültür University and Cagatay Catal of Qatar University has developed a stress-detection framework that combines synchronized physiological signals from a commercial wearable device with psychological survey data from university students. Their best model, a long short-term memory (LSTM) neural network trained on balanced data, classified students’ states into normal, stressed, or upset with 94.14 percent accuracy, outperforming a simpler deep neural network (DNN) baseline that reached 86.42 percent. The study, published as an open-access paper in the International Journal of Data Science and Analytics, offers a blueprint for non-invasive, real-time stress monitoring in high-stakes settings such as classrooms, clinics, and workplaces.
The motivation behind the work is stark. Stress-related disorders account for a substantial share of global healthcare costs, and uncontrolled stress has been linked to depression, anxiety, cardiovascular disease, metabolic disorders, immunosuppression, obesity, and cognitive deficits. Traditional assessment methods, however, are poorly suited to capturing stress as it unfolds. Self-report questionnaires depend on individuals’ subjective and often delayed assessments, while cortisol-based approaches require saliva, blood, or urine testing and cannot deliver continuous data. Acute stress, in particular, is problematic: the body’s rapid fight-or-flight response during an examination directly shapes performance outcomes, yet conventional tools cannot track it moment by moment. Wearable biosensors promise an objective, real-time alternative by measuring the autonomic nervous system’s responses, including heart rate variability, electrodermal activity, and skin temperature fluctuations.
For their study, the researchers drew on the BIOSTRESS dataset, compiled with ethical approval from Istanbul Kültür University, involving 48 undergraduate students, 28 men and 20 women, with an average age of 22.31. Data collection unfolded in four structured stages: demographic screening, survey completion, controlled biosignal measurement sessions, and real examination environments. In the controlled setting, participants completed a 10-minute Test of Academic Distress (TOAD) stress induction and a calm 15-minute session with a licensed psychologist designed to emulate test anxiety. Eight volunteers also wore the sensor during actual in-class examinations, providing physiological recordings under genuinely high-stakes conditions. The team used an Empatica E4 wristband to continuously record blood volume pulse at 64 Hz, electrodermal activity at 4 Hz, tri-axial accelerometer data at 32 Hz, and skin temperature at 4 Hz.
Turning these heterogeneous signals into usable training data required a careful preprocessing pipeline. Each recorded session was segmented into non-overlapping 30-second windows, with the first and last 10 seconds discarded to remove artifacts from session start and stop. Signals were time-aligned and interpolated to uniform sampling intervals, producing an upsampled set of 1,620,978 data points and a downsampled set of 103,812. Electrodermal signals were filtered with a Butterworth bandpass filter between 0.05 and 0.5 Hz to suppress motion artifacts and slow drift, and statistical features such as mean, standard deviation, skewness, kurtosis, and range were extracted from each segment. Accelerometer signals were bandpass filtered between 0.3 and 2.5 Hz, and descriptive statistics were computed per axis. Labeling relied on participants’ self-reports immediately after each session, which were mapped onto the 30-second segments with the help of synchronized logs and validated by the supervising psychologist, producing three classes: normal, stressed, and upset.
Class imbalance posed one of the study’s central technical challenges. The downsampled dataset contained 55,569 normal, 46,646 stressed, and only 1,597 upset labeled samples, a skew that can cripple machine learning models. The team applied the Synthetic Minority Over-sampling Technique (SMOTE), which generates new synthetic examples by interpolating between a minority sample and its nearest neighbors, after an 80-10-10 train-validation-test split to avoid data leakage. Crucially, SMOTE was applied only to the training subset, leaving validation and test sets untouched so that evaluation reflected the original data distribution. Feature dimensionality was further reduced using linear discriminant analysis, which finds a projection maximizing the ratio of between-class to within-class scatter, and features were scaled using standard and min-max normalization fitted on training data alone.
Two deep learning architectures of deliberately different complexity were then compared. The DNN was a feedforward network with dense layers of 256, 128, 64, and 32 units, ReLU activations, dropout regularization, and an Adam optimizer with a learning rate of 0.001 and a batch size of 32, trained to minimize mean squared error. The LSTM, by contrast, used two stacked recurrent layers with 128 and 64 units, designed to preserve temporal dependencies across sequences, an essential property for biosignals whose stress signature lies in how they change over time. Each 30-second segment contained 1,920 time steps sampled at 64 Hz, exactly the kind of ordered structure a recurrent network can exploit. Reduced-capacity variants, a two-layer sDNN and a single-layer sLSTM, were trained as well. Hyperparameters including layer counts, hidden units from 32 to 256, learning rates from 0.0001 to 0.01, batch sizes of 16, 32, and 64, and dropout rates between 0.1 and 0.5 were tuned through a structured manual grid search with early stopping.
The results told a clear story about the value of temporal modeling. On downsampled data without oversampling, the LSTM reached 90.11 percent accuracy with a loss of 0.95, while the DNN lagged at 81.29 percent. When trained on upsampled, SMOTE-balanced data, the LSTM achieved its best result of 94.14 percent accuracy with a loss of 0.42, whereas the DNN improved to 86.42 percent. Simplified models showed only marginal gains from upsampling, indicating that the benefits of data balancing become most apparent in architectures with greater representational capacity. Pairwise Wilcoxon signed-rank tests confirmed that the upsampled LSTM significantly outperformed both the DNN and sDNN models at the p < 0.01 level, although the overall improvement from upsampling across all models was not statistically significant (p = 0.125), a nuance the authors highlight as a caution against over-relying on synthetic data. Computational analysis showed the trade-off clearly: the LSTM required roughly 11.3 seconds per training epoch, about 1.8 times slower than the DNN’s 6.2 seconds, owing to backpropagation through time.
Beyond model performance, the statistical analysis of survey data yielded findings with social relevance. Female participants reported significantly higher stress scores than males, and smokers reported higher stress than nonsmokers (means of 3.7 versus 2.2, p < 0.05), a difference confirmed by Mann-Whitney U testing. Students with lower GPAs reported higher stress, and freshmen averaged higher stress scores than upperclassmen. Intriguingly, smokers reported less anticipatory test anxiety despite showing greater physiological stress, a contradiction the researchers attribute to the gap between acute bodily arousal and cognitive-emotional appraisal, and possibly to smoking being used as a coping behavior that nonetheless elevates sympathetic activation. The team also found a statistically significant positive correlation between participants’ TOAD scores and their average electrodermal activity, with high-anxiety individuals showing consistently elevated EDA levels. This agreement between subjective reports and objective physiology validates both the labeling protocol and the underlying premise that wearables can quantify felt stress.
The authors are candid about the study’s limitations. Only eight participants were recorded during real examinations, weakening statistical power for the highest-stress condition, and the cohort of 48 undergraduates from a single institution limits generalizability. SMOTE-generated samples may oversimplify the multidimensional dynamics of real physiological responses and do not inherently preserve temporal consistency, prompting the authors to suggest time-series-specific augmentation in future work. The framework, they emphasize, is designed for acute stress recognition rather than chronic stress monitoring, which would demand longitudinal datasets and adaptive learning strategies. Nonetheless, the practical implications are considerable. Because the system relies on signals continuously acquirable by commercially available devices like the Empatica E4, it is compatible with real-world deployment. The researchers argue that simple architectures may suffice for resource-constrained mobile or embedded applications, while LSTMs suit clinical decision support where fidelity matters most. Future work will compare the deep models against conventional machine learning approaches, including support vector machines and random forests, and explore sensor fusion with behavioral and contextual data. If such systems mature, they could enable personalized, real-time stress interventions, from wearable alerts before burnout sets in to adaptive learning platforms that respond to a student’s physiological state during an exam.
Subject of Research: Deep learning-based detection of acute psychosocial stress in university students using multimodal wearable biosignals and survey data
Subject of Research: Technology and Engineering
Article Title: Deep learning-based stress detection using multimodal biosignals
Article References: Kocaçınar, B., Çöpürkaya, Ç., Meriç, E., Erik, E. B., Akbulut, F. P., & Catal, C. (2026). Deep learning-based stress detection using multimodal biosignals. International Journal of Data Science and Analytics, 22(1), Article 290. https://doi.org/10.1007/s41060-026-01270-y
Image Credits: AI Generated
DOI: 10.1007/s41060-026-01270-y
Keywords: stress detection, deep learning, LSTM, multimodal biosignals, wearable sensors, electrodermal activity, blood volume pulse, SMOTE, test anxiety, physiological signals
Cite Scienmag News
APA MLA Chicago
Blake Davidson. (September 5, 2026). Multimodal biosignals power deep learning stress detection. Scienmag. https://scienmag.com/multimodal-biosignals-power-deep-learning-stress-detection/
Blake Davidson. “Multimodal biosignals power deep learning stress detection.” Scienmag, 5 September 2026, https://scienmag.com/multimodal-biosignals-power-deep-learning-stress-detection/. Accessed 5 September 2026.
Blake Davidson. “Multimodal biosignals power deep learning stress detection.” Scienmag. September 5, 2026. https://scienmag.com/multimodal-biosignals-power-deep-learning-stress-detection/
Copy citation Download RIS
Tags: AI in healthcareapplication of AI in mental health assessmentdeep learning models for physiological datadeep learning stress detectionexam anxiety detectionhigh-stakes stress monitoringlong short-term memory neural networks in health diagnosticsLSTM neural networks for stress classificationmachine learning accuracy in stress classificationmultimodal biosignal analysis for stress detectionmultimodal biosignalsnon-invasive physiological signal analysisnon-invasive stress assessmentopen-access research on biosignal-based stress detectionphysiological and psychological data fusion forphysiological signal analysispsychological survey data integrationpsychological survey data integration with biosignalsreal-time stress detection in high-stakes environmentsreal-time stress monitoringstress-related health disordersstress-related health disorders and economic impactwearable physiological sensorswrist-worn wearable sensors for stress monitoring
