Med-Stress: Medical Stress Monitoring

• Med-Stress is a research area focused on monitoring and detecting stress through wearable, contactless, and digital systems.
• It integrates multimodal signals such as physiological, neural, and behavioral data using fusion and representation learning techniques.
• Evaluation methods vary widely, emphasizing the need for robust, clinically translatable approaches in real-world settings.

Searching arXiv for recent and foundational papers on wearable and medical stress monitoring relevant to “Med-Stress.” “Med-Stress” (Editor’s term) denotes the body of work concerned with medical stress monitoring, detection, tracing, and management through wearable, contactless, and digital systems. In the supplied literature, the term spans acute stress event detection, perceived-stress estimation, chronic stress self-disclosure analysis, and just-in-time intervention workflows. Representative implementations range from wrist- and chest-worn physiological sensing to portable hdrEEG, thermal video, ultrawideband radar, multimodal office-behavior analysis, and social-media LLMs (Ninh et al., 2022, Kumar et al., 2020, Xu et al., 2024, Alqahtani et al., 29 Dec 2025).

1. Conceptual scope and stress constructs

Med-Stress research does not treat “stress” as a single operational target. One line of work focuses on acute stress, typically as a binary or multiclass state associated with laboratory or near-real-time physiological change. Examples include WESAD-style stress versus non-stress classification, cold-pressor versus warm-pressor discrimination, and real-time “Moments of Stress” derived from physiological events and self-confirmation in a mobile health intervention (Ninh et al., 2022, Kumar et al., 2020, Ta et al., 21 May 2025). A related line formalizes stress tracing as continuous prediction of a time-indexed stress sequence $S=\{s_1,s_2,\ldots\}$, emphasizing temporal evolution rather than isolated events (Xu et al., 2024).

A second line targets perceived or chronic stress, where the label is not a momentary autonomic event but a questionnaire-defined burden over days or months. In elderly cardio-oncology, the target is subject-level Perceived Stress Scale (PSS) at month 3 and month 6; in perceived-stress classification with EEG, GSR, and PPG, labels are derived from PSS-10 over the past 30 days; in social-media NLP, the task is identification of self-disclosures of chronic stress in English tweets (Kyprakis et al., 8 Apr 2026, Majid et al., 2022, Alqahtani et al., 29 Dec 2025). A third line frames stress as workload-related strain, using NASA-TLX or office-task conditions to model stress in sedentary knowledge work (Walambe et al., 2023).

This heterogeneity matters methodologically. Acute stress studies often emphasize short windows, autonomic reactivity, and event detection. Perceived-stress studies instead confront delayed, sparse, and subjective supervision. This suggests that Med-Stress is best understood as a family of stress-related inference problems linked by clinical or behavioral relevance, not by a single universal label definition.

2. Sensing substrates and observable correlates

Med-Stress systems are strongly multimodal. Wrist- and chest-worn platforms use combinations of electrodermal activity, blood volume pulse, skin temperature, heart rate, inter-beat interval, ECG, respiration, EMG, and accelerometry. The nurse dataset collected with Empatica E4 in a hospital includes EDA, HR, skin temperature, ACC, IBI, and BVP; subject-independent acute stress detection on WESAD uses EDA, BVP, and ST; context-aware fusion on WESAD distinguishes wrist sensors from chest sensors because the most informative contextual sensor differs by placement (Hosseini et al., 2021, Ninh et al., 2022, Rashid et al., 2023).

Other Med-Stress systems expand beyond standard wearables. Portable hdrEEG work links two single-channel biomarkers, ST4 and T2, to cortisol, HRV, pulse pressure, resilience, burnout, perceived stress, and trait anxiety, arguing that neural measures can add information not captured by endocrine, autonomic, or self-report measures alone (Maimon et al., 17 Sep 2025). Thermal-video work reconstructs ISTI, a cardiac sympathetic biomarker, from facial heat patterns and then classifies stress versus no-stress (Kumar et al., 2020). UWB radar work extracts heart-rate-related, respiratory-rate-related, and RF embeddings without body contact, then performs continuous stress tracing (Xu et al., 2024). Multimodal office monitoring adds facial expressions, posture, and computer interaction to physiology (Walambe et al., 2023). Social-media work shifts the sensing substrate entirely, treating language from Twitter and Reddit as a clinically adjacent signal of chronic stress disclosure (Alqahtani et al., 29 Dec 2025).

Modality class	Representative signals	Representative papers
Wearable physiology	EDA, BVP, ST, ECG, RESP, EMG, ACC	(Ninh et al., 2022, Rashid et al., 2023, Hosseini et al., 2021)
Neural or contactless physiology	EEG, hdrEEG, thermal ISTI, UWB radar HR/RR/RF	(Majid et al., 2022, Maimon et al., 17 Sep 2025, Kumar et al., 2020, Xu et al., 2024)
Behavioral or textual signals	face, posture, computer interaction, tweets	(Walambe et al., 2023, Alqahtani et al., 29 Dec 2025)

The literature repeatedly treats these channels as complementary rather than interchangeable. Cardiac and electrodermal signals index autonomic arousal; skin temperature contributes thermoregulatory information; accelerometry often functions as both predictor and confounder; neural signals provide a central correlate; and behavioral or textual channels expose stress expression at the level of action or self-disclosure.

3. Labels, datasets, and evaluation regimes

A defining feature of Med-Stress datasets is the weakness or partiality of supervision. In elderly oncology, one PSS score supervises many weekly smartwatch segments and ECG windows, producing a weakly supervised multiple-instance learning problem in which bags are variable-length and tied to a subject-horizon label rather than to instance-level annotations (Kyprakis et al., 8 Apr 2026). In the nurse dataset, event labels are generated by a hybrid pipeline: a pre-trained Random Forest proposes candidate stress episodes, end-of-shift surveys validate whether those periods were stressful, and the nurse provides stress level and contributing factors. The dataset therefore contains validated events, unvalidated model outputs, event times, durations, and contextual contributors, while unlabeled periods are not guaranteed to be true non-stress (Hosseini et al., 2021).

Naturalistic student monitoring uses direct self-report timestamps from smartwatch or app interaction: 54 students wore Apple Watch devices for 40 days, and the dataset contained 3,497 stress-event instances and 29,475 non-stress-event instances (Razavi et al., 2023). Real-time mHealth intervention work operationalizes stress moments through a hybrid physiological and subjective process in which algorithm-detected events are approved by the user and aggregated daily (Ta et al., 21 May 2025). At the other end of the spectrum, several benchmark studies rely on protocol-defined labels, such as TSST stress in WESAD, cold-pressor stress in thermal-video work, or cognitive-task stress in private BVP datasets (Kumar et al., 2020, Ali et al., 16 Mar 2025).

Evaluation protocols vary sharply in rigor. Subject-independent LOSO is common in stronger wearable studies, including oncology MIL, UniTS benchmarking on subject-level splits, TEANet, and subject-independent EDA/BVP/ST models (Kyprakis et al., 8 Apr 2026, Gabrielli et al., 2024, Ali et al., 16 Mar 2025, Ninh et al., 2022). By contrast, the WESAD transformer cross-modality study uses a random 85:15 split that includes all 15 participants in both train and test partitions; the paper itself therefore does not establish subject-exclusive generalization, despite reporting 99.73%–99.95% same-modality performance (Oliver et al., 26 Feb 2025). Ambulatory SMILE work makes the generalization problem explicit by reporting 90.77% internal accuracy and 91.24 F1 on an internal subset, but only 59.23% on challenge data, attributing the disparity to covariate shift (Dair et al., 2022).

This evaluation diversity has direct consequences for interpretation. Very high scores under participant-inclusive or internally recycled splits do not mean that stress monitoring is solved; the literature itself shows that subject independence, delayed labels, class imbalance, and distribution shift materially change the difficulty of the task.

4. Modeling paradigms

Two broad modeling traditions dominate Med-Stress. The first uses handcrafted physiological features with classical machine learning. Examples include XGBoost, Random Forest, ExtraTrees, SVM, LDA, and MLP on statistics derived from HR, HRV, EDA, GSR, temperature, and acceleration. In naturalistic student monitoring, XGBoost was the most reliable model, with reported AUC in the range of 0.64–0.66 in the primary comparison and a tuned-model test accuracy of 84.5% paired with AUC 0.57 (Razavi et al., 2023). In ambulatory ECG+GSR from the SMILE challenge, ExtraTrees with feature imputation yielded strong internal metrics but poor challenge generalization (Dair et al., 2022). In subject-independent acute stress detection from consumer-grade wrist signals, a simple multimodal neural network over engineered EDA, BVP, and ST features reached 94.50% mean accuracy and 94.16% balanced accuracy, outperforming prior WESAD baselines while maintaining a low standard deviation (Ninh et al., 2022).

The second tradition uses representation learning directly on raw or transformed signals. TEANet performs binary stress detection from raw 30-second BVP windows and reports 92.51% accuracy on RUET SPML and 96.94% on WESAD under LOSO, with augmentation specifically designed to rebalance the minority class (Ali et al., 16 Mar 2025). UniTS recasts stress detection as anomaly detection from HR and HRV sampled every 10 seconds from 60-second ECG or BVP windows, reporting F1 of 0.869 on DREAMER, 0.878 on MAHNOB-HCI, 0.834 on WESAD-ECG, and 0.856 on WESAD-BVP (Gabrielli et al., 2024). StressNet reconstructs ISTI from thermal facial video and reaches average precision 0.842 when using predicted ISTI for stress classification, versus 0.902 with ground-truth ISTI (Kumar et al., 2020). DST uses UWB radar-derived HR, RR, and RF embeddings, then reports an average 6.31% increase in detection accuracy over the best baseline across three datasets (Xu et al., 2024).

A third strand emphasizes fusion and adaptation rather than raw model depth. SELF-CARE learns a noise context from ACC for wrist devices or EMG for chest devices, selects branch classifiers specialized to different sensor subsets, and combines them with a Kalman-filter late-fusion layer; it reports 86.34% and 94.12% accuracy for wrist-based 3-class and 2-class stress classification, and 86.19% and 93.68% for chest-based classification (Rashid et al., 2023). In elderly oncology, multimodal wearable streams are converted into visual representations and processed by Tiny-BioMoE plus attention-based multiple instance learning, explicitly matching the weakly supervised setting in which one PROM score corresponds to many unlabeled windows (Kyprakis et al., 8 Apr 2026).

Taken together, these paradigms show that Med-Stress has not converged on a single canonical architecture. Feature-engineered models remain competitive in ambulatory physiology; weak supervision has made MIL important for clinical PROM-based problems; anomaly detection is used to support personalization and clinician control; and adaptive fusion is a response to sensor unreliability and context change.

5. Application domains and representative systems

Med-Stress has diversified across application domains rather than remaining confined to laboratory affective computing. In cardio-oncology, wearable stress estimation is framed as an adjunct to cardiotoxicity surveillance in an elderly multicenter breast cancer cohort of 387 patients. The model predicts perceived stress at month 3 and month 6 from smartwatch activity and sleep plus chest ECG, achieving global $R^2=0.24$ and Pearson $r=0.42$ at month 3 and $R^2=0.28$ and Pearson $r=0.49$ at month 6 under LOSO evaluation (Kyprakis et al., 8 Apr 2026). In occupational healthcare, the nurse dataset provides approximately 1,250 hours of physiological and accelerometry data gathered during COVID-era hospital shifts, with contextual stress surveys tied to detected events (Hosseini et al., 2021). A subsequent ensemble study built on a public nurse dataset and reported 93.7% macro F1 and 0.992 ROC-AUC on temporally held-out data, while also acknowledging unresolved issues in label specification, comfort, false alerts, and subject-independent evaluation (Sinhal et al., 10 Jul 2025).

In college mental health, two distinct Med-Stress directions appear. One is naturalistic event detection: 54 students wore watches for 40 days, and stress windows were built around self-reported high-stress taps (Razavi et al., 2023). The other is full closed-loop intervention: a 12-week randomized controlled trial of mHELP enrolled 125 students and analyzed 117 completers, with the treatment arm receiving real-time monitoring, breathing and focus exercises, counseling access, and other app-delivered supports. The primary outcome, “Moments of Stress,” declined substantially in the treatment group, whereas no significant between-group differences were observed for GAD-7, PHQ-8, or PSS (Ta et al., 21 May 2025).

Workplace stress in sedentary knowledge work is addressed through multimodal behavior fusion. Face, posture, computer interaction, and physiological inputs are combined in early and late fusion networks using SWELL-KW, with the paper reporting approximately 96% test accuracy for binary stress detection—96.09% in the abstract and 96.67% in the main table—and a NASA-TLX regression loss of 0.036 (Walambe et al., 2023). A plausible implication is that Med-Stress can extend beyond physiology-only pipelines when the environment is instrumented and privacy constraints are acceptable.

Contactless monitoring broadens the domain further. Thermal-video work targets privacy-preserving remote physiology through ISTI reconstruction, and UWB radar work explicitly argues for stress tracing without wearables (Kumar et al., 2020, Xu et al., 2024). At the opposite extreme, social-media NLP addresses chronic stress disclosure at population scale rather than physiology, with StressRoBERTa reaching 82% F1 on SMM4H 2022 Task 8 and 81% F1 on Dreaddit (Alqahtani et al., 29 Dec 2025). These systems are not direct clinical diagnostics, but they show that Med-Stress has expanded into ambient sensing and public-text surveillance as adjacent screening technologies.

6. Limitations, source validity, and future directions

A recurring limitation in Med-Stress is the gap between controlled benchmarks and realistic deployment. Laboratory datasets such as WESAD enable high performance but narrow definitions of stress, whereas real-world cohorts introduce weak labels, missing data, confounding activity, and subject heterogeneity. The ambulatory ECG+GSR study makes this explicit through the drop from 90.77% internal accuracy to 59.23% on challenge data (Dair et al., 2022). The student smartwatch study likewise reports modest discrimination, with AUC near 0.64–0.66, and notes inconsistencies between abstract- and table-level metrics (Razavi et al., 2023). The oncology MIL study characterizes its own results as moderate rather than diagnostic-grade, which is consistent with the mismatch between sparse PSS labels and dense wearable streams (Kyprakis et al., 8 Apr 2026).

Another limitation is cohort specificity. Elderly women with breast cancer at cardiotoxicity risk, healthy university students, hospital nurses during COVID-19, healthy adults in laboratory stress paradigms, and English-speaking social-media users are not interchangeable populations (Kyprakis et al., 8 Apr 2026, Hosseini et al., 2021, Alqahtani et al., 29 Dec 2025). This suggests that Med-Stress models are still heavily domain-bound. The literature also records unresolved translational issues: motion artifacts in wearable PPG, comfort and compliance problems in continuous wear, high false-alert burden, missing modality robustness, privacy risk in face or computer logging, and the absence of strong clinician-facing interpretability in many systems (Gabrielli et al., 2024, Sinhal et al., 10 Jul 2025, Walambe et al., 2023).

Source validity is itself a nontrivial issue. One arXiv record associated in the supplied material with “Personalized Stress Monitoring using Wearable Sensors in Everyday Settings,” (Tazarv et al., 2021), is described in the same material as an IEEEtran demo template rather than a scientific stress-monitoring paper. It therefore does not support claims about wearable physiology, personalized labeling, or machine-learning results. This source-level problem is a reminder that Med-Stress research depends not only on algorithmic rigor but also on careful document verification.

Future directions in the supplied literature converge on a few themes: leakage-aware temporal alignment and weak-label learning for sparse PROM settings; adaptive or personalized baselines rather than static population classifiers; context-aware or selective fusion under changing sensing conditions; lower-burden contactless sensing; and integration of monitoring with intervention, escalation, or clinician review (Kyprakis et al., 8 Apr 2026, Rashid et al., 2023, Xu et al., 2024, Ta et al., 21 May 2025). This suggests that the next phase of Med-Stress will likely be defined less by single benchmark gains than by robustness across populations, transparent evaluation, and clinically coherent linkage between stress sensing, decision support, and longitudinal care.