Stress and Well-being Dataset Overview

• Stress and well-being datasets are collections of multimodal data—including surveys, wearables, and digital traces—that operationalize diverse stress and wellness dimensions.
• They use varied methods such as self-reports, sensor recordings, and online narratives to capture both acute and chronic stress indicators.
• These datasets inform computational models and cross-study comparisons, advancing our understanding of stress dynamics and overall well-being.

Searching arXiv for papers on stress and well-being datasets to ground the article. arXiv search query: stress well-being dataset multimodal longitudinal survey wearable social media Stress and well-being datasets are research corpora and measurement resources that operationalize stress, affect, sleep, wellness dimensions, or broader psychosocial burden through text, self-report, wearable physiology, passive digital traces, longitudinal tabular records, or synthetic simulation. In recent arXiv work, the category includes crisis narratives in Dari, real-world physiological recordings tied to cannabis use and perceived stress, daily synthetic simulations of work-life dynamics, Reddit corpora for stress expression, multimodal interview and classroom protocols, student survey datasets, and even an annual socioeconomic index intended to track population well-being (Baktash et al., 28 Mar 2026, Azghan et al., 24 Mar 2025, Husseini, 28 Dec 2025, Turcan et al., 2019, Shakeel et al., 13 Jul 2025, Trindade et al., 2020). The resulting landscape is methodologically heterogeneous: some resources target acute stress induction, some capture chronic strain, some foreground interpretability and explanatory spans, and some are explicitly not empirical human datasets.

1. Conceptual scope and dataset types

Within the literature, “stress and well-being dataset” does not denote a single canonical format. AFSTRESS is a survey-based, multi-label corpus of self-reported stress narratives in Dari, created to support computational, social-science, and psychological analysis during an ongoing humanitarian crisis (Baktash et al., 28 Mar 2026). CAN-STRESS is a real-world multimodal dataset of Empatica E4 recordings and event-level self-reports designed to examine cannabis use, stress, and physiological response in daily life (Azghan et al., 24 Mar 2025). FLOW is a synthetic, feedback-driven longitudinal dataset intended as a controlled experimental environment for work, lifestyle, and well-being dynamics rather than as a proxy for observed human populations (Husseini, 28 Dec 2025). Dreaddit and Holistix treat stress or wellness as phenomena expressed in online discourse, using Reddit and Beyond Blue forum posts respectively (Turcan et al., 2019, Shakeel et al., 13 Jul 2025).

Two recurring boundary cases clarify the term’s scope. Nora describes a runtime monitoring and logging system that stores stress, emotion, sentiment, health checks, and session progress, but it does not release a standalone benchmark dataset with dataset size, train/dev/test splits, or annotation protocol (Winata et al., 2021). The TILES-2018 Sleep Benchmark is primarily a wearable sleep dataset for hospital workers; it supports stress- and well-being-oriented research indirectly through the larger TILES protocol and participant metadata, but the released benchmark itself is sleep-centered rather than a dedicated stress corpus (Feng et al., 4 Jul 2025).

This heterogeneity implies that dataset choice is inseparable from construct definition. Some resources operationalize stress as self-reported perceived strain, some as protocol condition labels, some as wellness-dimension class membership, some as psychophysiological response, and some as long-horizon socioeconomic movement. A plausible implication is that “stress detection” results are only comparable when the underlying construct, annotation source, and time scale are also comparable.

2. Collection paradigms and observable modalities

The main collection paradigms are text elicitation, wearable sensing, passive behavioral logging, structured surveys, and simulation. They differ in whether they prioritize ecological validity, temporal continuity, interpretability, or experimental control.

Dataset	Primary data source	Core target
AFSTRESS	Dari survey narratives	12 binary emotion/stressor labels
CAN-STRESS	Empatica E4 + event self-reports	perceived stress around daily events
Dreaddit	Reddit five-sentence segments	binary stress classification
Holistix	Beyond Blue forum posts	six wellness dimensions + spans
ForDigitStress	interview audio/video/PPG/EDA	time-continuous stress and emotions
VitaStress	wrist wearable laboratory protocol	baseline vs stress; multi-class stress
Stress Bytes	URL/app traces + monthly surveys	PSS-10 stress over time
FLOW	rule-based synthetic simulation	daily work/lifestyle/wellbeing dynamics

Text datasets span solicited narratives and naturally occurring social disclosure. AFSTRESS contains 737 responses collected through an online Dari survey in 2026, with demographic questions, a free-text narrative, and a checklist of applicable labels (Baktash et al., 28 Mar 2026). Dreaddit scraped posts from ten subreddits across abuse, anxiety, financial, PTSD, and social domains, then labeled five-sentence segments rather than whole posts (Turcan et al., 2019). Holistix extracted 2,000 raw Australian Beyond Blue posts, cleaned them to 1,420 posts focused on mental distress, and annotated both dominant wellness class and supporting text spans (Shakeel et al., 13 Jul 2025).

Wearable and multimodal datasets span ambulatory, semi-naturalistic, and controlled settings. CAN-STRESS records ACC at 32 Hz, BVP at 64 Hz, EDA at 4 Hz, HR at 1 Hz, IBI, and TEMP at 4 Hz during a 24-hour free-living day (Azghan et al., 24 Mar 2025). The IoT-enabled everyday biosignal study uses a Samsung Gear Sport smartwatch to capture PPG, acceleration, gyroscope, gravity, and EMA stress labels in 2-minute windows every 15 minutes for up to 4 times per hour (Tazarv et al., 2021). ForDigitStress combines room audio, close-talk audio, HD video, skeleton data, facial points, head position, facial action units, eye video, PPG, and EDA during a digital job interview (Heimerl et al., 2023). MultiPhysio-HRC adds EEG, ECG, EDA, RESP, EMG, voice, webcam video, and facial action units in industrial human-robot collaboration scenarios (Bussolan et al., 1 Oct 2025). VitaStress uses a Corsano CardioWatch to derive heart rate, RR intervals, respiration rate, EDA, temperature, and accelerometry, with processed data resampled to one sample every 30 seconds (Schreiber et al., 14 Aug 2025).

Passive digital behavior datasets shift the observable from physiology to behavioral context. Stress Bytes links 47,100,701 URL visits, 13,553,645 app visits, sociodemographics, and repeated monthly PSS-10 measurement across a seven-month panel of German internet users (Belal et al., 21 May 2025). Earlier daily stress recognition work uses Android call logs, SMS logs, Bluetooth proximity scans every 5 minutes, daily weather conditions, and Big Five personality traits to infer evening self-reported daily stress (Bogomolov et al., 2014).

Synthetic and macro-level resources occupy a distinct methodological niche. FLOW simulates 1,000 individuals from January 2024 to December 2025 at daily resolution, yielding 731,000 observations in the core daily tables (Husseini, 28 Dec 2025). The Socioeconomic Well-Being Index aggregates thirteen annual socioeconomic factors over 1986–2016 into an annual return series intended as an early-warning signal for deteriorating national well-being (Trindade et al., 2020).

3. Annotation regimes and formal target structures

Ground truth in this area is notably plural. AFSTRESS uses 12 binary labels—5 emotions and 7 stressors—with survey-validated, culturally grounded checklists rather than an exhaustive psychological taxonomy (Baktash et al., 28 Mar 2026). Its mean label cardinality is 5.54 and label density is 0.462, formalized as

$$\text{Label Cardinality} = \frac{1}{N}\sum_{i=1}^{N} |Y_i|$$

and

$$\text{Label Density} = \frac{1}{N L}\sum_{i=1}^{N} |Y_i|.$$

These statistics indicate that individual narratives are typically annotated with many simultaneously active labels rather than a single dominant state.

Crowdsourced and expert-guided text annotation follow different logics. Dreaddit annotators chose “Stress,” “Not Stress,” or “Can’t Tell” for five-sentence segments, with the crucial instruction to judge whether the author was expressing both stress and a negative attitude toward it; the final labeled set contained 3,553 instances, 52.3% of which were stressful, and inter-annotator agreement was Fleiss’s $\kappa = 0.47$ (Turcan et al., 2019). Holistix frames the task as six-class post classification with span-level evidence, developed under guidance from a social NLP researcher and a senior clinical psychologist; a second annotator reviewed 20% of entries, and Fleiss’ Kappa was reported as $\kappa = 75.92\%$ (Shakeel et al., 13 Jul 2025).

Protocol-driven multimodal datasets often combine task labels with questionnaires and continuous annotation. ForDigitStress uses frame-by-frame annotation in NOVA by two experienced psychologists, integrates participant self-reports, and reports Cohen’s $\kappa > 0.7$ for all labels (Heimerl et al., 2023). VitaStress treats the induced condition as the primary label and collects Self-Assessment Manikin ratings of valence, arousal, and dominance after the neutral baseline and each stress stimulus (Schreiber et al., 14 Aug 2025). MultiPhysio-HRC uses repeated STAI-Y1, NASA-TLX, SAM, and baseline NARS questionnaires; for classification, continuous STAI-Y1 and NASA-TLX scores are discretized per subject into low, medium, and high classes using thresholds based on subject mean $\mu$ and standard deviation $\delta$ (Bussolan et al., 1 Oct 2025).

In ambulatory settings, labels are often sparse relative to the sensor stream. The SMILE-based ambulatory ECG/GSR study relies on one self-reported stress label per hour on a 0-to-6 scale, while features are extracted from 5-minute sliding windows with 4-minute overlap, creating an explicit label-resolution mismatch for minute-level prediction (Dair et al., 2022). The IoT-enabled bio-signal study likewise emphasizes selective EMA prompting because constant querying would burden participants and reduce response rate (Tazarv et al., 2021).

A common misconception is that stress labels are interchangeable across datasets. The literature does not support that view: labels may denote a protocol phase, an hourly subjective score, a post-hoc wellness dimension, a multi-label stressor profile, or a synthetic state variable. This suggests that cross-dataset transfer requires explicit harmonization rather than nominal task matching.

4. Representative datasets and empirical findings

AFSTRESS is notable for combining low-resource NLP with substantive crisis analysis. It is the first multi-label corpus of self-reported stress narratives in Dari, with 737 responses and 12 binary labels. Structural stressors dominate: uncertain future appears in 62.6% of responses and education closure in 60.0%, exceeding emotional states. The strongest emotion–stressor co-occurrence is hopelessness with uncertain future, with Jaccard similarity $J = 0.388$ and 329 joint instances. The paper interprets this as evidence consistent with learned helplessness and hopelessness theory, and it identifies an emotion sequence $$\text{Stress} \rightarrow \text{Anxiety} \rightarrow \text{Sadness} \rightarrow \text{Burnout} \rightarrow \text{Hopelessness}$$ aligned with emotional cascade theory (Baktash et al., 28 Mar 2026).

CAN-STRESS emphasizes ecological validity. It includes 82 participants, split into 39 cannabis users and 43 non-users, each wearing an Empatica E4 for a continuous 24-hour period. Participants logged sleep, exercise, cannabis use, and perceived stress on a 1-to-10 scale. The paper reports higher mean EDA, higher EDA peaks per minute, and higher mean heart rate for users than for non-users, and it extracts 60-second EDA and HR segments around stress-labeled events to show increasing physiological arousal across no-stress, low-stress, medium-stress, and high-stress categories (Azghan et al., 24 Mar 2025).

Dreaddit established a benchmark for lengthy, domain-diverse social media stress classification. The full corpus contains 187,444 Reddit posts averaging 420 tokens, while the labeled subset contains 3,553 segments from 2,929 posts. Its design choice—segmenting posts into contiguous five-sentence chunks—preserves context while making supervision feasible, and it also supports future work on stress localization (Turcan et al., 2019). Holistix shifts the target from stress presence to holistic wellness dimensions, with 1,420 Australian mental-health forum posts labeled across physical, emotional, social, intellectual, spiritual, and vocational aspects, plus span-level evidence for the assigned label (Shakeel et al., 13 Jul 2025).

Workplace and educational settings motivate several domain-specific resources. The TILES-2018 Sleep Benchmark releases over 6,000 unique sleep recordings from 139 hospital employees over 10 weeks, plus baseline PSQI survey data and participant demographics, making it valuable for sleep-related pathways of occupational well-being (Feng et al., 4 Jul 2025). EmpathicSchool records 20 participants across nine classroom-like tasks, including presentation preparation, presentation delivery, an IQ test/Stroop task, calm music, amusement, and breathing exercise, while combining facial video, landmarks, HR, EDA, temperature, ACC, BVP, and IBI (Hosseini et al., 2022). Student survey datasets add another axis: one questionnaire-based study collected data from approximately 843 students aged 18 to 21 in Chhattisgarh, India using 28 questions across emotional well-being, physical health, academic performance, relationships, leisure, and stress/non-stress levels (Singh et al., 2024).

Synthetic and aggregate resources pursue reproducibility rather than direct observation. FLOW provides users.csv, daily_logs.csv, weekly_summaries.csv, interventions.csv, and daily_all.csv, with rule-based feedback loops linking workload, stress, sleep, mood, lifestyle behaviors, and body weight (Husseini, 28 Dec 2025). SWBI compresses thirteen annual socioeconomic factors into an annual return series, then models its tail risk with ARMA-GARCH dynamics and generalized hyperbolic innovations (Trindade et al., 2020).

5. Baseline models, benchmark performance, and generalization behavior

Benchmark results vary sharply by modality, dataset size, and label regime. In AFSTRESS, character TF-IDF with Linear SVM achieved the best reported performance, with Micro-F1 = 0.663 and Macro-F1 = 0.651, outperforming ParsBERT and XLM-RoBERTa. The same paper reports that per-label threshold tuning improved Micro-F1 by 10.3 percentage points over the default $t = 0.5$ threshold for the logistic-regression baseline, from 0.554 to 0.657 (Baktash et al., 28 Mar 2026). In Dreaddit, the strongest non-neural model was a logistic regressor using domain-specific Word2Vec embeddings and selected high-correlation features, reaching F1 = 0.7980, while BERT-base reached 0.8065 F1; the authors emphasize the logistic model’s interpretability and lower training cost (Turcan et al., 2019). Holistix reports MentalBERT as the best six-way classifier, with Accuracy = 0.74 and stronger explanation alignment than logistic regression under LIME-based comparison metrics (Shakeel et al., 13 Jul 2025).

Multimodal laboratory datasets often report higher within-protocol performance. ForDigitStress formulates binary stress recognition and finds that a simple feed-forward neural network using all modalities with early PCA achieves Accuracy = 88.3% and F1-score = 88.1%, while HRV is the strongest single modality (Heimerl et al., 2023). VitaStress reports 89% accuracy and F1-score 0.82 for binary baseline-vs-stress classification with Random Forest under leave-one-subject-out cross-validation, and 79% accuracy with F1-score 0.74 for a ternary task using kNN (Schreiber et al., 14 Aug 2025). EmpathicSchool finds that multimodal top-10 features perform best for stress recognition, with Random Forest reaching Recall $$\text{Label Density} = \frac{1}{N L}\sum_{i=1}^{N} |Y_i|.$$0, Precision $$\text{Label Density} = \frac{1}{N L}\sum_{i=1}^{N} |Y_i|.$$1, and F1 $$\text{Label Density} = \frac{1}{N L}\sum_{i=1}^{N} |Y_i|.$$2 (Hosseini et al., 2022).

Survey-tabular datasets can yield very high reported accuracies. A context-aware student framework evaluated on an 843-sample, 26-feature “Stress and Well-being” dataset reports 99.530% accuracy with a stacking classifier, while SVM with PCA reached 99.052% as the strongest individual model (Ovi et al., 1 Aug 2025). A separate workshop-derived college questionnaire dataset reports that SVM achieved 95% accuracy, 93% precision, 97% recall, and 94% F1-score on stress vs non-stress classification (Singh et al., 2024).

External robustness is much weaker than many in-domain scores imply. The ambulatory SMILE study reports 90.77% internal accuracy and 91.24 F1-score for an ExtraTrees classifier using ECG and GSR features with imputation, yet challenge performance drops to 59.23% accuracy on hidden data; the paper attributes the gap to covariate shift, missing data, and coarse hourly labels (Dair et al., 2022). A separate synthesis study reaches about 85% accuracy on unseen WESAD only after harmonizing multiple public datasets, generating protocol-aware synthetic subjects, and combining XGBoost with an ANN (Vos et al., 2022). This directly contradicts the notion that strong single-dataset validation is sufficient evidence of deployment readiness.

6. Methodological tensions, caveats, and research significance

The field is organized around several persistent tensions. The first is ecological validity versus experimental control. CAN-STRESS, Stress Bytes, the IoT-enabled bio-signal study, and daily mobile-phone stress recognition all prioritize ordinary-life behavior, but they inherit sparse labels, motion artifacts, attrition, and uneven compliance (Azghan et al., 24 Mar 2025, Belal et al., 21 May 2025, Tazarv et al., 2021, Bogomolov et al., 2014). VitaStress, ForDigitStress, EmpathicSchool, and MultiPhysio-HRC provide clearer task boundaries and richer synchronization, but they remain laboratory or semi-controlled protocols with limited sample sizes and stressors defined partly by experiment design (Schreiber et al., 14 Aug 2025, Heimerl et al., 2023, Hosseini et al., 2022, Bussolan et al., 1 Oct 2025).

The second tension is between construct breadth and benchmark tractability. AFSTRESS explicitly models structural stressors, emotional states, chronic stress, learned helplessness, and emotional cascades in a single multi-label framework (Baktash et al., 28 Mar 2026). Holistix broadens the target further to six wellness dimensions, while FLOW couples stress to sleep, mood, physical activity, diet, and body weight over two years (Shakeel et al., 13 Jul 2025, Husseini, 28 Dec 2025). By contrast, many wearable benchmarks reduce the task to binary stress classification. This simplification is operationally useful, but it omits the multi-causal and multi-dimensional structure that several datasets were designed to preserve.

The third tension concerns what a dataset is allowed to claim. FLOW is fully synthetic, not calibrated to a specific population, and should not be used for clinical, diagnostic, or policy conclusions (Husseini, 28 Dec 2025). SWBI is a single annual socioeconomic return series rather than an individual-level mental-health corpus (Trindade et al., 2020). Nora stores stress-related runtime logs but does not release a new benchmark dataset (Winata et al., 2021). TILES is best interpreted as a sleep-centered occupational well-being proxy rather than a rich standalone stress-and-anxiety release (Feng et al., 4 Jul 2025). These distinctions matter because methodological claims often travel farther than the data-generating process warrants.

The cumulative significance of this literature lies in the diversification of stress and well-being measurement. Low-resource language resources such as AFSTRESS extend computational mental-health research beyond English (Baktash et al., 28 Mar 2026). Region-specific annotation efforts such as Holistix add wellness-dimension granularity and explanatory evidence (Shakeel et al., 13 Jul 2025). Real-world passive sensing studies such as Stress Bytes and CAN-STRESS connect subjective stress to behavioral and physiological context at scale (Belal et al., 21 May 2025, Azghan et al., 24 Mar 2025). Workplace datasets such as MultiPhysio-HRC and TILES connect well-being to human-robot collaboration and shift work (Bussolan et al., 1 Oct 2025, Feng et al., 4 Jul 2025). A plausible implication is that future progress will depend less on a single universal benchmark than on principled cross-dataset reasoning: explicit construct alignment, calibrated transfer, interpretability, and careful separation of acute stress induction, chronic strain, and broader well-being.