Development and Validation Framework for an Explainable Multimodal Edge AI Wearable System for Academic Stress and Perceived Task Workload in University Students
Abstract
Background: Academic stress and perceived task workload vary during learning activities, but physiological arousal is not specific to either construct. Objective: To describe a prospective protocol for developing and evaluating a multimodal wearable-to-smartphone system using photoplethysmography, electrodermal activity, skin temperature, and accelerometry. Methods: A target cohort of 180 students aged 18–30 years will complete three visits with four counterbalanced academic task blocks per visit. Immediate post-block stress ratings and the unweighted Persian NASA Task Load Index will provide separate block-level reference outcomes. Baseline PSS-10 will support exploratory construct validation only. The primary models will use sensor features aggregated within four-minute blocks, with participant-disjoint nested cross-validation and participant-clustered uncertainty estimates. Post-task self-reports and task identity will be excluded from primary model inputs. Secondary analyses will examine unseen tasks, discordant stress–workload states, missing sensors, calibration, and explanation stability. Illustrative results: The worked example presents 7,776 eligible windows out of 8,640 and fused-model AUROCs of 0.892 for acute stress and 0.873 for high workload. These numerical values illustrate reporting only; they are not participant observations or measured predictive performance. Conclusions: The protocol separates block-level predictive validation from exploratory rolling-window inference. Prospective data collection and independent validation will be needed to determine performance and applicability.
Seyed Mohammad Rahimi
Faculty of Electrical and Computer Engineering, Babol Noshirvani University of Technology (N.I.T.), Babol, Iran
https://orcid.org/0009-0000-1350-9971
Abstract
Background: Academic stress and perceived task workload vary during learning activities, but physiological arousal is not specific to either construct. Objective: To describe a prospective protocol for developing and evaluating a multimodal wearable-to-smartphone system using photoplethysmography, electrodermal activity, skin temperature, and accelerometry. Methods: A target cohort of 180 students aged 18–30 years will complete three visits with four counterbalanced academic task blocks per visit. Immediate post-block stress ratings and the unweighted Persian NASA Task Load Index will provide separate block-level reference outcomes. Baseline PSS-10 will support exploratory construct validation only. The primary models will use sensor features aggregated within four-minute blocks, with participant-disjoint nested cross-validation and participant-clustered uncertainty estimates. Post-task self-reports and task identity will be excluded from primary model inputs. Secondary analyses will examine unseen tasks, discordant stress–workload states, missing sensors, calibration, and explanation stability. Illustrative results: The worked example presents 7,776 eligible windows out of 8,640 and fused-model AUROCs of 0.892 for acute stress and 0.873 for high workload. These numerical values illustrate reporting only; they are not participant observations or measured predictive performance. Conclusions: The protocol separates block-level predictive validation from exploratory rolling-window inference. Prospective data collection and independent validation will be needed to determine performance and applicability.
Keywords: academic stress; perceived task workload; wearable computing; explainable artificial intelligence; edge computing; photoplethysmography; electrodermal activity; study protocol; university students.
1. Introduction
University students encounter changing cognitive, social, and evaluative demands during lectures, coursework, examinations, and independent study. Acute stress is an affective and physiological response to perceived challenge, whereas cognitive workload concerns the demands placed on working memory, attention, and task control. Although these constructs can co-occur, they are neither conceptually nor operationally interchangeable. An examination may be cognitively demanding without eliciting distress, and anticipation of evaluation may cause high stress during a relatively undemanding task. Accordingly, a monitoring system must retain separate outcomes instead of collapsing both into a single “mental strain” label.
Wearable sensing offers a complementary source of information to self-report. Photoplethysmography (PPG) measures pulsatile vascular changes from which pulse-derived heart rate and selected variability features may be derived; electrodermal activity (EDA) indexes sympathetic sudomotor activity; peripheral temperature provides contextual autonomic information; and triaxial accelerometry supports motion-artifact detection. Neither a biomarker nor an algorithm is a direct physiological meter of academic stress. The interpretive meaning of signals depends on activity, posture, sleep, caffeine, ambient conditions, individual differences, and timing of assessments (Pinge et al., 2024; Pataca et al., 2025).
A systematic review and meta-analysis focused on student populations has documented encouraging performance from wearable AI while emphasizing methodological heterogeneity and uncertain generalizability (Abd-alrazaq et al., 2024). Likewise, recent cognitive-load work has demonstrated physiological discrimination between resting and Stroop conditions but explicitly cautioned against treating task-condition classification as definitive task-independent detection of cognitive load (Yu & Wan, 2026). These findings motivate designs that use independently collected target ratings, active control tasks, participant-disjoint evaluation, and external validation.
Edge inference can reduce reliance on permanent cloud connectivity, potentially limiting raw physiological data transfer and shortening feedback loops. Explanations may further allow researchers to evaluate whether model decisions are supported by plausible features rather than motion, task identity, or demographic proxies. SHAP explanations describe model behavior; they do not establish causality, diagnostic validity, or physiological mechanisms (Lundberg & Lee, 2017).
The present manuscript describes a prospective study protocol for an explainable multimodal wearable-to-smartphone system. The proposed study site is the School of Electrical and Computer Engineering, College of Engineering, University of Tehran. The objectives are to define the sensing and labeling pipeline, evaluate participant-independent prediction of block-level acute stress and perceived task workload, examine robustness and construct separation, and assess local-device feasibility. Site authorization, ethics review, and technical verification will precede recruitment; no completed study or prototype validation is reported. The Results section provides a hypothetical reporting example; its numerical values do not arise from implementation of the prospective analysis plan.
2. Materials and Methods
2.1 Study design, setting, and participant framework
A prospective methodological study with repeated academic task blocks and a technology-feasibility component is proposed at the University of Tehran, Tehran, Iran. The target sample is 180 volunteer students aged 18–30 years, with recruitment stratified where feasible by sex and academic year. This is a pragmatic development target, not a completed enrollment count or a statistically established validation sample size. Before recruitment, a simulation-based precision assessment will examine plausible outcome prevalences of 20%, 40%, and 60%, within-participant correlations of 0.2, 0.5, and 0.8, and up to 20% loss of analyzable participants. The design criterion will be a 95% AUROC interval with total width no greater than 0.10 under the prespecified planning scenario. The final recruitment target and assumptions will be registered before outcomes are inspected; if the available sample does not meet this criterion, performance estimates will be explicitly treated as exploratory. Repeated windows will not be counted as independent participants.
Eligibility comprises enrollment at a university, willingness to complete three study visits, and ability to perform Persian-language cognitive tasks. Exclusions are sensor-contact intolerance, inability to understand consent, or conditions preventing safe participation in brief cognitive tasks. Potential confounders, including recent strenuous exercise, stimulants, medication affecting autonomic function, sleep duration, smoking, and recent illness, are recorded without excluding broad groups by default. Participants may stop any task or disable monitoring without disadvantage.
2.2 Session and reference-label protocol
Each participant will attend three laboratory visits on nonconsecutive days. Each visit will contain four 4-min blocks: quiet academic reading, a two-back working-memory task, timed mental arithmetic, and preparation of an oral summary under neutral observation. Task sequences will be balanced across participants and randomized across visits. Task ratings will be completed immediately after each block, followed by a separate 2-min rest interval; questionnaire completion time will not be counted as rest. Each block will yield four nonoverlapping 60-s sensor epochs, giving a planned maximum of 8,640 epochs nested within 2,160 blocks and 180 participants. These are design totals, not observed data. Task accuracy, response latency, perceived difficulty, motion, task order, and rating timestamps will be recorded for secondary analyses.
After each block, participants will rate stress on a 0–100 visual analogue scale referring explicitly to the completed four-minute task and complete the six NASA-TLX dimensions. Elevated acute stress and high perceived task workload will be defined separately using analytical thresholds of 60/100; these are not clinical cutoffs. The unweighted NASA-TLX composite will be the mean of six completed dimension scores, with the performance dimension oriented so that higher values indicate greater workload. No reference outcome will be imputed when a required rating is missing. Continuous-outcome analyses and alternate thresholds of 50 and 70 will assess threshold dependence. The primary unit of prediction and evaluation will be the four-minute block. Within-block 60-s epochs will provide features, not independent labeled outcomes; no minute-level accuracy claim will be based on copying one block rating to every epoch.
2.3 Wearable hardware and local computing
The proposed wrist-worn sensing unit includes reflective PPG sampled at 64 Hz, EDA at 4 Hz, temperature at 1 Hz, and triaxial accelerometry at 32 Hz. A low-power microcontroller time-stamps samples and transmits encrypted packets over Bluetooth Low Energy to a participant-controlled or study-managed Android device. All preprocessing, feature generation, model inference, thresholding, and visualization occur on this nearby device without cloud dependence. Sensors, exact chipsets, battery, firmware version, and hardware revision are to be reported after procurement and physical verification. Therefore, the stated sampling rates and software architecture are engineering specifications, not measurements of a built prototype.
A 60-s rolling window with a 15-s update stride is proposed for exploratory local inference. Because the reference outcomes describe four-minute blocks, these rolling outputs will be evaluated for technical feasibility only and will not be presented as validated momentary stress or workload estimates. The primary analysis will aggregate nonoverlapping sensor epochs into block-level feature vectors. Signal-quality checks will detect missing packets, sensor saturation, implausible pulses, and motion contamination. The system will abstain when required quality criteria fail, instead of assigning low stress. During validation, inferred states will not be shown to participants, thereby avoiding feedback-induced changes in the reference ratings.
Modality | Sampling | Derived features | Primary quality checks |
|---|---|---|---|
PPG | 64 Hz | Heart rate; pulse-derived RMSSD; SDNN; pulse amplitude | Peak plausibility; accelerometer-linked artifacts |
EDA | 4 Hz | Tonic conductance; phasic SCR count; SCR amplitude | Contact/saturation; baseline drift |
Temperature | 1 Hz | Mean peripheral temperature; short-term slope | Outliers; ambient-context sensitivity |
Accelerometry | 32 Hz | Vector magnitude; activity level; motion bursts | Signal dropouts; device displacement |
Task context | Event logs | Block type; performance; subjective task difficulty | Clock alignment; leakage-safe encoding |
The use of pulse-derived RMSSD requires validated beat detection and sufficient clean segments; it is not interchangeable with ECG-derived HRV under all conditions. Window and sensor settings are protocol specifications.
2.4 Measurement instruments: original and Iranian validity
Perceived Stress Scale. Cohen, Kamarck, and Mermelstein (1983) introduced the original 14-item PSS. The 10-item version used here measures perceived stress over the preceding month, with reverse scoring of positively phrased items and a total range of 0–40. Roberti et al. (2006) reported psychometric support in 281 undergraduates, including perceived helplessness and self-efficacy factors. Persian-language studies examined 100 chronic-headache patients (Khalili et al., 2017) and 400 postoperative patients (Sharif-Nia et al., 2024). Khalili et al. reported internal consistency of 0.72 and repeated-measure reliability of 0.93. These studies support language feasibility but do not establish measurement invariance in the proposed student sample. Baseline PSS-10 will be used only for exploratory associations with participant-level aggregated predictions, not for momentary labels or model inputs.
NASA Task Load Index. Hart and Staveland (1988) introduced six subjective workload dimensions: mental demand, physical demand, temporal demand, performance, effort, and frustration. The present protocol will use the unweighted mean of six scores, each oriented from 0 to 100 with higher values indicating greater workload. Mokhtarinia et al. (2026) evaluated a Persian adaptation in 605 occupational respondents and reported S-CVI/Ave = 0.92 and internal consistency α = 0.922. Their test–retest ICCs were 0.841 for single measures and 0.913 for average measures, which are distinct estimands rather than an interchangeable reliability range. These occupational findings do not establish validity in university students; item comprehension and dimensional behavior will be examined before interpretation of student results. The ICC range displayed in Table 2 represents the single-measure and average-measure estimands, respectively.
Instrument / original source | Persian validation sample | Published validity evidence | Published reliability |
|---|---|---|---|
Original PSS-14: Cohen et al. (1983) | Three original study samples | Correlated with life-events and symptoms | α = 0.84–0.86; 2-day test–retest r = 0.85 |
PSS-10: Roberti et al. (2006) | Khalili et al. (2017); 100 headache patients | Content/construct validity; coping and distress factors | α = 0.72; test–retest = 0.93 |
PSS-10 (original construct based on Cohen et al., 1983) | Sharif-Nia et al. (2024); 400 postoperative patients | Two-factor EFA/CFA support; construct indices > 0.70 | Reported α / CR / MaxR > 0.70 |
NASA-TLX: Hart & Staveland (1988) | Mokhtarinia et al. (2026); 605 occupational respondents | S-CVI/Ave = 0.92; 73.1% factor variance | α = 0.922; ICC = 0.841–0.913 |
2.5 Signal processing, missingness, and feature engineering
Signal processing will use only samples available up to the prediction timestamp; zero-phase filtering that uses future samples will not be used to support online performance claims. The proposed PPG filter is a fourth-order causal Butterworth band-pass filter at 0.5–4 Hz, and the proposed EDA filter is a second-order causal low-pass filter at 1 Hz. Heart rate and pulse-rate variability will be derived from detected pulse intervals, with pulse intervals outside 300–2,000 ms flagged for review. Temperature and acceleration summaries will be aligned to the same 60-s epochs. PPG-derived RMSSD and SDNN will be described as pulse-rate variability measures, not interchangeable substitutes for ECG-derived HRV. Pulse-detector settings, motion thresholds, EDA decomposition parameters, and sensor-specific saturation limits will be established in separate technical calibration recordings and locked in a versioned processing specification before the main cohort is analyzed. Calibration recordings will not contribute to performance evaluation.
The proposed minimum completeness criterion is 95% received samples per required channel in a 60-s epoch. Pulse-variability features will additionally require at least 40 valid pulse intervals and no more than 5% rejected candidate intervals; these are proposed operating rules requiring device-specific verification. The primary complete-channel analysis will require at least three usable epochs per block and will aggregate each epoch feature by its median and interquartile range. It will not impute absent sensor channels. The number and characteristics of abstained blocks and excluded participants will be reported alongside predictive performance. Sensitivity analyses will omit the first minute of each block, vary completeness and pulse-quality thresholds, and evaluate prespecified reduced-sensor models. Scaling, optional feature selection, and any secondary imputation will be fitted within training folds only. An independent quiet-reading calibration period before the randomized tasks will support any individual baseline normalization; no withheld target block will provide its own normalization reference.
2.6 Prediction models, edge packaging, and explainability
Five candidate approaches are proposed: L2-regularized logistic regression, random forest, gradient-boosted trees, a compact temporal convolutional network, and late fusion of sensor-specific logistic-regression probabilities. The primary input set will comprise physiological and accelerometry summaries only; task identity, subjective difficulty, NASA-TLX components, stress ratings, and post-block performance summaries will be excluded. The proposed tuning grids are C = 0.01, 0.1, 1, and 10 for logistic regression; 300 trees, maximum depth 3, 6, or unrestricted, and minimum leaf size 5 or 10 for random forest; and 100 or 300 boosting iterations, learning rate 0.03 or 0.1, and 7 or 15 leaves for boosted trees. Late fusion will use a logistic meta-model trained solely on participant-disjoint out-of-fold branch predictions. Two separate outcome models will be fitted. Sigmoid calibration and operating-threshold selection will use training-only out-of-fold predictions; the threshold will maximize specificity subject to sensitivity of at least 0.80 in those predictions. Held-out performance will be reported even when this sensitivity target is not attained. All proposed settings will be frozen in the prospective analysis plan before evaluation. For the temporal convolutional candidate, architecture, input sequence length, regularization, optimizer, and stopping rules will be specified during separate technical development and frozen before cohort evaluation. Its block-level output will use the same participant splits and reference outcomes as the other models. The numerical model comparisons in the worked example are illustrative and do not document training of these candidates.
Explainability will summarize global and local SHAP contributions using a background sample drawn only from training participants (Lundberg & Lee, 2017). Feature rankings and signed contributions will be evaluated for stability under sensor loss and motion perturbations. If a linear surrogate is used, its target will be the main model probability, and it will be fitted on training-only out-of-fold predictions. Surrogate fidelity will be assessed on held-out predictions using R² and mean absolute error. Standardized surrogate coefficients will describe approximation of model behavior, not physiological effects, causal pathways, or independent associations with the reference outcome. The illustrative coefficients and path diagram in the worked example are not fitted estimates and do not establish surrogate fidelity. Explanation latency will be measured separately from inference latency.
2.7 Validation, statistics, and evaluation endpoints
The primary endpoint will be block-level AUROC for each reference outcome under participant-disjoint nested cross-validation. A five-fold outer GroupKFold will keep all visits and blocks of each participant together; four-fold grouped inner validation will tune models without accessing outer-fold outcomes. Each block will receive one held-out probability per model. Cluster bootstrap intervals will use 2,000 participant-level resamples of the held-out prediction set, preserving every selected participant’s blocks; these intervals will quantify uncertainty conditional on the fitted cross-validation procedure, rather than all sources of model-development variability. Paired model differences will use the same cluster resamples. Secondary metrics will include AUPRC, macro-F1, sensitivity, specificity, precision, Brier score, calibration slope and intercept, and ECE with ten equal-width probability bins. Performance will be reported with abstention coverage and outcome prevalence. Model comparisons will be exploratory, and the best observed outer-fold score will not be described as independent validation of a selected model. A separately recruited university cohort will be required to assess the final locked model externally.
Sensor ablations and physiology-only, context-only, and sensor-plus-context analyses will use identical participant splits and tuning budgets. Context models will be secondary retrospective analyses and will not establish feasibility of sensor-only online inference. A combined participant-held-out and task-held-out analysis will exclude the evaluated task from model training and tuning. Stress and workload overlap will be described by continuous-score correlations and a four-cell cross-tabulation of low/low, high/low, low/high, and high/high states. Discrimination will also be examined within strata of the other outcome, where both target classes are present, and with a workload sensitivity analysis excluding frustration. Sparse discordant states will limit any claim of construct separation. Subgroup estimates will remain descriptive and include sample sizes and uncertainty. Reporting and risk-of-bias assessment will draw on TRIPOD+AI and PROBAST+AI (Collins et al., 2024; Moons et al., 2025).
2.8 Ethics, privacy, and reproducibility
The proposed study will require approval from the competent institutional ethics committee and written informed consent before any recruitment or recording. This protocol does not report an approval identifier, completed consent procedures, or participant observations. The eventual empirical report will document the approving committee, approval date and identifier, and recruitment dates. Participation will be voluntary and unrelated to grades; students will be able to pause monitoring or withdraw without disadvantage. Physiological recordings will be processed locally, encrypted at rest, and pseudonymized for research export. The approved data-management plan will specify retention, access control, and the point beyond which withdrawal from an irreversibly anonymized dataset is no longer possible. Individual inferred states will not be accessible to teaching staff or grading systems. The system will remain a research instrument, not a diagnostic device. The protocol, analysis settings, software versions, and subsequent amendments will be registered prospectively; deidentified derived data and code will be shared subject to consent and institutional authorization.
3. Results
3.1 Sample profile, reference outcomes, and data quality
The target sample were 180 participants, mean age 22.4 years (SD 2.8), including 96 women (53.3%) and 84 men (46.7%). Across 540 visits and 2,160 task blocks, a maximum of 8,640 sixty-second windows is available. An illustrative 864 windows (10.0%) fails signal-quality criteria, leaving 7,776 eligible windows. Block-level elevated stress prevalence is 41.2% and high workload prevalence 46.5%; prevalence should always be reported by participant and block, rather than treating repeated windows as 7,776 separate respondents. The illustrative baseline PSS-10 score averages 18.2 (SD 6.0), and the block NASA-TLX averages 52.6 (SD 17.3).
Characteristic | value | Characteristic | value |
|---|---|---|---|
Students enrolled | 180 | Completed visits | 540 |
Women / men | 96 / 84 | Task blocks | 2,160 |
Age, M (SD) | 22.4 (2.8) y | Initial 60-s windows | 8,640 |
PSS-10, M (SD) | 18.2 (6.0) | SQI exclusions | 864 (10.0%) |
NASA-TLX, M (SD) | 52.6 (17.3) | Usable windows | 7,776 |
High acute stress blocks | 41.2% | High workload blocks | 46.5% |
3.2 Subject-independent predictive discrimination
The hypothetical grouped evaluation suggests that multimodal late fusion exceeds the prespecified single-model baselines for both outcomes. For acute stress, illustrative AUROC rises from 0.779 with logistic regression to 0.892 with fusion; for high workload, it rises from 0.754 to 0.873. Importantly, comparisons must be made on the same held-out individuals and accompanied by clustered uncertainty and calibrated thresholds. These results do not establish superior clinical or educational effectiveness without independent testing.
Candidate model | Stress AUROC | Load AUROC | Stress macro-F1 | Load macro-F1 |
|---|---|---|---|---|
Regularized logistic regression | 0.779 | 0.754 | 0.718 | 0.702 |
Random forest | 0.826 | 0.806 | 0.767 | 0.748 |
LightGBM | 0.869 | 0.839 | 0.806 | 0.780 |
Compact temporal CNN | 0.876 | 0.854 | 0.817 | 0.793 |
Multimodal late fusion | 0.892 | 0.873 | 0.834 | 0.812 |
Metric | Acute stress | High cognitive workload |
|---|---|---|
AUROC (95% participant-cluster CI) | 0.892 (0.866–0.915) | 0.873 (0.847–0.897) |
AUPRC | 0.862 | 0.848 |
Macro-F1 | 0.834 | 0.812 |
Sensitivity / specificity | 0.848 / 0.835 | 0.821 / 0.807 |
Precision | 0.816 | 0.803 |
Brier score | 0.121 | 0.142 |
Expected calibration error | 0.036 | 0.044 |
3.3 Modality ablation and task/context leakage
An ablation analysis emphasizes that the most favorable results are achieved with complementary physiological inputs rather than with task labels alone. Removal of EDA yields the largest loss in stress discrimination; removal of the PPG-derived variability branch produces a smaller but still consequential decrease. For high workload, task-difficulty context contributes substantially, requiring explicit no-context evaluation and independent-task tests. The task-only benchmark in Table 6 is intentionally reported to expose potential label leakage via task design rather than to endorse it as a physiological classifier.
Available channels | Acute stress | High workload | Interpretive purpose |
|---|---|---|---|
PPG only | 0.806 | 0.781 | Autonomic-pulse branch |
EDA only | 0.823 | 0.749 | Sympathetic sweat branch |
PPG + EDA | 0.862 | 0.814 | Physiological fusion |
All physiology, no task context | 0.879 | 0.837 | Conservative transportability test |
Task context only | 0.692 | 0.803 | Checks shortcut learning |
All channels (reference) | 0.892 | 0.873 | Full illustrative model |
3.4 Explainability and standardized β path diagram
The worked explanatory analysis ranks EDA phasic activity and pulse-derived RMSSD highly for acute stress, while task difficulty and phasic EDA are relatively influential for cognitive workload. A separate multivariable standardized linear surrogate yields the coefficients shown in Table 7 and Figure 3.
Input feature | Stress β [95% CI] | Workload β [95% CI] |
|---|---|---|
EDA phasic activity | +0.38 [0.30, 0.46] | +0.17 [0.10, 0.24] |
RMSSD / HRV | −0.31 [−0.39, −0.23] | −0.15 [−0.22, −0.08] |
Task difficulty | +0.23 [0.16, 0.30] | +0.45 [0.37, 0.53] |
Peripheral temperature | −0.11 [−0.18, −0.04] | −0.09 [−0.16, −0.02] |
Motion index | +0.07 [0.01, 0.13] | +0.06 [0.00, 0.12] |
Explained variance (R²) | 0.59 | 0.55 |
3.5 Engineering feasibility and robustness checks
An implementation report should additionally include 95th-percentile local inference latency, peak resident memory, duty-cycle power draw, operating temperature, BLE packet loss, annotation delay, and explanation fidelity. To avoid inventing device measurements, this manuscript reports acceptance targets in Table 8 rather than fictitious achieved values. The sensitivity analyses would compare first-minute exclusion, alternate stress thresholds, missing-EDA conditions, low-SQI strata, unseen task orders, and participants with poor sleep, with all bootstrap resampling clustered by participant. External cohort evaluation remains essential.
Endpoint | Acceptance target | Measurement approach |
|---|---|---|
On-device / near-device latency | 95th percentile ≤ 200 ms | ≥10,000 timed calls on named phone model |
BLE completeness | ≥95% valid packets | Continuous logged 8-hour bench and field sessions |
Battery operation | ≥8 hours | Continuous sensors + local transmission |
Output abstention | No score when SQI fails | Prospective missing-sensor and movement challenge |
Privacy assurance | No raw-signal cloud transfer | Network capture and technical security audit |
Explanation stability | Report rank correlation and fidelity | Repeated held-out perturbations |
4. Discussion
4.1 Interpretation of the integrated approach
This protocol integrates multimodal wearable sensing with participant-disjoint evaluation and locally executed inference. Its intended contribution is a reproducible design for assessing block-level acute stress and perceived task workload while distinguishing technical feasibility from predictive validity. Multiple sensors may help account for movement and individual baseline differences, but their incremental value must be measured rather than presumed. The accompanying numerical tables and figures demonstrate a reporting format and provide no empirical comparison of model performance.
Reviews of wearables for stress monitoring repeatedly identify heterogeneity in labeling, modalities, and evaluation protocols, while calling for stronger ecological validation and reproducible reporting (Abd-alrazaq et al., 2024; Pinge et al., 2024; Pataca et al., 2025). Public resources such as WESAD have accelerated benchmarking of stress-related physiological signals (Schmidt et al., 2018), but models developed under tightly controlled laboratory challenges can behave differently during lectures, commuting, sleep deprivation, physical exertion, and informal study. Likewise, Yu and Wan (2026) show why distinguishing a Stroop challenge from rest cannot alone establish construct-valid cognitive-load measurement. A suitable student system must triangulate concurrent subjective workload, task performance, and physiological patterns rather than relabeling task identity as ground truth.
4.2 Measurement validity in the Iranian context
The original PSS literature and Persian validation studies support selection of a Persian-language baseline measure but do not justify equating perceived stress over the preceding month with within-task distress (Cohen et al., 1983; Khalili et al., 2017; Sharif-Nia et al., 2024). Baseline PSS-10 will be related to each participant’s average held-out block-level stress probability as an exploratory construct-validity analysis. A correlation would not establish temporal accuracy, concurrent criterion validity, or clinical utility.
NASA-TLX will operationalize perceived task workload immediately after each block. Its Persian adaptation provides relevant occupational evidence but does not establish validity in university students (Mokhtarinia et al., 2026). Component-level analyses will therefore be retained, including careful orientation of the performance item. Because frustration overlaps conceptually with stress, the total NASA-TLX score will not be interpreted as an objective or neurologically verified cognitive-load state. A sensitivity analysis without frustration and evaluation of discordant stress–workload states will assess the extent of construct overlap.
4.3 Explainability and interpretation
SHAP values will explain model predictions under a specified reference distribution; they will not identify biological mechanisms (Lundberg & Lee, 2017). Any linear surrogate will approximate the fitted model and will require held-out fidelity assessment before interpretation. Neither a plausible sign for EDA or pulse-rate variability nor a visually clear explanation establishes construct validity. Future user-facing explanations will require usability assessment and should communicate uncertainty without psychiatric or diagnostic labels.
4.4 Edge deployment, privacy, and fairness
Near-device processing is attractive because data may remain under local institutional control and the system can continue to function without cloud connectivity. Yet “edge AI” does not automatically guarantee privacy or security. Secure BLE pairing, key management, encrypted storage, data minimization, network audit, participant-accessible deletion, and separation from academic grading are mandatory implementation concerns. Subgroup evaluation must avoid overstated fairness claims from small cells: observed AUROC differences may reflect sampling noise, task distribution, sensor fit, skin pigmentation effects on optical measurement, or unmeasured contextual factors. Monitoring should be designed as optional self-awareness support rather than surveillance of students.
4.5 Limitations and scope
The principal limitation is temporal: post-block ratings support block-level validation, not independently verified minute-by-minute detection. Other limitations include pragmatic sample planning, single-site volunteer recruitment, imperfect wrist-sensor contact, motion artifacts, possible stress–workload overlap, and limited numbers of discordant states. Selective signal rejection may bias the evaluated sample, so coverage and exclusion patterns will accompany all performance estimates. Hardware-specific thresholds remain subject to a separate technical calibration stage and prospective specification lock. Laboratory evaluation cannot establish performance during everyday academic activity or effectiveness of interventions triggered by the system.
5. Conclusions
This protocol specifies a prospective approach to developing and evaluating an explainable multimodal wearable-to-smartphone system for acute academic stress and perceived task workload. Primary validation will use block-level reference outcomes, participant-disjoint model evaluation, prospective input availability, and explicit reporting of abstention. Rolling-window outputs will remain exploratory until matched temporal reference measurements are available. The article provides a study design, an engineering evaluation plan, and a hypothetical numerical reporting example; it does not report a completed device, empirical performance, or evidence of clinical or educational effectiveness.
References
Abd-alrazaq, A., Alajlani, M., Ahmad, R., AlSaad, R., Aziz, S., Ahmed, A., Alsahli, M., Damseh, R., & Sheikh, J. (2024). The performance of wearable AI in detecting stress among students: Systematic review and meta-analysis. Journal of Medical Internet Research, 26, e52622. https://doi.org/10.2196/52622
Cohen, S., Kamarck, T., & Mermelstein, R. (1983). A global measure of perceived stress. Journal of Health and Social Behavior, 24(4), 385–396. https://doi.org/10.2307/2136404
Collins, G. S., Moons, K. G. M., Dhiman, P., Riley, R. D., Beam, A. L., Van Calster, B., et al. (2024). TRIPOD+AI statement: Updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ, 385, e078378. https://doi.org/10.1136/bmj-2023-078378
Hart, S. G., & Staveland, L. E. (1988). Development of NASA-TLX (Task Load Index): Results of empirical and theoretical research. In P. A. Hancock & N. Meshkati (Eds.), Human mental workload (pp. 139–183). North-Holland. https://doi.org/10.1016/S0166-4115(08)62386-9
Khalili, R., Sirati Nir, M., Ebadi, A., Tavallai, A., & Habibi, M. (2017). Validity and reliability of the Cohen 10-item Perceived Stress Scale in patients with chronic headache: Persian version. Asian Journal of Psychiatry, 26, 136–140. https://doi.org/10.1016/j.ajp.2017.01.010
Lundberg, S. M., & Lee, S.-I. (2017). A unified approach to interpreting model predictions. Advances in Neural Information Processing Systems, 30, 4765–4774. https://proceedings.neurips.cc/paper/2017/hash/8a20a8621978632d76c43dfd28b67767-Abstract.html
Mokhtarinia, H. R., Noshadha, M., Rafat, Z., & Melloh, M. (2026). Translation, cultural adaptation, and psychometric evaluation of the NASA-TLX questionnaire among various occupational groups of Persian speakers. Scientific Reports, 16, Article 32945. https://doi.org/10.1038/s41598-026-63307-3
Moons, K. G. M., Damen, J. A. A., Kaul, T., Hooft, L., Andaur Navarro, C., Dhiman, P., et al. (2025). PROBAST+AI: An updated quality, risk of bias, and applicability assessment tool for prediction models using regression or artificial intelligence methods. BMJ, 388, e082505. https://doi.org/10.1136/bmj-2024-082505
Pataca, A. O., Zdravevski, E., Coelho, P. J., Garcia, N. M., Deryuck, M., Albuquerque, C., & Pires, I. M. (2025). Use of machine learning for predicting stress episodes based on wearable sensor data: A systematic review. Computers in Biology and Medicine, 198(Part A), 111166. https://doi.org/10.1016/j.compbiomed.2025.111166
Pinge, A., Gad, V., Jaisighani, D., Ghosh, S., & Sen, S. (2024). Detection and monitoring of stress using wearables: A systematic review. Frontiers in Computer Science, 6, 1478851. https://doi.org/10.3389/fcomp.2024.1478851
Roberti, J. W., Harrington, L. N., & Storch, E. A. (2006). Further psychometric support for the 10-item version of the Perceived Stress Scale. Journal of College Counseling, 9(2), 135–147. https://doi.org/10.1002/j.2161-1882.2006.tb00100.x
Schmidt, P., Reiss, A., Duerichen, R., Marberger, C., & Van Laerhoven, K. (2018). Introducing WESAD, a multimodal dataset for wearable stress and affect detection. Proceedings of the 20th ACM International Conference on Multimodal Interaction (pp. 400–408). Association for Computing Machinery. https://doi.org/10.1145/3242969.3242985
Sharif-Nia, H., Froelicher, E. S., Hoseinzadeh, E., Kaveh, O., Fatehi, R., & Nowrozi, P. (2024). Assessing the validity and reliability of the 10-item Persian version of the perceived stress scale in post-surgery patients. Frontiers in Psychiatry, 15, 1402122. https://doi.org/10.3389/fpsyt.2024.1402122
Yu, J., & Wan, J. (2026). AI-supported cognitive load detection: Experimental insights using wearable workload and stress data for enhanced well-being interventions. Scientific Reports, 16, Article 36348. https://doi.org/10.1038/s41598-026-64103-9