Mental stress (MS), an ever-present aspect of life, occurs when external demands exceed an individual’s available physiological and psychological coping resources [1,2]. While short-term exposure to stress may enhance focus and performance [3], frequent or prolonged exposure can adversely affect health, contributing to psychiatric conditions [4,5] and cardiovascular disease [6]. Consequently, accurate and timely detection followed by effective stress management is essential to mitigate these adverse health outcomes.

Traditional stress assessment methods, such as cortisol analysis [7], are burdensome. In contrast, self-report questionnaires, such as the Perceived Stress Scale [8], administered retrospectively or through ecological momentary assessment, offer greater convenience but are limited in temporal granularity [9]. Moreover, they may be susceptible to recall and reporting bias and less reliable among individuals with alexithymia. Effective stress detection, by contrast, should be minimally invasive and support continuous monitoring to capture early warning signs of stress [10].

Given that stress modulates the activity of the autonomic nervous system (ANS) [11], physiological biomarkers such as the electrocardiogram (ECG) and electrodermal activity have been widely proposed for automated stress monitoring. Furthermore, these signals are increasingly accessible through modern wearable devices, making real-time stress tracking more feasible. However, physiological signals often exhibit subtle and complex patterns that can be difficult to analyze and interpret using traditional statistical methods [1].

Machine learning (ML) has emerged as a powerful tool for analyzing high-dimensional data, excelling at uncovering patterns and complex relationships [12]. Consequently, ML has increasingly been used in recent studies to detect stress responses from the biomarkers mentioned above [13-16]. However, most existing studies exhibit one of the following limitations: (1) small cohorts; (2) a narrow range of mental stressors investigated; and (3) restrict MS detection to seated baseline (BL) conditions. This restriction leaves it unclear whether ML models can distinguish stress-induced ECG signals from those triggered by everyday movements, which often produce similar physiological responses. Moreover, because individuals encounter a wide range of novel stressors in daily life, models must generalize to stimuli absent from their training data. This aspect has been largely overlooked in prior work. The few studies that investigate generalization [17-19] typically train their ML models on one dataset and test them on another. Although informative, this approach conflates differences in participant demographics, sensor hardware, and stressor types, making it hard to isolate which factor drives any performance drop.

In this study, we address these gaps by developing and rigorously evaluating two ML models, namely logistic regression (LR) and extreme gradient boosting (XGBoost), on a large ECG dataset sampled at 1000 Hz. The dataset, collected from a controlled laboratory experiment at the Vrije Universiteit Amsterdam and published by van der Mee et al [20], included 127 participants who performed both diverse mental-stress tasks and everyday activities (eg, dishwashing and walking). In this study, we define “mental stress” as the physiological response elicited by cognitively or emotionally challenging tasks. Subsequently, we formulated a binary MS versus no-stress classification task, where the no-stress condition includes seated BLs, recovery periods, and low-to-moderate-intensity physical activities. This design not only improves ecological validity but also tests each model’s ability to distinguish stress‐induced ECG changes from those driven by physical exertion, without relying on additional sensors (eg, accelerometers). To assess generalization, we evaluate our models on unseen participants and on “novel” stressors, using a leave-one-stressor-out protocol, where we withhold each mental stressor during training and assess model performance on this specific stressor. We further investigate the robustness of our models to lower sampling frequency and the reduction of the feature set. Together, these experiments provide key insights for transitioning ECG-based MS detection from laboratory settings to real-world contexts.

The structure of the following section is in line with the TRIPOD+AI (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis–Artificial Intelligence) statement [21,22]. Sections of the guideline not applicable to this research were omitted for brevity.

This study is a retrospective analysis of pseudonymized data collected from a controlled laboratory experiment conducted at the Vrije Universiteit Amsterdam and published by van der Mee et al [20]. The experimental conditions are shown in Table 1. In addition to ECG signals, subjective affect was assessed after each experimental condition using self-reported measures of positive and negative affect, as measured by the Maastricht Questionnaire [23]. We refer to the original study by van der Mee et al [20] for detailed explanations of the experimental procedure.

As outlined in the study by van der Mee et al [20], participants were required to be Dutch-speaking, employed or enrolled in school, and aged 18-48 years. Exclusion criteria included a BMI above 30, high cholesterol, diabetes, liver or thyroid disease, and use of medications that affect the ANS, such as antidepressants or anticholinergics. The final sample included 127 participants (71 female and 56 male) with a mean age of 23.3 (SD 5.6) years. We refer the reader to the original study for additional details on recruitment and population. The data used in this analysis were pseudonymized.

For this research, the different experimental conditions, as shown in Table 1, were categorized into the following labels:

ECG signals are susceptible to various forms of noise and artifacts, which can obscure underlying cardiac activity and hinder accurate analysis. Thus, data preprocessing is crucial to ensure the quality and consistency of the input for the subsequent analysis.

In the first stage of data preprocessing, ECG segments unrelated to the laboratory conditions of interest were removed. These segments included the first and last minutes of the experimental setup, which captured the experimental setup and ECG lead removal, as well as any designated “short break” conditions. Compared with the original study, we excluded high-intensity physical activities (eg, treadmill running and stair climbing), as these activities are easily distinguishable from MS by, for instance, considering the maximum heart rate (HR) and could inflate reported model performance.

In the subsequent step, a 0.5 Hz high-pass Butterworth filter (order=5) was used to remove BL wander and slow drifts in the signal BL caused by respiration, movement, or electrode issues, as well as a power line filter (50 Hz) to attenuate noise from electrical power sources. Both filtering steps were implemented using the preprocessing pipeline provided by the NeuroKit2 package (version 0.2.7) [24]. While filtering addresses signal quality issues, some segments may remain unreliable due to factors such as temporary electrode detachment caused by leads accidentally falling off during the experiment. To identify and address these segments, a signal quality index (SQI) was computed for each QRS complex. Segments with an SQI below 25% were deemed unreliable and discarded. This threshold was chosen based on visual inspection, where segments with SQI values below this level often exhibited flat lines, rendering the identification of key peaks unreliable.

As illustrated in Table 2, the 30-second sliding-window approach with a 10-second shift yielded 31,408 observations.

Due to the controlled, high-quality conditions of the laboratory-based ECG data collection and the preprocessing pipelines used in this study, missing values were scarce. Out of 31,408 analysis windows (Table 2), only 50 (0.16%) windows contained missing data, all of which were attributed to the TWA feature, which requires an adequate number of detected T-peaks for reliable computation. We imputed these missing values using a k-nearest neighbors approach with a k value of 5, where neighbors were selected using Euclidean distance across all 55 standardized (as described below) features, and the missing values were imputed using the mean values of the 5 nearest neighbors. Both the imputer and the scaler were fitted exclusively on the training set to prevent data leakage.

Although the class distribution between the MS and nonstress states (BL, LPA, and MPA) is relatively balanced (45.68% MS vs 54.32% nonstress; Table 2), the leave-one-stressor-out analyses exhibit more pronounced imbalance, as specific mental stressors are left out during training, resulting in a class distribution of approximately 33.33% MS versus 66.67% nonstress examples (when PASAT or TA stimuli are left out). To overcome the class imbalance, we applied the synthetic minority oversampling technique (SMOTE) [76] exclusively to the training set to upsample the minority class, that is, the MS condition. SMOTE is an oversampling method that creates new synthetic instances of the minority class by interpolating between neighboring minority examples. To maintain a consistent training pipeline, we also used SMOTE in the main analysis, although it had a negligible impact on model performance in the more balanced MS versus nonstress setting (AUROC with vs without SMOTE: 0.741 vs 0.742 for XGBoost; 0.724 vs 0.724 for LR).

We framed our primary outcome as a binary classification task: detecting MS episodes across all nonstress states, including seated BLs, recovery periods, and LPA to MPA. In this way, our nonstress class reflects both desk-bound environments, such as office work in front of a computer, as well as more mobile contexts, such as those found in the teaching or nursing profession. By incorporating both dynamic and seated non-MS conditions, our models must learn to distinguish ECG changes induced by MS from those resulting from resting or physical activity—an inherently more complex task than handling either condition alone.

Given that this study represents a retrospective analysis of precollected data, ethical approval was not required under the Dutch Medical Research Involving Human Subjects Act (WMO) [77]. However, the study from which the experimental data were obtained [20] was approved by the Medical Ethical Committee of the Vrije Universiteit Amsterdam Medical Center (METc VUmc #2017.374, ABR #NL62442.029.17). All participants provided written informed consent before the start of the experiment and were either compensated with research credits (for student participants) or with a €50 (US $57.92) gift voucher. All data were pseudonymized.

Before evaluating our models, we first examine 2 fundamental ECG biomarkers, namely mean HR and AVNN reactivity, to illustrate the intrinsic challenge of detecting MS in everyday routine activities. Mean HR was selected for its simplicity of calculation from an ECG signal. In contrast, HR reactivity (or its inverse, AVNN reactivity) was chosen for its link to cardiac sympathetic and parasympathetic tone and its reliability in recognizing MS in experimental conditions [34]; however, because it relies on a controlled BL measurement, it cannot be defined in unconstrained, fully ambulatory settings.

Figure 1 shows the kernel density estimates of mean HR for five experimental conditions: seated BL, recovery (sitting), low- and moderate-intensity activity, MS, and high-intensity activity. As expected, increasing physical activity leads to a corresponding increase in mean HR. Crucially, the distribution for MS (blue) overlaps extensively with both low- and moderate-intensity activity (green and yellow) and even with the resting peaks (BL and recovery in gray). This pronounced overlap underscores the difficulty of disentangling MS-induced changes from those arising during everyday activity or recovery based on a simple measurement such as HR.

Figure 2 shows the AVNN reactivity, calculated as the difference between AVNN during each experimental stressor and the seated BL, across all MS conditions. Although some individuals exhibit positive AVNN reactivity, represented by the mean change from the seated BL, was significantly negative across all tasks, indicating an overall decline in AVNN compared with BL (PASAT: t₁₂₆=−7.66, P<.001; PASAT [repeat]: t₁₂₆=−12.02, P<.001; RAVEN: t₁₂₆=−3.19, P=.002; SSST: t₁₂₆=−11.92, P<.001; TA: t₁₂₆=−8.20, P<.001; TA [repeat]: t₁₂₆=−13.09, P<.001 for 2-sided paired t tests). However, the magnitude of this change varied considerably across experimental conditions. For instance, TA repeat elicits the most considerable mean AVNN reduction, whereas RAVEN matrices produce a more modest decline.

Taken together, these figures demonstrate that neither mean HR nor HR reactivity (or its inverse AVNN reactivity) alone can reliably discriminate MS from everyday activity or account for heterogeneity across stress tasks and individuals. This motivates our extraction of 55 time-, frequency-, nonlinear, and morphological ECG features, and the use of ML models to develop an MS classifier that detects physiological patterns characteristic of MS.

The large interindividual variability in ECG responses underscores that MS reactions are person-specific, likely influenced by subjective perceptions of stress. In Multimedia Appendix 3, we also show changes in both positive and negative affect relative to the sitting BL. As expected, positive affect significantly decreased across all mental stressors (PASAT: t₁₂₆=−13.25, P<.001; PASAT (repeat): t₁₂₆=−13.45, P<.001; RAVEN: t₁₂₆=−7.96, P<.001; SSST: t₁₂₆=−4.65, P<.001; TA: t₁₂₆=−7.71, P<.001; TA [repeat]: t₁₂₆=−5.00, P<.001 for 2-sided paired t tests). Likewise, negative affect significantly increased compared with the sitting BL (PASAT: t₁₂₆=−12.00, P<.001; PASAT [repeat]: t₁₂₆=−6.57, P<.001; RAVEN: t₁₂₆=−6.09, P<.001; SSST: t₁₂₆=−6.67, P<.001; TA: t₁₂₆=−8.31, P<.001; TA [repeat]: t₁₂₆=−3.21, P=.002 for 2-sided paired t tests). Nonetheless, apparent individual differences emerged, highlighting the challenge of building models that generalize across both stressors and individuals.

We next examined feature importance and model calibration, that is, how well the predicted probabilities represent the true probability of the outcomes. Figure 4 shows the 10 most important predictors for mental-stress detection in the XGBoost model according to the SHAP values. Notably, 5 out of these features originate from the frequency domain, highlighting the importance of spectral features in stress detection. Two heart-rate fragmentation indices (PSS and IALS) and fuzzy entropy also rank highly, with higher values of these nonlinear features pushing XGBoost’s output towards the MS class.

Interestingly, maximum HR and NN20, 2 simple time-domain features, are also important for predicting MS, demonstrating the relevance of time-domain metrics.

In Multimedia Appendix 9, we present the corresponding SHAP values for the LR model, along with its feature rankings based on the absolute values of the model coefficients. Fuzzy entropy emerges as the most important predictor of MS, underscoring its significance.

IALS again appears to be a crucial feature for MS detection, thereby confirming the importance of heart-rate fragmentation indices and underscoring the results obtained with the XGBoost model. Among time and frequency domain features, total and maximum power in the HF band, SD of HR, and PNN20 are also of high importance for the LR model.

The corresponding model-explainability results for the MS versus the seated nonstress condition (Multimedia Appendix 9; Figures 3-5) also highlight the importance of fuzzy entropy, as well as both the high-frequency and very high-frequency components, as predictive features.

In the health care domain, accurate probability estimates are crucial for informed decision-making [66,70]. For this reason, we assessed model calibration using the Brier score and the expected calibration error with 10 bins.

LR achieves a bootstrapped mean Brier score of 0.215 (95% CI 0.199‐0.232), while XGBoost scored 0.209 (95% CI 0.194‐0.224), thereby falling just above our <0.20 cutoff for well-calibrated probabilities. However, the AUROC for the XGBoost model is 0.615 (95% CI 0.586‐0.643), and 0.603 (95% CI 0.572‐0.633) for the LR, indicating substantial relative improvements over the dummy classifier BL that predicts the majority class (MS). Figure 5 shows the calibration curves for both classifiers, indicating slight underconfidence at low prediction probabilities (0.05‐0.25).

Generating the full set of 55 features is time-consuming and increases memory demands, which are both undesirable for resource-constrained wearable devices. Moreover, a large feature set can obscure model interpretability. To assess the trade-off between model simplicity and predictive performance, Figure 6 reports the AUROC as the number of features is progressively reduced. Notably, performance remains surprisingly robust, with over 95% of the original AUROC retained for both models, even with only 10 features. Interestingly, XGBoost outperforms the LR model across all feature sets. In Multimedia Appendix 10, we report the corresponding results for the AUPRC score, which are in line with the previously stated insights. The full list of selected features for each feature set is provided in Multimedia Appendix 11. As noted in the Methods section, feature selection was conducted using a single validation split. The resulting feature rankings should therefore be interpreted with caution.

Every day, life exposes individuals to a wide variety of mental stressors, making it practically infeasible to train a model on every possible type of stimulus. Thus, a model’s ability to perform well under unseen mental stressors is essential for real-world deployment [17]. Table 3 reports the AUROC for each model under a leave-one-stressor-out scheme. For each held-out stressor, the model is trained on all remaining stressors plus the nonstress condition and then tested on its ability to discriminate the held-out stressor from the nonstress condition. Note that if a mental stressor had a repeat condition in the experimental protocol (eg, TA and PASAT), both stressors were left out during training of the ML models. To ensure these results accurately reflect the difficulty of generalizing to an unseen stressor, rather than just the effect of having fewer stressor types and training examples, we also report, for each left-out mental stressor, the BL AUROC on the remaining (known) stressors.

As shown in Table 3, model performance on held-out stressors generally declines compared with stressors included in the training set, as expected. Nevertheless, all models achieve AUROC well above the 0.50 chance level on most unseen stressors, confirming their ability to generalize to new mental stimuli. The exception was SSST, for which the 95% CI overlaps chance performance. Notably, the magnitude of performance drop to unseen mental stressors varies across the different stimuli. While both the SSST and RAVEN are most difficult to classify correctly as MS, the opposite holds true for the TA condition. The fact that LR, XGBoost, and the RF model show comparable generalization capability suggests that differences in generalizability to novel stressors arise from the underlying ECG response, rather than from model complexity.

Multimedia Appendix 12 presents AUPRC results that closely mirror the AUROC findings. SSST and RAVEN are the hardest to detect when excluded from model training. Although model performance is generally worse for the left-out stressor, the TA stressor (and its repeat condition) shows the opposite pattern: the AUPRC score when TA is held out exceeds that of other mental stressors included in training. This suggests that TA produces particularly distinctive and readily identifiable stress patterns.

While our leave-one-out evaluation results suggest that our ML models can overall generalize to novel mental stressors, the generalizability to unseen mental stressors varied significantly. In particular, the sharp contrast between TA and SSST, 2 mental stressors that elicit strong physiological responses as measured by AVNN reactivity (Figure 2), is intriguing. We therefore examined how each model performed on individual stressor types. Table 4, therefore, reports AUROC scores stratified by mental-stressor condition. Again, model performance varies by mental stressor: TA achieves AUROC>0.75, while others (eg, RAVEN and SSST) yield only moderate performance, which suggests that certain MS stimuli produce more distinguishable ECG responses (as compared with the nonstress class) than others. Importantly, this observation holds for LR, XGBoost, and RF alike, demonstrating that these differences are not related to model complexity. The results for the AUPRC, as presented in Multimedia Appendix 13, align with these findings.

While the relatively poor model performance on the RAVEN stressor likely stems from its weaker elicited physiological responses (AVNN reactivity is displayed in Figure 2), the similarly low performance on SSST is more surprising. One possible explanation is that SSST elicits a fundamentally different physiological response compared with the other mental stressors, which could account for the poor generalization of the ML models to the SSST stressor observed in Table 3. To test this hypothesis, we retrained our LR and XGBoost classifiers separately on each stressor, that is, training and testing exclusively on one condition at a time. Multimedia Appendix 14 summarizes these per-stressor results, showing that both models outperform random BLs across all mental stressors. Notably, SSST shows marked improvement: despite both models struggling to detect it in earlier analyses, it now achieves an AUROC of 0.73 for LR and 0.76 for XGBoost, comparable to other mental stressors. This indicates that the SSST task elicits detectable MS responses. However, when SSST is excluded during training, AUROC drops sharply when tested on it (Table 3). Together, this indicates that while our ML models can learn to discriminate between SSST and nonstress when trained on SSST stressors alone, the patterns that the ML models learn from other mental stressors do not generalize to SSST. Interestingly, a similar pattern is also observed for the RAVEN stressor, further supporting this interpretation.

Given that MS can elicit ECG responses similar to those induced by physical activity (Figure 1), our ML models may conflate metabolic demand with MS. We therefore examined condition-specific sensitivity and specificity for light (LPA) and MPA (at 1000 sampling rates, 55 features).

The XGBoost model achieved good sensitivity for MS (0.800; 95% CI 0.698‐0.889), with the LR also yielding acceptable performance (0.772; 95% CI 0.656‐0.877). For the LPA, XGBoost achieved a specificity of 0.787 (95% CI 0.662‐0.903), whereas LR performed comparably, with a bootstrapped mean specificity of 0.794 (95% CI 0.685‐0.895). However, our models performed worse for MPA (LR: 0.444, 95% CI 0.314‐0.576; XGBoost: 0.418, 95% CI 0.299‐0.542), suggesting that even with a comprehensive feature set of 55 features, distinguishing stress-induced ECG responses from those elicited by MPA remains challenging.

We developed ML models to distinguish MS from a composite nonstress BL encompassing various everyday activities (eg, walking at a normal pace, recovering from MS, or vacuum cleaning). Using 55 features derived from the ECG signal, XGBoost achieved an AUROC of 0.741 and AUPRC of 0.706, while LR performed comparably (AUROC: 0.724; AUPRC: 0.692) at 1000 Hz. Both models achieved high sensitivity (XGBoost: 0.800; LR: 0.782) but low specificity (XGBoost: 0.512; LR: 0.509), with this pattern being particularly pronounced for MPA. Our results, therefore, highlight a fundamental challenge in distinguishing stress-induced cardiac responses from those driven by physical exertion when relying solely on ECG features. The ML models used in our study also demonstrated robustness to ECG signal downsampling, retaining more than 93% of performance using only 10 features.

While the nonlinear model (XGBoost) numerically outperformed our linear model (LR), the performance differences were small (<0.02 difference in AUROC and AUPRC at 1000 Hz), suggesting that complex nonlinear models offer only marginal benefits relative to linear BLs. One potential explanation is the so-called Rashomon effect [80], which posits that there exist many equally well-performing models for specific datasets—a phenomenon often observed in high-stakes applications and tabular datasets [81]. High outcome variability, likely present in our dataset, as physiological responses to mental stressors are both participant- and stressor-specific (Figure 2 displays the HRV response), increases the likelihood that a simple model will perform on par with a more complex one [82]. The fact that we do not apply time-series models to the raw ECG dataset using, for instance, deep learning approaches and instead work with tabular data also likely further contributes to this phenomenon.

Our findings reveal a clear trade-off between window size and model performance, as increasing the window size led to model improvements. This can be explained by more reliable features, as shortening the window size has been shown to affect the quality of HRV features [25]. Deep learning approaches, such as convolutional or recurrent neural networks, can partially overcome this trade-off, as they learn feature representations directly from the ECG signals and have been shown to achieve promising performance, even using shorter window sizes than those considered in our study [83,84]. However, these models are often considered “black boxes” and typically require more data for training than the ML models considered in our study. Although recent advances in time-series explainability methods, such as those by Crabbé and Schaar [85] and Enguehard [86], have been proposed to overcome the challenge of interpreting deep learning time-series models, these approaches typically highlight only key segments in the time-series signal driving the model prediction. However, interpreting raw time series is challenging for humans. In contrast, our ML models use features based on well-studied physiological constructs, thereby facilitating the interpretation of model predictions using methods such as SHAP.

The feature importance analysis revealed the importance of fuzzy entropy for discriminating MS responses from other nonstress states. This finding aligns with prior work, which indicates that entropy-based measures, capturing the irregularity of the ECG signal, increase during periods of MS [87]. Notably, 2 HR fragmentation indices, PSS and IALS, also emerged as important features, despite being less commonly used in ML models for detecting MS. Our findings further highlight the importance of frequency-based features, consistent with a comprehensive review [28] covering 37 studies. In their studies, the authors identified changes in the low- and high-frequency components of the ECG signal as the most reported marker of MS.

Additionally, our analysis identified the VHF (as opposed to the UHF) band as an important predictor. Although the physiological origin of the VHF band is less well understood than that of the LF or HF bands, its predictive value likely reflects stress-induced hyperventilation, a well-documented physiological response to various stressors [35-37]. While the VHF band remained an important feature for distinguishing MS from a seated BL (Multimedia Appendix 9; Figures 3-5), suggesting a physiological contribution under sedentary conditions, we acknowledge that this band might also represent motion artifacts or electromyogram noise during active conditions. The low specificity observed during MPA may partially reflect this confound: If VHF power increases during physical activity, due to motion artifacts rather than vagal modulation, the model may misclassify physical exertion as MS. Further research on this band under stress is needed, particularly using accelerometry or respiratory measurements to distinguish stress-related changes from motion artifacts.

Our leave-one-stressor-out experiments demonstrated that our ML models could generalize to most held-out stressors, though performance varied by stressor. While SSST and Raven proved to be the most difficult stressors to recognize, the opposite was observed for TA: both models detected MS evoked by TA and its repeat condition, even when they had never been trained on it. This suggests that the TA stressor produces a strong and distinct physiological ECG response that can be easily distinguished from nonstress states. Whether this is related to differential psychological processes engaged by these tasks is an open question. The TA evokes frustration and a fear of punishment, whereas the SSST is social-evaluative and invokes a fear of exclusion, similar to Raven and mental arithmetic tasks, which also involve ego threats. Building on this, our findings imply that robust, generalizable models require training on a diverse set of various distinct mental stressors, or, when constrained, focus on mental stressors that evoke a strong physiological response, such as TA.

Although our results demonstrate that the ML models can detect MS with high sensitivity, distinguishing stress-evoked ECG responses from those induced by MPA remains especially challenging (specificity LR: 0.444; XGBoost: 0.418), even with a comprehensive set of 55 handcrafted features. Because deep learning models can learn more nuanced temporal ECG patterns, a possible path forward is to apply time-series deep learning directly to raw ECG data. Moreover, incorporating additional sensing modalities (eg, accelerometer or electrodermal activity) may help the model separate cardiac responses driven by metabolic demand from those driven by psychological stress. We therefore consider deep learning–based multimodal approaches an important direction for future research.

Importantly, our results suggest potential for translation to photoplethysmography (PPG)-based settings, as the most predictive features in our ML models are derived from beat-to-beat dynamics, specifically, the R-peaks and interbeat intervals, rather than ECG waveform morphology (eg, T-wave amplitude). Notably, the XGBoost model maintains most of its performance with 10 features (AUROC: 0.729) or 5 features (AUROC: 0.716; Figure 6). These feature subsets consist of PPG-compatible features (HR statistics and frequency-domain features; Multimedia Appendix 11), thereby supporting the possibility that our results may generalize to PPG signals. Consistent with this potential, Taoum et al [88] demonstrated that HRV metrics derived from PPG agreed with the ECG counterpart for 37 out of 48 features for recordings of 5 minutes and 30 seconds, and further for 16 features in 1-minute and 30-second windows. However, since we train on 30-second ECG epochs, future work must determine whether PPG devices can reliably extract those intervals in similarly short windows. Encouragingly, our models remain robust even when the ECG is downsampled to 125 Hz, supporting deployment on lower-resolution PPG wearables. Yet, PPG signals are more susceptible to motion artifacts than research-grade ECG [89], which may challenge the direct transfer of our feature pipeline. Given these considerations, validating and adapting our models for real-world PPG recordings, particularly under movement conditions and within 30-second windows, is therefore an important direction for future research and essential before our approach can be deployed on PPG-based wearables.

Despite the results and insights presented in the paper, a few limitations warrant attention.

First, our cohort comprises Dutch-speaking young working adults (aged 18‐48 years). While our analytical framework is broadly applicable, physiological stress responses are highly individual and may vary across age groups, cultures, and health statuses.

Additionally, activity patterns might change over the lifespan [102], further limiting generalizability. Importantly, our results do not directly generalize to older populations, given that HRV, HR reserve, and other time-domain features generally decline with age [103], potentially making the classification problem more difficult. Given that distinguishing MS from cardiac responses evoked by physical activity proved very challenging despite our comparable young cohort, we anticipate that this challenge will persist (or even become more complicated) when considering an older population. An interesting direction for future work is therefore to validate our models in more diverse populations, such as older adults or non-Western cohorts, to ensure robustness and generalizability of our findings.

Second, we used global z-standardization across the training set to enable generalization to unseen users without requiring per-user calibration data. While this approach is attractive for real-world deployment, it does not account for the law of initial values [104], that is, it ignores interindividual variability in physiological responses to mental stressors present in our study (Figure 2). Our global normalization scheme may therefore bias model predictions against specific physiological phenotypes, for instance, by increasing false-positive rates among individuals with elevated resting HR or reduced BL HRV. Future work could examine the extent to which global versus individual-aware normalization strategies systematically affect downstream model performances across different BL physiological characteristics.

Third, we rely solely on task labels rather than participants’ self-reports to assign MS conditions, which prevents us from confirming that every labeled stress window corresponds to actual perceived stress. Incorporating self-assessments of perceived stress could have therefore reduced label noise, particularly given the pronounced interindividual variability in physiological responses (Figure 2 and Multimedia Appendix 3), which may partly account for the observed failure to generalize to the SSST stressor. Furthermore, instead of treating stress as a binary variable (stress or nonstress), one could use perceived ratings to classify the specific level or intensity of perceived MS [105].

Fourth, a related limitation of this study concerns the labeling of recovery periods. Physiological recovery from stress is gradual rather than instantaneous, suggesting that segments labeled “non-stress” during recovery periods likely retain residual stress physiology, constituting class-dependent label noise [106]. Empirically, label noise has been shown to adversely affect model performance across various ML approaches, including deep learning methods [107-109]. Our sensitivity analysis supports this observation: excluding recovery periods improved AUPRC for LR and XGBoost (see “Model Performance under Simplified Conditions”). We retained recovery periods as nonstress for comparability with prior stress-detection literature, which commonly assigns labels based on the protocol, and to provide conservative performance estimates. Yet, an alternative approach could either treat recovery as a buffer period excluded from training and evaluation or apply methods such as importance reweighting [110] or weighted surrogate loss [111].

Fifth, while our study uses ECG data to detect MS and downsamples the signal to mimic the hardware constraints of consumer-grade devices, such wearables typically rely on PPG. The most predictive features in our models are derived from R-peaks and beat-to-beat dynamics, which can also be extracted from PPG, suggesting that our approach may be transferable. However, this hypothesis remains to be validated empirically. Future work should therefore assess the extent to which our findings can be generalized to PPG-based settings.

Last, the MS tasks in our experimental setup were performed while participants remained seated. In everyday life, however, individuals often experience MS while standing or moving, for example, a teacher navigating a lively classroom. Future work should therefore extend the evaluation to simultaneous stress-and-activity scenarios to fully validate ambulatory MS detection, as done by Sun et al [63], Hosseini et al [112], and Kaczor et al [113].

This study evaluated ML models for distinguishing MS from routine physical activities using a single-sensor ECG. Both LR and XGBoost achieved acceptable discriminative performance (AUROC >0.7), with XGBoost providing only marginal benefits over the linear BL (Δ<0.02), suggesting that simple, interpretable models can perform competitively. Performance remained robust when downsampling from 1000 to 125 Hz and reducing the feature set to 10, thereby supporting lightweight deployment. However, our findings reveal a critical limitation of single-sensor ECG approaches: although both models detected MS with acceptable to good sensitivity (XGBoost: 0.800; LR: 0.782), specificity was low (XGBoost: 0.512; LR: 0.509), particularly for MPA. This indicates that stress-induced cardiac responses cannot be reliably distinguished from those driven by physical exertion using ECG features alone. Generalization to unseen stressors was also stressor-dependent, with models performing well above chance for most stressors but near chance for the social-evaluative stimulus (SSST). Future studies should validate our findings in ambulatory settings where MS and physical activity can co-occur and explore both multimodal approaches and the use of PPG via wearables as a practical alternative to ECG for continuous stress monitoring.