Prenatal stress assessment using heart rate variability and salivary cortisol: A machine learning-based approach.

OBJECTIVE: To develop a machine learning algorithm utilizing heart rate variability (HRV) and salivary cortisol to detect the presence of acute stress among pregnant women that may be applied to future clinical research.
METHODS: ECG signals and salivary cortisol were analyzed from 29 pregnant women as part of a crossover study involving a standardized acute psychological stress exposure and a control non-stress condition. A filter-based features selection method was used to identify the importance of different features [heart rate (HR), time- and frequency-domain HRV parameters and salivary cortisol] for stress assessment and reduce the computational complexity. Five machine learning algorithms were implemented to assess the presence of stress with and without salivary cortisol values.
RESULTS: On graphical visualization, an obvious difference in heart rate (HR), HRV parameters and cortisol were evident among 17 participants between the two visits, which helped the stress assessment model to distinguish between stress and non-stress exposures with greater accuracy. Eight participants did not display a clear difference in HR and HRV parameters but displayed a large increase in cortisol following stress compared to the non-stress conditions. The remaining four participants did not demonstrate an obvious difference in any feature. Six out of nine features emerged from the feature selection method: cortisol, three time-domain HRV parameters, and two frequency-domain parameters. Cortisol was the strongest contributing feature, increasing the assessment accuracy by 10.3% on average across all five classifiers. The highest assessment accuracy achieved was 92.3%, and the highest average assessment accuracy was 76.5%.
CONCLUSION: Salivary cortisol contributed a significant increase in accuracy of the assessment model compared to using a range of HRV parameters alone. Our machine learning model demonstrates acceptable accuracy in detection of acute stress among pregnant women when combining salivary cortisol with HR and HRV parameters.

Introduction

Stress is a body’s physiological response to one or more stimuli that have disrupted its mental or physical equilibrium [1]. In contrast to the environments in which our stress reaction system evolved, the growing mental burden from societal expectations and daily workload in the present day contributes to pervasive, chronic stress [2]. In a comprehensive survey conducted in 2017, up to 60 percent of employees across 35,000 organizations worldwide reported high levels of stress, and 35 percent reported constant but manageable stress levels [3]. Prolonged stress is usually associated with depletion of overall health [4], resulting in a higher likelihood of disease, including psychological illnesses, heart disease, asthma, obesity, and diabetes [5]. Many studies have also shown that stress can lead to various maladaptive health behaviors such as smoking, poor sleep, and unhealthy eating habits [6].

Maternal stress in the context of pregnancy may be particularly harmful due to the transmission of stress signals to the developing fetus. Pregnant women may encounter various stressors that exceed those of everyday life for non-pregnant people including worries about the wellbeing of themselves and their developing baby, concerns about delivery, and financial or other socioeconomic stressors associated with a growing family. The Covid-19 pandemic in the last two years is another factor that significantly contributes to elevated maternal stress [7]. These stressors have been shown to be associated with preterm birth, low birth weight, risk of gestational hypertension, and adverse health and behavioral outcomes in offspring [8-10]. In order to design effective interventions that may help to mitigate the adverse effects of maternal stress on pregnancy and offspring health outcomes, there is first a need to understand the optimal approach for assessing stress in pregnancy.

Stress assessment with high accuracy is challenging as stress is related to many different factors. The most common method of assessing current or recent psychological stress is through self-report questionnaires [11]. Although questionnaires and interviews are practical and enable researchers to collect subjective information from a large number of participants, these methods suffer from multiple disadvantages including recall bias, social desirability bias, and ignoring some questionnaire items. Studies that aim to capture changes in stress by administering repeat questionnaires over a short time span can cause survey fatigue [12]. Furthermore, inter-individual variability in how questionnaire items are interpreted may reduce the validity of some questionnaires [13]. In standard clinical practice with pregnant women, stress assessment surveys are not typically used due to time constraints and the various limitations of these methods as noted above. As research on objective stress characterization using real time, non-invasive technology evolves, there is potential for these techniques to be translated to clinical settings so that clinicians and/or their pregnant patients may benefit from momentary feedback on patients’ physiological state. In turn, this could facilitate implementation of simple, effective stress management techniques that could reduce the adverse effects of stress on health outcomes for mother and baby.

Objective stress assessment methods using ubiquitous sensing and machine learning technologies are therefore preferable in human studies to more reliably capture stress reactivity or changes in stress over time. Statistical methods can measure stress according to one or two physiological features at a time. However, machine learning approaches can assess stress objectively by fusing multiple physiological features, and even multiple modalities, thereby increasing efficiency and accuracy. Psychological stress first activates the autonomic nervous system (ANS), which can be detected through fluctuations in biological signals such as stress hormones or heart rate activity [14].

Cortisol, a glucocorticoid stress hormone, is one of the primary biomarkers of psychological stress [15]. During pregnancy, cortisol plays a crucial role in supporting fetal development [16]. Maternal cortisol levels may be increased up to four times compared to non-pregnant individuals [17], which has a positive influence on fetal neural development [18]. However, excessive prenatal stress results in fetal exposure to heightened circulating cortisol levels, which has negative consequences for neurodevelopmental outcomes and may contribute to preterm delivery [19]. Although cortisol is readily detectable in saliva in the unbound form and salivary cortisol levels correlate strongly with serum levels [20], assessing stress via cortisol measurement in pregnancy has several challenges. As cortisol levels increase with advancing gestation, studies must aim to standardize the gestational timing of sample collection across subjects. Moreover, stress assessment relying on self-obtained saliva collections in free-living situations may be inaccurate as cortisol is very sensitive to changes across time of day and time of awakening [21].

Heart rate variability (HRV) is another important avenue for stress research but has been understudied in the context of pregnancy. HRV is a measure of cyclical variations in beat-to-beat intervals. It contains time-domain, frequency-domain as well as nonlinear parameters. Greater variation in beat-to-beat intervals indicates lower stress activity in the ANS. Therefore, lower levels of time-domain HRV features are indicative of higher stress [22]. Certain patterns in frequency-domain HRV parameters (e.g., low-frequency band and high-frequency band) are associated with stress exposure [23]. The ratio of the low-frequency band and the high-frequency band is also associated with self-reported maternal depression and stress in pregnancy [24]. HRV can be easily measured using non-invasive wearable technology, making it an attractive physiological measure in both lab-based and free-living research settings. However, little is known about the reliability of HRV parameters as a measure of stress and stress reactivity in pregnancy, or how HRV performs compared to cortisol levels to characterize stress.

The aim of this study is to develop an objective, multi-feature algorithm to detect the presence of acute stress among pregnant women following standardized stress exposure, using HRV parameters, salivary cortisol, and machine learning with the help of a filter-based feature selection method. We also determine the contribution of including salivary cortisol measures to increase the accuracy of the assessment model. To the best of our knowledge, this research is the first effort to leverage the association between cortisol and HRV in building objective stress assessment methods.

Methods

Study design

This is a secondary analysis of cortisol and HRV data collected from a cross-over study that aimed to assess the effects of superimposed psychological stress on the postprandial metabolic response to a standardized breakfast meal during pregnancy. The study was a cross-over design involving two visits to the clinical research facility, during which participants underwent either a control non-stress task or an acute psychosocial stress challenge task, with a 1 to 2 week washout period between visits. The visit order was the same for all participants, with the control task on the first visit and stress task on the second, in order to reduce the likelihood of higher anticipatory stress when returning for visit 2 if the stress task was undertaken on visit 1. The University of California, Irvine (UCI) institutional review board approved this study, and all participants provided written informed consent.

Participants and recruitment

Pregnant women were recruited to the study between February 2018 and March 2020. Women were eligible if they were of Hispanic ethnicity, aged 18–40 years, had a pre-pregnancy BMI 25.0–34.9 Kg/m2, carrying a singleton pregnancy, less than 30 week’s gestation, non-diabetic with a normal result on the standard glucose challenge test at 24–28 weeks, non-smoker, and fluent in either English or Spanish. Study exclusion criteria were non-Hispanic ethnicity, aged <18 or >40 years, pre-pregnancy BMI <25.0 or ≥35.0 Kg/m2, carrying more than one baby, >30 week’s gestation, diabetes (including gestational diabetes mellitus), hypertension, preeclampsia, diagnosis or treatment of any other condition that may disrupt metabolic, endocrine or immune function, diagnosis of a current psychiatric disorder or use of psychotropic medications, unwillingness to attend two study visits or consume the milkshake drink on each visit. The parent study restricted eligibility to women of Hispanic ethnicity to have a homogenous cohort of participants known to experience higher levels of socio-cultural stressors. Homogeneity among participants is important to limit inter-individual variability that could influence metabolic biomarkers that are primary outcome measures for the parent study. Recruitment took place in-person at UCI Health-affiliated obstetric clinics in Orange County, California, or by passive measures through the dissemination of brochures or recruitment emails.

Data collection

The data collection protocol is shown in Fig 1. Participants arrived at the clinical research facility in the morning (8–9 am) following an overnight fast. Their cell phone and any other electronic devices in their possession were stored in a secure location for the duration of the visit to avoid potential distractions that may influence stress levels. Participants were asked to report their time of awakening on the morning of each visit. On visit 1, usual stress levels over the past month were assessed by the Perceived Stress Scale [25].

Fig 1

Data collection protocol for each study visit.

A baseline saliva sample was collected before consuming a milkshake drink, which was the same at both visits. Participants then stood upright while an Actiheart electrocardiograph monitor (CamNtech Ltd.) was placed on the chest to continuously measure heart rate and inter-beat interval. A second saliva sample was collected at 15 minutes after the baseline sample, followed by the task period (15-minute duration), and post-task saliva sample collection immediately after the task (30 minutes post-baseline sample). Subsequent saliva sample collections occurred at 45-, 60-, 90-, and 120-minutes post-baseline. Only water was provided and no other food or drink was consumed until the end of the study visit. Apart from the 20-minute period from the time of Actiheart placement until the end of the task, participants remained comfortably seated throughout the visit with neutral reading material.

The 15-minute control non-stress task during visit 1 was performed with a friendly research staff member already familiar with the participant. This involved a relaxed conversation on topics such as recent or upcoming holiday plans, movies, TV shows, or books they enjoyed or selecting baby names. The researcher ended the task period by announcing the time for the next saliva sample collection.

The Trier Social Stress Test (TSST) was used during visit 2 as a standardized, validated performance task designed to activate the hypothalamic-pituitary-adrenal (HPA) axis and the sympathetic nervous system to evoke a measurable physiological stress response [26]. This laboratory-based psychosocial stressor incorporates elements of unfamiliarity, uncontrollability, and a threat to self-esteem to induce a short-term psychological and physiological stress response. The TSST has been previously administered among pregnant populations without any adverse effects [27]. It consists of 5 minutes of speech preparation, 5 minutes of speech performance under the stern observation of an evaluative committee while being video-taped, and a 5-minute complex, mental arithmetic task. This task was performed in a different room to other study procedures and with two research personnel who were unknown to the participant to increase sensations of unfamiliarity.

Saliva samples were collected using Salimetrics oral swabs, which were placed under the tongue for 2 minutes at each collection time point. Saturated swabs were inserted into a Salivette tube (Sarstedt) and centrifuged at 1500g for 15 minutes to extract saliva, which was then aliquoted and stored at -80°C. Saliva samples were assayed for cortisol at the Institute for Interdisciplinary Salivary Bioscience Research, UCI, using a commercially available enzyme immunoassay (Salimetrics, LLC). The assay uses 25 microliters of saliva per determination, has a lower limit of detection of 0.003 μg/dl, a standard curve range from 0.012 to 3.0 μg/dl, and average intra- and inter-assay coefficients of variation less than 5% and 10%, respectively. Assays were performed in duplicate and the average of the duplicate was used in statistical analyses. Outliers were assessed by detecting data points greater than 3SD from the mean on samples from both visits and outlying values were excluded. To account for inter-individual variation in cortisol values influenced by time of awakening on the mornings of the study visits, we used normalized cortisol variables, adjusted for time (minutes) interval from awakening until time of arrival at the visit, as the input for statistical analysis and the assessment model.

Data processing and analysis

Actiheart data

The Actiheart device provides Inter-beat Intervals (R-peak intervals) extracted from processed ECG signals. We used the 5-minute time windows of the Inter-beat Interval (IBI) signal to calculate the heart rate (HR, heartbeats per minute), time-domain parameters (i.e., RMSSD, AVNN, SDNN, and pNN50), and frequency-domain parameters (i.e., low frequency (LF), high frequency (HF), and LF/HF). The HRV parameters are briefly described in Table 1. Further details about the heart rate and HRV features calculation have been previously described [4]. The abnormal IBI and HRV values generated by motion artifacts were removed before proceeding with the analysis according to the removal criteria described in another study utilizing HRV measures [28]. The removal criteria are based on the normal range of HR and IBI values.

Table 1

A summary of HRV parameters.

Parameter	Units	Description
IBI	ms	Inter-beat interval; the time interval between two successive heartbeats
RMSSD	ms	The root mean square of successive differences between adjacent NN intervals
AVNN	ms	The average value of NN intervals
SDNN	ms	The standard deviation of normal NN intervals
pNN50	-	The proportion of the number of pairs of successive NN intervals differing more than 50 ms divided by the total number of NN intervals
LF	ms2	Power of the low-frequency band of the IBI signal (i.e., 0.04 Hz—0.15 Hz)
HF	ms2	Power of the high-frequency band of the IBI signal (i.e., 0.15 Hz—0.4 Hz)
LF/HF	-	The ratio of LF to HF

ms, milliseconds; NN, time (normalized) between two detected heartbeat detections.

Since the frequency of cortisol measurement was less than the measurement of the HRV parameters, we matched the cortisol value measured at the beginning of a 15-minute interval to the HRV parameter values measured at the three subsequent 5-minute time intervals. For instance, we matched the three 5-minute windows of HRV data between the 15- and 30-minute saliva collection time points to a single cortisol value (i.e., the value measured at 15 minutes). We characterized the acute stress exposure period during visit 2 as the 30-minute window starting from the initiation of the TSST (i.e., between the 15–45 minute saliva collection time points). Data collected at all time points during visit 1 were considered for the non-stress status.

Statistical analysis

We first visually compared the pattern of HR, HRV features, and cortisol values across time between two visits for each participant using line graphs, and the average values for each measure at each visit per participant using bar graphs. The area-under-the-curve (AUC) values of HR, each HRV feature, and cortisol for the two study visits were calculated, using the formula shown below (1), where C(t) represents the value HR or HRV or cortisol values in time t.

We used the paired-samples t-test to assess differences in the mean AUC values between visits using a p-value of <0.05 to indicate significance. We used Python and Python libraries, including Scipy [29] and Panda [30] to program the statistical analysis functions.

Feature selection

To maintain generalizability and avoid overfitting during the objective stress assessment, we implemented a feature selection method. Feature selection reduces the computational complexity, training time, and overfitting degree. Moreover, it improves the accuracy of the classification. Feature selection methods can be summarized into three categories: filter-based methods, wrapper-based methods, and embedded methods. We used a filter-based feature selection method which determines the relationship between multiple input features and target labels statistically. In other words, it can evaluate and filter out features that will be used in the classification model. Compared to the other two methods, the filter-based method is computationally cheaper and has less risk of overfitting [31].

In our filter-based methods, Gini impurity gain is added to find the most informative features for the assessment model. A decision-tree-based random forest classifier (max depth = 16) [32] is used to output the feature importance vector. Inside the decision tree model, each node is a condition on one of the features, and these nodes separate the data into two different sets. Data with the same labels are separated into the same set in an optimal scenario. The splitting condition depends on the impurity of the features chosen in each node. During the training process, the contribution to the decrease in the impurity of each feature is computed. Finally, the importance of each feature is ranked according to this measurement.

Machine learning-based predictive models for stress assessment

We developed a set of machine learning-based algorithms to detect the presence of acute stress exposure among pregnant women using HRV parameters and salivary cortisol. This is a binary classification between stress versus non-stress. Five different classification methods were implemented, including AdaBoost [33], XGBoost [34], Random Forest [32], Support Vector Machine (SVM) [35], and K-nearest-neighbor (KNN) classifiers [36]. The AdaBoost classifier is a meta-estimator that begins by fitting a classifier on the original dataset and then fits additional copies of the classifier on the same dataset but where the weights of incorrectly classified instances are adjusted such that subsequent classifiers focus more on difficult cases. XGBoost classifier is an optimized distributed gradient boosting library designed to be highly efficient, flexible, and portable. Random Forest classifier is an ensemble learning algorithm that fits a number of decision tree classifiers on various sub-samples of the dataset and uses averaging to improve the predictive accuracy and control over-fitting. Support Vector Machine is a margin-based classification technique used for the classification of linear as well as non-linear data. K-nearest-neighbor method uses k number of nearest data points and predicts the result based on a majority vote. We used the Scikit-learn library to create our classification models [37].

To accurately evaluate the performance of our classification models in terms of generalizability, the 10 folds Leave-one-subject-out cross-validation method [38] was used. Cross-validation is one of the most reliable algorithms that is used to objectively estimate the accuracy of a machine learning model on unseen data. It achieves this by training a model using different subsets of data and obtaining the average accuracy on the rest of the data as a test. In our study, among each iteration of our Leave-one-subject-out cross-validation method, 28 out of the total 29 participants’ data were included for model training. For testing, only the remaining participants’ data points were used. The final assessment accuracy of the model was obtained by averaging the accuracy of all constructed models.

Results

Thirty-three participants completed the study and of these, HRV and cortisol data at both study visits were available for 29 participants. The mean PSS score was 13.4±3.7 out of a potential range of 0–40. PSS values from 14–26 indicate moderate stress and values from 27–40 indicate high stress levels. Thus, on average, this population of pregnant women had borderline moderate levels of perceived stress in their daily lives.

Within-subject comparisons of parameters

We first visually compared the HR, HRV parameters, and cortisol in 5-minute time windows separately for each participant across the two studies visits. The comparison results can be summarized into three distinct types, representing inter-individual variability in the physiological response to acute stress exposure. Seventeen participants display an obvious difference in HR, HRV parameters, and cortisol between the stress and non-stress conditions. One example of this type of response is shown in Fig 2. Eight participants do not display a clear difference in HR and HRV parameters but do display a notable difference in cortisol, as shown in Fig 3. In the remaining 4 participants, no obvious difference in any of the parameters including cortisol could be detected. Fig 4 presents an example of this pattern of physiological non-responsiveness to the stress exposure. This variability in stress reactivity creates challenges for standard statistical techniques to detect the presence of stress.

Fig 2

Comparison of the HR, HRV parameters, and cortisol between visit 1 and visit 2 in 5-minute segments for a sample participant.

This participant demonstrated an obvious difference in HR, HRV parameters, and cortisol.

Fig 3

Comparison of the HR, HRV parameters, and cortisol between visit 1 and visit 2 in 5-minute segments for a sample participant.

This participant did not demonstrate a clear difference in HR and HRV parameters but did have a significant difference in salivary cortisol.

Fig 4

Comparison of the HR, HRV parameters, and cortisol between visit 1 and visit 2 in 5-minute segments for a sample participant.

This participant did not demonstrate any clear difference among all the features, including salivary cortisol.

We also visually compared the average value for all HRV features and cortisol between study visits for each participant (Fig 5). Among the 29 participants, we see that 19 participants have higher heart rate, 12 participants have higher RMSSD, 17 participants have higher SDNN, 13 participants have higher AVNN, and 12 participants have higher pNN50 on exposure to stress (visit 2) versus non-stress (visit 1) conditions. As for frequency-domain features, 14 participants have higher LF band, 17 participants have higher HF band, and 12 participants have higher LF/HF ratio in visit 2 compared to visit 1. Finally, 23 participants have a higher average cortisol value in visit 2 compared to visit 1, which is the most obvious feature. In summary, for most participants, average values for HR, SDNN, HF, LF/HF ratio, and cortisol tend to be higher under acute stress, while other time- and frequency-domain parameters tend to be lower compared to the non-stress condition upon visual inspection.

Fig 5

Comparison of average values for HR, HRV parameters, and cortisol between visit 1 and visit 2 for all 29 participants.

Results of the paired t-test comparing the group mean AUC data for 9 features between visits are summarized in Table 2. Only the AUC for heart rate is statistically significantly different between visits, such that on average, participants had higher heart rate throughout the stress visit versus the non-stress visit.

Table 2

Within-subject comparison of AUC values for HR, HRV parameters, and cortisol between stress and non-stress conditions.

Features	Visit 1, non-stress condition (Mean±SD)	Visit 2, stress condition (Mean±SD)	t value	df	P-value
HR	1726.31 ± 209.03	1836.13 ± 263.57	-3.285	28	0.003
RMSSD	649.13 ± 309.35	633.53 ± 285.59	0.369	28	0.715
SDNN	971.17 ± 289.14	1015.08 ± 317.29	-1.276	28	0.213
AVNN	13195.91 ± 2156.73	13219.98 ± 2540.54	-0.149	28	0.883
pNN50	1.91 ± 2.10	1.76 ± 2.10	0.634	28	0.531
LF	5606.18 ± 3400.39	5352.93 ± 3115.48	0.586	28	0.562
HF	5268.97 ± 4880.39	5523.71 ± 5213.20	-0.412	28	0.683
LF/HF	34.05 ± 23.96	34.98 ± 26.93	-0.262	28	0.796
Cortisol (standardized for time of awakening)	0.07 ± 1.05	0.04 ± 1.00	0.207	28	0.837

Objective stress assessment results

We built an objective stress assessment model using the 9 features mentioned above based on machine learning methods. In order to reduce the computational complexity and save training time, we implemented a feature selection method using Gini impurity gain.

We first conducted the feature selection on HR and HRV parameters only, excluding cortisol. According to the feature selection results, 5 out of 8 features were selected as the best combination. The selected features in the order of importance are AVNN, HR, SDNN, LF, and LF/HF, indicating that AVNN is the most important feature for accurately detecting stress reactivity in the absence of cortisol. HR represents the frequency of heart beats and AVNN represents the average length of heartbeat intervals. Although AVNN and HR appear highly related from a statistical perspective, they represent distinct HRV features and thus, both contribute value to the model. When cortisol is added into the feature selection process, the order of importance among the features is cortisol, SDNN, AVNN, LF/HF, pNN50, and LF. Cortisol makes the largest contribution to the accuracy of stress detection. Among all the HRV features, SDNN contributes most to the assessment accuracy in the presence of cortisol, and HR is no longer selected as an important feature.

Following feature selection, we built the stress assessment model using the top 5 features selected in the absence of cortisol using 5 different machine learning models. The assessment results are shown in Table 3. The AdaBoost classifier has the best performance with an assessment accuracy of 67.11%. The model was then repeated using the 6 features selected from 9 parameters including cortisol. AdaBoost remained the best classifier with an increased assessment accuracy of 76.51%, which is a 9.4% improvement compared to the stress assessment model without cortisol.

Table 3

Stress assessment results.

	AdaBoost	XGBoost	Random Forest	Support Vector Machine	K Nearest Neighbor
5 selected features in the absence of cortisol	67.11%	65.43%	64.04%	64.14%	60.15%
6 selected features in the presence of cortisol	76.51%	74.1%	73.55%	74.65%	73.27%%
All 9 features (including cortisol)	73.09%	72.21%	72.08%	73.1%	69.5%

The performance of the AdaBoost classifier with six selected features using the Leave-one-subject-out cross-validation method is shown in Fig 6. The highest accuracy is 92.31% and the lowest accuracy is 42.51%. For other classifiers, the accuracy improvement after adding cortisol as a feature is 8.67% for the XGB classifier, 9.51% for the Random Forest classifier, 10.51% for the SVM classifier, and 13.12% for the KNN classifier. That is an average 10.24% increase after introducing cortisol into the stress assessment model. The stress assessment results using all 9 features without the feature selection are also shown in Table 3. There is a slight accuracy drop for all the classifiers besides the increase in computational complexity and training time.

Fig 6

The performance of the AdaBoost classifier with 6 selected features using the Leave-one-subject-out cross-validation method.

Discussion

Principle results

This study demonstrates wide inter-individual variability in bio signal responses to acute psychological stress, including HR, HRV, and salivary cortisol values, in a cohort of pregnant women. Despite this variability, using a machine learning-based algorithm, we identified the optimal combination and classification of stress bio signals that are capable of assessing the presence of stress with almost 77% accuracy. Salivary cortisol contributed a significant increase in accuracy of the assessment model compared to using a range of HRV parameters without cortisol.

The fluctuation of HR, time-domain HRV, and frequency-domain HRV parameters across stress and non-stress conditions was evident for more than half of the study participants (17 out of 29), which suggests that these parameters in isolation may only be moderately reliable signs of stress exposure. In comparison, a noticeable visual difference in cortisol values across the two visits was observed for 25 out of 29 participants, although on average, the difference in cortisol AUC was not statistically significant between the stress versus non-stress visits. This points to the potential unreliability of using standard statistical methods and salivary cortisol alone as an objective measure of acute stress. Indeed, salivary cortisol is a more sensitive measure of HPA-axis function than serum cortisol since it is present only in the unbound, biologically active form in saliva [39]. Cortisol binding proteins in plasma increase markedly during pregnancy thereby reducing concentrations of the active hormone. Although both HRV and salivary cortisol levels reflect ANS activation, HRV parameters are affected by the sympathetic adrenal medullary (SAM) axis rather than the HPA-axis which triggers cortisol release. These different modes of activation may contribute to the observed differences in fluctuations of these bio signals under stress.

HR and HRV parameters can be measured easily using a chest band, a smartwatch, or a ring [28] in everyday life or in laboratory settings making it a feasible approach for most research studies. However, our results suggest that among pregnant women, HRV parameters alone may not be adequate as an acute stress assessment tool but may benefit from other bio signals to help distinguish between stress and non-stress states more accurately. Furthermore, ECG signals may be influenced by factors other than psychological stress. The results of the paired t-test in our study indicated a significant difference in HR between the stress and non-stress conditions. Besides the stimulation of stress, increased HR on the stress visit may also be related to the fact that women walked a short distance to a separate room for the TSST, while on the non-stress visit, they remained in the same room but stood up during the task period. These additional movements may have increased their HR to some extent around the time of stress exposure. Meanwhile, no significant difference was found in HRV parameters suggesting that standard statistical methods may not be sufficient to detect average HRV differences across a group. It is also important to note that 4 participants failed to demonstrate any notable difference in salivary cortisol, HR, or HRV between the stress and non-stress conditions. More advanced analytical approaches, such as machine learning-based feature selection, may be required in such cases to optimize the data and distinguish between stress and non-stress situations.

We implemented the machine learning-based feature selection using the Gini impurity index to reduce the computational complexity and avoid overfitting. Moreover, we present the most important features of the objective stress assessment model when cortisol is added by repeating the feature selection. The feature selection results differ from the statistical analysis because this method makes decisions according to the contribution of stress assessment accuracy. If the accuracy is not improved when a feature is added, these features will not be used further in the assessment model.

We also tested the performance of our model to reliably assess stress using five different machine learning methods. The assessment accuracy increased on average 10% across all five classifiers after adding cortisol as a feature. The significant difference in cortisol values for most participants helps the model to better distinguish between stress and non-stress states. Because of participants’ different HR, HRV, and cortisol reactions to stress exposure, the best performance model, AdaBoost classifier, achieved assessment accuracy higher than 90% and lower than 30% with different participants by using the Leave-one-subject-out cross-validation method. When we use all nine features, the assessment accuracy decreases slightly. The remaining three features add more computational complexity, resulting in model overfitting in the training process, and a poor performance during testing. The AdaBoost classifier achieved the highest assessment accuracy of almost 77% with the selected six features. These results demonstrate that our algorithm can assess an acute stress state for pregnant women using a combination of HRV parameters and salivary cortisol values with acceptable accuracy.

Comparison with previous studies

To the best of our knowledge, this is the first study to develop a machine learning-based algorithm to detect the presence of acute stress exposure among pregnant women using a combination of HRV features and salivary cortisol values. Ilyumzhinova et al. [40] reported the association between changes in salivary cortisol and RMSSD (one of the time-domain HRV parameters) among black pregnant women exposed to the TSST. They found that higher self-reported levels of stress related to experiences of discrimination among participants were associated with lower levels of cortisol reactivity and higher levels of RMSSD following the TSST, which demonstrates how baseline maternal factors may influence an individual’s stress responsivity. However, this study did not evaluate other HRV parameters in response to the TSST, nor did they utilize a combination of cortisol and HRV features to characterize the presence or absence of stress.

Previous studies that have attempted to objectively characterize stress in healthy, non-pregnant subjects report varying degrees of success. Bakker and Pechenizkiy proposed a stress assessment system using galvanic skin response (GSR), a measure of skin conductance that is affected by sweat gland activity under stress [41]. The authors concluded that using this bio signal in isolation was insufficient to determine the presence of stress with high accuracy. Sun et al. [42] used a physical activity protocol to induce stress and collected ECG, GSR, and accelerometer data for 30 minutes to implement a stress detection system. This study detects mental stress affected by physical activities. Han et al. [43] assessed stress among healthy people using 25 features extracted from ECG, Photoplethysmograph (PPG), and GSR signals. They achieved an accuracy similar to ours at the cost of huge computational complexity. In another study [44], a stress detection tool was developed using ECG, GSR, respiration rate, blood pressure, and peripheral capillary oxygen saturation (SpO2) using a waist strap. Although this study included multiple bio signals, the highest accuracy achieved by the Leave-one-subject-out cross-validation method was 89.3%, which is lower than the performance of our model (93.55%). None of these studies have pregnant participants, and their methods were all focused on healthy subjects. We identified only one published study that built a stress detection system for pregnant women [45]. This study implemented the TSST to induce stress and collected HR, ECG, GSR signals during the experiment. With more than 20 features extracted from these three signals, they achieved an average stress assessment accuracy of 70%, and their best Leave-one-subject-out accuracy was 81%. Our model performed better in both measurements using only six features that included salivary cortisol.

Future research is required to increase the accuracy of stress assessment by incorporating more bio signals such as GSR in combination with cortisol and HRV and applying more advanced machine learning methods. Although saliva samples can be conveniently collected by a cotton swab in free-living and human laboratory settings, this method is more invasive for participants than wearable sensors and performing cortisol assays adds a level of complexity.

Strengths and limitations

The crossover design is a strength of this study as it facilitates intra-individual comparison of physiological responses to stressful and non-stressful tasks. Also, the strict eligibility criteria concerning the gestational age at study assessments, ethnicity, and pre-pregnancy BMI range help to minimize inter-individual variation and thereby increase the validity of the results. We chose to study a cohort of Hispanic pregnant women as the Hispanic population in the U.S. may be subject to higher levels of baseline stress than non-Hispanic people due to various sociodemographic and acculturative factors. While our cohort perceived borderline moderate levels of stress over the past month, it is unknown whether such baseline stressors contribute to higher or lower stress responses to the controlled stress task compared to non-Hispanic women. Furthermore, our findings cannot necessarily be generalized to other racial/ethnic groups, or to non-pregnant or free-living subjects encountering everyday stressors. However, data collection in free-living situations may be subject to more motion artifact noise in physiological signals in comparison with data collection in a controlled environment and mostly stationary state (especially for ECG data which is worn on the participant’s chest).

We also acknowledge that the use of a laboratory-based standardized stress protocol does not necessarily reflect real-world stressful events, and indeed not all participants in this study demonstrated a physiological stress response to the task. However, the TSST is the most widely used, validated protocol for stress research in laboratory settings and the machine learning approach helps to overcome the issue of non-responders to the stressor by utilizing those data to train the model with greater accuracy to detect the presence of stress. Introduction of meal ingestion as part of the parent study could also be considered a limitation for the present analysis, as it is unknown to what degree the act of consuming a meal may have moderated the stress response. However, as the meal type and time of ingestion was standardized across all subjects and occurred on both visits, we assume that any potential influence of the meal on the physiological stress response is also standardized across participants. While we adjusted cortisol values for time of awakening at each visit, it is also possible that stressful experiences occurred for some participants before arriving to their study visits, which could have elevated their baseline cortisol values. We were unable to account for these potential sources of external stress in the data.

Other limitations of this study relate to the data processing procedures. The Acticheart only provided the IBI data and not the raw ECG signal, which may have facilitated the calculation of more accurate HRV parameters. Removal of corrupted IBI data due to motion artifacts resulted in some gaps in our HRV feature comparisons. Most of these gaps occurred at the very beginning of data collection, in the 5 minutes before participants underwent the stress or non-stress task period. This resulted in inadequate baseline HRV measurements from which to compute reactivity scores for each parameter, which may have been a more robust summary measure than the AUC values on which we relied for the statistical analysis. As noted above, salivary cortisol values were only available at set intervals according to the timing of sample collections, while HRV parameters are continuously measured and data reported in 5-minute intervals. As a result, the cortisol value in three consecutive 5-minute segments was assumed to be the same for the purpose of the feature selection models. If it would be possible to measure cortisol every 5 minutes, there is potential to further increase the accuracy of our stress assessment model.

Conclusion

In summary, our results demonstrate wide variability in objectively measured stress reactivity signals among pregnant women exposed to a standardized laboratory stress task. HR, SDNN, HF, LF/HF ratio, and cortisol tend to be higher under acute stress, while other time- and frequency-domain parameters tend to be lower. By adding cortisol as a feature alongside HR and HRV parameters, our model can detect pregnant women’s stress status with acceptable accuracy. Therefore, our objective stress assessment algorithm provides a framework for future clinical research studies to consider a combination of physiological stress signals and machine learning-based approaches to assess stress more accurately in human subjects. The potential translation of this work to the assessment of chronic stress in humans requires further research.

30 May 2022

28 Jun 2022

We thank the reviewers for their thorough review of this manuscript and helpful feedback. We have addressed each comment below and annotated the relevant changes made to the manuscript.

In response to the query of data availability from reviewer #1, we have uploaded the study data to an online repository which may be accessed at this link during the review process: https://datadryad.org/stash/share/8SKG9TIJOiVgeV90rKvNBI34O2x6eyNgjjgxt4Om8CE

REVIEWER #1:

This is an interesting topic and the authors have done a good job with the manuscript. However, my main concern is in regard to the rationale behind the study. After reading the manuscript I’m still not sure what the benefit is of undertaking machine learning to identify stress in this population and how this might be practically useful.

A few comments:

Abstract

Comment #1: From reading the aims and methods sections, I’m not actually clear what data the machine learning is based on.

Response: We thank the reviewer for their comments. The machine learning algorithms use the physiological data collected from pregnant women including Heart Rate, Heart Rate Variability (RMSSD, AVNN, SDNN, pNN50, LF Band, HF Band, and LF/HF Ratio), and salivary cortisol. Salivary cortisol is considered the primary stress hormone but it is not always convenient to measure in free-living subjects. Meanwhile, heart rate variability (HRV) offers a convenient alternative measure of stress physiology using non-invasive wearable technology. However, there are limited studies that consider the reliability of HRV as an indicator of prenatal stress, and the combination of cortisol and HRV assessment in pregnancy has not yet been tested. Therefore, in this study, we applied machine learning methods to assess the presence of acute stress in pregnant women using these multimodal parameters simultaneously.

We have updated the methods description in the abstract for greater clarity on the input features (line 28-30):

“A filter-based features selection method was used to identify the importance of different features [heart rate (HR), time and frequency-domain HRV parameters and salivary cortisol] for stress assessment and reduce the computational complexity.”

We have also added the following sentence to the introduction to justify the rationale for the machine learning approach (line 103-106):

“Statistical methods can measure stress according to one or two physiological features at a time. However, machine learning approaches can assess stress objectively by fusing multiple physiological features, and even multiple modalities, thereby increasing efficiency and accuracy.”

Comment #2: I’m also not clear on the importance of this research – could the objective be a little clearer in terms of application / background

Response: We have clarified the objective statement and its application to read as follows (line 23-25):

“To develop a machine learning-based algorithm utilizing heart rate variability and salivary cortisol to detect the presence of acute stress among pregnant women that may be applied to future clinical research”

Introduction

Comment #3: Some of the language is a little informal e.g., ‘wear and tear’, ‘people nowadays’ etc.

Response: We have revised these sentences to read as follows (lines 66-67 and 69-70):

“...the growing mental burden from societal expectations and daily workload in the present day contributes to pervasive, chronic stress”

“Prolonged stress is usually associated with depletion of overall health…”

Comment #4: I’m not sure what a stress pulse survey is. Is this scientific?

Response: Thank you for the comment. The stress pulse survey was conducted by the ComPsych company. ComPsych is the pioneer and the world's largest provider of employee assistance programs, servicing more than 60,000 organizations and 130 million individuals throughout the U.S. and 190 countries. Therefore, the survey is a robust reflection of the global working population. We have removed the term “Stress Pulse” from the manuscript text and replaced it with the word “comprehensive” to avoid any confusion among readers (line 67-69).

Comment #5: The stressors described on line 76 seem to have been chosen at random. Surely racial discrimination is not pregnancy specific? Also what is the difference between perceived stress and stress?

Response: Thank you for highlighting this valid point. We have rewritten the sentence to emphasize the additional stressors unique to pregnancy, as follows (line 75-78):

“Pregnant women may encounter various stressors that exceed those of everyday life for non-pregnant people including worries about the wellbeing of themselves and their developing baby, concerns about delivery, and financial or other socioeconomic stressors associated with a growing family.”

Comment #6: I don’t buy the argument that stress must be assessed during pregnancy. Why is it important to measure physiological stress rather than via self-report? Surely the practicality of self-report in a medical context outweighs the drawbacks, and is it likely that cortisol and/or HRV will be measured in a clinical context? This section could be stronger.

Response: We appreciate the reviewer’s comment on this point. We have added the following sentences to the Introduction to justify the importance and applicability of objective stress measures in the context of pregnancy (line 93-100):

“In standard clinical practice with pregnant women, stress assessment surveys are not typically used due to time constraints and the various limitations of these methods as noted above. As research on objective stress characterization using real time, non-invasive technology evolves, there is potential for these techniques to be translated to clinical settings so that clinicians and/or their pregnant patients may benefit from momentary feedback on patients’ physiological state. In turn, this could facilitate implementation of simple, effective stress management techniques that could reduce the adverse effects of stress on health outcomes for mother and baby.”

Methods

Comment #7: Why were participants excluded if they were not of Hispanic ethnicity? This is a limitation in terms of generalizability

Response: We agree with the reviewer that the ethnicity criteria is a limitation on generalizability for this study, and this has been acknowledged in the limitations section of the discussion. The decision to restrict to Hispanic ethnicity was based on the parent study design and outcome measures. We have added the following sentences to the methods section to provide further clarity in this regard (line 162-166):

“The parent study restricted eligibility to women of Hispanic ethnicity in order to have a homogenous cohort of participants known to experience higher levels of socio-cultural stressors. Homogeneity among participants is important to limit inter-individual variability that could influence metabolic biomarkers that are primary outcome measures for the parent study.”

Results

Comment #8: Figure 2 – 4 are very repetitive. It may be more useful to present data on the whole rather than for individual participants.

Response: We appreciate this comment and while we understand that figures 2-4 might appear to present repetitive data, we wish to clarify the important differences and value that each figure holds for the reader. We have updated the text in this part of the results as follows (line 299-308):

“The comparison results can be summarized into three distinct types, representing inter-individual variability in the physiological response to acute stress exposure. Seventeen participants display an obvious difference in HR, HRV parameters, and cortisol between the stress and non-stress conditions. One example of this type of response is shown in Figure 2. Eight participants do not display a clear difference in HR and HRV parameters but do display a notable difference in cortisol, as shown in Figure 3. In the remaining 4 participants, no obvious difference in any of the parameters including cortisol could be detected. Figure 4 presents an example of this pattern of physiological non-responsiveness to the stress exposure. This variability in stress reactivity creates challenges for standard statistical techniques to detect the presence of stress.”

By displaying these three figures of distinct patterns in stress reactivity, we demonstrate to the readers that individuals tend to differ significantly in response to stress and these varying reactions have a significant impact on our algorithm’s stress assessment accuracy.

Comment #9: There are a number of statements about variables being ‘higher’ or ‘lower’ under certain conditions. On what basis are these judgements made? It does not seem that any statistical or numerical data is presented.

Response: We understand that this comment refers to the results text that relates to Figure 5. These descriptive comparisons are based on visual inspection of differences in AUC values for each parameter, as distinct from the statistical testing of differences that is described in the following paragraph with Table 2. We have updated the manuscript accordingly to ensure that the reader is clear where we refer to higher or lower measures on visual inspection versus the subsequent statistical testing by paired t-tests (line 326-328):

“In summary, for most participants, average values for HR, SDNN, HF, LF/HF ratio, and cortisol tend to be higher under acute stress, while other time- and frequency-domain parameters tend to be lower compared to the non-stress condition upon visual inspection.”

Comment #10: Information missing for t test data (e.g., t value, df)

Response: These values have been added to Table 2 (page 18)

Discussion

Comment #11: Is it possible that some of the participants simply didn’t find the test stressful? This may explain why HR fluctuations etc were not seen

Response: We agree with the reviewer that it is likely that some participants did not find the TSST stressful and this is reflected in the visualization of HRV and cortisol features displayed in figure 4. In the updated cortisol AUC data presented in Table 2 (values normalized for time of awakening), we also do not see a statistically significant difference in total cortisol output on average between the stress and non-stress conditions. The value in the machine learning approach is that it makes use of this variation in stress responses across participants to train the model to better detect true stress exposure with greater accuracy than standard statistical methods. We have added the following to the limitations section to acknowledge this limitation but also the value it provides for this analysis (line 493-498):

“We also acknowledge that the use of a laboratory-based standardized stress protocol does not necessarily reflect real-world stressful events, and indeed not all participants in this study demonstrated a physiological stress response to the task. However, the TSST is the most widely used, validated protocol for stress research in laboratory settings and the machine learning approach helps to overcome the issue of non-responders to the stressor by utilizing those data to train the model with greater accuracy to detect the presence of stress.”

Comment #12: Given the reliability of cortisol as a measure of stress as discussed, I’m still unclear on the aim of this study. Why would we need to use machine learning to identify stress using cortisol outcomes? Why not just use the cortisol outcomes themselves?

Response: We appreciate this comment. On this resubmission, we wish to highlight that our revised data better supports the rationale to use HRV data in addition to cortisol for assessing the presence of stress. On recommendation of reviewer 2 (comment #4), we have generated normalized cortisol values that account for the influence of time of awakening at each cortisol measurement point. We re-computed the cortisol AUC using these updated values and note that on the paired t-test, we no longer see a statistically significant difference in cortisol AUC between the stress and non-stress visits (Table 2). Thus, we cannot say definitively that cortisol alone is sufficient to discriminate a stressful exposure, even in the controlled lab-based setting utilized in our study. We have further expanded on this point in the discussion (lines 386-390):

“In comparison, a noticeable visual difference in cortisol values across the two visits was observed for 25 out of 29 participants, although on average, the difference in cortisol AUC was not statistically significant between the stress versus non-stress visits. This points to the potential unreliability of using standard statistical methods and salivary cortisol alone as an objective measure of acute stress.”

Further, cortisol is not always convenient to measure in free-living subjects due to the need to collect saliva samples, a mildly invasive protocol, and to account for diurnal variations. These challenges, especially as they relate to pregnancy, have been outlined in lines 117-120. Conversely, HRV is increasingly feasible to measure both in controlled and free-living settings with non-invasive wearable technology. We have added a sentence to the introduction explaining this point (line 128-130):

“HRV can be easily measured using non-invasive wearable technology, making it an attractive physiological measure in both lab-based and free-living research settings. However, little is known about the reliability of HRV parameters as a measure of stress and stress reactivity in pregnancy, or how HRV performs compared to cortisol levels to characterize stress.”

While HRV may be a preferred measure in free-living research studies over more invasive and complicated saliva collections for cortisol, our study demonstrates that the sensitivity for detecting stress using HRV in the absence of cortisol requires further work. HRV is a continuous measure and helps to complement cortisol which has temporal resolution. In this way, HRV helps to fill the gaps between cortisol measurements and as machine learning research develops in this field, HRV could be used to help interpolate cortisol responses as an indicator of stress. This point is explained in the original version of the discussion (lines 471-477):

“Further, cortisol cannot be easily measured continuously but only at specified time points of sample collections. Thus, it would benefit future research in the field of stress to computationally predict an individual’s pattern of cortisol response to stress using other physiological bio signals measured continuously by wearable sensors, such as HR, HRV parameters, GSR, and respiration rate. This could help reduce the need to measure salivary cortisol in future research, thereby increasing efficiency and reducing the burden for participants and researchers.”

Comment #13: Unclear what is meant by ‘discrimination stress’ line 403. Was this in the study referred to?

The referenced study measured maternal stress related to experiences of racial discrimination. We have updated the sentence for clarity as follows (line 440-442):

“They found that higher self-reported levels of discrimination stress related to experiences of discrimination among participants were associated with lower levels of cortisol reactivity and higher levels of RMSSD following the TSST”.

Comment #14: Line 445 – was baseline stress measured in the chosen population? If not I’m not sure this argument is very strong.

Response: Baseline perceived stress levels were assessed in this cohort at the beginning of the study. A statement of this measure has been added to the methods under Data Collection (line 173-175):

“On visit 1, usual stress levels over the past month were assessed by the Perceived Stress Scale.”

We have reported the average stress scores in the beginning of the results section (line 294-296)

“The mean PSS score was 13.4±3.7 out of a potential range of 0-40. PSS values from 14-26 indicate moderate stress and values from 27-40 indicate high stress levels. Thus, on average, this population of pregnant women had borderline moderate levels of perceived stress in their daily lives.”

We have also expanded on the relevance of this measure in the discussion in the Strengths and Limitations section (line 485-487):

“While our cohort perceived borderline moderate levels of stress over the past month, it is unknown whether such baseline stressors contribute to higher or lower stress responses to the controlled stress task compared to non-Hispanic women.”

REVIEWER #2:

Comment #1: Technical soundness:

The medical results of this experiment are limited to pregnant women of Hispanic origin who were overweight before pregnancy. There seem to be minor flaws in the collection of data, as described below. The attempt to build an objective stress assessment model is interesting and commendable, with an acceptable accuracy (average 77%), although visual assessment of graphs classified 86% of cases correctly. Medically, it is more interesting to objectively detect and evaluate chronic stress exposure rather than acute stress.

Response: We agree with the reviewer regarding the importance of chronic stress evaluation for translational science purposes. However, this paper provides groundwork for future studies to consider whether HRV in free-living settings alone is adequate to detect the presence of stress or if the combination of cortisol and HRV measures may be a preferred combination to increase accuracy. Whether this combination is also applicable to chronic stress assessment remains to be determined, and we have added a sentence at the end of the conclusion paragraph to highlight this gap for future studies in the field (line 530-531):

“The potential translation of this work to the assessment of chronic stress in humans requires further research.”

Comment #2: This is a secondary analysis of a cross-over study which primarily intended to study the effect of psychological stress on postprandial metabolic response to a standardized meal. The metabolic data are not accounted for in the paper. Supposedly they are subject for a different publication.

Strictly speaking, therefore, the intervention in this experiment is primarily the ingestion of a standardized meal (which normally triggers the parasympathetic autonomic system) and, secondly, a stress test task (which activates the sympathetic autonomic system). The superimposed triggers of the autonomic nervous system of the experiment complicate the interpretation of data and might explain some of the heterogeneity noted. The authors might elaborate on the effect of the dual intervention in their text.

Response: We thank the reviewer for this valuable comment. The parent study required meal ingestion to address the primary research question regarding the metabolic effects of the interplay of nutrition and stress. We acknowledge that the meal ingestion may limit interpretation of results in the present paper as it is unknown from our data how the physiological stress measurements may have differed if participants remained in the fasting state for the duration of each visit. However, given that all participants consumed the same meal at approximately the same time on each visit, we would assume that any potential stimulation of the parasympathetic response is similar across subjects. Thus, the difference in the physiological stress response between visit 1 and visit 2 is standardized in this cohort, allowing the machine learning algorithm to systematically detect this difference in physiological measures. We have added a discussion around this point in the Strengths and Limitations section (line 498-503):

“Introduction of meal ingestion as part of the parent study could also be considered a limitation for the present analysis, as it is unknown to what degree the act of consuming a meal may have moderated the stress response. However, as the meal type and time of ingestion was standardized across all subjects and occurred on both visits, we assume that any potential influence of the meal on the physiological stress response is also standardized across participants.”

Comment #3: The experiments in the study seem to have been conducted rigorously, except for the fact that visit 2 (TSST visit), other than introducing the intended psychological stress, inadvertently added physical movements (the women had to walk to a separate room) which might have increased heart rate and therefore influenced also the HRV parameters. The authors comment appropriately on this in the discussion section (line 368).

Response: We agree that this added movement during visit 2 could have influenced the HR and HRV features and have acknowledged this in the discussion, as noted by the reviewer.

Comment #4: Another concern, which is not commented on, is that the time of awakening of the participants are missing and might have been different on the two occasions. Time of awakening is crucial to the assessment of cortisol levels, since the time from the morning peak values of cortisol would have influenced the baseline cortisol values, and might not be comparable between the two occasions (although the participants arrived at the research centre at about the same time). If most of them rose later in the morning on visit 2, for example, they would indeed display higher cortisol levels from start. It would therefore probably have been more correct to analyse changes from baseline of cortisol rather than absolute cortisol levels and AUC. The level of stress in the morning in the participants’ everyday home environment, before arriving at the research centre, are not taken into account and might also be of importance for baseline cortisol levels

Response: We agree with the reviewer that time of awakening and earlier morning stressors could have influenced cortisol levels during the study visits and we regret this oversight in the original submission. We did ask participants to report their time of awakening on the morning of each study visit. In this revision, we have computed the time interval in minutes from time of awakening until time of arrival at each visit and regressed each participant’s cortisol values against this time interval variable at each visit to generate standardized cortisol values (i.e., adjusted for time of awakening). We then repeated the machine learning models and statistical analysis and updated the results accordingly in the manuscript. We note that the paired t-test for difference in cortisol AUC between visits is no longer significant using the standardized cortisol values (Table 2), however, the outcome of the machine learning model remains virtually unchanged (Table 3).

We have added a sentence in the methods to state that time of participant awakening was recorded (line 173-174):

“Participants were asked to report their time of awakening on the morning of each visit.”

In the methods section, we have added a description for how the cortisol values were adjusted (line 213-216):

“To account for inter-individual variation in cortisol values influenced by time of awakening on the mornings of the study visits, we used normalized cortisol variables, adjusted for time (minutes) interval from awakening until time of arrival at the visit, as the input for statistical analysis and the assessment model.”

Unfortunately, we are unable to adjust the cortisol values for any potential stressful events that occurred before arriving to the visits, and have acknowledged this as a limitation in the discussion (line 503-307):

“While we adjusted cortisol values for time of awakening at each visit, it is also possible that stressful experiences occurred for some participants before arriving to their study visits, which could have elevated their baseline cortisol values. We were unable to account for these potential sources of external stress in the data.”

Comment #5: A technical weakness, which the authors already point out in the text (line 454), is that the Actiheart registration did not provide a raw ECG-signal, which risks to make the HRV data less accurate.

Line 202: Based on what criteria were Actiheart artifacts removed (since no visual inspection of ECG was possible)?

Response: We wish to clarify that the abnormal IBI and HRV values generated by motion artifacts were removed before proceeding with the analysis. We did not have access to the raw ECG signal. The Actiheart only removed part of the motion artifacts but we found that some IBI data was still affected by noise. Therefore, we removed the abnormal IBI and HRV variables according to the IBI removal criteria introduced in a previous HRV validation study. The methodology has been updated and the corresponding study cited (line 224-227):

“The abnormal IBI and HRV values generated by motion artifacts were removed before proceeding with the analysis according to the removal criteria described in another study utilizing HRV measures []. The removal criteria are based on the normal range of HR and IBI values.”

Comment #6: Statistics

The time t used for calculation of the AUC seems to be the same for cortisol and for HRV-parameters (namely 120 minutes). This probably suits cortisol dynamics which are quite slow, but does not match HRV dynamics, since parasympathetic withdrawal and reinstatement are almost immediate in effect, and the researchers will therefore in the same lapse of time capture instances of stress (with lower HRV), but possibly also a period of relaxation/ relief after stress, which might translate into higher HRV, thus leveling out the AUC on visit 2 (particularly of the parasympathetic indices). This might explain why the researchers found no statistical difference in HRV parameters when comparing AUCs. It probably does not affect the objective stress assessment model, however.

Response: Thank you for this important insight. We acknowledge that the ideal approach would be to compute HRV reactivity scores (i.e. change from baseline to peak stress moment at ~45 mins) for the various HRV parameters, and compare these reactivity scores between visits in the paired t-tests. Unfortunately there was very limited baseline HRV data collection performed before starting the stress/non-stress task in this study (often less than 5 minutes) and for some subjects, several of these baseline data points had to be removed due to noise. While we acknowledge this as a limitation, we agree with the reviewer that the machine learning approach overcomes this challenge by using all data available from the ECG rather than relying on suboptimal summary measures. We have added this limitation to the discussion section (line 511-515):

“Most of these gaps occurred at the very beginning of data collection, in the 5 minutes before participants underwent the stress or non-stress task period. This resulted in inadequate baseline HRV measurements from which to compute reactivity scores for each parameter, which may have been a more robust summary measure than the AUC values on which we relied for the statistical analysis.”

Comment #7: This reviewer has no deeper insight in programming or in machine learning-based algorithms and can therefore only comment on the quality of medical data used for feature selection. To this reviewer’s knowledge all HRV parameters are influenced by HR, are highly interdependent and are not normally distributed. HR and AVNN are inherently redundant - it is surprising the feature selection allows both, considering it is constructed to avoid overfitting?

Response: We thank the reviewer for this comment. During the feature selection process, the algorithm tries all different combinations of features and the contribution to the decrease in the impurity of each feature is computed. The importance of each feature is ranked according to this contribution. As we can see from the feature selection results, a combination of 5 features including AVNN, HR, SDNN, LF, and LF/HF are selected when cortisol is not added. AVNN and HR are the first two features which means they each make a bigger contribution to discriminate stress vs. non-stress than the remaining three features. HR represents the frequency of heart beats and AVNN represents the average length of heartbeat intervals. Although AVNN and HR seem highly related from a statistical perspective, they represent different HRV features and thus, both add value to the model. However, after cortisol is added, HR is no longer selected as an important feature but AVNN remains. We have added an explanation about these features to the results section for further clarity (line 346-349):

“HR represents the frequency of heart beats and AVNN represents the average length of heartbeat intervals. Although AVNN and HR appear highly related from a statistical perspective, they represent distinct HRV features and thus, both contribute value to the model”.

Comment #8: All data supporting the conclusions seem to be fully available in the Supporting information file. (This is a secondary analysis of a cross-over study which primarily intended to study the effect of psychological stress on postprandial metabolic response to a standardized meal. The metabolic data are not accounted for in the paper. Supposedly they are subject for a different publication.)

Response: In this resubmission, we have uploaded our final dataset of ECG and cortisol data from all participants at each visit to the online data repository, DRYAD. The dataset in DRYAD will be publicly available upon acceptance of the manuscript.

This work is presented in intelligible standard English.

Comment #9: Line 108: Statement should be backed up by a reference.

Response: A reference has been added to support this sentence (line 120).

Comment #10: Line 172 and 178: TSST reference?

Response: References have been added to support these statements (lines 194 and 198).

Comment #11: Line 216: “,” instead of “.” as in one sentence.

Response: This has been updated (line 242).

Comment #12: Line 453: Actiheart instead of Actiheat.

Response: This has been updated (line 502).

Comment #13: Figure 5: Dots should not be connected by lines, and it should be clarified that unlike Figure 2, 3 and 4, x is no longer time, but represent discrete individuals.

Response: We have updated Figure 5 and now present the data in bar chart format for improved interpretability. Each x-axis is labeled such that each set of bars represents unique participants.

Submitted filename: Responses to Reviewer comments.docx

Click here for additional data file.

26 Aug 2022

Prenatal Stress Assessment using Heart Rate Variability and Salivary Cortisol: A Machine Learning-Based Approach

PONE-D-22-11525R1

Dear Dr. Lindsay,

We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements.

Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication.

An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at http://www.editorialmanager.com/pone/, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to-date. If you have any billing related questions, please contact our Author Billing department directly at authorbilling@plos.org.

If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org.

Kind regards,

Enzo Pasquale Scilingo, Ph.D.

Academic Editor

PLOS ONE

Additional Editor Comments (optional):

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation.

Reviewer #1: All comments have been addressed

**********

2. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #1: Yes

**********

3. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #1: Yes

**********

4. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #1: Yes

**********

5. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #1: Yes

**********

6. Review Comments to the Author

Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)

Reviewer #1: (No Response)

**********

7. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: No

**********

1 Sep 2022

PONE-D-22-11525R1

Prenatal Stress Assessment using Heart Rate Variability and Salivary Cortisol:  A Machine Learning-Based Approach

Dear Dr. Lindsay:

I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department.

If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact onepress@plos.org.

If we can help with anything else, please email us at plosone@plos.org.

Thank you for submitting your work to PLOS ONE and supporting open access.

Kind regards,

PLOS ONE Editorial Office Staff

on behalf of

Professor Enzo Pasquale Scilingo

Academic Editor

PLOS ONE