In recent years, perinatal maternal care has become an increasingly important part of public health in society (1). Accelerometers have been widely used to ascertain sleep patterns through recording movement of the body throughout the night, which may be usable to monitor perinatal maternal health. Unlike traditional methods such as polysomnography (PSG), accelerometers allow estimation of sleep–wake conditions because the body tends to remain stationary when falling asleep. The motion amplitude during sleep becomes distinctively smaller than that in the waking state (2, 3). While the accuracy of accelerometer-based evaluations of sleep conditions remains limited, and accelerometers have not been listed by the American Academy of Sleep Medicine as one of the four types of devices that can evaluate sleep (4), numerous studies have investigated and summarized the validity of using accelerometers to measuring sleep objectively. Research to date has supported to some extent the validity of using accelerometers and actigraphy devices (e.g., fitbit, Jawbone, MovBand) for measuring sleep outside of the laboratory setting (5, 6). Various studies have examined the accuracy of such devices in patients with sleep problems and found no significant differences in sleep onset latency (SOL), wakefulness after sleep onset, and total sleep time between PSG and actigraphy measurements in subjects with or without sleep problems (7–10). Another study assessing the validity of the MovBand 3 against previously validated medical sleep monitors reported that it provided a valid and reliable assessment of sleep conditions, including the number of awakenings, deep sleep, light sleep, and physical activity (4). Therefore, ample evidence has been presented to confirm the stability and accuracy of such devices.

Regardless of the abovementioned limitations, accelerometers offer a significant advantage in terms of ease of use in daily life. For example, the sleep patterns of perinatal women can be monitored using accelerometers throughout the pregnancy and postpartum periods. Perinatal women tend to have difficulties sleeping. Many studies have reported that pregnant women experience significant sleep disruption and sleep disorders (11–13), and it is generally known that sleep disturbances occur during the third trimester of pregnancy (14–16). Sleep disorders can be related to mental health problems during the perinatal period, including postpartum depression (PPD) (17). Pregnancy is marked by considerable physiological changes and a multitude of symptoms, many of which are likely to disrupt sleep. It is also accompanied by dramatic hormonal changes, which have a significant potential to impact sleep quality (18). Given the impact sleep has on physical and mental well-being, assessing sleep quality throughout pregnancy is crucial. Therefore, to recognize and evaluate the sleep quality of pregnant women and the risk of PPD, it is important to gain a better understanding of the characteristics of sleep–wake conditions during pregnancy.

There are several types of sleep in a sleep cycle: shallow sleep (stages 1 and 2) and deep sleep (stages 3 and 4), which are categorized as non-rapid eye movement (NREM) sleep, and rapid eye movement (REM) sleep (19–21). In the typical sleep process, the sleep cycle, including the wake condition, shallow sleep, deep sleep, and REM, is repeated several times during the night, and sometimes, each cycle is spanned by short-term wake conditions. There are several types of sleep disturbances, including difficulty falling asleep, maintaining sleep, and awakening in the early morning. To develop countermeasures to sleep problems, it is important to consider the sleep cycle and these types of sleep disturbances. Even though accelerometers cannot detect REM sleep precisely, they present the benefit of allowing us to grasp the overall sleep cycle patterns throughout sleep with only limited effort on the part of study participants.

It is noteworthy that the arousal levels in wake conditions before and after sleep can differ. The arousal level in the wake condition before sleep can be related to difficulty falling asleep, and that in the wake condition after sleep can be related to sleep quality. Therefore, it may be worthwhile to differentiate between the wake conditions before and after sleep.

Heart rate can be a useful marker for assessing sleep cycles because it is known to differ between wake and sleep conditions (22). Some investigators have reported correlations between sleep and autonomic nervous system (ANS) activity. Various studies have indicated that the parasympathetic nervous system (PNS) activity is principally influenced by the circadian system, whereas the sympathetic nervous system (SNS) activity is principally influenced by the sleep system (23). It is generally known that the cardiac PNS is activated during NREM sleep (24–26). In contrast, the cardiac PNS is inactivated during REM sleep (27, 28). Heart rate variability (HRV) is the physiological phenomenon of variations in the time interval between heartbeats. HRV can reflect numerous physiological factors and is always used as a measure of ANS activity (29–31). Time domain features include the coefficient of variation R-R interval (CVRR), the standard deviation of all NN intervals (SDNN), the square root of the mean squared differences of successive NN intervals (RMSSD), the number of interval differences of successive RR intervals greater than 50 ms (NN50), and the proportion dividing NN50 by the total number of RR intervals (pNN50). Frequency domain features include low frequency (LF), high frequency (HF), and the ratio of LF to HF (LF/HF). Determining the influence of sleep on HRV is of considerable interest because of the relevance of sleep stages, HRV, and ANS. Many studies have reported a correlation between different sleep conditions and HRV. Some investigators (32–34) have used changes in HRV to analyze and consider sleep quality, sleep disorders, and mental diseases. Prediction of whether someone is asleep or awake based on HRV has also been attempted.

HRV may be related to not only the sleep–wake conditions, but also the physiological conditions underlying sleep disturbances, which are considered to be related to the dysregulation of the SNS and PNS. The perinatal period introduces a myriad of changes, such as sleep disturbances characterized by insomnia symptoms and poor sleep quality, which are highly prevalent during pregnancy and can increase depressive symptomatology and PPD (35). Regarding the correlation between the ANS, HRV, and sleep, prior studies have investigated whether accounting for sleep disturbances may explain some of the heterogeneity in the association between HRV and depression (36–39).

Machine learning algorithms are a widely used technology that have become one of the core technologies of artificial intelligence and data science. Regarding HRV and other types of clinical information, many scientists have also used machine learning to establish prediction models for different sleep conditions. In addition, several studies (40–42) have shown that the k-nearest neighbor (k-NN) can be used as a sleep condition classifier for different sleep conditions. Some studies (40, 43) have indicated that support vector machine (SVM) is an appropriate method for discriminating between the different sleep conditions. Various studies (41, 42, 44) have also reported that artificial neural network (ANN) is a suitable method for distinguishing the different sleep conditions. Other studies (41, 42, 44–48) have also shown that the random forest (RF) could solve the sleep condition recognition problem, which was considered arousal and valence prediction based on physiological signals. Some studies (49, 50) have confirmed that a deep learning algorithm, long short-term memory (LSTM), was an useful method for predicting different sleep conditions. Mendez (51) set HRV as an important feature and carried out a hidden Markov model (HMM) as a classifier. On this basis, the HMM was used to classify the different sleep stages (NREM and REM). However, since the publication of these previous studies, various new machine learning algorithms have been developed, including RF, gradient boosting trees, stochastic gradient descents, extreme gradient boosting, and ANNs. The application of these algorithms may be beneficial, allowing more efficient prediction of sleep conditions based on HRV.

This study aimed to indicate the extent to which the sleep–wake condition and differentiation between the wake conditions before and after sleep could be predicted in pregnant women based on heart rate-relevant information as indicators of ANS functioning using various algorithms, considering the difference in the arousal level during the wake conditions before and after sleep. Thirteen methods [k-NN, SVM, logistic regression (LR), RF, Naïve Bayes (NB), decision tree (DT), gradient boosting tree (GBT), stochastic gradient descent (SGD), extreme gradient boosting (XGBoost), ANN, convolutional neural network (CNN), LSTM, and gated recurrent unit (GRU)] were applied to predict three types of sleep–wake conditions in pregnant women (“wake,” “shallow sleep,” and “deep sleep”) followed by prediction of four types of sleep–wake conditions (“wake condition before sleep,” “wake condition after sleep,” “shallow sleep,” and “deep sleep”), and the accuracy of each prediction model was subsequently evaluated.

In this study, significant differences among sleep–wake conditions were found for all nine HRV indicators [CVRR, SDNN, RMSSD, NN50, pNN50, LF, HF, LF/HF, and LF/(LF + HF)]. Post hoc analysis indicated that CVRR, SDNN, NN50, pNN50, LF, LF/HF, and LF/(LF + HF) were significantly lower and RMSSD and HF were significantly higher during deep than during shallow sleep. These findings support previous studies that reported significant differences in CVRR, SDNN, RMSSD, LF, HF, and LF/HF among different sleep–wake conditions (120, 121). It is noteworthy that many previous studies, with several exceptions, ignored NN50 and pNN50. Furthermore, several studies investigating NN50 and pNN50 reported finding no significant correlations between these two HRV indicators and sleep–wake conditions (122–124). Contrary to these previous findings, the present study found that NN50 and pNN50 significantly differed among sleep–wake conditions. The major difference between the present and previous studies was the target participants: pregnant women in the present study, compared with non-pregnant women and men in previous studies.

NN50 is the number of interval differences of successive RR intervals greater than 50 ms, and pNN50 is the proportion dividing NN50 by the total number of RR intervals, both of which reflect vagal tone (125–128). Previous studies have revealed that vagal tone is altered during pregnancy. In general, vagal tone and reactivity are reported to decrease as pregnancy progresses (129–131). While CVRR, SDNN, LF, LF/HF, and LF/(LF + HF) are useful SNS markers for differentiating sleep–wake conditions, NN50 and pNN50 may not be distinguishable between different sleep–wake conditions among subjects without pregnancy. As pregnancy conditions alter the vagal tone, NN50 and pNN50 may constitute useful factors that vary according to and allow differentiation between sleep–wake conditions (significantly lower NN50 and pNN50 during wake conditions compared with sleep conditions).

In the present study, we investigated whether there were differences in HRV indicators between the wake conditions before and after sleep as an unprecedented trial. Interestingly, the same HRV indicators [CVRR, SDNN, NN50, pNN50, LF, LF/HF, and LF/(LF + HF)] showed significantly lower values during the wake condition before sleep than during that after sleep. In addition, HF was significantly higher during the wake condition before sleep than during that after sleep. These results suggest that there may be apparent differences in arousal levels and ANS conditions between the wake conditions before and after sleep, similar to the differences among the sleep–wake conditions.

Unlike previous studies, which applied a limited number of different algorithms to predict sleep–wake conditions based on HRV indicators, the present study conducted comprehensive evaluations of widely used machine and deep learning algorithms. Among the 13 Python-based machine and deep learning algorithms, 12, except NB, provided high AUCs (0.82–0.88) and accuracy (0.78–0.81) for predicting sleep–wake conditions based on HRV indicators. As for the predictions of four conditions (deep sleep, shallow sleep, wake condition before sleep, and wake condition after sleep), different algorithms showed a wider range of AUCs and accuracy than did those in the prediction of three conditions (deep sleep, shallow sleep, and wake conditions). GRU showed the highest AUC and accuracy (0.86 and 0.79), respectively, followed by LSTM (0.78 and 0.73), RF (0.76 and 0.71), GBT (0.76 and 0.71), XGBoost (0.76 and 0.71), CNN (0.75 and 0.71), k-NN (0.75 and 0.67), ANN (0.74 and 0.70), DT (0.73 and 0.69), SVM (0.73 and 0.67), SGD (0.72 and 0.70), LR (0.72 and 0.67), and NB (0.61 and 0.42) in predicting the four conditions. These results suggested that the differences in ANS conditions among the three types of sleep–wake conditions were clear enough to be equally predictable by most algorithms.

In contrast, the differences in ANS conditions among the four types of sleep–wake conditions were not as evident as those among the three types of sleep–wake conditions. Therefore, a specific algorithm was needed to obtain high AUCs and accuracy for predicting the four conditions. Deep learning (GRU and LSTM) and ensemble learning methods in machine learning (RF, GBT, and XGBoost) may be suitable for differentiating between the wake conditions before and after sleep in predictions for the four conditions.

In previous studies conducted to predict sleep–wake conditions based on HRV indicators, one or several algorithms were tested in each study, as summarized in Table 3. In contrast, all previous studies testing deep learning algorithms used only one algorithm. Among the studies testing multiple algorithms, those carried out to test machine learning algorithms, including RF, consistently indicated that RF showed the highest prediction accuracy. The studies testing machine learning algorithms, including NB, indicated that NB showed lower prediction accuracy compared with the others. It is reasonable that the accuracy of NB (132–134) would be the lowest among these methods because it is a simple algorithm that has the advantage of predicting a binary classification. Therefore, NB is not suitable for predicting multiple classifications, as conducted in the present research for prediction of three types of sleep–wake conditions using nine HRV indicators.

GRU showed the highest AUC and accuracy of 0.86 and 0.79, respectively, in predicting the four conditions. As mentioned in the Methods section, we used GRU with a multi-head attention prediction model instead of GRU only. GRU with a multi-head attention prediction model can effectively improve the task processing effect. It calculates the probability weight of different data, improves the quality of hidden layer feature extraction, and reduces the problem of information loss in feature extraction. A multi-head attention mechanism is widely used in various sequence prediction tasks. A number of previous studies (135–137) reported that a model with an attention mechanism achieved high accuracy. We conducted a preliminary test using GRU without a multi-head attention mechanism, which resulted in a lower AUC and accuracy compared with GRU with a multi-head attention mechanism and LSTM (data not shown). GRU can automatically classify different sleep–wake conditions based on HRV indicators, and thus outperforms conventional methods because it considers the complex and cyclic characteristics of sleep and wakefulness.

RF-based evaluation of important features in the prediction indicated that pNN50 was the most important feature for predicting sleep–wake conditions and differentiating between the wake conditions before and after sleep, followed in order by RMSSD, NN50, SDNN, CVRR, HF, and LF. pNN50, RMSSD, NN50, SDNN, CVRR, HF, and LF may be sensitive SNS-based biomarkers for predicting sleep–wake conditions and differentiating between the wake conditions before and after sleep. Previous studies have indicated that RMSSD, SDNN (138), CVRR (139), HF (122, 139–141), and LF (139, 140) are useful features for predicting sleep stages, which supports our results. Although some previous studies have examined pNN50 and NN50, these features did not help differentiate between sleep stages (123, 124, 142). The present study noted that pNN50 and NN50 were important features for predicting sleep stages among pregnant women, suggesting that these features may be related to the ANS characteristics during pregnancy conditions. pNN50 and NN50 were identified as important features for predicting sleep–wake conditions and allow differentiation between the wake conditions before and after sleep among pregnant women probably because these features differed significantly among the sleep–wake conditions and the wake conditions before and after sleep.

The three cross-validation methods with automatic hyper-parameter tuning, including GridSearchCV, RandomizedSearchCV, and Bayesian optimization searches (117, 143), were also tested. Among these, RandomizedSearchCV achieved the highest accuracy with the shortest time consumption. GridSearchCV can ensure the accuracy of parameters within the specified parameter range by traversing all possible parameter combinations. However, this is very time-consuming and provides low accuracy in the case of large datasets and multiple parameters. Regarding hyper-parameter optimization, RandomizedSearchCV search is more effective than GridSearchCV. Some previous studies have indicated that Bayesian optimization searching provides the best accuracy (117, 144), whereas others (145) have indicated that the limitations of Bayesian optimization sometimes makes its search effectiveness unstable and not significantly better than RandomizedSearchCV. Compared with RandomizedSearchCV, the drawback of Bayesian optimization is the greater consumption of computational resources, which may result in its taking longer to escape the local optimum, and its deployment in a distributed system (146). Memory consumption, training time, power consumption, and parallelism are essential for deep learning, while the feature of borrowing ideas from previous results prevents the Bayesian method from direct parallelization. Although many recent developments have solved this problem, it is still not as natural as RandomizedSearchCV, which is easy to combine with early stopping strategies. Such a combination could be expected to vastly improve the efficiency of narrowing down the search space (145). Regarding the performance of the model, while KCV and LOOCV showed almost equivalent performance, KCV was selected because LOOCV was time-consuming, taking longer to fit a dataset compared with KCV.

A verification study with alternative usages of JMP and WEKA, as well as Welch’s method on Python to extract the HRV based on RR intervals, assured the validity of the present findings, which showed replicated prediction accuracies with slight differences due to the variability in parameter regulations.

The major finding of the present study in the clinical context is that vagal tone appears to be an important factor for differentiating between sleep–wake conditions, specifically among pregnant women. The findings also suggest that pregnancy conditions alter vagus nerve conditions in different ways between sleep and wake conditions. Another major finding in the clinical context is that HRV indicators can be useful for differentiating between not only sleep–wake conditions, but also the wake conditions before and after sleep. The wake condition before sleep may reflect a drowsy state, whereas that after sleep may reflect an alert state. By observing HRV indicators, drowsiness and readiness to falling asleep can be objectively evaluated. This information may be useful in interventions to improve sleep health in pregnant women.

This research has several limitations. First, the sample size was relatively small (N = 154). In the future, the accuracy of the model should be verified with a greater quantity of data, and the most suitable algorithms should be selected. Second, sleep stages were defined based on the information collected through body motion sensors without recording EEGs; therefore, there was a limitation in terms of the accuracy of sleep stage determination. Third, REM sleep was not taken into account in the present study because it is not detectable by an accelerometer. Fourth, the types of machine learning and deep learning were limited. Transformer was not applicable because of the small number of sleep condition observations. After collecting more data, an advanced method such as transformer may be applicable in order to optimize prediction models and parameters in the future. Fifth, concerning wake conditions, among the two types of wake conditions, we arbitrarily extracted the 30 min of wakefulness before sleep and that immediately after waking up in the morning (i.e., the wake condition after sleep). In the future, different durations of observations, such as 15, 45, or 60 min (both before sleep and immediately after waking) could alternatively be applied to more completely investigate the nature of wake conditions. Finally, as this study focused on the heart rate of pregnant women, the possibility that the fetal heart rate might have interfered with the observed maternal heart rate remains. However, unless an ultrasound transducer is placed correctly over the fetal heart, the fetal heart rate signal cannot be reliably acquired. The disposable electrode placed on the maternal heart in this study rarely detects the fetal heart rate. Therefore, the observed heart rate should accurately reflect the maternal heart rate without considerable interference from the fetal heart rate.

This research tested 10 machine and three deep learning algorithms to predict sleep–wake conditions based on HRV indicators. Most of the tested algorithms except NB could provide a suitably accurate prediction of three types of sleep–wake conditions (wake, shallow sleep, and deep sleep). GRU was the most accurate method for predicting four sleep conditions (including differentiation between two wake conditions): the wake conditions before and after sleep, shallow sleep, and deep sleep. Moreover, pNN50, RMSSD, NN50, SDNN, CVRR, HF, and LF were important features for predicting sleep–wake conditions and differentiating between the wake conditions before and after sleep.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving human participants were reviewed and approved by the ethics committee of Graduate School of Medicine, Tohoku University. The patients/participants provided their written informed consent to participate in this study.

XL took major role in the data analysis. CO, ZY, YT, KI, NS, NH, and HT also contributed to the data analysis. CO, NW, TS, TaN, HU, KM, MI, TO, FN, NF, JS, ShK, MY, NY, and HT contributed to the acquisition of data. CO, ToN, TT, SO, and GT contributed to the data management. XL, CO, KI, NS, NK, SaK, RK, YH, MH, YK, JS, TH, MS, SF, MN, ShK, NH, and HT contributed to the interpretation of the data. XL and HT drafted the manuscript. KI, NS, NK, SaK, SF, NH, and HT critically revised the manuscript for important scientific content. XL, CO, NW, FN, ShK, and HT made substantial contributions to the conception and design of the study. All authors contributed to the article and approved the submitted version.

This work was supported by a grant from the Strategic Research Program for Brain Sciences from the Japan Agency for Medical Research and Development (AMED) under Grant No. JP16dm0107099, the Tohoku Medical Megabank Project from the Ministry of Education, Culture, Sports, Science and Technology (MEXT) of Japan and AMED under Grant Nos. JP17km0105001, JP21tm0124005, and JP22tm0124005, and the Tohoku University Advanced Research Centre for Innovations in Next-Generation Medicine.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.