The salivary cortisol classification based on the heart rate variability

Introduction

The stress system has been defined as having an important role in the survival of an organism. The perception of stress by the brain leads to a series of comprehensive events that are initiated in the brain and body. These events serve as a coping mechanism for stress, thereby enabling the biological system to return to a normal state. The prolonged exposure to stress and persistent tension created in today’s daily life results in decreased recovery opportunities. These circumstances result in persistent stress within the society [1]. Mild acute stress is associated with an enhanced efficiency of various body organs, including the nervous and cardiovascular systems, as well as other organs. Stress can lead to decreased efficiency or increased susceptibility to developing several disorders, such as cardiovascular disease. Stress can also lead to deficiencies in immunity, fertility, metabolism, memory, and analytical and decision-making abilities. These outcomes are associated with an inability to react correctly in different situations [2]. The activation of the sympathetic axis, the result of acute stress and hypothalamus activation, is accompanied by an increase in the heart rate and blood pressure [3]. According to studies, in addition to the concentration of cortisol hormone used as an active stress indicator, electrophysiological changes in bio-signals, including electrocardiography and electroencephalography, are also valid indicators in stress diagnosis. The HRV, also referred to as pulse rate variability, is a measure of the activity of the autonomic nervous system that is associated with psychological stress. The component of the HRV that exhibits low frequency reflects sympathetic activity control and exhibits a positive correlation with mental stress. A decrease in HRV is observed during physical or mental stress, which is indicative of a decrease in parasympathetic activity and an increase in sympathetic activity. This is also a risk factor for arrhythmia and heart attacks [2], 4].

Several studies have reported that the HRV indexes have a good predictive power of 85–92 % of perceived stress. The conscious part of stress that is perceived by the individual is classified in these studies and the unconscious part is neglected [5], 6]. Many studies have reported that the activation of the stress system is not necessarily perceived by a person, or at least not a part of it. It extends to a subconscious level. Certain physiological indicators of stress, such as the level of cortisol concentration and the nonlinear and spectral characteristics of HRV, may undergo modifications subsequent to stress [7], [8], [9]. This suggests that HRV indexes that survey cortisol concentration levels, the major marker of stress system activity, more reliably correspond to a stress response rather than self-reporting emotional state and it has not been noted in As per research conducted on the correlation between stress and cortisol secretion, as well as its impact on electrophysiological biomarkers such as HRV, it has been demonstrated that machine learning can aid in the development of an algorithm for categorizing cortisol. The objective of this investigation is to delineate the stress system of an individual in terms of cortisol concentration levels by means of machine learning.

Materials and methods

MLP [14]

The multilayer perceptron (MLP) is a neural network that links multiple layers in a directed graph, thereby ensuring that the signal path only traverses a single direction through the nodes. This algorithm is a category of feedforward artificial neural networks (Figure 1).

Figure 1:

Feedforward neural network.

Until the outcome calculation is complete, the forward movement is repeated in this network. Multilayer perceptron networks include a set of weights that must be adjusted for neural network training and learning. The training phase encompasses the modification of parameters, or weight, and the model is biased towards error minimization. The training encompasses the process of adjusting parameters, weight, and model biases in order to minimize error. The back propagation algorithm is utilized to adjust the weight and bias rate in relation to error. In order to attain the optimal performance of this algorithm, the dataset shall adjust the hyper-parameters such as learning rate, non-linear activation function, repeat number, and implemented optimization type.

Results

A dataset from the Baqiyatallah University of Medical Sciences was used for the training of models. The mean and SD of variables were showed in Table 2. All datasets were trained and evaluated using a 10-fold cross-validation technique. To evaluate the performance of the trained model, F1, accuracy, and confusion matrix criteria were used.

As mentioned in the method section, three classes of cortisol concentration are the optimal range (class 2): 5–15 ng/mL, and other classes (class 1 and 3) are the non-optimal range: less than 5 ng/mL and more than 15. Table 1 depicts the F1 rate and accuracy of machine learning models that have been trained using the three algorithms, namely SVM, XBG, and MLP. The highest level of precision is attributed to XGB.

Figure 2 depicts the confusion matrixes derived from the evaluation of each of the three models. The SVM algorithm was able to place 98 % of data accurately in class 1, 96 % of data accurately in class 2, and 95 % of data accurately in class 3, on average, in 10 categories. The XBG algorithm was able to place 100 % of data accurately in class 1, 99 % of data accurately in class 2, and 98 % of data accurately in class 3, on average in 10 categories. Furthermore, the MLP algorithm could place 92 % of data accurately in class 1, 95 % in class 2, and 91 % in class 3, on average, in 10 categories.

Figure 2:

Charts show the confusion matrix were derived from the implementation of SVM, XBG, and MLP.

The rates of F1 and accuracy are approximately equal in each one of the three algorithms which shows the appropriate function of the machine learning models created. The confusion matrices indicate that models could predict the samples of classes 1, 2, and 3.

Discussion

The objective of the study was to establish cortisol as a reliable biological indicator of stress, utilizing HRV indices, a topic that had not been extensively explored in the literature. The current study demonstrated that the frequency and non-linear HRV indexes could separate the optimized cortisol level from the non-optimized cortisol level 1 (lower and upper than the normal range) in healthy adult men. The outcomes of this model were acceptable. The model constructed by the XGBoost algorithm demonstrated superior performance compared to MLP and SVM. The results of this study demonstrate that SVM, XBG, and MLP algorithms have equal F1 and accuracy rates and the same standard deviation.

Until now, no study similar to this has been conducted. According to statistical studies, it has been demonstrated that there exists a significant disparity in saliva cortisol levels with HRV indexes in stressful situations such as TSST, as compared to those in resting states [9]. Furthermore, these indexes have demonstrated a significant disparity in patients compared to healthy individuals [7], 9]. The classification studies in the field of stress indicate the presence or absence of stress based on the individuals report of HRV indexes [5], 6]. Stress initiates a variety of modifications in the brain centers and pathways that are documented [15], [16], [17] as well as changes in the autonomic control of the body. Alizadeh et al. conducted an evaluation in 2020 to determine the relative efficacy of band frequencies of brain activity and complexity features in classifying cortisol concentration in optimal and non-optimal states. The findings revealed that EEG features meticulously predicted the cortisol level through the utilization of Artificial Neural Networks, attaining an accuracy of 94.1 % [18].

The stress system does not always correspond to the awareness level perceived by the individual. Although the individual reports no stress in the recovery stage of stress, the cortisol level and some HRV indexes have not returned to normal levels [8]. This finding cautions that synchronizing the activity of the brain stress system with a conscious comprehension of stress and its classification based on biological and electrophysiological indicators such as HR may not yield precise information regarding the precise state of the brain stress system and its impact on biology. In a review of Korean researchers conducted in 2017, 37 studies were included in a meta-analysis study. After conducting a thorough examination of various HRV indicators in both the time domain and frequency, it has been determined that HRV is highly sensitive to changes in the stress-related activity of the autonomic nervous system, including but not limited to changes in the sympathetic nervous system and peripheral nervous system In the majority of studies, the HRV variables underwent modifications in response to the stress resulting from the various methodologies. The most common factor reported concerning changes in HRV variables was the decrease in parasympathetic activity, characterized by PHF decrease and PLF increase. The complexity of the HRV signal is also reduced by measuring the SDs of the Poincare plot of HRV, even after recovery time [7], [8], [9]. HRV is associated with the activity of a flexible network of brain structures that are dynamically organized in response to environmental challenges. As per the findings of brain imaging studies, it has been suggested that the presence of HRV may be associated with a decreased perception of threat, which is attributed to the activity of certain regions of the cerebral cortex, such as the ventromedial prefrontal cortex [19].

In light of the studies, the HRV response to stress in an individual with depression has strongly declined. PHF, PLF, and LF/HF in depressed individuals are opposite to those in healthy ones. Furthermore [20], individuals suffering from post-traumatic stress disorder (PTSD) exhibit a lower relative power in PHF and PLF, as well as in R-R signal intensity and SD, in comparison to individuals who are in good health. This decrease was particularly lower in PHF [21]. The surveying of PTSD studies revealed a total of 784 articles, of which 22 were incorporated into the conclusive analysis. The saliva cortisol level in PTSD patients was significantly lower when compared to the control group (p=0.022, 0.04; 95 % CI 0.53, SMD=0.028) [22]. Depressed individuals have a higher cortisol level than healthy individuals, and depression intensity has a direct relationship with the level of cortisol [23]. Individuals who experience a brain stroke at rest exhibit a higher level of cortisol than those who are healthy, and they do not exhibit an increase in cortisol levels subsequent to stressful circumstances such as TSST. These individuals exhibit a lower level of PLF in comparison to those who are in good health [7], 8].

The practice of self-reporting stress classification based on the HRV indexes has been extensively utilized. A study conducted in 2014 demonstrated that the HRV time and frequency variables could distinguish between the presence or absence of stress, moderate average stress, and high stress, based on the self-report of individuals with a power of 92 % [24]. Another study conducted in 2014 indicated that the prediction of cognitive stress, utilizing facial markers, heart rate, LF/HF ratio, and respirations per minute, could be achieved with a predictive power of 85 % [25]. In a study conducted in 2015, it was demonstrated that utilizing a combination of time, frequency, and non-linear HRV indicators, a radial basis function classification method coupled with an SVM was capable of distinguishing between drivers who were under severe stress and those who were in a normal situation, with a statistical significance of 83 % [26].

Conclusions

Based on the findings of the surveys conducted, it can be inferred that the machine learning model employed in this study, which is based on a collection of HRV linear and non-linear indexes with a high level of prediction, is favored over stress studies due to its high estimation power of 99 % and comprehensive estimation of the primary indicator of the stress system, cortisol concentration levels, as opposed to the self-report system of stress. This study is applicable to artificial intelligence methods used to evaluate the stress system state in unconsciousness state.