Machine learning approaches to the determinants of women’s vasomotor symptoms using general hospital data

Background To analyze the determinants of women’s vasomotor symptoms (VMS) using machine learning. Methods Data came from Korea University Anam Hospital in Seoul, Korea, with 3298 women, aged 40–80 years, who attended their general health check from January 2010 to December 2012. Five machine learning methods were applied and compared for the prediction of VMS, measured by a Menopause Rating Scale. Variable importance, the effect of a variable on model performance, was used for identifying major determinants of VMS.


Background
Vasomotor symptoms, referring to hot ashes and sweating, are major symptoms of peri-menopausal and post-menopausal women and a main cause of their hospital visit [1]. Once it was considered to be a temporary symptom but now it is reported to exert a lasting effect on their quality of life [2]. A decline in hormone concentration is considered to be one of its risk factors, disrupting brain neurotransmission and hypothalamic thermoregulation [3]. Other candidates for its risk factors include age, body mass index, race, smoking and depressive symptoms [4][5][6]. However, these results were not consistent and more research is to be done in this direction. Based on increasing evidence, moreover, vasomotor symptoms are expected to associate with major chronic diseases such as cognitive impairment, cardiovascular disease, diabetes mellitus and sleep disorder [7][8][9]. More research is needed to analyze women's vasomotor symptoms and their major determinants.
This study is the rst attempt to demonstrate that machine learning provides an invaluable decision support system to predict women's vasomotor symptoms and analyze their determinants. Existing literature did not address which variables are more important for the prediction of vasomotor symptoms.
This was because previous studies were based on an unrealistic assumption of ceteris paribus, "all the other variables staying constant". Machine learning methods such as the random forest are free from such an unrealistic assumption of "all the other variables staying constant" [10,11]. As addressed in the section of Results below, the performance of the random forest is much better and much more stable than that of logistic regression, a popular statistical approach in conventional studies. Moreover, data in this study are larger than those in previous studies − 3298 women and 104 independent variables on the thyroid function, the lipid pro le, the liver function, in ammation markers, insulin resistance, the monocyte, cancer antigens, the lung function and other information.

Participants and Variables
Data came from Korea University Anam Hospital in Seoul, Korea, with 3298 women, aged 40-80 years, who attended their health check from January 2010 to December 2012. The dependent variable was vasomotor symptoms measured by a Menopause Rating Scale item on hot ashes and sweating (MRS-1). In the early 1990s the MRS was introduced as a standardized scale to measure women's aging symptoms and health-related quality of life. The MRS consists of 11 items such as vasomotor symptoms (MRS-1), heart discomfort (MRS-2), sleep problems (MRS-3), and each item has a scoring system from 0 (no symptoms) to 4 (very severe symptoms) [12] And 104 independent variables were included in this study such as age (years), menopause age (years), thyroid stimulating hormone (µU/mL), monocyte (%) and triglyceride (mg/dL) (See Tables 1 and 2 in Results for detail).

Analysis
Five machine learning approaches were used for predicting vasomotor symptoms, the dependent variable of this study: linear regression, random forest and arti cial neural networks (ANNs) with one, two and three hidden layers [13]. Based on linear regression, a linear line is found in a way that it minimizes, among an in nite number of linear lines, the sum of the squares of errors (Errors are gaps between actual and predicted values of the dependent variable). A decision tree consists of (1) internal nodes (each meaning a test on an independent variable), (2) branches (each denoting an outcome of the test) and (3) terminal nodes (each representing a value of the dependent variable). A random forest creates many training sets, trains many decision trees and makes a prediction with a majority vote ("bootstrap aggregation"). An ANN includes one input layer, one, two or three hidden layers, and one output layer.
Neurons in a previous layer combine with "weights" in the next layer (Here, the weights are numerical values showing how much effects neurons in a previous layer have on neurons in the next layer). This operation is done in the order of weights in a layer next to the input layer, its following layer, and so on. This process is called the feedforward algorithm. Then, these weights are adjusted based on how much contribution they made to the errors of the ANN (Here, errors are gaps between actual and predicted values of the dependent variable). This operation is done in the order of weights in the output layer, its previous layer, and so on. This process is called the backpropagation algorithm. These algorithms are repeated until a certain standard is achieved for the accurate prediction of the dependent variable [13].
Data on 3298 participants were divided into training and validation sets with a 75:25 ratio. The models were built (or trained) based on the training set with 2474 observations then the models trained were validated based on the validation set with 824 observations. The mean squared error (MSE), the average of the squares of errors among 824 observations, was introduced as a criterion for validating the models trained. Here, errors are gaps between actual and predicted values of the dependent variable, vasomotor symptoms. Variable importance from the random forest, a mean-impurity gap between a complete model and a model excluding a certain variable, was adopted for identifying major determinants of vasomotor symptoms (mean impurity, or the degree of data being mixed at a node on average, is proportional to the MSE). The greater "mean-impurity increase" is de ned as the greater variable importance. R-Studio was employed for the analysis on April 2020.

Ethics statement
This retrospective study complied with the tenets of the Helsinki Declaration and was approved by the Institutional Review Board (IRB) of Korea University Anam Hospital on 20 Jan 2020 (2020AN0031). Informed consent was waived by the IRB.

Results
Descriptive statistics for continuous and categorical variables in this study are summarized in Tables 1 and 2, respectively. The mean score of MRS 1 (vasomotor symptoms), age and menopause age of 3298 participants were 0.8093, 52 (years) and 49 (years), respectively. Their mean thyroid stimulating hormone, monocyte and triglyceride were 2.98 (µU/mL), 6.47 (%) and 104.20 (mg/dL), respectively. The MSEs of the ve machine learning models are shown in Table 3. A single 75:25 split of the training and validation sets would reduce the validity and generalizability of the results. For this reason, the random split and the statistical analysis were repeated 3 times and their average MSE was calculated for each of the ve statistical methods, i.e., linear regression, random forest and ANNs with one, two and three hidden layers. The random forest was the best model for predicting vasomotor symptoms. Its average MSE (0.9326) was much smaller than those of linear regression and the ANNs (12.4856, 1.5576, 1.5184 and 1.5833, respectively). Linear regression was the worst model and it registered a great variation in terms of MSE, from 1.0119 (Run 2) to 29.5104 (Run 3). Based on variable importance from the random forest, the most important determinants of vasomotor symptoms were age, menopause age, thyroid stimulating hormone, monocyte and triglyceride, as well as gamma glutamyl transferase, blood urea nitrogen, cancer antigen 19 − 9, C-reactive protein and lowdensity-lipoprotein cholesterol (Run 1 in Table 4, Fig. 1). Indeed, the following determinants ranked within the top 20 in terms of variable importance: cancer antigen 125, total cholesterol, insulin, free thyroxine, forced vital capacity, alanine aminotransferase, forced expired volume in one second, height, homeostatic model assessment for insulin resistance and carcinoembryonic antigen. The ndings of linear regression present useful information about the effect of a major determinant on vasomotor symptoms. For example, vasomotor symptoms will decrease by 0.03 if menopause age increases by 1 year. Likewise, vasomotor symptoms will increase by 0.01 if thyroid stimulating hormone increases by 1 unit (µU/mL). It is to be noted, however, that the results of linear regression are based on an unrealistic assumption of ceteris paribus, "all the other variables staying constant". In this context, the ndings of linear regression are to be considered as just supplementary information to the variable importance from the random forest.

Discussion
The results of this study are consistent with previous ndings on the associations of vasomotor symptoms with age [4], the lipid pro le [14], the liver function [9], in ammation markers [15] and insulin resistance [16,17]. This study provides the following additional information for existing literature as well. Firstly, menopause age was the second most important determinant of vasomotor symptoms and their linkage was found to be negative in this study. Little study has been done on this topic and only a few independent suggestions have been made on a negative association between menopause age and cardiovascular disease [18] and a positive linkage between cardiovascular disease and vasomotor symptoms [19,20]. In a similar context, one would expect a negative relationship between menopause age and vasomotor symptoms, and this study supports this expectation given the high ranking of menopause age from the random forest in this study (2nd ). More effort should be made for identifying their underlying mechanism. Secondly, thyroid stimulating hormone was a top 3 determinant of vasomotor symptoms and their association was reported to be positive for the participants aged 40-80 years in this study. According to the previous results, however, their linkage was (1) neutral for pre-or peri-menopausal women aged 42-52 years [21] and (2) negative for euthyroid menopausal women aged 49-59 years [22]. Age and thyroid status would be important mediating variables here and more examination is to be done on a relationship among thyroid stimulating hormone, vasomotor symptoms and their mediating variables.
Thirdly, the ranking of the monocyte was fourth among the most important determinants of vasomotor symptoms in this study. It was reported that the ratio of monocyte to high-density-lipoprotein cholesterol has positive linkages with the risks of coronary artery disease [23], metabolic syndrome [24] and polycystic ovary syndrome [25]. One possible explanation for these results is that the higher monocyte ratio is associated with systematic in ammation [25][26][27], which leads to the higher risks of the diseases above. This explanation can be extended to an association between the monocyte and vasomotor symptoms. More effort is to be made in this direction and this study would be a good starting point.
Fourthly, it was unexpected that cancer antigen 19 − 9, cancer antigen 125 and carcinoembryonic antigen were among top 20 determinants of vasomotor symptoms in this study. There were a couple of studies on the effect of hormone replacement therapy on these cancer antigens among post-menopausal women [28,29]. But no examination has been done and more investigation is needed on a direct relationship between vasomotor symptoms and these cancer antigens. In a similar vein, forced vital capacity and forced expiratory volume in one second ranked fteenth and seventeenth among the most important determinants of vasomotor symptoms in this study, respectively. There were a few studies on the impact of hormone replacement therapy on these lung-function indicators among post-menopausal women [30,31]. However, no literature is available and more research is needed on a direct relationship between vasomotor symptoms and these lung-function indicators.
This study is the rst machine-learning study to predict women's vasomotor symptoms and analyze their determinants. Based on the results of this study, the random forest could address which variables are more important for the prediction of vasomotor symptoms, while its performance was much better and much more stable than that of logistic regression, a popular statistical approach in conventional studies.
In addition, data in this study are larger than those in previous studies − 3298 women and 104 independent variables. The ndings of this study demonstrate that machine learning provides an invaluable decision support system for the prediction of vasomotor symptoms. However, this study had some limitations. Firstly, this study adopted a cross-sectional design. Expanding data with a longitudinal design is expected to improve the accuracy of machine learning signi cantly. Secondly, this study did not consider possible mediating effects among variables. Thirdly, data came from a single center in this study. Expanding data with a multi-center design will be a good topic for future research.

Conclusions
In conclusion, for preventing women's vasomotor symptoms, preventive measures would be needed regarding the thyroid function, the lipid pro le, the liver function, in ammation markers, insulin resistance, the monocyte, cancer antigens and the lung function. Machine learning provides an invaluable decision support system to predict women's vasomotor symptoms and analyze their determinants.

Abbreviations
Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. 2020R1C1C1013579).

Availability of data and materials
The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.
Ethics approval and consent to participate Data will be made available on request Consent for publication Not applicable Figure 1 Top 20 variables from random forest variable importance