Development, promotion, and application of online OvAge calculator based on the WeChat applet: Clinical prediction model research.

In reproductive medicine, a woman’s fertility potential can be used to measure the likelihood of a successful pregnancy using assisted reproductive technology. Domestic and foreign scholars in the field of reproduction have proposed models of female ovarian scoring and staging systems for predicting pregnancy outcomes, such as What’s My Fertility, Stages of Reproductive Aging Workshop (STRAW), OvAge, and the AFA (anti-Müllerian) Hormone level-follicle-stimulating hormone level-age) models [ 1 – 5 ]. These models can help women of childbearing age to effectively understand their fertility status so that they can reasonably plan for pregnancy. Due to confounding factors such as age, genetics, environment, psychology, society, lifestyle, iatrogenic infection, immunity, etc., the age of the ovaries of women of the same chronological age is also very different [ 6 – 11 ]. Moreover, the fertility and ovarian quality of women vary from country to country due to their differences in national conditions, regional characteristics, and living habits. Therefore, it is important to localize the existing female ovarian scoring model for predicting pregnancy outcome and transform it into a tool suitable for Chinese women [ 12 ]. Currently, indicators commonly used in clinical to assess ovarian age and pregnancy outcomes include age, anti-Müllerian hormone (AMH), atrial follicle count (AFC), follicle-stimulating hormone (FSH), clomiphene citrate stimulation test (CCST), inhibin B (INHB) or a combination of the above [ 13 – 17 ]. The ideal scoring system and model should be universally accepted, repeatable, easy to understand, and applicable to the general population. We selected the ovarian age calculator OvAge for localization improvement to make it more suitable for ovarian age prediction of Chinese women. Initially, we used two traditional methods of recruiting subjects. In the first stage, we recruited women who visited the gynecology and reproductive medicine department of local hospital and obtained their clinical data. Data collected in this way were relatively reliable but time consuming. In the second stage, we recruited healthy women through CRO (Contract Research Organization) and monitored their relevant indicators. The obtained data are relatively reliable but time consuming, costly, and subjects did not cover all women. Therefore, we wanted to explore a method of collecting clinical data that could compensate for the shortcomings of the two methods. With the rapid development of the internet and 5G, people’s lives are inseparable from the internet. In the current open network environment, WeChat has become a phenomenon-level application software in China [ 18 ]. Taking advantage of WeChat’s characteristics of fast information dissemination and strong interactivity, we developed an online OvAge calculator based on the WeChat applet by optimizing the existing OvAge model. To obtain the estimated value of ovarian age, the subject could directly search for the "ovarian age calculator" applet in WeChat and then enter the test interface. Subjects only need to enter the required values in the applet to obtain their predicted ovarian age or ovarian age and fertility assessment results and obtain professional medical advice. This study used data collected in three ways to optimize the OvAge model for predicting ovarian age to make it more suitable for Chinese women and demonstrated the feasibility of this optimization through a series of data analyses. This study can provide some ideas for the localization improvement of the model and provide more data support for the further optimization of the OvAge age calculator.

According to the OvAge calculator in the literature, during phase I data collection in this study, we collected the chronological age of 113 women (16 healthy individuals, 45 PCOS, and 52 DOR patients) with five indicators related to the formula ( Table 1 ). DOR: Diminished ovarian reserve. PCOS: Polycystic ovary syndrome. AMH: Anti-Müllerian hormone. FSH: Follicle-stimulating hormone. PRL: Prolactin. AFC: Antral follicle counting. BMI: Body mass index. n = 16: 16 sample data for healthy women. n = 52: 52 for DOR women. n = 45: 45 for PCOS women. We used the OvAge calculator in the literature to process the Phase I data and verified their biological age with the predicted ovarian age. The chronological age of PCOS patients was significantly older than the ovarian age calculated by OvAge (p<0.001). The chronological age of DOR patients was significantly younger than the ovarian age calculated by OvAge (p<0.05). There was no significant difference between the chronological age of healthy individuals and the ovarian age calculated by OvAge, but the P value was close to the critical value, suggesting a trend of difference ( Table 2 ). The above results indicate that the chronological age for healthy people is not completely consistent with the ovarian age calculated by OvAge in the literature. n = 16: 16 sample data for healthy women. n = 52: 52 for DOR women. n = 45: 45 for PCOS women. For healthy people, the chronological age was not exactly the same as the ovarian age calculated by OvAge in the literature, suggesting that this calculator is not completely suitable for Chinese women. To establish a more suitable ovarian age assessment method for the Chinese population, we tried to optimize the OvAge formula. Age was the dependent variable, and the independent variables for modeling were BMI, P, T, E 2 , PRL, LH, FSH, and AMH. We considered normal linear regression, a combined generalized linear model with logarithm and identity link functions, and modeled it with the independent variable of age. We used stepwise linear regression based on the AIC value for variable selection and constructed a regression model for the phase II data, which are represented as mod1a, mod2a, and mod3a (Mod1 indicates that the dataset conforms to the normal distribution. Mod2 indicates that the dataset is a nonnormal Poisson distribution, and the GLM with the conventional log-link is selected. Mod3 indicates that the dataset is a nonnormal Poisson distribution and the GLM with the identity-link is selected. The “a” indicates that the selected dataset is from dataset II, and “b” indicates that the selected dataset is from dataset II and dataset III. “a” indicates that the selected dataset is from dataset II. Mod1a, mod2a, and mod3a only contain PRL and AMH variables. PRL was not statistically significant, whereas AMH was statistically significant ( Table 3 ). *** P < 0.001; ** P < 0.01; * P 0.05. According to the results of prediction errors Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Mean Absolute Error (MAE) ( Table 4 ), the smaller the MSE, RMSE, and MAE, the better the prediction. It was concluded that mod3a was the best fit to phase II data. MSE: mean squared error. RMSE: root mean squared error. MAE: mean absolute error. We used the OvAge calculator from the literature and the established mod3a formula to test the phase I and phase II datasets to compare the chronological age with the predicted ovarian age ( Table 5 ). The Mod3a calculation results show that there was no significant difference between the chronological age of healthy women and the ovarian age predicted by OvAge, and P2 was less than P1, indicating that the difference between the predicted results for healthy people was more significant than that of the OvAge calculator in the literature. The chronological age of PCOS patients was not significantly different from the ovarian age calculated by OvAge (p>0.05). The chronological age of DOR patients and the ovarian age calculated by OvAge were also not significantly different (p>0.05). This finding shows that the accuracy of Mod3a’s prediction of ovarian age was worse than that of the literature OvAge calculator. n = 68: 68 sample data for healthy women. n = 52: 52 for DOR women. n = 45: 45 for PCOS women. To further optimize the OvAge calculation formula we established by expanding the sample size of the test data, we tentatively developed an online OvAge calculator based on the WeChat applet using the mod3a formula ( Fig 2 ). A subject self-report system based on the WeChat applet-online OvAge calculator. a. QR code of online OvAge calculator; b. Main interface; c. Data Acquisition Module; d. Questionnaire module; With the help of the online OvAge calculator in the WeChat applet, 9000+ sample data were collected in only two weeks, and data from 97 healthy people were screened out as model optimization dataset III. (A) QR code of online OvAge calculator. (B) Main interface. (C) Data Acquisition Module. (D) Questionnaire module. As mentioned earlier, we used the data from phase II and phase III, and based on mod1a, mod2a, and mod3a, the models were optimized to mod1b, mod2b, and mod3b (b means the selected data came from the sum of phase II data and phase III data). We corrected the differences between datasets from different sources by introducing a new variable New (New = 0 represents phase II data, New = 1 represents phase III data) and considered adding new variables and their corresponding interaction terms with other variables to this model. However, the results of variable selection based on the stepwise linear regression of the AIC value showed that no significant variable effects related to the new variables were found, which means that there was no difference in AMH and PRL between the phase II and phase III data( Table 6 ). At the same time, the prediction errors of MSE, RMSE, and MAE were the lowest in mod2b ( Table 7 ), indicating that mod2b was the best formula after the second optimization. The model was estimated as: OvAge = exp 3.5254 − 0.0001 * PRL − 0.0231 * AMH . *** P < 0.001; ** P < 0.01; * P 0.05. MSE: mean squared error. RMSE: root mean squared error. MAE: mean absolute error. We used the literature formula, the best formula from the first optimization (mod3a), and the best formula from the second optimization (mod2b) to test with the phase I+II dataset ( Table 8 ). The test results show that for the disease population (DOR/PCOS), P1<P3<P2, and the accuracy of the calculation results of the ovarian formula after two optimizations was not as good as that of the literature formula, but the accuracy of the calculation results of the second optimization formula was significantly improved. At the same time, for healthy people, P3>P1>P2, the calculation results of the first optimization formula were not as accurate as the literature formula, but after the second optimization, the accuracy of the calculation result was better than the literature formula. Mod2b had higher accuracy in predicting the ovarian age of the normal population, which was better than Mod3a and literature formulas. At the same time, the accuracy of predicting the ovarian age of the disease population was better than mod3a, but it was still not as good as the literature formula. n = 68: 68 sample data for healthy women. n = 52: 52 for DOR women. n = 45: 45 for PCOS women.

This research combines the existing OvAge model to develop an OvAge online calculator based on the WeChat applet. Then, we collected basic data and gynecological examination reports from Chinese women through an online application, aiming to explore new methods for subject self-report and provide more data support for further optimization of the OvAge model. This approach provides more support for the development and promotion of the online OvAge calculator, creates a new method of network-based recruitment of clinical subjects, and provides new ideas for clinical research.

In data collection for this study, we collected the AMH, FSH, luteinizing hormone (LH), estradiol (E 2 ), progesterone (P), testosterone (T), prolactin (PRL), and vitamin D levels of subjects from Day 2 to Day 4 of the menstrual cycle. At the same time, we collected pelvic or vaginalultrasound images on the same day of hormone testing and recorded the subjects’ height and weight. One special point is that pelvic or vaginal ultrasound was dispensable for subjects collected from the WeChat applet. This process is divided into three stages: (1) Phase I data collection from May 1 to June 30, 2020. We collected clinical data from subjects from the Department of Gynecology and Reproductive Medicine of Jiangsu Provincial Hospital of Traditional Chinese Medicine, including health, PCOS, and DOR populations. Within two months, we recruited 113 women (16 healthy individuals, 45 PCOS, and 52 DOR populations). The study was approved by the China Clinical Trials Registry (Registration Number ChiCTR2000037522) and the Medical Ethics Committee of Jiangsu Hospital of Traditional Chinese Medicine (Application Number 2019NL-152-02). The experimental design fully considered the principles of safety and fairness, and the study did not cause any harm or risk to the subjects. Recruitment was based entirely on the principle of voluntary and informed consent, and the privacy of volunteers was protected as much as possible. In addition, there were no conflicts of interest in this study, (2) Phase II data collection occurred from July 1st to August 31st, 2020, and the subjects were recruited through a contract research organization (CRO) company for a fee. A total of 52 healthy women were included. All tests were carried out in the laboratory of Jiangsu Provincial Hospital of Traditional Chinese Medicine, (3) Phase III data collection occurred due to the low collection efficiency in the first two stages. We tried a brand new online OvAge calculator with a subject self-report system based on a WeChat applet. We clarified the functions and requirements of the OvAge calculator and handed it over to a professional software development company to complete using JavaScript and WXML (WeiXin markup language) development languages, and the basic software version 1.0.0 launched on September 23rd, 2020. In two weeks, more than 9,000 sample data points were collected using the online OvAge calculator in the WeChat applet. After screening, 97 cases of healthy population data were available. The age range of patients participating in this study through the above three recruitment methods was 15–55 years old. The training set object used to build the model was healthy women. The inclusion criteria for healthy women were a history of spontaneous conception, intact ovaries, and normal menstruation, with an average interval of 21 to 35 days. The exclusion criteria were estrogen or progesterone use or breastfeeding two months before registration, pregnancy, history of female infertility, endometriosis, ovarian cystic mass, history of ovarian surgery, PCOS, gynecological malignancy, history of radiotherapy or chemotherapy, autoimmune diseases, chronic systemic metabolic and endocrine diseases including hyperandrogenemia, hyperprolactinemia, diabetes, thyroid disease, hypogonadotropic hypogonadism, and drug usage that may cause irregular menstruation. In addition, the test set also included PCOS and DOR populations. PCOS was diagnosed according to Rotterdam criteria, and there were at least 2 of the following 3 characteristics: hypomenorrhea/amenorrhea, clinical and/or biochemical hyperandrogenemia, and polycystic ovarian morphology (PCOM) [ 19 ]. The diagnosis of DOR has not been standardized, and there is no international standard. Combined with the current guidelines of the National ART Surveillance System (NASS) and Society for Assisted Reproductive Technology (SART), the diagnostic criteria used in this study were age ≥ 40 years old, basic FSH ≥ 12 mIU/mL, AFC ≤ 5~7, and AMH ≤ 1.1 ng/mL. Meeting two or more of the above four items was determined to be DOR [ 20 ]. In this study, subjects from gynecology and reproductive medicine clinics, CRO companies, and OvAge WeChat applets were screened based on the above criteria, and past medical history was collected through professional questionnaires. R software (version 4.0.2) was used to analyze a dataset of 262 subject records in the first to third phases. The whole analysis consists of 4 main steps ( Fig 1 ). Flow chart of optimizing the OvAge model. First, we divide the data collected in the three stages into three datasets: training set, test set, and prediction set. The training set includes phase II and phase III data, which include 52 and 97 observation data points, respectively. The prediction set included phase I data, including healthy controls (HC,16 subjects), DOR (52 subjects), and PCOS (45 subjects). We checked each dataset to ensure data quality, including missing values and consistency. For missing data, discard the corresponding observations and then perform modeling. To describe the relationship between the new response variable, OvAge, and the set of inputs, which are the independent variables, we applied generalized linear model (GLM) theory to our research for the training set. The GLM provided a unified method for modeling all types of response variables, such as continuous, binary, and ordinal response or variables in the form of proportions. It has been hypothesized that in the healthy population, OvAge is equal to chronological age, which is not continuous and nonnormal. Therefore, a Poisson distribution was chosen as the random component of the GLM for modeling the expected value of OvAge. Identity and logarithm were chosen as candidate link functions. For Ockham’s razor (law of parsimony) and avoiding collinearity problems due to the possible nonindependence of predictor variables, stepwise selection, a conventional variable selection method, was applied using the Akaike information criterion (AIC). We first constructed regression models for only Phase 2 data. Next, we used the same variables in the models for only Phase 2 data to fit corresponding models for both Phase 2 and 3 data. Considering all possible interactions among covariates to assess any improvements in AIC value and accuracy, the best model was rebuilt. Interaction terms in the model provide interactive effects and were considered when the main effects were significant. The final model was chosen by comparing pairs of models for only Phase 2 data and both Phase 2 and 3 data with the assessment of prediction indices, mean squared error (MSE), root mean squared error (RMSE), and mean absolute error (MAE), for the test set of 19 observations. (iv) The final generalized linear model was applied to the healthy controls (HCs), DOR, and PCOS datasets in the prediction set. The hypothesis was that, in the case of HCs, the ovarian age would be equal to the chronological age. In the case of DOR, the ovarian age would be higher than the chronological age, and in the case of PCOS, the ovarian age would be lower than the chronological age.