POvaStim: An online tool for directing individualized FSH doses in ovarian stimulation

Abstract

Individualized controlled ovarian stimulation (COS) is a milestone for treatment of infertility. 1 The two acknowledged time points for individualized ovarian stimulation are, one, the beginning of each new treatment cycle when the starting dose is selected and, two, during a given COS cycle when dose adjustment is performed. A few algorithms for directing the follicle-stimulating hormone (FSH) dose have been developed. The number of oocytes retrieved (NOR) during ovarian stimulation was used as the outcome variable, and then the outcome variable and independent variables are stratified and a dose suggested by considering their experience. 2 , 3 La Marca et al. proposed a new idea of ovarian sensitivity using the NOR per unit of the starting dose of FSH. 4 For the first time, they included a dose variable into the outcome measurement and built a model that could predict specific individual doses. They used serum anti-Müllerian hormone (AMH), serum basal FSH, and age to predict the ovarian sensitivity, with a squared coefficient of determination (r 2) of 0.3. However, the outcome variable is the ratio of the actual NOR per starting dose, both of which are unknown predictors before ovarian stimulation and must be assumed before the dose can finally be predicted. They then defined the assumed NOR as nine, and then the starting dose of FSH could be predicted. 4 Although this study has great innovative value for predicting the starting dose of FSH compared with previous models, fixing the NOR as nine may not be individualized enough for all patients undergoing COS. We had previously established two models (models 1 and 2) for predicting the NOR using either day 2 available predictors or day 6 available predictors. 5 We have now taken the idea of ovarian sensitivity proposed by La Marca et al. 4 and combined our previously established NOR prediction models to establish and validate another two new models (models 3 and 4) for predicting the starting (day 2) and adjusting (day 6) doses of FSH during COS. This cohort study was prospective and observational and performed at Peking University Third Hospital. The data were the same as we used in our previous paper for predicting the NOR. 5 Briefly, a total of 669 GnRH antagonist COS cycles were collected without data selection from April to September 2020. After excluding the data with incomplete records, 621 COS cycles were finally analyzed. Model 3 for predicting the starting dose OF FSH using menstrual cycle day 2 available predictors The outcome variable was the ratio of model 1’s predicted NOR 5 to the actual daily dose of FSH. Model 1 was previously built to predict the NOR using day 2 available predictors of AMH, antral follicle counts (AFC), FSH, and age, with main effects (contributions) of 90.2%, 3.6%, 1.2%, and 0.3%, respectively. 5 Predictive variables initially included in model 3 for predicting the starting dose of FSH (model 3) were age, body mass index, cause of infertility, AFC, day 2 levels of AMH, FSH, Luteinizing hormone (LH), estrodiol (E2), testosterone, androstenedione (A4), and inhibin B. The distribution of the outcome variable was tested for normality. The result indicated a skewed distribution (Shapiro-Wilk test, W = 0.8691) (Figure 1A) that approximated to a log-normal distribution, therefore logarithmic transformation of the data was considered. After the logarithmic transformation, the distribution of the outcome variable was closer to a normalized distribution (Shapiro-Wilk test, W = 0.9907) (Figure 1B), thus the ratio of the log-transformed outcome variable was deemed as the outcome variable in the subsequent analysis. The linear relationship between each predictor and the outcome variable was determined separately. Most of the predictors showed a linear relationship with the outcome variable; the exception was AMH, which showed a nonlinear relationship (Figure 1C). After the logarithmic transformation, the goodness of fit of AMH was significantly better, from an r 2 of 0.657 to 0.864 before and after transformation, respectively. In the subsequent analysis, only AMH was analyzed in the logarithmic form; none of the other independent variables were transformed. Figure 1 Model building process of models 3 and model 4, predicting the starting and adjusting doses of FSH using menstrual cycle day 2 and day 6 available predictors, respectively (A and B) Distribution of the outcome variable prior to and after data transformation. The x and y axes in (A) show the outcome variables, namely the ratio of model 1’s predicted NOR to the average daily dose of FSH. (C) The independent variable of AMH before and after data transformation. (D and E) Model 3 building process, which is used for predicting the starting dose of FSH using day 2 available predictors. The red line in model 3 demonstrated in (D) and (E) represents the optimal number of predictors selected by the software automatically. (F and G) Scatterplots displaying the relationships between the predicted and actual outcome variables of model 3 in the training set and validation set. (H) Contributions of each independent variable in model 3 for predicting the starting dose of FSH. (I) Contributions of each independent variable in model 4 for predicting the day 6 adjusting dose of FSH. All the predictors were screened using LASSO regression, which is a method we have used previously. 5 First, the data were divided randomly into training (70%) and validation (30%) sets. The best subset method was used for variable selection, and the variable screening process is shown in Figures 1D and 1E. When four variables of logarithmic-transformed basal AMH, AFC, basal FSH, and age were included, the scaled −log L (β) value in the validation set no longer decreased, and thus model 3 was established. The performance of model 3 was visualized in a scatterplot that showed the relationship between the predicted outcome variable and the actual outcome variable in the training set and validation set (Figures 1F and 1G). The r 2 of model 3 in the training and validation set were 0.911 and 0.923, while the square root of the variance of the residuals (root-mean-square error [RMSEs]) of model 3 were 0.237 and 0.224 in the training and validation sets, respectively. The contributions of the four predictors in model 3 evaluated by main effects and total effects are shown in Figure 1H; AMH contributed the most. Model 4 for predicting the adjusting dose of FSH using menstrual cycle day 6 available predictors Using the same statistical method we used for building model 3, we established model 4 using the ratio of model 2’s predicted NOR 5 to the actual daily dose of FSH as outcome variable. Model 2 was previously built to predict the NOR; the predictors in model 2 include day 6 available predictors of Δinhibin B (day 6 minus day 2), basal AMH, AFC, and age. 5 Predictive variables initially included for predicting the adjusting dose of FSH were all the predictors used in model 3 as well as the Δ levels of AMH, LH, E2, testosterone, A4, and inhibin B. The generalized r 2 of model 4 were 0.922 and 0.909 and RMSEs of 0.236 and 0.231 in the training and validation sets, respectively. The final predicting variables included in model 4 were Δinhibin B level, AMH, AFC, and age, with main effects and total effects indicated in Figure 1I; Δinhibin B contributed the most. Online tool based on models 3 and 4 Ovarian sensitivity, defined as the ratio of the predicted NOR to the actual daily dose of FSH, is used as the dependent variable in models 3 and 4. In these two models, only the daily dose of FSH is an unknown predictor and thus can be predicted. Model 3 predicts the daily dose based on the available predictors on day 2 and thus can be used for predicting the starting dose of FSH. Model 4 predicts the daily dose based on the available predictors on day 6 and thus can be used for predicting the day 6 adjusting dose. The performances of the two models have high r 2 of over 0.9, which means that our algorithms could explain more than 90% of ovarian sensitivity, which, to our knowledge, make them the most powerful models in directing FSH doses. The algorithms have been developed into an easy applicable online tool for free use (POvaStim, http://121.43.113.123:8004). POvaStim may contribute to the improvement of pregnancy outcomes and the reduction of the incidences of cycle cancellation and ovarian hyperstimulation syndrome, which needs to be verified through the randomized controlled trial studies in the future.