Prediction model of gonadotropin starting dose and its clinical application in controlled ovarian stimulation

The outcome of in vitro fertilization (IVF) is not only related to in vitro conditions in laboratory and personnel operations [1‐3] but also closely related to the condition of oocytes. Controlled ovarian stimulation (COS) is one of the essential procedures in the IVF treatment cycle that determines the quality and quantity of oocytes. The multiple follicular growth in COS is the direct effect of exogenous Gn, which leads to supranormal circulating concentrations and recruitment of follicles whose follicle stimulating hormone (FSH) sensitivity threshold is exceeded [4, 5]. In general, the more exogenous doses of Gn, the more oocytes will be retrieved in the COS process. However, in actual clinical practice, the ovary response to Gn is complicated and can not be simply characterized by a linear relationship. Besides, growing evidence shows that an optimal number rather than a maximal oocyte retrieved is the preferred outcome of the COS process [6]. With the same dose of Gn, a poor ovarian response will lead to low quality or quantity of oocytes, resulting in a high cycle cancellation rate and low pregnancy rate [7]. In contrast, a hyper-ovarian response may contribute to high hormone levels with life-threatening side effects, such as ovarian hyperstimulation syndrome (OHSS), which will eventually lead to a higher fresh embryo transfer cancellation rate [8].

In the repeated IVF cycles, the choice of starting Gn dose is mainly based on the response observed in previous attempts, while in the first IVF cycle, the selection of the dose is primarily based on empirical methods and refers to a woman’s basic clinical characteristics such as anti-Müllerian hormone (AMH), antral follicle count (AFC), age, body mass index (BMI), etc. [9‐11]. However, this strategy lacks unified standards and is suspectable to the clinician's expertise and knowledge. Till now, there is no exact recommendation standard for the dose of Gn around the world [12]. Therefore, in the context of treatment strategies aimed at optimal oocyte retrieval, customizing the dose of Gn for patients, especially the Gn starting dose (GSD), is a challenging problem faced by clinicians in the COS process.

Nowadays, some works have been conducted on building complex models based on multiple ultrasound-derived and biochemical indexes to dictate the GSD in IVF cycles [5, 13‐15]. These methods, although helpful, lack comprehensive and quantifiable evaluation of the predictive performance, limiting its application in clinical practice. With the development and the increasing integration with the medical domain, machine learning is becoming more and more popular to mine the medical experience and knowledge hidden in the clinical data so as to facilitate medical practitioners to make good and scientific clinical decisions and understand the hidden principle/theory behind the phenomenon and medical data [6, 16‐21].

Based on the patient's clinical data and machine learning methods, this study aims to establish a GSD prediction model that customizes the GSD and yields a good COS outcome. Besides, a comprehensive and quantitative evaluation was conducted to determine the influence of each factor on the GSD. Eventually, the proposed model is served as a scientific auxiliary tool that helps clinicians tailor the GSD in actual clinical application.

This study built a regression GSD model that can accurately predict the GSD value for the first time. Based on the prediction results of the model, a comprehensive and quantitative evaluation of the influencing factors is conducted. Compared with the traditional method that purely relies on the clinician's experience, the proposed GSD model in this study is an effective auxiliary decision-making tool for customizing the GSD in an IVF treatment cycle, as validated by both historical clinical data and real clinical applications.

Regarding tailoring the GSD in an IVF treatment cycle, some research has been done based on the indicators related to the ovarian reserve function. A prospective study of 145 "standard" patients conducted by Poporic-Todororic et al. [13] found that AFC and ovarian interstitial blood flow assessed by total power Doppler score could serve as guidance for the determination of the GSD. However, ovarian interstitial blood flow is difficult to detect in clinical practice. Marca et al. [5] proposed a GSD calculation model based on some ovarian reserve markers (age, AMH, and bFSH), which can only give rough suggestions on the GSD. Based on age, BMI, bFSH, AFC, and other indicators, Olivennes et al. [14] recently proposed the “CONSORT” algorithm to calculate the GSD. However, it has limited clinical application because it can not reduce the incidence of OHSS and the potential IVF cycle cancellation [14, 23]. Recently, Zhu et al. [15] established a GSD calculation model based on the most common ovarian reserve biomarkers. Still, it is challenging to apply to patients with a high or low ovarian response. To sum up, these studies have the problem of limited categories of biomarkers, and they lack a comprehensive and quantifiable standard for the influence factors on GSD, therefore, limiting the above methods' application in clinical practice.

Based on two machine learning methods of ANN and SVR, this study constructed two regression models for predicting the GSD. Compared with the SVR model, the ANN model has much better modeling performance. For a patient, the GSD is essentially a highly nonlinear function of its influencing factors, which is affected by continuous variables (such as age, AFC, AMH, bFSH, E₂, LH, and so on) as well as discretized ones (e.g., infertility type and therapeutic regimen). Since the ANN method can approach any type of nonlinear model in theory and can deal with different kinds of inputs (both continuous and discretized), the ANN-based model's modeling performance is supposed to be better than that of the SVR-based model. The ANN-based GSD model constructed in this study has good prediction performance: the regression coefficient of all samples is R = 0.951, and 77.49% of the modeling error of samples is within ± 30 units. Therefore, the ANN-based GSD model is chosen and utilized in the COS process for tailoring the GSD in the IVF treatment cycle.

Based on the proposed GSD model, a comprehensive and quantitative evaluation of the influencing factors is conducted. For the GSD, the top four significant factors are: age (\({miv}^{*}\)=1), AMH (\({miv}^{*}\)=-0.841), AFC (\({miv}^{*}=\)-0.499) and BMI (\({miv}^{*}\)=0.44). That is, the older the patient and the higher the BMI value, the greater the GSD required in the IVF treatment cycle; and the higher the values of AMH and AFC, the lower GSD required. Among these factors, age is the most important one affecting the GSD. In the existing research, the importance of these factors has also been confirmed. The work [24] validates that age is the most critical factor affecting ovarian reserve function and an independent predictor of ovarian reactivity. With the increase of age, the ovarian reserve capacity becomes poor, the number and quality of follicles and responsiveness of the ovary to drugs are reduced [25, 26], and the GSD required should increase accordingly. Regarding the AMH and AFC, they can not only reflect the number of follicle recruitment but also predict low ovarian response and high response [4, 5, 27, 28]. AMH is widely used in the clinic because it is not affected by the state of the hypothalamus and oral contraceptives [29, 30]. AFC can directly reflect the reserve level of follicles, and the low-level AFC is related to low ovarian response and high pregnancy failure [28, 31]. However, due to potential error in ultrasound measurement, the accuracy of predicting ovarian reserve via AFC is limited. Therefore, when evaluating the ovarian function and tailoring the GSD, AFC is suggested to work together with other factors, such as AMH, age, etc. This also confirms our evaluation result that AMH substantially influences GSD more than AFC. What’s more, the higher the BMI, the more the GSD requires, and the proportion of OHSS complications in light-weight women is higher even using a very low GSD due to the influence of the pharmacokinetic process (drug clearance and drug dose distribution) [11, 32‐34].

Regarding the therapeutic regimen, clinicians normally formulate the controlled ovarian stimulation therapeutic regimen according to their own clinical experience by considering factors like age, AFC, AMH, economic situation and complications. At present, the GnRH agonist long regimen is the most commonly used COS regiment in IVF/ICSI treatment in patients with normal ovarian response [35]. The GnRH antagonist regimen is best used for known or suspected high responders, including women with PCOS, as it reduces the risk of OHSS [36]. Although the therapeutic regimen is a discretized influencing factor, it has a high effect on the GSD value (shown in Table 1), i.e., the therapeutic regimen is a variable of the GSD prediction model. For each therapeutic regimen, as listed in Table 1, there could be a corresponding GSD per the prediction model.

In addition to the evaluation index \({miv}^{*}\), the Pearson/Spearman correlation coefficient R_p is calculated that reflects the univariate importance of an influencing factor on GSD, as listed in Table 1. The comparison between the \({miv}^{*}\) and R_p value is shown in Fig. 8. From the comparison, we know that the trend (symbol) of \({miv}^{*}\) is the same as that of R_p (except for E₂). For both \({miv}^{*}\) and R_p, the top three factors are age, AMH and AFC; however, their ranking are different. The descending ranking of R_p is AMH (R_p = -0.392), AFC (R_p = -0.357) and age (R_p = 0.301), while that of \({miv}^{*}\) value is age (\({miv}^{*}\) = 1), AMH (\({miv}^{*}\) = -0.841), and AFC (\({miv}^{*}\) = -0.499). Since the proposed GSD model is a regression model, the \({miv}^{*}\) value calculated based on it can evaluate the influencing factors with considering the comprehensive influencing factors simultaneously, which have more scientific merit than Pearson/Spearmen coefficient R_p that is achieved by conducting the univariate analysis between each influencing factor and the GSD. Therefore, the comprehensive evaluation conducted in this work has better scientific merit. The existing work also partially supports our evaluation result [24].

Finally, some clinical applications of the proposed model have been conducted. Clinical application results show that our GSD model is a good mathematical expression and summary of clinicians' experience and knowledge, with the calculated results being very similar to the decision of senior clinicians (average deviation only 14.08 units). We can also see from Fig. 11 that the closer the actual NOR is to 16 (the optimal NOR in this research [6]), the smaller the average deviation could be, i.e., our GSD model can promote the NOR to be close to the optimal number of oocytes. This phenomenon follows the setting of the adaptive adjustment weight of samples in Adaptive adjustment weight of a sample. Besides, the relative average deviation of patients with NOR = 16 should be the smallest in theory; but in fact, due to the inevitable modeling error and the subjectivity and inconsistency of clinicians when deciding on GSD, the average deviation of patients with NOR = 16 is slightly higher than these with NOR = 15 and 17. In general, the closer the NOR is to 16, the smaller the average deviation could be. The results shown in Fig. 11 are indirect reflections that our model can make the NOR close to 16, and there is no direct evidence and no percentage of NOR close to 16. This is indeed a limitation of our current work.

Based on the proposed GSD model and 10 influencing factors of patients, we can get the recommended GSD in the COS process, which could serve as a good reference and basis for clinicians to tailor the GSD in the IVF treatment cycle. At the same time, the recommended GSD can make the potential NOR close to the optimal number 16.

In the future, we will investigate the effect of predicted GSD on more types of IVF-ET outcomes, like the NOR and pregnancy conditions. Also, other machine learning models considering more influencing factors will be built that could yield more types of prediction, like the outcomes of NOR and pregnancy conditions. Finally, the proposed GSD prediction model will have some further clinical applications, especially in comparing the COS result with or without the proposed GSD model.

This research proposed a machine learning-based GSD model that can accurately predict the GSD in the COS process of the IVF treatment cycle. With considering the comprehensive influencing factors of GSD, the model is constructed that could precisely estimate the GSD and simultaneously could possibly make the NOR close to the optimal number. Based on our model, a comprehensive quantitative evaluation of the influencing factors is conducted. The proposed model is a good summary of the clinician’s knowledge and experience as hidden in the historical clinical data and thus can assist the clinicians in customizing GSD in a scientific and impersonal way.