Our findings showed that increasing the dose of gonadotropins administered during ovarian stimulation for IVF was not associated with improved efficacy in terms of the number of retrieved MII oocytes, particularly among patients predicted to produce low (1–3) or intermediate (4–8) MII oocyte counts. Similar outcomes were obtained when analyzing gonadotropin dosing either as a cumulative dose or considering the starting and continuation doses separately.

Consistent with our findings, previous studies have shown a negative correlation between the total gonadotropin dose used in assisted reproductive technology and the number of retrieved oocytes [21]. Patients receiving 1001–2000 IU of total gonadotropin during ovarian stimulation yielded the highest number of retrieved oocytes, whereas doses exceeding 5000 IU significantly reduced this outcome by as much as half. Our results showed that patients with low (1–3 oocytes) and intermediate (4–8 oocytes) MII prediction did not benefit from increasing gonadotropin doses. Interestingly, patients with the highest MII predictions (9–12 oocytes) achieved better outcomes with a mid-range cumulative gonadotropin dose compared with the lowest 1050 IU and highest 2100 IU doses.

The subgroup of patients with the lowest MII prediction is of particular concern; these patients often receive high gonadotropin doses to stimulate follicle development. However, our findings challenge the assumption that maximizing the dose is as effective as generally believed, which aligns with previous studies showing that increased gonadotropin doses did not enhance the ovarian response in predicted poor responders [29, 30].

The gonadotropin dosing regimen for patients with high MII oocyte prognosis should also be carefully considered, as this subgroup is at a heightened risk of ovarian hyperstimulation syndrome. In our study, the most beneficial effect on the retrieved MII oocyte count obtained in this group was observed when a daily FSH dose of 225 or 300 IU was administered. These patients benefited more when the total dose of gonadotropins administered during stimulation was in the range of 1275–1800 IU, suggesting that, in patients with a high predicted MII, the choice of a continuation dose of gonadotropin may be of greater importance. As noted in [21], about 80% of patients in their study received FSH doses outside the optimal range, raising significant concerns regarding patient safety. Furthermore, several studies have shown that increasing gonadotropin dosage negatively affects clinical pregnancy rates, live birth rates [31, 32], and the production of fertilized or good-quality embryos [31, 33]. Recently, the shift towards milder stimulation approaches has been widely discussed [34]. Stimulation with lower FSH doses (⩽150 IU/day) has been shown to offer similar IVF success rates compared to conventional high-dose regimens, but with a reduced risk of ovarian hyperstimulation [35]. Our findings align with this emerging trend, highlighting the diminishing role of high-dose gonadotropin treatments and the growing importance of milder protocols.

Comparisons of natural and high-dose gonadotropin-stimulated IVF cycles in humans demonstrated that gonadotropin stimulation can influence oocyte maturity, fertilization, and cleavage-stage embryo morphology [36]. However, the authors did not differentiate between FSH doses separately. Elevated daily gonadotropin doses are associated with a higher proportion of immature oocytes [37].

The lack of increased MII oocyte retrieval with higher doses of FSH during ovarian stimulation can be attributed to various factors. Individual responses to gonadotropins can greatly differ due to follicular sensitivity, which can be influenced by genetic variations in genes encoding hormone receptors, such as FSHR and LHCGR. Certain sequence variants in these genes have been linked to altered ovarian responses, affecting retrieved oocyte counts, stimulation duration, and FSH consumption [38]. Specific genotypes were associated with an increased ratio of FSH dosage to the collected oocyte count [39], whereas others might benefit from increased doses [40, 41]. To better understand these factors’ influence on patient responses and their impact on gonadotropin dosing efficacy, further research is warranted.

An additional aspect revealed by our study is the use of a machine learning tool to predict the number of mature oocytes, serving as an aid in decision-making regarding gonadotropin dosage. Our research uncovers another important aspect of optimizing ovarian stimulation: using a machine learning tool to estimate the number of mature oocytes as a factor in determining the prescribed dosage of gonadotropins. We demonstrated that an ML tool can accurately predict the number of MII oocytes retrieved after stimulation. The most important features for this prediction are AMH and AFC, which are commonly used to determine gonadotropin doses [42, 43]. However, by including less influential features, we captured their cumulative effects on prediction. As many of these features are interrelated, incorporating them into the model enhances its predictive accuracy.

Utilizing ML tools to predict oocyte numbers presents several advantages for guiding gonadotropin dosing strategies in IVF, thus enabling personalized treatment plans. Analyzing multiple variables, such as patient characteristics, biomarkers, and historical data, has the potential to lead to more informed clinical decisions and reduce trial-and-error approaches. ML models can help optimize gonadotropin dosages to obtain maximum efficacy while minimizing risks, such as OHSS, potentially improving resource use, and lowering treatment costs. However, the clinical integration of ML tools will require rigorous validation, continuous monitoring, and multidisciplinary collaboration between data scientists, clinicians, and reproductive health experts to ensure the reliability of these predictive models and patient safety.

The strength of this study lies in the utilization of our previously developed machine-learning model, which provides highly accurate predictions of the number of MII oocytes [25]. The choice to utilize the ML model was driven by its ability to capture complex nonlinear relationships between multiple clinical factors, such as age, BMI, previous stimulation outcomes, and conditions, such as endometriosis, PCOS, and other causes of infertility. These factors collectively enhance the accuracy of MII oocyte predictions, surpassing the capabilities of traditional statistical methods, particularly in handling the intricacies of our large and complex dataset. Furthermore, by using predictions from our machine learning model to create study groups, we ensured that patients within the same group had highly similar characteristics encompassing various fertility-related features. This approach allowed us to establish a direct relationship between patient prognosis, gonadotropin dosage, and stimulation outcome, as all other patient characteristics were already encompassed within the model-based prediction. Our study also supports the idea of using machine learning models to precisely predict stimulation outcomes for optimizing gonadotropin dosing selection, as recently reported [31]. These authors demonstrated that employing the machine learning model for dose selection resulted in higher numbers of mature oocytes, fertilized embryos, and usable blastocysts, while reducing the amount of starting and total FSH used.

A limitation of the method presented in our study is related to model bias. Regression models tend to be biased towards the mean value in the studied population, so the trained model is inclined to generate lower predictions for patients with the highest expected MII and inflate the number of MII predicted for patients with the lowest expectations. Additionally, gynecologists are more inclined to assign the highest doses to patients with low MII oocyte retrieval expectancy; therefore, the size of the group of patients with a predicted MII oocyte count < 4 who received a dose of 300 IU/day was significantly larger than that of the group receiving lower doses administered under the same MII predictions. Another limitation is that gynecologists who prescribe higher doses may possess additional knowledge that was not recorded in the data. Our data should be interpreted with caution because our analysis was limited to patients with MII oocyte predictions of up to 12 oocytes. The use of higher doses of gonadotropins may lead to ovarian hyperstimulation syndrome [44], which is particularly possible in patients with higher predicted MII counts. Therefore, further studies should include such patients to determine whether our observations apply to those with higher MII prognoses. Additionally, to avoid bias, our study excluded stimulations with extreme AMH or AFC values. This limits the generalizability of our model to outliers that may occur in some patients. Therefore, a dedicated model and analysis of such cases are required.

A potential limitation of our study was the heterogeneity of the study population, arising from the division of stimulations into nine distinct gonadotropin-dosing regimens. Some dosing regimen groups had relatively small numbers of participants. To mitigate potential biases, we excluded groups with fewer than 30 individuals from certain analyses and separately examined the effects of starting and continuation FSH doses in the larger groups. The inclusion of diverse protocols reduced the risk of selection bias, thereby ensuring that the study captured a wide range of patient responses. The diversity in stimulation protocols created a rich dataset for exploring dose–response relationships, allowing for a nuanced analysis of how varying gonadotropin dosages impact ovarian stimulation outcomes and provide a more comprehensive understanding of the treatment effect. Moreover, this study was based on a dataset obtained in a real-world clinical setting where diverse conditions are commonly encountered. To verify our findings, we suggest that a prospective study be performed where a prediction is first performed and then different doses of gonadotropins are administered to patients with the same prediction. It would also be worth determining how such doses affect other IVF success outcomes, such as the number of high-quality embryos obtained or live birth rate.

Another limitation is the distinctive characteristics of the study population, which is marked by a relatively high prevalence of genetic defects and endocrine disorders. This may constrain the generalizability of our findings to populations with different prevalence rates of these characteristics. However, our primary focus was to evaluate the impact of gonadotropin dosing on ovarian stimulation outcomes, with genetic factors considered integral to the overall patient profile. Thus, these factors were included as covariates in the analysis to account for their potential influence on the stimulation outcomes. Further research is important to study diverse populations to reflect real-world clinical scenarios in which patients may present with various genetic and endocrine conditions.

In conclusion, our results show that increasing gonadotropin doses for ovarian stimulation did not enhance the efficiency of MII oocyte retrieval beyond the predicted number. In the future, the application of ML models may enhance gonadotropin dosing precision, thereby avoiding unnecessary increases in medication and improving treatment cost efficiency.