Assessment of the impact of male factor infertility on the outcomes of assisted reproductive technology programs using machine learning techniques

According to the Federal Service for State Statistics, since 2016 there is a high natural population decline and the birth rate decline in Russia [1, 2]. One of the most effective and current ways to overcome infertility and improve the demographic situation is using assisted reproductive technology (ART) programs. Since 2013, implementing ART programs has become possible at the expense of compulsory health insurance (CHI), that has become an additional positive incentive to increase the access to infertility treatment in various regions of Russia [3, 4]. The structure of infertility in married couple is largely region-specific [5]. For example, the maximum frequency of male reproductive health disorders per 100 thousand male population is observed in the North Caucasian and Ural Federal Districts, where their rate is 3 times higher than the national average rate. On the contrary, in the Far Eastern Federal District, the rate of male infertility is 3 times lower than the average rate in the Russian Federation [6].

It is noted that the number of patients who seek infertility treatment at an older reproductive age has been increasing over the last years [7]. It is generally accepted that one of the determining factors that influences the effectiveness of ART is the age of women, while on the contrary, men are able to remain fertile for many years. However, it should be noted that with age there is a decline in testosterone levels and the parameters of spermatogenesis in men, and the frequency of miscarriages in their female partners increases. There is an increased number of de novo mutations in sperm (4–5% growth annually), and chromosomal aneuploidies occur more often [8]. Interpretation of spermogram parameters in men during dynamic observation can also be ambiguous; the significance and influence of some parameters on the effectiveness of ART treatment remains controversial.  Danis R.B. et al. reported that assessment of fertility potential using spermogram parameters varies greatly, and identification of the most significant parameter to predict natural pregnancy or using ART remains uncertain. Also, the study showed that determination of the percentage of morphologically healthy sperm in the ejaculate has a low predictive value [9].

To reconstruct relationship between certain characteristics of the ejaculate using a large set of different parameters for obtaining not only the prognosis of treatment effect using ART, but also identification of the most significant factors that determine the resulting prediction, a large number of mathematical models have been proposed, including the use of artificial intelligence (AI) and machine learning (ML). ML enables to specify the algorithms for input data and the use of statistical analysis to predict output information in the process of emergence of new data. One of the promising research areas in ML and male factor infertility is the use of AI to predict blastulation rate, embryo quality and the outcome of infertility treatment using ART [10]. You J.B. et al. reported that from the practical point of view, optimization of sperm selection using ML to improve the effectiveness of ART programs, seems to be an urgent task. However, the issue of interpreting the most important spermogram parameters for the use of their values ​​to build accurate predictive models using ML, remains controversial [11].

It should be noted that the classical mathematical models cannot automatically select the optimal characteristics, and the researcher decides independently according to his point of view to select the most significant characteristics for model building. In addition, if there are nonlinear relationships, the model cannot be able to capture the nonlinearity. Given the presence of both categorical and quantitative data in the spermogram available for analysis, the likelihood of nonlinearity is extremely high. Thus, the use of ML-based systems makes it possible to process complex dependencies that are resistant to noise, and obtain good indicators of data generalization and highlight the most significant characteristics [12]. Moreover, the analyzed information can be processed in the presence of information imperfections or incomplete data to obtain comparable predictive ability. Given the fact that the study of influence of various spermogram parameters on prediction of treatment outcomes seems extremely relevant in terms of optimization of preparation of men for IVF cycle, the purpose of this study was to determine the significance and influence of each parameter characterizing the quality of the ejaculate on pregnancy rate, as well as the influence of these parameters on the embryonic stage of ART programs using linear regression and ML.

Material and methods

The retrospective study included 1021 married couples ages 21–40 years, who sought infertility treatment using ART. Written informed consent was obtained from each couple to process their personal data.  Inclusion criteria in the study were infertility caused by tubo-peritoneal, male or combined factors, chronic anovulation or diminished ovarian reserve, as well as normal karyotype of the spouses, ovarian stimulation according to the protocol with a gonadotropin releasing hormone (GnRH) antagonist, the standard protocol of luteal phase support after embryo transfer, obtaining the patients’ oocytes on the day of transvaginal puncture (TVP), single-embryo transfer. Exclusion criteria included abnormalities of the uterus, abnormal karyotype, and the use of donor oocytes or sperm.

The patients participating in the study underwent ovarian stimulation according to the protocol of GnRH antagonist administration from day 2 or 3 of the menstrual cycle. When the follicle reached ≥17 mm in diameter, HCG was administered to patients to trigger final oocyte maturation (800 patients), or in case of risk of ovarian hyperstimulation syndrome, the trigger was replaced with GnRH agonist (138 patients), or dual trigger was used for final oocyte maturation (83 patients). Transvaginal follicle puncture (TFP) was performed 35–36 hours after triggering ovulation, followed by oocyte collection and assessment of oocyte quality. Subsequently, fertilization of the obtained oocytes was done using in vitro fertilization (IVF) (5.6% of patients), intracytoplasmic sperm injection (ICSI) into an oocyte (81.9%) and physiological ICSI (PIXI) (12.5%). High fertilization rate after the ICSI procedure was associated with lower spermogram values on the day of TFP. All stages of culturing were done in COOK multi-gas incubators (Ireland) in 25 μl drops of culture media under oil (Irvine Sc., USA) at Prof. B.V. Leonov Department of Assisted Technologies in Infertility Treatment, Academician V.I. Kulakov National Medical Research Center for Obstetrics, Gynecology and Perinatology. On day 5 after fertilization, the embryo was transferred into the uterine cavity using soft Wallace catheter (Germany) or Cook catheter (Australia). The remaining embryos of appropriate quality were cryopreserved for further use in cryoprotocol. Standard luteal phase support and managing the care of patients after embryo transfer was provided. On day 14 after embryo transfer, beta subunit human chorionic gonadotropin (beta-HCG) levels were evaluated. When beta- hCG levels were positive, 21 days after embryo transfer the patients underwent pelvic ultrasound to diagnose clinical pregnancy. Further management of pregnancy was carried out individually.

The study analyzed the number of the best-quality, good-quality and average-quality blastocysts, the number of embryos that stopped developing, the morphological assessment of embryo quality using Gardner's embryo grading system, the frequency of clinical pregnancy depending on spermogram parameters on the day of TFP (sperm concentration, the percentage of progressive motility and non-progressive motility, sperm immotility and morphologically healthy sperm).

Statistical data processing and the algorithm for building decision tree

Microsoft Excel, version 15.0 with appropriate license was used for statistical analysis of the obtained data. The Shapiro–Wilk test was used to check the normality of the sample. In interpreting statistical analysis results, the significance level of p-value equal to 0.05 was considered to be critical. Multivariate analysis was used to analyze datasets in all models in the study. Dependent variables (response variables, that is, the characteristics, which were influenced by independent variables) and independent variables (which are represented as “Characteristics” in the Tables, section “Results”) were used in regression analysis. The applicability of linear regression analysis was checked. The normality of distribution of residuals and the equality of their variances were checked, and only uncorrelated features were selected. The Pearson correlation coefficient was used for testing non-correlation. Regression equation is the following:

 =

where b –  the parameters of the model;