The prediction of semen quality based on lifestyle behaviours by the machine learning based models

Abstract

Purpose

To find the machine learning (ML) method that has the highest accuracy in predicting the semen quality of men based on basic questionnaire data about lifestyle behavior.

Methods

The medical records of men whose semen was analyzed for any reason were collected. Those who had data about their lifestyle behaviors were included in the study. All semen analyses of the men included were evaluated according to the WHO 2021 guideline. All semen analyses were categorized as normozoospermia, oligozoospermia, teratozoospermia, and asthenozoospermia. The Extra Trees Classifier, Average (AVG) Blender, Light Gradient Boosting Machine (LGBM) Classifier, eXtreme Gradient Boosting (XGB) Classifier, Logistic Regression, and Random Forest Classifier techniques were used as ML algorithms.

Results

Seven hundred thirty-four men who met the inclusion criteria and had data about lifestyle behavior were included in the study. 356 men (48.5%) had abnormal semen results, 204 (27.7%) showed the presence of oligozoospermia, 193 (26.2%) asthenozoospermia, and 265 (36.1%) teratozoospermia according to the WHO 2021. The AVG Blender model had the highest accuracy and AUC for predicting normozoospermia and teratozoospermia. The Extra Trees Classifier and Random Forest Classifier models achieved the best performance for predicting oligozoospermia and asthenozoospermia, respectively.

Conclusion

The ML models have the potential to predict semen quality based on lifestyles.

Introduction

Semen analysis is used to detect the fertility capacity of men in andrological practice and it is the first recommended test for an infertility diagnostic work-up [1]. It is recommended that the semen sample is given close to the laboratory [1] but giving a sample at the hospital could sometimes cause embarrassment and some men may hesitate to provide a sample to learn their fertility capacity. As a result, home semen analysis kits and smartphone-based semen analyzers have started to be used by men [2, 3] to minimize the embarrassment, but these methods have not been used worldwide and their accuracy and cost-effectiveness are have not been fully discussed [2, 3].

According to our knowledge, there have not been any nomograms or predictors using conventional statistical methods to detect semen quality, but it is well known that lifestyle behaviors could affect semen quality [4]. With the increasing use of artificial intelligence in medicine, a limited number of studies have recently been published that aimed to predict semen quality based on lifestyle behaviors using machine learning algorithms [5,6,7,8,9,10,11,12,13,14,15,16].

In this study, we aimed to find the machine learning (ML) method that has the highest accuracy in predicting the semen quality of men based on basic questionnaire data about lifestyle behavior.

Material-methods

This retrospective-designed study was conducted after ethical approval was obtained (Eskisehir City Hospital, Non-Interventional Clinical Research Ethics Committee; Date: 16/02/2024; Decision Number: ESH/GOEK 2024/77). The medical records of the men whose semen had been analyzed for any reason between August 2021 and January 2023 at the Eskisehir City Hospital were collected. The exclusion criteria were: Aged < 18 or ˃50 years, diagnosis of azoospermia, low semen volume (less than 1.5 mL), abnormal genetics, history of any type of testicular or genitourinary tract or pelvic surgery, recurrent or subclinical varicocele, cryptorchidism, small-sized testis (normal testicular volume is 12.5–19 cc), treated cancer, vascular problem, hematologic illness, systemic disease, genitourinary system infection, or hormonal problems. After the application of the exclusion criteria, the men who also had data about their lifestyle behavior on file were included in the study. This data included details of their Body Mass Index (BMI), smoking and alcohol consumption, coffee intake, physical activity, sauna usage, cell phone usage, and the wearing of tight-fitting underwear as described previously [17, 18]. To ensure strict selection, ex-alcoholics and ex-smokers, passive smokers, and those who only participate in the other lifestyle factors irregularly were also excluded from the study. The lifestyle factors were coded ‘1’ if the BMI was ≥ 25, he smoked every day, drank any amount of alcohol, drank more than 3 cups of coffee a day, did not do any type of exercise regularly, regularly wore Tight underwear, went to a sauna-Turkish Bath regularly, or had a mobile phone ≥ 10 years during the 3-month window before semen collection. If a man who had a BMI of < 25, did not smoke, did not drink alcohol, did not drink more than 3 cups of coffee a day, exercised regularly, did not wear tight underwear, did not go to a sauna, or had used a cell phone < 10 years, the lifestyle factors were coded ‘0’. After collecting the data about the lifestyle behaviors, all semen analyses of the men included in the study were evaluated according to the WHO 2021 guideline. All semen analyses were categorized as normozoospermia (normal semen characteristic value) or abnormal. If oligozoospermia (sperm concentration < 16 × 10⁶/ml of semen) and/or asthenozoospermia (motility < 30% spermatozoa with progressive motility), and/or teratozoospermia (morphologically normal spermatozoa < 4%) [1] had been detected in a semen sample, these results were categorized abnormal. All results were then grouped as normozoospermia, oligozoospermia (sperm concentration < 16 × 10⁶/ml of semen), asthenozoospermia (motility < 30% spermatozoa with progressive motility), or teratozoospermia (morphologically normal spermatozoa < 4%). The 4 groups were analyzed separately by statistical methods and the ML algorithms were applied to each group.

The Shapiro–Wilk test was used to test the normality of data distribution. Continuous variables were expressed as mean ± standard deviation, median (minimum–maximum), and categorical variables were expressed as counts (percentages). Comparisons of normally distributed continuous variables between the materials/groups were performed using the student’s t-test. Comparisons of non-normally distributed continuous variables between the groups were performed using the Mann–Whitney U Test. Comparisons of categorical variables between the groups were performed using the Yates Chi-Square test and the Monte Carlo Chi-Square test. A two-sided P value < 0.05 was considered statistically significant.

The study was designed according to the principles of ML. The Extra Trees Classifier, Average (AVG) Blender, Light Gradient Boosting Machine (LGBM) Classifier, eXtreme Gradient Boosting (XGB) Classifier, Logistic Regression, and Random Forest Classifier techniques were used as ML algorithms. 70% of the data was used for training and the remaining 30% for testing. In the tests conducted with these models, the model success rates were determined based on accuracy, sensitivity, and specificity values with confusion matrix metrics and the area under curve (AUC) graph in the receiver operating characteristic (ROC) curve analysis. A confusion matrix, which contains information on actual and predicted classifications performed by a classification system and the performance of such systems, is generally assessed using the data in the matrix. Independent variables that significantly affect each group’s dependent variable were selected by the permutation feature importance method, which is based on a decrease in the model score when a single variable value is randomly shuffled (1).

Results

Seven hundred thirty-four men who met the inclusion criteria and had data about lifestyle behavior were included in the study. As seen in Table 1, 356 men (48.5%) had abnormal semen results, 204 (27.7%) showed the presence of oligozoospermia, 193 (26.2%) asthenozoospermia, and 265 (36.1%) teratozoospermia according to the WHO 2021. Smoking, regularly wearing tight underwear, and regularly going to a sauna / Turkish Bath were statistically significant between having normal and abnormal semen results (p = 0.001, 0.003 and 0.038, respectively). While asthenozoospermic males had a statistically significance difference in the age parameter (p = 0.013), teratozoospermic males had a statistically significance difference in age, smoking, and alcohol use parameters (p = 0.025, 0.001, and 0.034, respectively) (Table 2).

Table 1 Data of patients included in the study

Table 2 Comparison of the data of the patients included in the study by grouping them according to spermiogram results, oligozoospermia, asthenozoospermia, and teratozoospermia conditions

Among the six models, the AVG Blender model had the highest accuracy (61.2%) and AUC (58.4%) for predicting normozoospermia. The Extra Trees Classifier, Random Forest Classifier, and AVG Blender model achieved the best performance for predicting oligozoospermia, asthenozoospermia, and teratozoospermia with an accuracy of 75.5%, 69.6%, and 64.4%, respectively with an AUC of 80%, 74%, and 69.2%, respectively (Table 3 and Fig. 1). Age and smoking were the most significant featured factors for all-predictive models. Table 4 shows the confusion matrices of the algorithms, detailing the number of true positive, false positive, true negative, and false negative cases in the predicted results.

Table 3 Prediction results of machine learning algorithms

Fig. 1

Impact factors of the variables and the receiver operating characteristic (ROC) curve analysis of the study. A AVG Blender of the normozoospermia, B Extra Trees Classifier of the oligozoospermia, C Random Forest Classifier of the asthenozoospermia, D AVG Blender of the teratozoospermia (1: Age, 2: Body Mass Index, 3: Smoking (any amount), 4: Alcohol intake (any amount), 5: Coffee intake (> 3 cups/day), 6: Don’t exercise regularly, 7: Mobile phone ≥ 10 Years, 8: Tight-fitting underwear (regularly), 9: Sauna (regularly)) (AVG Blender: Average Blender, LGBM: Light Gradient Boosting Machine)

Table 4 Confusion matrix showing the results of the classification processes

Discussion

Infertility is a major cause of stress for couples because the diagnosis and treatment are often thought to be very complex and the failure to find the cause of infertility makes the situation even worse. For this reason, healthcare professionals have started to use clinical tools to make accurate decisions for diagnosis and/or treatment to minimize the uncertainty for couples. With the increasing use of artificial intelligence in medicine, machine or deep learning-based tools have become increasingly used in reproductive medicine.

A semen analysis is the first laboratory method used for assessing male reproductive health. However, many men feel embarrassed about giving a semen sample, even though it is the most important tool. Attempts have been made to use artificial intelligence-based clinical decision-making tools to at least have an idea of semen quality instead of going to an andrology clinic. The ML-based algorithms used for this purpose have been used with questionnaire-based information about lifestyle behavior to predict semen results [5,6,7,8,9,10,11,12,13,14,15,16]. The authors of this study chose these factors to develop a prediction model because the separate and cumulative effects of lifestyle on semen quality have long been known [17,18,19]. GhoshRoy et al. reviewed the studies about lifestyle and environmental factor-based analyses for determining semen quality [20]. In this review, all the classifiers and performance parameters of these algorithms that have been used were summarized.

In our study, we collected the lifestyle behavior data from the questionnaire form that had been prepared previously [17]. All semen analyses were evaluated with the recently published WHO 2021 guideline. All the responses from the questionnaire forms and semen analyses were gathered as a dataset. After balancing and validating this dataset, we tried to find the ML model that could predict male fertility capacity with the highest accuracy and AUC. We found that AVG Blender had the highest performance in predicting normozoospermia and the LGBM Classifier could be used for predicting the oligozoospermic, teratozoospermic, or asthenozoospermic semen analysis results.

There are limited studies on the prediction of semen parameters based on modifiable lifestyle factors by AI methods. However, various non-validated questionnaire forms were used to collect the data. When these forms were analyzed, it could be seen that different parameters were identified as risky and/or modifiable lifestyle behavior without sufficient evidence. The second major limitation of these studies is that WHO guidelines for semen analysis have been updated. The reference limits of semen parameters were revised in the 6th version of the WHO guidelines, therefore the results of these studies have become invalid and the algorithms recommended by these previous analyses should be reconducted. As with the previous reports in the literature, our outputs will become invalid and the algorithms will need to be re-run if the reference limits of semen parameters are updated by later versions of the WHO guidelines. Another limitation of this current study is that our data was obtained from a single center, and we did not validate the ML models with data from external infertility clinics. Another limitation of the prediction of semen quality based on lifestyle behaviors is that various AI methods have been employed and the reports lack information about the development of the models, the various parameters that have been used to find the highest performance to predict the semen quality, and that the health professionals related to infertility may not have the knowledge to fully understand the process and the results.

Conclusion

The ML models used in this study have the potential to predict semen quality based on lifestyles. Studies with larger training datasets obtained from standardized and validated questionnaire forms about lifestyle behavior should be designed and the AI methods should be developed with a wide range of performance parameters. Furthermore, extensive information should be reported about the construction methods of the models to enable clinicians and couples to easily understand the results.