This study was approved by the Ethics Committee of Ruijin Hospital, School of Medicine, Shanghai Jiao Tong University (No. 2019-185), and all participants recruited signed informed consent forms. As shown in the patient recruitment flowchart (Figure 1), from October 2019 to September 2021, 6,951 men examined in our center were recruited. Participants with a BMI < 32 and without chromosome abnormalities were included. The exclusion criteria were as follows: prostatic inflammation and organic injury, seminal tract obstruction, cancer, hypospadias, low testosterone levels, varicocele, mumps, cryptorchidism, diabetes, microdeletion of the Y chromosome azoospermia factor (AZF), hyperlipidemia, hypertension, and sexually transmitted diseases. Ultimately, 6,388 men were included after dropping 563 patients who met the exclusion criteria. According to the different evaluations intended, the routine seminal assay including semen volume, sperm concentration, and sperm motility, was performed for all participants, while sperm morphology tests and DNA fragmentation index (DFI) examination was performed in 3,018 and 2,209 participants, respectively. Each participant completed a baseline questionnaire before the semen analysis, and cases with missing questionnaire responses were excluded. Thus, the final dataset included 5,109 men whose semen volume, sperm concentration, and sperm motility were analyzed. Furthermore, sperm morphology and DFI analyses were performed in 2,511 and 1,812 participants, respectively.

The questionnaire comprised 13 items including 10 pertaining to habitual lifestyles and three general conditions including age, abstinence period, and date of questionnaire completion; the details are listed in Supplementary Table 1. Sleeplessness and intensity of sports activity were analyzed using the Insomnia Severity Index (Supplementary Table 2) and modified Physical Activity Questionnaire (Supplementary Table 3) (19), respectively.

Semen samples were collected in sterile plastic container by asking the participants to masturbate. The participants were asked to void urine and wash their hands and external genitalia before masturbating to provide the sample. The sample collected was placed in a water bath maintained at 37°C for 30–60 min for liquefaction. Semen volume was measured by weighing, assuming a semen density of 1.0 g/ml. Sperm concentration (spermatozoa N/mL) and motility (%) were evaluated using a computer-aided sperm analysis system. DFI was determined by flow cytometry after staining with acridine orange, and sperm morphology was investigated using the Diff-Quick staining method. Reference values from the World Health Organization semen analysis manual were used to assess semen characteristics (20), and values below the lower threshold provided in the WHO manual were defined as abnormal. Besides, the threshold of DFI 30% was applied to classify normal (DFI < 30%) or abnormal (DFI ≥ 30%) groups according to a previously published article (21).

The algorithm used in this study was extreme gradient boosting (XGBoost). The feature importance was calculated by the gain method from the XGBoost python library, which worked by averaging training loss reduction caused by feature utilization for each splitting. The input variables were the information collected from the questionnaire of each patient, and the output variables were the semen quality parameters. The input variables were considered categorical variables (Supplementary Table 1), and the output variables were considered dichotomous variables according to the criterion described above. The six semen quality parameters were independent indicators; the XGBoost model was developed using different hyperparameters, separately, to improve the accuracy of the algorithm.

Cross-validation was performed to adjust the parameters. First, a relatively high “learning_rate” was used and the optimum “n_estimators” was selected for this “learning_rate”. Secondly, the parameters “max_depth” and “min_child_weight” were adjusted for the selected “learning_rate” and “n_estimators.” Owing to the unbalanced category of the dataset, the training dataset was oversampled, and the “scale_pos_weight” was always equal to 1. Then, the learning rate was reduced. Next, “max_depth” was adjusted to simplify the XGBoost model according to the results obtained for the test dataset. The clean dataset used for XGBoost was randomly split into training and test datasets in a ratio of 70:30. Hyperparameter details are described in Table 1. Lastly, k-fold cross-validation with k = 10 was performed to evaluate machine learning models.

Descriptive statistics were used to summarize general demographics. The correlations among 13 questionnaire items were evaluated by Pearson's correlation coefficients. For continuous variables, data are expressed as mean ± SD for normally distributed data or median (Interquartile range, IQR) values for non-parametric data. For categorical variables, data are expressed as percentages.

Univariate and multivariable logistic regression was used to identify the factors related to semen quality. For each independent variable, odds ratios (ORs) and 95% confidence intervals (CIs) were estimated. Collinearity analyses were performed before the logistic regression analysis, and the model's goodness-of-fit was graphically evaluated (ROC curves). The response variables were categorized per the method used for the XGBoost algorithm, and stepwise regression was applied for all multivariate logistic regression analyses. Moreover, k-fold cross-validation with k = 10 was performed to evaluate the accuracy of the model.

The univariate and multivariable logistic regression analyses were performed with SAS version 9.4 (SAS Institute, Cary, NC, USA), and k-fold cross-validation was performed with k = 10 by using the package for R (version 4.1.2). Other statistical analyses were performed using SPSS 23.0. P < 0.05 was considered statistically significant.

Figure 2 shows the proportions of participants corresponding to the subgroups for the following questionnaire items: (1) season of semen examination; (2) age; (3) abstinence period; (4) smoking status; (5) alcohol consumption; (6) staying up late; (7) sleeplessness; (8) consumption of pungent food; (9) intensity of sports activity; (10) sedentary lifestyle; (11) work in hot conditions; (12) sauna use in the last 3 months; and (13) exposure to radioactivity.

Among the 5,109 males, the median semen volume, sperm concentration, total sperm count, rapid progressive motility of the sperm, progressive motility of the sperm, and total sperm motility were 3.3 ml (95% CI: 3.40–3.49 ml), 68.1 × 10⁶/ml (95% CI: 79.26–82.84 × 10⁶/ml), 214.5 × 10⁶ (95% CI: 262.28–275.23 × 10⁶), 23.0% (95% CI: 22.76–23.50%), 47.8% (95% CI: 45.50–46.74%), and 60.4% (95% CI: 56.46–57.85%), respectively. The median normal sperm morphology among 2,511 men was 6.0% (95% CI: 6.24–6.51%), and the median DFI of 1,915 men was 14.4% (95% CI: 17.29–18.39%). In addition, 18.2% of the participants showed abnormal sperm morphology (morphologically normal forms, <4.0%, n = 2,511) and 13.9% had high DFI (≥30%, n = 1,812).

We trained XGBoost with the input of the 13 items and achieved 60.7–70.3% accuracy, 55.4–72.5% sensitivity, and 39.9–70.4% specificity for the test set (Table 2).

The AUC of the XGBoost model for semen volume was 0.648 (Figure 3A) and the following cross-validation showed that the AUC of the model was 0.617. The feature importance plotted via XGBoost showed that the maximum score was for smoking status followed by abstinence period and staying up late (Figure 3B). Logistic regression analyses (Figure 3C) revealed that smoking status, abstinence period, sedentary lifestyle, and age were predictive markers of semen volume. The AUC of the combined markers (AUC = 0.655) was higher than that of the individual markers (AUC = 0.465, 0.563, 0.523, and 0.457, respectively), and the following cross-validation based on the multivariate regression analysis showed that the AUC of the model was 0.539. The maximum odds ratio was related to smoking status (OR = 4.69), indicating it to be the most important predictor (Supplementary Table 4). Abstinence period, the second most important factor as revealed by XGBoost, was significantly associated with semen volume in the logistic regression analysis (Supplementary Table 4). Besides, as shown in Figure 3D, the OR per the regression analysis indicated that men who smoked more than 20 cigarettes/day were more likely to have a lower semen volume (OR: 4.69, 95% CI: 3.39–6.49, P < 0.001). However, males who smoked <10 cigarettes/day were less likely to have a lower semen volume (OR: 0.67, 95% CI: 0.48–0.93, P < 0.05) than non-smokers. Men who practiced abstinence for more than 7 days or had a sedentary lifestyle (≥5 h/day) were less likely to have a lower semen volume (OR: 0.63, 95% CI: 0.46–0.87, P < 0.01 and OR: 0.81, 95% CI: 0.65–1.00, P < 0.05, respectively).

The AUC of the XGBoost model for sperm concentration was 0.661 (Figure 4A), and the cross-validation showed that the AUC of the model was 0.674. The feature importance plotted using XGBoost showed that the maximum important score was for smoking status, followed by age and season of semen examination (Figure 4B). The AUCs of the logistic regression analyses (Figure 4C) revealed that smoking status, age, intensity of sports activity, and consumption of pungent food were predictive markers of sperm concentration. The AUC of the combined marker (AUC = 0.680) was higher than those of individual markers (AUC = 0.457, 0.540, 0.519, and 0.489, respectively), and the cross-validation based on the multivariate regression analysis showed that the AUC of the model was 0.547. The maximum odds ratio was observed for smoking status (OR = 6.97), indicating it is the most important predictor (Supplementary Table 5). Age, the second-most important factor revealed by XGBoost, also showed significant association with sperm concentration via logistic regression assay (Supplementary Table 5). Besides, as shown in Figure 4D, males who smoked more than 20 cigarettes/day were more likely to have lower sperm concentrations than non-smokers (OR: 6.97, 95% CI: 5.18–9.37, P < 0.001), but smokers were less likely to have lower sperm concentrations than non-smokers when they smoked < 10 cigarettes/day (OR: 0.13, 95% CI: 0.07–0.22, P < 0.001). Older men (>35 years) were less likely to have lower sperm density (OR: 0.72, 95% CI: 0.57–0.91, P < 0.01) than younger men (<30 years).

The AUC of the XGBoost models for progressive sperm motility was 0.697 (Figure 5A), and the cross-validation showed that the AUC of the model was 0.698. The feature importance plotted using XGBoost showed that smoking status was the most important factor, followed by abstinence period and alcohol consumption (Figure 5B). The AUCs of the logistic regression analyses (Figure 5C) revealed that smoking status, abstinence period, alcohol consumption, age, exposure to radioactivity, and working in hot conditions were predictive markers of progressive sperm vitality. The AUC of the combined marker (AUC = 0.705) was slightly higher than that of other markers, and the cross-validation based on the multivariate regression analysis showed that the AUC of the model was 0.696, which was similar to that of XGBoost (AUC = 0.697).

The top-two odds ratios were observed for smoking status and abstinence period (OR = 11.16 and 2.05), indicating their importance in predictions (Supplementary Table 6). Alcohol consumption, which was identified as the third-most important by XGBoost, also showed a significant association with progressive sperm motility in the logistic regression assay (Supplementary Table 6).

Moreover, as shown in Figure 5D and Supplementary Table 6, males who smoked more than 20 cigarettes/day were more likely to have lower progressive sperm motility (OR: 11.16, 95% CI: 7.82–15.93, P < 0.001) than non-smokers, but smokers were less likely to have lower progressive sperm motility than non-smokers when they smoked <10 cigarettes/day (OR: 0.07, 95% CI: 0.05–0.11, P < 0.001). Males who maintained abstinence for more than 7 days were more likely to show lower progressive sperm motility (OR: 2.05, 95% CI: 1.63–2.57, P < 0.001).

The AUC of the XGBoost models for total sperm vitalities was 0.660 (Figure 6A), and the cross-validation showed that the AUC of the model was 0.686. The feature importance plotted via XGBoost showed that smoking status played the most important part, followed by working in hot conditions and abstinence period (Figure 6B). The AUCs of the logistic regression analyses (Figure 6C) revealed that smoking status, working in hot conditions, abstinence period, season of semen examination, alcohol consumption, consumption of pungent food, age, and exposure to radioactivity were predictive markers of total sperm vitality. The AUC of the combined marker (AUC = 0.700) was slightly higher than that of the other markers, and the cross-validation based on the multivariate regression analysis showed that the AUC of the model was 0.749. The maximum odds ratio was observed for smoking status (OR = 10.35), indicating it was the most important predictor (Supplementary Table 7). Moreover, working in hot conditions and abstinence period, two of the top-three important factors revealed by XGBoost, also significantly affected total sperm motility in the regression analysis (Supplementary Table 7).

As shown in Figure 6D and Supplementary Table 7, males who smoked more than 20 cigarettes/day were more likely to have lower total sperm motility than non-smokers (OR: 10.35, 95% CI: 7.35–14.56, P < 0.001), but smokers were less likely to have a lower total sperm motility than non-smokers when they smoked <10 cigarettes/day (OR: 0.06, 95% CI: 0.03–0.10, P < 0.001). Males who worked under hot conditions were less likely to show low total sperm motility (OR: 1.63, 95% CI: 1.20–2.21, P < 0.05). Moreover, males who maintained abstinence for more than 7 days were more likely to have lower total sperm motility (OR: 1.72, 95% CI: 1.37–2.17, P < 0.001).

The AUC of the XGBoost model for sperm morphology was only 0.506 (Figure 7A), and the cross-validation showed that the AUC of the model was 0.520. The feature importance plot created using XGBoost showed that smoking status was the maximum important factor (Figure 7B). The AUCs of the logistic regression analyses (Figure 7C) revealed that smoking status was a predictive index for sperm morphology with a poor AUC (0.539), and the cross-validation based on the multivariate regression analysis showed that the AUC of the model was 0.543. As shown in Figure 7D and Supplementary Table 8, males who smoked more than 20 cigarettes/day were more likely to have abnormal sperm morphology than non-smokers (OR: 3.0, 95% CI: 1.76–5.12, P < 0.001), but the trend did not appear for males who smoked <10 cigarettes/day (OR: 0.53, 95% CI: 0.39–0.73, P < 0.001).

The AUC of the XGBoost model for DFI was 0.686 (Figure 8A), and the cross-validation showed that the AUC of the model was 0.697. The top three important features affecting total sperm vitality were age, abstinence period, and smoking status (Figure 8B). The AUCs of the logistic regression analyses (Figure 8C) revealed that age, abstinence period, smoking status, and staying up late were predictive markers of sperm DFI. The AUC of the combined marker (AUC = 0.725) was higher than that of the other individual markers (AUC = 0.661, 0.598, 0.466, and 0.443). The cross-validation based on the multivariate regression analysis showed that the AUC of the model was 0.648. The top-two odds ratios appeared for age and abstinence period (OR = 5.47 and 3.61), indicating their important predictive roles. Smoking status, the third important factor revealed by XGBoost, was also shown to significantly affect sperm DFI in regression analysis (Supplementary Table 9).

Besides, as shown in Figure 8D and Supplementary Table 9, older males (>35 years) and those maintaining abstinence for more than 7 days were more likely to have higher DFI (OR: 5.47, 95% CI: 3.41–8.76, P < 0.001 and OR: 3.61, 95% CI: 2.27–5.75, P < 0.001) than younger males (<30 years old) and those maintaining abstinence for <4 days, respectively. Males who smoked <10 cigarettes /day were less likely to have a high DFI (OR: 0.42, 95% CI: 0.27–0.66, P < 0.001) than non-smokers. Nevertheless, when they smoked more than 20 cigarettes/day, the odds ratio of having high a DFI increased (P > 0.05).

Considering the close relationships among the variables, Spearman rank correlation analysis was performed. As shown in Supplementary Table 10, significant positive correlations were observed between sedentary lifestyle and age, staying up late, sleeplessness, consumption of pungent food, and exposure to radioactivity, with correlation coefficient (ICC) values of 0.043, 0.078, 0.056, 0.061, and 0.438, respectively. Meanwhile, sedentary lifestyles showed negative correlations with smoking status, intensity of sports activity, and working in hot conditions (ICC = −0.088, −0.119, and −0.134, respectively). Positive correlations were observed between staying up late and smoking status, alcohol consumption, sleeplessness, consumption of pungent food, sedentary lifestyle, working in hot conditions, sauna use in the last 3 months, and exposure to radioactivity (ICC = 0.185, 0.238, 0.310, 0.342, 0.078, 0.087, 0.067, and 0.034; P < 0.05), but staying up late showed negative correlations with age, abstinence period, and intensity of sports activity (ICC = −0.074, −0.055, −0.067; P < 0.05). Exposure to radioactivity showed positive correlations with staying up late, consumption of pungent food, and a sedentary lifestyle (ICC = 0.034, 0.046, and 0.438; P < 0.05), but showed negative correlations with smoking status, alcohol consumption, and working in hot conditions (ICC = −0.142, −0.028, and −0.109).