Scholarlink

Background <p>Y chromosome microdeletions (YCMs) are the second most prevalent genetic determinant of male infertility, but validated predictive models are limited.</p> Methods <p>We retrospectively enrolled 4096 male infertility patients (October 2013–October 2023) for model development and randomly divided them into training (80%, n = 3276) and testing (20%, n = 820) sets. A second, later-period validation cohort of 333 patients from the same institution (October 2024–October 2025) was used to evaluate model performance over time. A multi-stage variable selection strategy, including univariate analysis, LASSO regression, stepwise regression, and best subset selection, was applied to identify key predictors. Six predictive models, incorporating logistic regression, random forest, and XGBoost algorithms, were developed and validated using a three-tier framework assessing discrimination, calibration, and clinical utility.</p> Results <p>Variable selection identified sperm concentration (AUC = 0.703, P &lt; 0.001) and follicle-stimulating hormone (FSH; AUC = 0.618, P = 0.012) as the key predictors of YCMs. Among six models evaluated by tenfold cross-validation, the dual-variable XGBoost model combining FSH and sperm concentration achieved near-maximal discrimination (mean AUC 0.766 ± 0.037, sensitivity 0.766, specificity 0.653, accuracy 0.661), a modest but statistically significant improvement over the sperm concentration-based logistic regression model (AUC 0.705 vs 0.766; + 8.6%; DeLong P &lt; 0.001), though with lower sensitivity. In the later-period cohort, the model achieved an AUC of 0.762 (95% CI: 0.644–0.879), sensitivity 0.833 and specificity 0.702. Decision curve analysis suggested clinical net benefit across the 5–30% threshold range.</p> Conclusions <p>The dual-variable XGBoost model, incorporating FSH and sperm concentration, demonstrates moderate predictive performance for YCMs, evaluated by tenfold cross-validation and a second, later-period validation cohort. This prescreening tool using routine clinical data may help prioritize genetic testing and improve diagnostic efficiency in male infertility management. However, the relatively low positive predictive value and the wide later-period validation confidence interval indicate that prospective multicenter validation is warranted before broader potential clinical application.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Predicting Y chromosome microdeletions in male infertility patients using machine learning models: a retrospective clinical study

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月5日

查看全文
  • Qi-Tao Zhan,
  • Shui-Bo Shi,
  • Wei-Kang Chen,
  • Yuan-Yuan Zhou,
  • Feng‑Bin Zhang,
  • Jing‑Ping Li

摘要

Background

Y chromosome microdeletions (YCMs) are the second most prevalent genetic determinant of male infertility, but validated predictive models are limited.

Methods

We retrospectively enrolled 4096 male infertility patients (October 2013–October 2023) for model development and randomly divided them into training (80%, n = 3276) and testing (20%, n = 820) sets. A second, later-period validation cohort of 333 patients from the same institution (October 2024–October 2025) was used to evaluate model performance over time. A multi-stage variable selection strategy, including univariate analysis, LASSO regression, stepwise regression, and best subset selection, was applied to identify key predictors. Six predictive models, incorporating logistic regression, random forest, and XGBoost algorithms, were developed and validated using a three-tier framework assessing discrimination, calibration, and clinical utility.

Results

Variable selection identified sperm concentration (AUC = 0.703, P < 0.001) and follicle-stimulating hormone (FSH; AUC = 0.618, P = 0.012) as the key predictors of YCMs. Among six models evaluated by tenfold cross-validation, the dual-variable XGBoost model combining FSH and sperm concentration achieved near-maximal discrimination (mean AUC 0.766 ± 0.037, sensitivity 0.766, specificity 0.653, accuracy 0.661), a modest but statistically significant improvement over the sperm concentration-based logistic regression model (AUC 0.705 vs 0.766; + 8.6%; DeLong P < 0.001), though with lower sensitivity. In the later-period cohort, the model achieved an AUC of 0.762 (95% CI: 0.644–0.879), sensitivity 0.833 and specificity 0.702. Decision curve analysis suggested clinical net benefit across the 5–30% threshold range.

Conclusions

The dual-variable XGBoost model, incorporating FSH and sperm concentration, demonstrates moderate predictive performance for YCMs, evaluated by tenfold cross-validation and a second, later-period validation cohort. This prescreening tool using routine clinical data may help prioritize genetic testing and improve diagnostic efficiency in male infertility management. However, the relatively low positive predictive value and the wide later-period validation confidence interval indicate that prospective multicenter validation is warranted before broader potential clinical application.