Scholarlink

Objective <p>Cervical cancer, a global public health burden, is linked to persistent high-risk HPV infection. Multiple HPV co-infections (≥ 2 genotypes) elevate high-grade cervical lesion risk, but current tools lack accuracy for coinfection risk assessment. This study developed a vaginal microenvironment-based machine learning (ML) model to predict multiple HPV co-infections for early risk stratification.</p> Methods <p>A retrospective cohort of 1,682 HPV-infected patients (Jan 2024–Jan 2025, General Hospital of Northern Theater Command) was split into training (70%) and validation (30%) sets. Variables included age, BMI, cervical lesions, 11 vaginal biomarkers, and 23 HPV genotypes. LASSO regression was performed to screen 14 candidate variables, followed by multivariate logistic regression to identify independent risk factors. For variables with rare events, Firth regression was used to validate their robustness. Nine ML algorithms were trained/validated using ROC-AUC, DCA, and SHAP.</p> Results <p>LASSO identified 9 core predictors, with 7 independent factors: cervical lesions, leukocytosis, fungal hyphae, Trichomonas vaginalis, sialidase, leukocyte esterase, and H<sub>2</sub>O<sub>2</sub>. These 9 core predictors identified by LASSO were incorporated into the 9 machine learning models. The NNET model performed best: AUC = 0.876 (95% CI 0.852–0.900), sensitivity = 96.9%, specificity = 75.3%. DCA confirmed clinical net benefit at 1–85% risk thresholds. SHAP analysis highlighted age, H<sub>2</sub>O<sub>2</sub>, and cervical lesions as key predictors.</p> Conclusion <p>A ML model integrating vaginal biomarkers and clinical features was developed, with NNET demonstrating better predictive accuracy for clinical decision support.This model enables early risk stratification for multiple HPV co-infections, guiding personalized screening intervals or therapeutic interventions for high-risk subgroups.</p> Graphical Abstract <p></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

Vaginal Microenvironment-Driven neural network model for predicting multiple HPV co-infections: a machine learning clinical decision framework

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年10月14日

查看全文
  • Li XiaoNan,
  • Chen MeiHan,
  • Hou Dan

摘要

Objective

Cervical cancer, a global public health burden, is linked to persistent high-risk HPV infection. Multiple HPV co-infections (≥ 2 genotypes) elevate high-grade cervical lesion risk, but current tools lack accuracy for coinfection risk assessment. This study developed a vaginal microenvironment-based machine learning (ML) model to predict multiple HPV co-infections for early risk stratification.

Methods

A retrospective cohort of 1,682 HPV-infected patients (Jan 2024–Jan 2025, General Hospital of Northern Theater Command) was split into training (70%) and validation (30%) sets. Variables included age, BMI, cervical lesions, 11 vaginal biomarkers, and 23 HPV genotypes. LASSO regression was performed to screen 14 candidate variables, followed by multivariate logistic regression to identify independent risk factors. For variables with rare events, Firth regression was used to validate their robustness. Nine ML algorithms were trained/validated using ROC-AUC, DCA, and SHAP.

Results

LASSO identified 9 core predictors, with 7 independent factors: cervical lesions, leukocytosis, fungal hyphae, Trichomonas vaginalis, sialidase, leukocyte esterase, and H2O2. These 9 core predictors identified by LASSO were incorporated into the 9 machine learning models. The NNET model performed best: AUC = 0.876 (95% CI 0.852–0.900), sensitivity = 96.9%, specificity = 75.3%. DCA confirmed clinical net benefit at 1–85% risk thresholds. SHAP analysis highlighted age, H2O2, and cervical lesions as key predictors.

Conclusion

A ML model integrating vaginal biomarkers and clinical features was developed, with NNET demonstrating better predictive accuracy for clinical decision support.This model enables early risk stratification for multiple HPV co-infections, guiding personalized screening intervals or therapeutic interventions for high-risk subgroups.

Graphical Abstract