绝经后女性良性阵发性位置性眩晕复发的预测：一项基于机器学习的临床研究

文献信息

中文摘要

背景： 良性阵发性位置性眩晕（BPPV）是一种常见的周围性前庭疾病，复发率高，严重影响患者的生活质量。多项研究已证实，绝经后女性由于内分泌紊乱及其他生理特点，是BPPV发病和复发的高危人群。然而，目前缺乏高效的复发预测模型，传统评估方法效能有限，难以满足个体化干预的需求。本研究旨在构建并验证用于预测绝经后女性BPPV复发的机器学习模型，识别关键危险因素，为临床早期识别高危人群提供依据。
方法： 我们回顾性分析了2023年1月至2025年12月期间在四川省三家医院确诊并成功治疗的BPPV患者数据。患者按7：3的比例分为训练集和验证集。采用LASSO回归和Boruta算法筛选特征。构建了六种机器学习模型：逻辑回归（LR）、极端梯度提升（XGBoost）、支持向量机（SVM）、多层感知机（MLP）、梯度提升机（GBM）和元集成方法。比较模型预测性能，并通过SHapley加性解释（SHAP）分析评估模型可解释性。
结果： 通过LASSO和Boruta双重算法筛选，确定了四个模型选择的绝经后女性BPPV复发预测因子：骨质疏松、血清钙、维生素D和雌二醇。在评估的六种机器学习算法中，元集成模型的AUC点估计值最高，为0.857（95%置信区间[CI]，0.779-0.928），整体性能协调一致。随后的SHAP分析阐明，血清维生素D、钙和雌二醇水平降低，加之存在骨质疏松，是驱动高复发风险预测的最关键因素。
结论： 元集成模型能够有效预测绝经后女性BPPV复发风险。骨质疏松、血清钙、维生素D和雌二醇这四项指标为临床早期识别复发高危人群提供了客观依据，有助于制定个体化随访策略和干预方案。

英文摘要

BACKGROUND: Benign paroxysmal positional vertigo (BPPV) is a common peripheral vestibular disorder with a high recurrence rate that significantly impairs patients' quality of life. Multiple studies have confirmed that postmenopausal women, due to endocrine disorders and other physiological characteristics, represent a high-risk population for both BPPV onset and recurrence. However, efficient recurrence prediction models are currently lacking, and traditional assessment methods have limited efficacy, making it difficult to meet the needs of individualized intervention. The aim of this study was to construct and validate machine learning models for predicting BPPV recurrence in postmenopausal women, identify key risk factors, and provide evidence for early identification of high-risk populations in clinical practice.
METHODS: We retrospectively analyzed data from BPPV patients diagnosed and successfully treated at three hospitals in Sichuan Province between January 2023 and December 2025. Patients were divided into training and validation sets at a 7:3 ratio. Features were screened using LASSO regression and the Boruta algorithm. Six machine learning models were constructed: Logistic Regression (LR), Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), Multilayer Perceptron (MLP), Gradient Boosting Machine (GBM), and Meta-Ensemble method. Model predictive performance was compared, and model interpretability was assessed through SHapley Additive exPlanations (SHAP) analysis.
RESULTS: Through dual algorithmic screening, LASSO, and Boruta delineated four model-selected predictors for BPPV recurrence in postmenopausal women: osteoporosis, serum calcium, vitamin D, and estradiol. Among the six machine learning algorithms evaluated, the Meta-Ensemble model yielded the highest AUC point estimate of 0.857 (95% confidence interval [CI], 0.779-0.928), delivering a harmonized overall performance. Subsequent SHAP analysis elucidated that decreased serum levels of vitamin D, calcium, and estradiol, compounded by the presence of osteoporosis, were the most critical factors driving high recurrence risk predictions.
CONCLUSION: The Meta-Ensemble model can effectively predict BPPV recurrence risk in postmenopausal women. The four indicators of osteoporosis, serum calcium, vitamin D, and estradiol provide objective evidence for early clinical identification of high-risk populations for recurrence, facilitating the development of individualized follow-up strategies and intervention protocols.