基于机器学习预测突发性聋伴高血压患者预后

A machine learning-based study to predict poor prognosis in patients with sudden deafness with hypertension

摘要:

杨瑾1 米彦芳[1]:目的:分析突发性聋伴高血压患者发生不良预后的影响因素,基于4种机器学习算法构建预测模型,比较不同模型的预测效果。 方法:选取2023年2月1日-2024年5月31日就诊于郑州大学第二附属医院耳鼻喉科的合并高血压的突发性聋患者为研究对象。通过医院电子病历系统收集患者临床资料,根据治疗后听力恢复情况分为有效组和无效组,采用单因素分析、最小绝对收缩和选择算子(LASSO 回归)和Boruta算法筛选预测变量。以8:2的比例将患者随机分为训练集和验证集。采用4种机器学习算法［logistic 回归(LR)随机森林(RF)、极端梯度提升(XGBoost)和支持向量机(SVM)］构建预测模型并进行验证,比较4种模型的受试者工作特征曲线下面积(AUC)、敏感性、特异性等指标,采用Delong检验比较各模型在测试集中的AUC,并确定最佳模型。? 结果: 232例患者纳入研究,共筛选出7个与无效预后密切相关的变量,包括听力损失程度、听力图类型、糖尿病、高血压病程等。对4种模型的预测性能进行验证,发现Xgboost模型(AUC=0.787,95%CI:0.642-0.931)预测性能最佳,4种模型ROC曲线之间AUC的差异无统计学意义。? 结论: 合并高血压的突发性聋患者不良预后的发生风险受患者初始听力水平、听力图类型、糖尿病、高血压病程、吸烟及高尿酸血症等因素影响。4种机器学习模型均具有良好的预测性能,其中XGboost模型表现预测性能最优。

Abstract:

Objective: To analyse the factors influencing the occurrence of poor prognosis in patients with sudden deafness with hypertension, to construct a prediction model based on four machine learning algorithms, and to compare the predictive performance of different models. Methods: The study design was retrospective. Sudden deafness patients with combined hypertension who attended the Department of Otorhinolaryngology, Second Affiliated Hospital of Zhengzhou University during February 2023 to May 2024 were selected as study subjects. Patients" clinical data were collected through the hospital"s electronic medical record system. Patients were divided into effective and ineffective groups according to hearing recovery after treatment, and predictor variables were examined using one-way analysis, least absolute shrinkage and selection operator (LASSO regression), and Boruta algorithm. Patients were randomly assigned to training and validation sets in an 8:2 ratio. Based on the training set data, four machine learning algorithms [logistic regression (LR), random forest (RF), extreme gradient boosting (XGBoost) and support vector machine (SVM)] were used to construct the prediction models; based on the validation set data, the area under the curve (AUC) of the subjects" working characteristic, sensitivity, specificity, accuracy and F1 score of the four models were compared. The Delong test was used to compare the AUC of each model in the test set and to determine the best model.? Results: A total of 232 patients were included in the study. Univariate analysis, LASSO regression and the Boruta algorithm identified a total of 7 variables (the degree of hearing loss, audiogram type, duration of diabetes mellitus, and the course of hypertension,etc.) that were strongly associated with poor prognosis must be taken into account. The predictive performance of the four models was validated based on the validation set data, and it was found that the Xgboost model (AUC=0.787, 95% CI: 0.642-0.931) exhibited the most favourable predictive performance, and the difference in AUC between the ROC curves of the four models was not statistically significant.? Conclusion: The risk of poor prognosis for sudden deafness in combination with hypertension is influenced by a number of factors, including initial hearing level, audiogram type, diabetes mellitus, duration of hypertension, smoking, and hyperuricemia, among others. The four machine learning models demonstrated satisfactory predictive performance, with the XGboost model exhibiting the most optimal predictive performance.