Department of Otolaryngology-Head and Neck Surgery, Institute of Otolaryngology, Chinese PLA General Hospital, Beijing, China.
Medical Support Center, Chinese PLA General Hospital, Beijing, China.
Clin Otolaryngol. 2018 Jun;43(3):868-874. doi: 10.1111/coa.13068. Epub 2018 Feb 20.
Sudden sensorineural hearing loss (SSHL) is a multifactorial disorder with high heterogeneity, thus the outcomes vary widely. This study aimed to develop predictive models based on four machine learning methods for SSHL, identifying the best performer for clinical application.
Single-centre retrospective study.
Chinese People's liberation army (PLA) hospital, Beijing, China.
A total of 1220 in-patient SSHL patients were enrolled between June 2008 and December 2015.
An advanced deep learning technique, deep belief network (DBN), together with the conventional logistic regression (LR), support vector machine (SVM) and multilayer perceptron (MLP) were developed to predict the dichotomised hearing outcome of SSHL by inputting six feature collections derived from 149 potential predictors. Accuracy, precision, recall, F-score and the area under the receiver operator characteristic curves (ROC-AUC) were exploited to compare the prediction performance of different models.
Overall the best predictive ability was provided by the DBN model when tested in the raw data set with 149 variables, achieving an accuracy of 77.58% and AUC of 0.84. Nevertheless, DBN yielded inferior performance after feature pruning. In contrast, the LR, SVM and MLP models demonstrated opposite trend as the greatest individual prediction powers were obtained when included merely three variables, with the ROC-AUC ranging from 0.79 to 0.81, and then decreased with the increasing size of input features combinations.
With the input of enough features, DBN can be a robust prediction tool for SSHL. But LR is more practical for early prediction in routine clinical application using three readily available variables, that is time elapse between symptom onset and study entry, initial hearing level and audiogram.
突发性聋（SSHL）是一种具有高度异质性的多因素疾病，因此其预后差异很大。本研究旨在基于四种机器学习方法建立 SSHL 的预测模型，以确定最适合临床应用的模型。
单中心回顾性研究。
中国北京解放军医院。
2008 年 6 月至 2015 年 12 月共纳入 1220 例住院 SSHL 患者。
应用先进的深度学习技术——深度置信网络（DBN），结合传统的逻辑回归（LR）、支持向量机（SVM）和多层感知器（MLP），输入 149 个潜在预测因子得出的 6 个特征集，对 SSHL 的二分类听力结果进行预测。采用准确率、精确率、召回率、F1 评分和受试者工作特征曲线（ROC-AUC）下面积来比较不同模型的预测性能。
在原始数据集（包含 149 个变量）中，DBN 模型的预测能力最佳，总准确率为 77.58%，AUC 为 0.84。但经过特征筛选后，DBN 的性能有所下降。相比之下，LR、SVM 和 MLP 模型则表现出相反的趋势，当仅纳入 3 个变量时，其预测效能最大，ROC-AUC 范围为 0.79~0.81，随后随着输入特征组合数量的增加而逐渐下降。
在输入足够的特征时，DBN 可作为一种强大的 SSHL 预测工具。但在常规临床应用中，LR 更适合使用三个易于获得的变量进行早期预测，即症状出现到研究入组的时间间隔、初始听力水平和听力图。
