Abstract：
BACKGROUND: Vestibular migraine (VM) and Menière\'s disease (MD) are two common causes of recurrent spontaneous vertigo. Using history, video-nystagmography and audiovestibular tests, we developed machine learning models to separate these two disorders.
METHODS: We recruited patients with VM or MD from a neurology outpatient facility. One hundred features from six \"feature subsets\": history, acute video-nystagmography and four laboratory tests (video head impulse test, vestibular-evoked myogenic potentials, caloric testing and audiogram) were used. We applied ten machine learning algorithms to develop classification models. Modelling was performed using three \"tiers\" of data availability to simulate three clinical settings. \"Tier 1\" used all available data to simulate the neuro-otology clinic, \"Tier 2\" used only history, audiogram and caloric test data, representing the general neurology clinic, and \"Tier 3\" used history alone as occurs in primary care. Model performance was evaluated using tenfold cross-validation.
RESULTS: Data from 160 patients with VM and 114 with MD were used for model development. All models effectively separated the two disorders for all three tiers, with accuracies of 85.77-97.81%. The best performing algorithms (AdaBoost and Random Forest) yielded accuracies of 97.81% (95% CI 95.24-99.60), 94.53% (91.09-99.52%) and 92.34% (92.28-96.76%) for tiers 1, 2 and 3. The best feature subset combination was history, acute video-nystagmography, video head impulse test and caloric testing, and the best single feature subset was history.
CONCLUSIONS: Machine learning models can accurately differentiate between VM and MD and are promising tools to assist diagnosis by medical practitioners with diverse levels of expertise and resources.
METHODS: We recruited patients with VM or MD from a neurology outpatient facility. One hundred features from six \"feature subsets\": history, acute video-nystagmography and four laboratory tests (video head impulse test, vestibular-evoked myogenic potentials, caloric testing and audiogram) were used. We applied ten machine learning algorithms to develop classification models. Modelling was performed using three \"tiers\" of data availability to simulate three clinical settings. \"Tier 1\" used all available data to simulate the neuro-otology clinic, \"Tier 2\" used only history, audiogram and caloric test data, representing the general neurology clinic, and \"Tier 3\" used history alone as occurs in primary care. Model performance was evaluated using tenfold cross-validation.
RESULTS: Data from 160 patients with VM and 114 with MD were used for model development. All models effectively separated the two disorders for all three tiers, with accuracies of 85.77-97.81%. The best performing algorithms (AdaBoost and Random Forest) yielded accuracies of 97.81% (95% CI 95.24-99.60), 94.53% (91.09-99.52%) and 92.34% (92.28-96.76%) for tiers 1, 2 and 3. The best feature subset combination was history, acute video-nystagmography, video head impulse test and caloric testing, and the best single feature subset was history.
CONCLUSIONS: Machine learning models can accurately differentiate between VM and MD and are promising tools to assist diagnosis by medical practitioners with diverse levels of expertise and resources.
摘要：
背景：前庭性偏头痛（VM）和梅尼埃病（MD）是复发性自发性眩晕的两个常见原因。使用历史记录，视频眼震描记术和听前庭试验，我们开发了机器学习模型来区分这两种疾病。
方法：我们从神经科门诊机构招募了VM或MD患者。来自六个“功能子集”的一百个功能：历史，急性视频眼震描记术和四项实验室测试(视频头部脉冲测试，前庭诱发的肌源性电位，使用热量测试和听力图）。我们应用了十种机器学习算法来开发分类模型。使用三个“数据可用性”层进行建模，以模拟三个临床设置。“第1层”使用所有可用数据来模拟神经耳科诊所，“第2层”仅使用历史记录，听力图和热量测试数据，代表普通神经科诊所，和“第3层”单独使用历史，就像在初级保健中发生的那样。使用十倍交叉验证评估模型性能。
结果：将160例VM患者和114例MD患者的数据用于模型开发。所有模型都有效地将这两种疾病分为三层，准确率为85.77-97.81%。性能最好的算法(AdaBoost和随机森林)的准确率为97.81%(95%CI95.24-99.60)，层1、2、3的94.53%（91.09-99.52%）和92.34%（92.28-96.76%）。最好的特征子集组合是历史，急性视频眼震描记术，视频头脉冲测试和热量测试，最好的单一特征子集是历史。
结论：机器学习模型可以准确区分VM和MD，并且是具有不同专业知识和资源水平的医疗从业者协助诊断的有前途的工具。
方法：我们从神经科门诊机构招募了VM或MD患者。来自六个“功能子集”的一百个功能：历史，急性视频眼震描记术和四项实验室测试(视频头部脉冲测试，前庭诱发的肌源性电位，使用热量测试和听力图）。我们应用了十种机器学习算法来开发分类模型。使用三个“数据可用性”层进行建模，以模拟三个临床设置。“第1层”使用所有可用数据来模拟神经耳科诊所，“第2层”仅使用历史记录，听力图和热量测试数据，代表普通神经科诊所，和“第3层”单独使用历史，就像在初级保健中发生的那样。使用十倍交叉验证评估模型性能。
结果：将160例VM患者和114例MD患者的数据用于模型开发。所有模型都有效地将这两种疾病分为三层，准确率为85.77-97.81%。性能最好的算法(AdaBoost和随机森林)的准确率为97.81%(95%CI95.24-99.60)，层1、2、3的94.53%（91.09-99.52%）和92.34%（92.28-96.76%）。最好的特征子集组合是历史，急性视频眼震描记术，视频头脉冲测试和热量测试，最好的单一特征子集是历史。
结论：机器学习模型可以准确区分VM和MD，并且是具有不同专业知识和资源水平的医疗从业者协助诊断的有前途的工具。
