McKearney Richard M, Bell Steven L, Chesnaye Michael A, Simpson David M
Institute of Sound and Vibration Research, Faculty of Engineering and Physical Sciences, University of Southampton, Southampton, United Kingdom.
Ear Hear. 2022 May/Jun;43(3):949-960. doi: 10.1097/AUD.0000000000001151.
The primary objective of this study was to train and test machine learning algorithms to be able to detect accurately whether EEG data contains an auditory brainstem response (ABR) or not and recommend suitable machine learning methods. In addition, the performance of the best machine learning algorithm was compared with that of prominent statistical detection methods.
Four machine learning algorithms were trained and evaluated using nested k-fold cross-validation: a random forest, a convolutional long short-term memory network, a stacked ensemble, and a multilayer perceptron. The best method was evaluated on a separate test set and compared with conventional detection methods: Fsp, Fmp, q-sample uniform scores test, and Hotelling's T2 test. The models were trained and tested on simulated data that were generated based on recorded ABRs collected from 12 normal-hearing participants and no-stimulus EEG data from 15 participants. Simulation allowed the ground truth of the data ("response present" or "response absent") to be known.
The sensitivity of the best machine learning algorithm, a stacked ensemble, was significantly greater than that of the conventional detection methods evaluated. The stacked ensemble, evaluated using a bootstrap approach, consistently achieved a high and stable level of specificity across ensemble sizes.
The stacked ensemble model presented was more effective than conventional statistical ABR detection methods and the alternative machine learning approaches tested. The stacked ensemble detection method may have potential both in automated ABR screening devices as well as in evoked potential software, assisting clinicians in making decisions regarding a patient's ABR threshold. Further assessment of the model's generalizability using a large cohort of subject recorded data, including participants of different ages and hearing status, is a recommended next step.
本研究的主要目的是训练和测试机器学习算法，使其能够准确检测脑电图（EEG）数据中是否包含听觉脑干反应（ABR），并推荐合适的机器学习方法。此外，将最佳机器学习算法的性能与著名的统计检测方法的性能进行比较。
使用嵌套k折交叉验证对四种机器学习算法进行训练和评估：随机森林、卷积长短期记忆网络、堆叠集成和多层感知器。在一个单独的测试集上评估最佳方法，并将其与传统检测方法进行比较：Fsp、Fmp、q样本均匀分数检验和霍特林T2检验。这些模型在基于从12名听力正常的参与者收集的记录ABR和15名参与者的无刺激EEG数据生成的模拟数据上进行训练和测试。模拟使得数据的真实情况（“存在反应”或“不存在反应”）可知。
最佳机器学习算法（堆叠集成）的灵敏度显著高于所评估的传统检测方法。使用自助法评估的堆叠集成在不同集成规模下始终实现了高且稳定的特异性水平。
所提出的堆叠集成模型比传统的统计ABR检测方法以及所测试的其他机器学习方法更有效。堆叠集成检测方法在自动ABR筛查设备以及诱发电位软件中可能都具有潜力，可协助临床医生做出关于患者ABR阈值的决策。建议下一步使用大量受试者记录数据（包括不同年龄和听力状况的参与者）对该模型的可推广性进行进一步评估。
