Performance Evaluation of Deep Learning Techniques on Classifying Commercial Sensor-based Auditory Brainstem Response

To objectively evaluate the health condition of the auditory pathway, Auditory Brainstem Response (ABR) is used in the clinic. The clinical diagnosis is concluded by the audiologist according to the morphology and characteristic waves of ABR through manual screening. It is a time-consuming and subjective task that not only reduces the efficiency of the physicians but also introduces subjective error which might cause an incorrect diagnosis. Therefore, it is necessary to propose an objective method that can solve the abovementioned problems. Deep Learning (DL) technique develops fast and is widely used in many fields, including biomedical engineering. Hence, DL might help with the objective classification of ABR and release the physicians from the heavy workload. In this paper, we evaluated the classification performance of three commonly used DL methods, namely EEGNet, DeepConvNet and ShallowConvNet, on several metrics. The results showed that on the task of classifying normal and abnormal ABR signals, EEGNet had the best performance, with the accuracy of 0.9995 ± 0.0011, the precision of 0.9995 ± 0.0011, the recall of 0.9995 ± 0.0011, the F1-score of 0.9995 ± 0.0011, and the AUC of 1.0000 ± 0.0000. It is believed that DL can help with objective ABR classification and release the physicians' workload.