TY - JOUR
T1 - A comparison of supervised classification methods for auditory brainstem response determination.
AU - McCullagh, Paul
AU - Wang, Haiying
AU - Zheng, Huiru
AU - Lightbody, G
AU - McAllister, G
PY - 2007
Y1 - 2007
N2 - The ABR is commonly used in the Audiology clinic to determine and quantify hearing loss. Its interpretation is subjective, dependent upon the expertise and experience of the clinical scientist. In this study we investigated the role of machine learning for pattern classification in this domain. We extracted features from the ABRs of 85 test subjects (550 waveforms) and compared four complimentary supervised classification methods: Naïve Bayes, Support Vector Machine Multi-Layer Perceptron and KStar. The Abr dataset comprised both high level and near threshold recordings, labeled as 'response' or 'no response' by the human expert. Features were extracted from single averaged recordings to make the classification process straightforward. A best classification accuracy of 83.4% was obtained using Naïve Bayes and five relevant features extracted from time and wavelet domains. Naïve Bayes also achieved the highest specificity (86.3%). The highest sensitivity (93.1%) was obtained with Support Vector Machine-based classification models. In terms of the overall classification accuracy, four classifiers have shown the consistent, relatively high performance, indicating the relevance of selected features and the feasibility of using machine learning and statistical classification models in the analysis of ABR.
AB - The ABR is commonly used in the Audiology clinic to determine and quantify hearing loss. Its interpretation is subjective, dependent upon the expertise and experience of the clinical scientist. In this study we investigated the role of machine learning for pattern classification in this domain. We extracted features from the ABRs of 85 test subjects (550 waveforms) and compared four complimentary supervised classification methods: Naïve Bayes, Support Vector Machine Multi-Layer Perceptron and KStar. The Abr dataset comprised both high level and near threshold recordings, labeled as 'response' or 'no response' by the human expert. Features were extracted from single averaged recordings to make the classification process straightforward. A best classification accuracy of 83.4% was obtained using Naïve Bayes and five relevant features extracted from time and wavelet domains. Naïve Bayes also achieved the highest specificity (86.3%). The highest sensitivity (93.1%) was obtained with Support Vector Machine-based classification models. In terms of the overall classification accuracy, four classifiers have shown the consistent, relatively high performance, indicating the relevance of selected features and the feasibility of using machine learning and statistical classification models in the analysis of ABR.
M3 - Article
VL - 129
SP - 1289
EP - 1293
JO - Studies in Health Technology and Informatics - Medicine Meets Virtual Reality 12
JF - Studies in Health Technology and Informatics - Medicine Meets Virtual Reality 12
SN - 0926-9630
IS - Pt 2
ER -