This paper proposed a novel hearing loss detection method. Our method first used seven Hu moment invariants to extract features. Afterwards, we used support vector machine to act as the classifier. The 10x5-fold cross validation shows our method yielded an overall accuracy of 77.47± 1.17%. The sensitivities of healthy control, left-sided hearing loss, and right-sided hearing loss are 77.60± 5.72%, 77.60± 4.30%, and 77.20± 5.98%, respectively. In all, our method is effective in hearing loss identification.

Abstract-In order to develop a new hearing loss detection method, this paper proposed to combine wavelet entropy with feedforward neural network trained by genetic algorithm. The dataset contains 72 subjects-24 healthy controls, 24 left-sided hearing loss patients, and 24 right-sided hearing loss patients. The 10 runs of 8-fold cross validation showed that optimal decomposition level was 4, better than the results using decomposition level of 2, 3, and 5. Our method using 4-level decomposition yielded a sensitivity for healthy controls of 81.25±4.91%, a sensitivity for left-sided hearing loss of 80.42±5.57%, a sensitivity for right-sided hearing loss of 81.67±6.86%, and an overall accuracy of 81.11±1.34%.