Predicting hearing loss symptoms from Audiometry data using FP-Growth Algorithm and Bayesian Classifier

Abstract:This paper presents the results of applying machine learning algorithms to predict hearing loss symptoms given air and bone conduction audiometry thresholds. FP-Growth (frequent pattern growth) algorithm was employed as a feature extraction technique. The effect of extracting naïve Bayes classifier's vocabulary from patterns generated by FP-Growth algorithm was explored. Both multivariate Bernoulli and multinomial naïve Bayes models were used with and without the feature extraction. The results were validated with repeated random sub-sampling validation performed using 5 partitions with 10, 20, 30, 40 and 50 training examples respectively averaged over 10 iterations. The multivariate Bernoulli model with feature extraction is found to be more accurate in predicting hearing loss symptoms with average error rate of only 0, 0.5, 1, 1.75 and 5.4% for the partitions with 10, 20, 30, 40 and 50 training examples respectively compared to multinomial model with feature extraction. However, the two models with feature extraction produce better results than same models without feature extraction