Data mining of audiology data to find patients who might benefit from ITE hearing aids or tinnitus maskers

We describe our work on the data mining of hearing aid patient data to answer the following research questions: * Which factors influence the choice of ITE (in the ear) as opposed to BTE (behind the ear) hearing aids? * For patients diagnosed with tinnitus, which factors influence the decision whether to fit a tinnitus masker? Our data set is 180,000 patient records provided by the hearing aid clinic at James Cook University Hospital, Middlesbrough, UK. The records contain fixed vocabulary fields such as diagnosis, audiograms and short free text notes. The data set is unusual in that many of the patients were prescribed ITE hearing aids, which are not generally available on the National Health Service in the UK. Using PCA (principal component analysis) and the chi-squared test, we found that flat hearing loss audiograms were associated with ITE aids and audiograms with airbone gaps were associated with BTE aids. We also found that males tended to use ITE aids while females tended to use BTE aids. There was a positive association between ITE and age below 70. Patients with severe hearing loss tended to use BTE hearing aids while patients with a mild to moderate hearing loss tended to use ITE hearing aids. An analysis of the free text notes showed that ITE hearing aid types tended to use lacquer, had vents, required reshelling of ear impressions, had changes made to the hearing aid itself, were reviewed and the wearers were making progress. There was no association found for gender and the use of tinnitus maskers. The keywords ‘tinnitus’ and ‘masker’ were found together in the free text notes significantly more often in patients with mild to moderate hearing losses, and significantly less often when the patient wore a BTE hearing aid or was aged 54 or less. We combined our full set of data attributes (audiograms, gender, age, diagnosis and free text keywords), using logistic regression and a Naive Bayesian approach in two separate experiments for both of our two research questions. The resulting models can be used as the basis of decision support systems, where the inputs are hearing aid clinic data for new patients, and the outputs are either the estimated probability that the patient requires an ITE aid or that he or she requires a tinnitus masker.