Abstract
Several pathological or genetic factors can produce a deafness. To establish the origin and the degree of hearing loss, the clinician makes his diagnosis on the speech and pure-tone audiogram. Indeed the pure-tone audiogram is a good indication of the loss of the outer hair cells which are responsible for the cochlear amplification of the sounds. However it doesn't bring the knowledge about the loss of the comprehension speech, especially with loud environments. Even if the hearing aids help the patient to hear in calm places, it is not that clear in noisy ones. The idea is to detect new audiologic profiles which would be richer and more homogeneous according to the prosthetic base. The main purpose of this thesis is to develop new machine learning methods based on the cohort of Amplifon that will help the hearing care professional to adjust the hearing aids. The final goal is to propose to the patients, a personalised procedure when they first arrive at Amplifon. It will optimise their comfort during the change in environment. The first approach is to cluster the patients into similar audiometric groups which could be dealt as particular deafness (genetic origins for example). In this case, the following step would be to select the right number of groups by applying the best clustering algorithms. Another possibility is to attempt to get the scores of the prosthetic adaptation. Actually, it would assist the hearing care professional (after checking the patient results) to obtain a measure which quantifies the degree of the prosthetic satisfaction according to the hearing aid foreseen. Another secondary goal is to look for a prediction algorithm based on the available data which would be capable of computing this score. At last, if the prediction analysis presents poor performances, we will focus on improving the prosthetic adjustment. This final step is bound to the domain of the prescriptive analysis and it consists on integrating external elements to the cohort which matter for the patient (the aesthetic, ...).