 Hi, I am Raul Sanchez Lopez and I'm going to present hearing deficits and auditory profiling, data-driven approaches towards personalized audiology. What I'm going to present today is based on three studies that I have done with the supervisors of my PSD, so Torsten Dao, Sebastian Santuget, Mihał Feriskowski and Federica Bianchi, and also with Tobias Nier and in the last study with Will Bitmer. And these three studies are not actually in line with machine learning or big data. These are more in under the umbrella of knowledge discovery in databases, where basically what we do is to gain knowledge from the analysis of databases, rather than using predictions or aiming to have millions of observations like in big data. So what we want to investigate is related to hearing deficits. The sensory neural hearing loss is complex. And the hearing deficits are typically characterized by the loss of sensitivity, so the audiogram. But these are the consequences of auditory impairments. If there are problems in the auditory nerves, in the hair cells or in the street of ascularies, this will lead to a loss of sensitivity. On the other end, there are supra-threshold deficits, so even when the sound is audible, we can have deficits in terms of loudness perception, specular temporal resolution, binaural processing abilities, or speech perception. In audiology, the focus has been mainly in the loss of sensitivity. But in these studies, we want to shift the focus to the supra-threshold deficits. And the question for rehabilitation is, can we compensate for individual deficits? So in our view, compensating for individual deficits is really difficult. So we aim to have a precision audiology by a stratification. And I'm going to show you what I mean by that. So from a given listener, we would get a lot of information to create an auditory profile. And this auditory profile is not only based in the audiogram, but also in other tests. For example, the discrimination of tones, or speech perception, or the discrimination of complex sounds. We want to identify groups where the listeners are really similar to each other in terms of hearing deficits and really dissimilar to the other groups, and then provide targeted treatments for each of these groups. The first challenge that we face in this perspective is how to classify the listeners. And then we came up with the idea of profiling. And we use profiling for a certification. And I'm going to show you what I mean with this example. Imagine that we have a heterogeneous group of people. And then we can identify the two main sources of variability, in this case, things on the body and things on the face. And on top of that, we have archetypal people that they are in the corners of this bidimensional space. Then if we look again at our heterogeneous group, we can then form groups where each of the groups have listeners, or in this case, people that they are really similar to each other, and they are dissimilar to each other. And also, they are dissimilar and characterized by these two main sources of variability. If we apply this to audiology or to the hearing deficits, the two main sources of variability we call them distortions. And it's an abstract dimension called distortion type one and distortion type two. And we would have some listeners with a low degree of these two distortions and some other listeners with a high degree of these two distortions. And also, interestingly, we would have listeners with a high degree of one of the distortion and a low degree of the other distortions. If we can classify a given listener in one of these four groups, we would know better what are the hearing deficits. In these studies, we have the following research questions. Is it valid to classify the patients in these hypothesized auditory profiles? And which are the hearing deficits that characterize each of these four groups. And last but not least, does each auditory profile have different needs in terms of hearing rehabilitation. So, in order to explore this, we design a method for data driven profiling. And this is based on three stages. In the first stage, we try to identify the two main sources of variability for doing that we use principal component analysis. Then to identify these archetypal patterns and hidden patterns in the data we use archetypal analysis that is particularly useful to identify extreme action plugs. The profile identification or the clustering that we use was based on the similarity of each of individual observations to the archetypal patterns. We run a first study by analyzing the data of two previous studies. When we analyze these two data sets with the method, the method was promising and we could identify groups that they were doing what we wanted. So, finding similarity within the group, the similarity across the groups. But there were discrepancies in terms of the results and also this was before because there were different types of participants and auditory tests in each of the two studies that we reanalyze. To solve this, we run a new study. And here we have more control and we aim to have a reunion group of listeners and also more auditory tests. We invited 75 listeners with different hearing abilities from mild hearing losses to severe hearing losses. And then we use a test battery with potential for clinical implementation that was pinpointing in different aspects of auditory processing. Audibility, speech perception, binaural processing abilities, spectral temporal modulation, spectral temporal resolution, and loudness perception. We use the same method, but we learn from the previous study that individual data points can influence the data quite a lot. So we took these three steps and we run them iteratively with only part of the data. So we decimated the data in terms of tests and in terms of listeners and we run this 1000 times. With this, we computed the probability of each individual listener to belong or to be similar to the ones in the corner to the archetypal patterns. And then we can plot them in this b-dimensional space. So the higher the probability, the closer to the corners they are. These are the results and each of the numbers correspond to one individual listener. And as you can see, we can find some clusters in the data that they are quite close to the corners. To characterize each of them, we take these listeners that they were really close to each other, and then we plot them in this other plot. Here we have the archetypal patterns. In the y-axis, we have a percentile rank. The higher number means that there's a good performance and a lower means that there's a poor performance. We have a subset of the data. On the left, we have variables that are highly correlated to the distortion type 2 and on the right to the distortion type 1. If we look at the results of the profile A, we see a good performance in the most of the variables. While if we look at the results of profile C, there was a poor performance in most of them. For example, B had a good performance in the distortion type 2 and a poor performance in the distortion type 1, and with the profile D, we had the opposite. If we look at individual variables, the distortion type 1 was highly correlated to speech intelligibility in noise and poor temporal processing abilities. The distortion type 2 to loudness perception and poor spectral processing abilities. So the conclusions of this study was, again, that it was valid to use our method and that the auditory deficits that characterize each of the auditory profile were in line of speech intelligibility deficits in one of the dimensions and loudness perception related deficits in the other dimension. Let's have a still and solve the last question. Does each profile have different needs in terms of hearing rehabilitation? And before that, I'm going to show you this picture where we have imagined that we are in a situation like cocktail parties scenario, and we have two people with hearing loss with hearing aids. And for them, this situation is particularly difficult. The same hearing deficits and the same hearing aids. However, one of them belongs to the group of the satisfied hearing hearing aid users and the other to the unsatisfied or not so satisfied hearing aid users. There is evidence that the satisfaction with hearing aids is correlated to the amount of hearing healthcare. Even with the highest technology, the service that the patients get in the hearing care center is crucial. So the audiologist, the decisions that they take and their experience is crucial for a good hearing rehabilitation. In this study, we explore a profile based decision making where we make use of the lessons learned from before the hearing deficits and the auditory profiling. However, this was a retrospective study where we didn't have any information about some of the super threshold measures. But we look at the audiometrics from the previous study and we realize that the thresholds at low and high frequencies were correlated to the distortion type two and distortion type one. So we could do a simplification. And instead of using the bear, the better hearing rehabilitation profiles, we can use the teddy bear audiometric groups. And this was an idea from Bill Widman. So we want to stratified listeners based on sensory impairment, and we have data from questionnaires about activity limitations and participation restrictions. And our goal is to identify patterns of benefit from optimal hearing rehabilitation to suit optimal hearing rehabilitation, and then look at the other two to see what we have to improve in order to have an optimal rehabilitation. The data driven analysis method that we use was again based on dimensionality reduction, in this case using factor analysis, the certification of the data set based on the sensory impairments. And we use archetypal analysis now to identify archetypal patterns of benefit, and then to cluster the listeners into benefit profiles. We also use a random forest in order not to predict the benefit patterns, but to estimate what is the importance of each of the individual items of these questionnaires for each of the audiometric groups. Here you see the archetypal patterns of benefit for each of the four of the metric groups. And on the right, there is the predictor importance for the questionnaires SSQ 12 and HHQ. If now we look, for example, at the group C in gray, we see the importance for the entire data set. And in yellow, the ones for only the people that are in the group audiometric group C, and we see that they are really similar. However, in some other audiometric groups, there were some questions in these items that they had higher importance and some others that they had lower importance. If we look at this from a different perspective, we can have in this case speech understanding and hearing handicap for the benefit profile one that we call it the minor improvements. And what we have in the X axis is the four audiometric groups. If we want this to be improved and be close to the optimal, we would have the priority of in the case of the group A and D to do, for example, counseling in order to improve the hearing handicaps, but in profile B and C, they want to improve their abilities in speech and noise scenarios. If we look now at the benefit profile two major improvements, we can see that there are other priorities and it's a bit more complex, but there are also differences in each of the audiometric groups. As a general summary, we have proposed a data driven method for profile identification that allows to find meaningful patterns and a solid definition of these groups. The resulting auditory profiles were defined not only by the thresholds but also super threshold performance and they were in agreement with previous classifications like auditory phenotypes and at the noise and distortion model. So we saw the potential of the decision making based on data driven approaches. And also, we have proposed a hearing rehabilitation pathway based on hearing deficits, difficulties and handicaps. I would like to thank my co-authors, collaborators, and many people that I have discussed about all these results in the last five years, innovation funding, the Bayard partners, and especially the hearing science Scottish section where I spent my external stay. And that's all. Thank you for your attention.