 Hello everyone, my name is Enrico Migliorini and today I'm here to present to you the data study that I conducted in the context of the MOSA X project on how variation in cochlear implant performance is related to differences in map parameters. I'm going to assume that you have some familiarity with cochlear implants, neuro prosthetics used to restore hearing capabilities to the profoundly deaf. You may not know that the MOSA X project is aimed at investigating cases in which some CI recipients have unusually and inexplicably low scores in speech recognition. The aim of our study is to find out why these recipients have these low scores, if there is some way that we can address it, and if so, how we can do so. I am a computer scientist, so my part of the job is currently focusing on the fitting process, that is the process through which each implant is tailored to the recipient. And specifically, I am looking into the map, which is a configuration of each electrode that consists of several values. Two important ones are T and C levels, stands for threshold and comfort, and represent the minimum and maximum electrical stimulation that the electrode will deliver. Another important factor is the dynamic range, which is simply the difference between T and C levels, as well as electrode impedance and neural response thresholds, although the latest have not been investigated yet in the context of the study. Fitting is generally a time consuming process that is handmade by audiologists, and it requires some trial and effort in order to find maps that work well for the recipients. Part of my aim is to automate and improve the process, creating tools that may assist the audiologist and the recipient, find a map that is best for them. Now something about the dataset of the data study. This is data that was collected at the Radbud University Medical Center in Naimechen, and it is a large dataset consisting of more than 8000 records. However, it has been heavily filtered in order to get the data that we were looking for. We have selected 320 recipients, all post-lingually deafened, and all over 18 years of age. And we have selected for each of them the latest audiometry test that was on record. These were spanning data range that goes from 2005 to 2020. For each audiometry test, we also included the map that was in use at the time, and we investigated T and C levels of that map. But first, a word on previous literature. There is quite some investigation into patient characteristics and cochlear implant performance. I have selected here the work by Lazar and others just as a title of example, but there are dozens of them. And that work investigated characteristics such as duration of deafness, etiology, and age of the recipients. On the fitting, I can mention work such as Lazar and others from 2000, which investigated implant-specific rather than electrode-specific values, for example pulse rate, pulse width amplitude mapping function. More to the point of the electrode by electrode value, we have the research from the graph and others in 2020, which investigated ENC levels, dynamic range, and T and RT, and Kim and others who specifically investigated dynamic range, finding only weak association and weak correlation. Now, the approach that I took was quite different, because instead of calculating average levels and average dynamic range, I used principle component analysis. This is a dimensionality reduction technique that allowed me to condensate the 44 values that make up T and C levels for a cochlear implant into four values that are a linear combination of these 44. So for example, the first component, here on the right, you can see maps which differ from each other only by different amounts of the first component. We can see rather easily that this map look like they are all shifted by a constant amount, and that is why the first component has been called constant shift for us. The second one that turned out from my analysis was very similar to dynamic range, because as we can see here in the graph, the higher the second component, so the more towards red, the color of the line, then the wider the difference between T and C levels. The third component was tilt, so quite simply the slope of the curves, while the first component was a central curvature, like a second order factor in a curve. Now the first thing that we checked was for correlation, and while we found a weak correlation, both according to Spearman and Pearson's correlation measures, this was not significant after we applied Bonferroni-Holm correction, and even before applying it, the coefficient factor was very small, up to 0.12. This is consistent with the findings of Kim and others, who also found a very weak correlation between dynamic range and tilt and speech recognition score. But are there significant differences between the populations of the ones who perform better and the ones who perform more poorly in specifically the speech recognition in quiet test? They have split the data into two categories according to their scores, the top 33%, which are the ones that were performing better than 87%, and the bottom 33%, which is the ones that were performing worse than 76%. And here on the left you can see the average maps for the top third tile in blue and the bottom one in red. You can see how there are some visible differences, which I'm going to explain more in detail right now. When checking for PCA components, the second one, the one that was called dynamic range, was significantly different even after Bonferroni-Holm correction. You can see on the right the results of the Mann-Whitney-Wilcoxson test and the kernel density estimation for the second component in blue again, the top scoring population in red, the bottom scoring population. And as you can see, the central values, the means and the medians as well, are quite significantly different. Although there is a wide overlap, and therefore this component is not a reliable predictor of performance, there is still quite a large and significant difference in averages. Specifically, higher scores are associated with a wider dynamic range. When, in order to validate these results, I have actually afterwards calculated by hand some parameters, such as average T level and average C level. This did not result, did not show any significant difference between the two populations, but when calculating dynamic range as the difference between the above two levels, then we had again almost the same graph as the one for the second PCA component, as well as we could now quantify the difference precisely, and it was a difference of six current levels, from an average of 50 to an average of 56 current levels between the two populations. I also investigated T and C level tilts, the slopes, but that didn't return any significant result either. So what is the discussion of these results? The results mirror very closely the ones of the graph and others, so they are consistent in previous research, and this is interesting because we were not looking to replicate that study. We were not searching for the same parameters, and yet the study converged to the same conclusion when one of the PCA components turned out to align very precisely with dynamic range. This therefore validates the results from that study, which was conducted at Amsterdam VUMC with different data set in different conditions. Now what are the next steps for the investigation? The next important step is an investigation into the causal relationship that is linking better performance and wider dynamic range. Are there some factors which relate to both effects, some factors which influence, such as for example neural health, which influence both dynamic range and speech perception? Or is there the possibility of a therapeutic intervention, therefore by adjusting subjects maps in order to slowly force a wider dynamic range, we could improve their scores? These are the two directions in which we are going to continue the investigation. Thank you very much for listening, and I'm not sure how the questions will be handled in this online conference, but I'm assuming you will find a way to ask me anything you'd like, and I'll be of course very willing to answer you. Thank you for listening again.