 Good morning. I would like to talk to you today about time efficient hearing tests and the use in the fitting of hearing aids Now traditional methods for testing in audiology For example measuring the audiogram is to use an adaptive procedure Where the next stimulus is chosen according to the response from the previous one or two trials for example Increase the level for trial and if the client says no on trial in minus one and Decrease the level for trial and if the client says yes on trial in minus one and usually this is done one frequency at a time And it has only moderate efficiency. It takes longer than it needs to take So I want to talk about a method called Bayesian active learning also called machine learning Which can be used to do hearing tests more efficiently And with Bayesian active learning you use a model that's characterized by a certain number of parameters The object is to estimate the values of the parameters and The stimulus for the next trial is chosen to produce the greatest reduction in uncertainty about the parameters and prior knowledge can be taken into account and In particular the responses of all previous trials I'll take an into account when choosing the stimulus for the next trial and These methods are highly efficient and can be quicker and more accurate than traditional methods for determining the audiogram and other tests Now here's an illustration of a basic problem in estimating the the audiogram or other tests in general The there's inherent variability in the responses of the subjects So imagine that you present a signal with a fixed frequency and various different levels shown by the symbols here And when the subject says yes, that's indicated by a blue symbol when they respond No, that's indicated by a red symbol and you can see that there's a region Where there's some ambiguity? The responses are mixed up and if you present the sound repeatedly at one of these levels the subject will sometimes say yes And sometimes say no and we estimate that the threshold falls somewhere within this region But with a certain area of uncertainty And we also know that if you test at an adjacent frequency it's likely that the Threshold at that frequency is not very different from the response at the frequency you started with But the range of uncertainty Increases as the frequency gets further away from your initial test frequency So we talk about this as covariance across frequencies that there's a The thresholds tend not to vary greatly from one test frequency to the next and To deal with this we need a probabilistic model easy of the response We need to allow for the fact that the subjects don't always respond consistently And we need a good policy for choosing the stimuli for the next trial or the stimulus for the next trial And within Bayesian active learning This is typically done by using a Gaussian process to model the response So we assume that the response is comes from a Gaussian distribution for any given level of the signal But that there's a certain variability reflecting the randomness of the Gandic Gaussian process And now policy for choosing the stimulus for the next trial is based on the concept of mutual information Now, let's just first of all look at the issue of covariance Imagine that we start off knowing absolutely nothing about some quantity that I'll call a function of X So X is some variable could be frequency could be something else could be level But we we have something that's a function of that variable and we start by knowing nothing about it So this gray area is the region of uncertainty But then once we get an estimate for a given value of X the uncertainty near that value of X is reduced So not only at the value of X that we got the estimate for but at nearby ones And if we keep repeating that process We can gradually reduce the uncertainty and get a good overall estimate of where we want what the function is as a function of X So going back to the concept of mutual information the Mutual information measures the information that two variables X and Y share And basically mutual information is a measure of the amount of information that's obtained about X when observing Y so once we know why how much have we learned about X and Basically in Bayesian active learning the stimulus for the next trial is chosen to maximally reduce The expected uncertainty in the model after the response for that trial And I won't go into the technical details of how that's done because I don't understand them anyway And other people know much more about them than I do So here's an example of how we applied this to the measurement of an audiogram And in this first method that I'll describe on each trial three pulses of a tone are presented all at the same level So beep beep beep and the subject is asked did you hear the tone or not? And we use three pulses to help the subject distinguish the tone from any tinnitus that they might have and to Generally know what to listen out for Now before starting the the Bayesian active learning or machine learning We start by presenting a grid of stimuli with Optive space frequencies and various levels to get starting values for the parameters And then we switch to the active learning and evaluate the Gaussian process after each trial To choose the most informative level and frequency for the next trial and this Animation illustrates how it's done. The star will appear in different places in this grid And that indicates the stimulus for the next trial the combination of frequency and level and a yes answer is shown by a blue circle No answer is shown by a red square and you'll see how The estimate builds up over time So we start with the initial grid To Get an idea of the starting parameters of the model And then once the initial grid is over we start the Bayesian active learning and this thick line Indicates the current best estimate and these sinna gray lines indicate the boundaries of the uncertainty region And I won't go through that whole thing But after a relatively small number of trials you get a pretty good estimate Of what the audiogram is the hearing loss as a continuous function of frequency Now Another method that we've tried involves counting tones and we did this because we thought it might be even faster And with this method you present six tone pulses all of the same frequency But decreasing in level as illustrated here And you ask the subject on that trial how many pulses did you hear so enter a number from one to six And in theory this gives you more information for trial But at the cost of slightly increased variance and so the issue is does it actually need to a faster Evaluation time for a given accuracy and with this method there are three Variables that we have to consider the frequency for a particular trial the starting level and the step sizes of the level decrease Now I'll show you the results of an experiment we did Comparing four different methods for assessing the audiogram And one of these was standard audiometry Another was standard audiometry, but implemented on a pc fitted with audiometric headphones Then the yes no active learning method that I just described And finally the counting method that I just described so each of those was compared on 40 ears 19 hearing impaired 20 normal hearing and one nearly normal, but not quite And this shows comparisons of average thresholds with standard deviations And if you look at this left hand panel This shows the comparison between a standard audiogram obtained on an audiometer and an audiogram obtained on the pc In the using the same method and using audiometric headphones and as expected Or the mean differences are close to zero db So there's no significant difference And the error bars here just indicate the typical variability from one measurement to the next If you repeat a standard audiogram and it's quite common to get deviations of 5 db or so Between two audiograms measured on different occasions On the right here. It's sorry in the middle here We see a comparison of the standard method with the yes no machine learning method And here we're showing differences based on two estimates from the machine learning method One is based on the lowest level that was heard And the other was based on the 50% point of the Gaussian process And you can see that the 50% point actually gets pretty close To the standard method Where whereas the lowest level heard Gives a slight difference of four or five db Finally, we compare the counting method and the yes no method And on average these give very similar estimates And you'll notice that the range of of uncertainty or differences Which is a continuous function of frequency here Is smaller than when the standard methods are repeated So we get higher repeatability with these machine learning methods And now the question is what is the efficiency of these counting and yes no methods well The this graph shows The root mean square difference between the estimate after a given number of trials shown along the bottom here Relative to the final estimate of the complete end of the run And what you can see is that the counting method shown by the dash line gets within 5 db of the final estimate after 20 trials And within 3 db of the final estimate After 40 trials and that's quicker than the yes no method Which took about 30 trials to get within 5 db And about 50 trials to get within 3 db So we can get high accuracy And good repeatability with this yes no counting method after only about 20 or to 40 trials I want to move on now to the measurement of dead regions Now a dead region is a region of the cochlea where there are no or very few functioning inner hair cells synapses or neurons And no information about basilar membrane vibration in a dead region Is transmitted to the brain and that's illustrated here This is the cochlea of a young man who died in an accident And who had previously been exposed to intense sounds from rifle shots And these dark lines show the neurons that are leaving the cochlea And that would eventually form the auditory nerve And you can see that at the base of the cochlea That's normally tuned to high frequencies there are no neurons The neurons have degenerated probably As a result of lots of inner hair cells And this is then a high frequency dead region Okay now we can characterize the extent of the dead region In terms of the characteristic frequencies of the surviving inner hair cells and neurons Immediately adjacent to the dead regions Who we have a schematic cochlea The green area is functioning normally Or maybe with some hearing loss but not completely dead And then the black area shows the dead region And in this case the boundary occurs at a place in the cochlea tuned to 2500 hertz And so we would say that this dead region has an edge frequency of 2500 hertz Now basilar membrane vibration in a dead region is not detected So for example if there's a dead region at the basal end of the cochlea And you present a high frequency tone That tone will not be detected at its normal place in the cochlea But that tone may be detected if you make it intense enough If it produces sufficient vibration at an adjacent live region And detection of a tone at the wrong place in the cochlea Is called off frequency listening or off place listening And so tests for diagnosing dead regions are based on detecting Off frequency listening or off place listening And one test that's been widely used is the psychophysical tuning curve or PTC And to measure a PTC the signal is fixed in level and frequency And the level of a narrow band noise needed to mask the signal Is determined as a function of the mask or center frequency And when a dead region is absent the tip of the PTC usually lies at the signal frequency But when the tip of the PTC where the mask level is lowest Is shifted away from the signal frequency This indicates that the signal produces peak excitation in a dead region So the signal frequency falls in a dead region And the frequency at the tip of the PTC indicates the boundary of the dead region Now I want to describe next an efficient test A machine learning test for determining the edge frequency of a dead region And in this test we don't try to estimate the whole PTC We're simply trying to estimate the edge frequency But it's based on the concept of a PTC And in this test we start by performing Transforming frequency to the herb number scale Which has units counts And this is like a scale of distance along the basilar membrane And it's not too different from a logarithmic frequency scale But not exactly the same And imagine that we present the signal shown by the black line here With a certain level and frequency And the distribution of excitation produced along the basilar membrane By that signal is illustrated by the blue line But imagine that there's a dead region Indepated by this gray area So the signal is producing its peak excitation within the dead region Then it will be detected by off-frequency or off-place listening In this lower frequency region here Now suppose we add a masker indicated by this black rectangle And it produces a distribution of excitation along the basilar membrane Looking like this Then in this example the signal will be detectable Because the excitation it produces in the live region Is well above the excitation produced by the masker But if we increase the masker level Then we will reach a point where we just mask the signal excitation So what we're going to do here is to Choose different combinations of the masker level and the masker frequency To find the point where the signal is just masked And to estimate the edge frequency of the dead region And in the model, when we do this We have independent variables, things that we're adjusting Of masker level and frequency We get the responses of the subject And we use those to define parameters in a hearing model And the two variable parameters in that hearing model Are the edge frequency of the dead region And the extent to which the auditory filters are broader than normal Or factor by which the auditory filters are broader Because we know that hearing loss is often associated with a broadening But in this particular case We only need to calculate the broadening for the auditory filter That's tuned close to the edge frequency of the dead region So each of these is just one parameter So this illustrates how the intelligent dead region test works Just concentrate on this middle panel As for the audiogram test We start by presenting a grid of signals with fixed frequency on a variable level And there are two frequencies here And the pluses indicate that the subject responded yes The blue circles indicate that the subject responded no And this gives us starting values for our parameters And then we start the Gaussian process for estimating the edge frequency of the dead region So we go through a sequence of trials You can see where the responses are building up And quite quickly we get a pretty good estimate of the shape of the PTC And where the edge frequency is And we compared the edge frequency estimates to those obtained using fast PTCs So this was done with people who'd already been tested with PTCs And we knew they had dead regions And we found that this fast method gave very similar estimates To those obtained using the fast PTCs And the edge frequency estimates for three runs of the fast region test Showed only a small scatter So we had very good repeatability And we got precise estimates of the edge frequency in between 20 and 50 trials So this is again quite a fast test Okay, how do we apply this to the fitting of hearing aids? Well, we've developed a rule of thumb over the years That if the edge frequency is greater than three kilohertz Then you would fit the hearing aid as usual You don't change your fitting method from your normal one But if the edge frequency is below three kilohertz Then we recommend reducing the gain for frequencies above 1.7 FE And we've found in a number of studies that gives better results Than the standard fitting methods Okay, in the last part of the talk I just briefly want to discuss a method for estimating equal loudness contours An equal loudness contour is a map of the sound level As a function of frequency that's required to obtain a fixed loudness And these are often measured in normal hearing subjects And the task that we used was to adjust the length of a line Using the mouse to match the impression of loudness Of the tone that the person had just heard The line length was initially set to 10 pixels And possible line lengths ranged from 0 to 1,280 pixels Corresponding to 0 to 35 centimeters at a distance of 60 centimeters And the predicted loudness for a given combination of frequency and level Was represented by a Gaussian distribution on a logarithmic pixel scale With its mean corresponding to the most likely line length And the variance representing both the subject's variability and responses And the uncertainty of the Gaussian process And the whole process took 10 to 12 minutes To measure equal loudness contours for a range of fixed loudness levels And for a range of frequencies Now here are examples of results for normal hearing subjects We also have results for hearing impaired subjects But we'll only show these results for normal hearing subjects here The signals were presented only to one ear We've got four different loudness levels And the solid lines with the error bars Show the estimated equal loudness contours For these four different loudness levels Plotted as a function of frequency transformed to a cam scale again And for comparison The ISO standard equal loudness contours in the latest standard ISO PY322 are shown by the dashed lines And you can see that we got very close to those standard equal loudness contours For these normal hearing listeners So we can get accurate results Corresponding to those in the standard In a relatively short number of trials And this could be applied to hearing aids when adjusting the compression Nearly all hearing aids incorporate some form of frequency dependent amplitude compression But at present the amplitude compression required at a given frequency Is estimated from the audiometric threshold at that frequency And we know that that's not really adequate Because loudness recruitment can vary marginally for a fixed audiometric threshold And we think that equal loudness contours can be used to estimate the amount of loudness recruitment As a function of frequency and to set the compression ratio at each channel frequency Based on the estimated loudness recruitment at that frequency Using the equal loudness contours So to sum up I've described results for three tests Which can be performed rapidly using Bayesian active learning The audiogram, edge frequency of a dead region, unequal loudness contours And we think that these can lead to more personalized hearing aid fitting In a relatively short time And I'll stop there and invite your questions and comments Thank you