WO2003070322A1

WO2003070322A1 - Apparatus and method for neural response telemetry of a cochlear implant by automatic shape recognition

Info

Publication number: WO2003070322A1
Application number: PCT/FR2003/000577
Authority: WO
Inventors: Basile Charasse
Original assignee: Newmedic International; Centre National De La Recherche Scientifique (C.N.R.S.)
Priority date: 2002-02-22
Filing date: 2003-02-21
Publication date: 2003-08-28
Also published as: WO2003070322A8; FR2836386A1; AU2003222940A1

Abstract

The invention concerns an apparatus and method for neural response telemetry of a cochlear stimulation implant (1) comprising a plurality of implantable electrodes (2) designed to deliver a cochlear electric stimulation and to collect a neural response in the form of time-based measured response curves. Means for automatic shape recognition enable to classify automatically each response curve among a plurality of predetermined types of shapes. Said classification enables to determine and optimize adjustment of the operating parameters of the cochlear implant.

Description

APPARATUS AND METHOD FOR TELEMETRY OF NEURAL RESPONSE OF A COCHLEAR IMPLANT BY SHAPE RECOGNITION

AUTOMATIC

The invention relates to an apparatus and a method for telemetry of neural response of a cochlear stimulation implant comprising a plurality of implantable electrodes adapted for delivering cochlear electrical stimulations and for collecting the neural responses induced by these stimulations, corresponding to a potential. of electrical composite action generated by cochlear neurons.

Hearing aids with cochlear implants have been known and proven for a long time, they make it possible to restore hearing in patients with permanent deafness. After implantation, the adjustment of certain operating parameters of the prosthesis must be carried out to adapt it to the patient.

In particular, the electrical stimulation levels that will be delivered by the implant must be adjusted for each electrode. These adjustment levels, called the perception threshold T and the comfort threshold C, respectively represent the minimum electrical voltage inducing a sound sensation and the maximum electrical voltage borne by the patient. To do this, several methods are known and used in practice by the practitioners responsible for adjusting the implant.

The most common known method is a subjective method, in which the practitioner varies the stimulation voltage, and questions the patient about his or her sound perceptions. The practitioner must, using the patient's answers, determine if the patient perceives a sound sensation or if the sound sensation is too loud and uncomfortable. The answers given by the patient thus allow the clinician to determine and adjust the T and C thresholds of the latter. In practice, this method cannot be used in all cases and in particular with patients such as young children, who are unable to express their auditory impressions to the practitioner, or even to understand the task asked of them. . Other known methods aim to overcome these drawbacks and propose to determine the thresholds from objective electrophysiological measurements.

One known method of this type consists in measuring the electrical potentials evoked by electrical cochlear stimulation in the patient's skull. This method requires specific equipment (electrodes distributed over the skull), and is therefore cumbersome to implement. In addition, the measurement is very sensitive and can be disturbed by the patient's movements. This method is therefore not practical and cannot be used in young children.

US-5,758,651 describes another method, called neural response telemetry (NRT), making it possible to measure the neural response of the auditory nerve induced by electrical stimulation applied by the cochlear electrodes. This method is used clinically by practitioners to make adjustments to the cochlear implant. It has in particular the advantage of not requiring the attention or the immobility of the patient, which in particular allows its use on young children.

The neural response recorded by this method is the electrical composite action potential, called "ECAP", generated by the cochlear neurons of the auditory nerve and measured over time. It is therefore a measurement on the peripheral auditory system. This neural response can come in several forms which must be identified and analyzed by the practitioner. The standard form of neural response sought by the latter is a biphasic curve comprising a negative peak NI and a positive peak PI, making it possible to define the amplitude of the neural response corresponding to the difference in tensions between these two peaks.

By varying the stimulation voltage, the practitioner measures the variations in the amplitude of response of the auditory nerve. These variations in the neural response generally correspond to a line, called the growth line, whose interpolation and extrapolation provide

^"in particular a value, known as the T-NRT threshold, obtained at the intersection of this growth line with the abscissa axis (stimulation voltages). It has been demonstrated that it is possible, from the T-NRT threshold, to deduce the T and C thresholds of the patient with good reliability (cf. for example "The relationship between EAP and EABR Tresholds and levels used to program the nucleus 24 speech processor : data from adults "Carolyn Brown Michelle L. Hughes, Betty Luk, Paul J. Abbas, Abigail Wolaver, and Jonathan Gervais. 0196/0202/00 / 2102-0151 / 0 Ear &Hearing;" Estimation of psychophysical levels using the electrically evoked compound action potential measured with the neural response telemetry capabilities of cochlear corporation 's CI24M device ". Kevin H. Franck and Susan J. Norton 0196/020/01 / 2204-0289 / 0 Ear & Hearing). However, in practice, the implementation of this method raises many problems.

First, the clinician must obtain a neural response curve effectively comprising the NI and PI peaks in order to measure the amplitude of the response. However, to obtain such a response curve, a certain number of parameters linked to the measurement (in particular response acquisition time, amplification gain, stimulation voltage, etc.), called measurement parameters, must be adjusted. and optimized. This first approach turns out to be long, subjective, and the results strongly depend on the clinician's expertise. Secondly, when the response curves making it possible to plot the growth line for an electrode have been obtained, the clinician must distinguish among these which are usable from those which are not. Certain response curves may in particular have a poor signal / noise ratio and be difficult to identify. This selection is subjective and the results of this selection have a direct influence on the subsequent setting of the implant.

Third, the clinician must measure the peak voltage values on each response curve himself to determine the amplitude of response. This measurement, which is also subjective, directly influences the setting of the implant, and is long and tedious.

In addition, these operations must be repeated for all the electrodes. However, in practice, a cochlear implant conventionally has more than 20 electrodes. Given the time that clinicians can devote to setting the implant, in practice only 4 or 5 electrodes are tested and used to make the setting.

The adjustment of a cochlear implant with the NRT method is therefore not optimal. To date, it is long, subjective and incomplete.

In addition, it should be noted that adjustments and calibrations of the cochlear implant must be carried out regularly, since the physiological listening capacities of a patient can vary over time, in particular according to his age, and modifications. physiological that can be caused by the wearing of the hearing aid itself.

Furthermore, WO-0052963 describes a method and an apparatus in which particular parameters determined from the neural response are used to define an optimal operating mode or an optimum stimulation rate, and thus to best regulate the functioning of the cochlear implant. In particular, the response time of appearance of positive or negative peaks by the neural response is used. However, in practice, the analysis of these parameters, carried out by the clinician, comes up against the same problems as those described above.

The invention therefore aims to overcome these drawbacks by proposing an apparatus and a method for telemetry of neural response thanks to which the adjustments of the operating parameters of a hearing aid comprising a cochlear implant can be carried out quickly, more simply than in the state of technique, more objectively and reliably.

More particularly, the invention aims to allow a more complete adjustment of the cochlear implant, that is to say for more electrodes, in a period of time compatible with clinical constraints.

The invention also aims to propose an apparatus and a method for automating the NRT method - in particular in the determination of the T and C thresholds - so as to allow its implementation by inexperienced practitioners.

The invention also aims to make it possible to obtain an adjustment of the cochlear implant as objective as possible, which is neither linked to the know-how of the practitioners, neither in the reliability of their interpretation, nor in the patient's capacity for expression and understanding.

To do this, the invention relates to a neural response telemetry device comprising: - a cochlear stimulation implant comprising a plurality of implantable electrodes adapted to deliver cochlear electrical stimulations and to collect neural responses corresponding to the composite action potential electrical generated by the cochlear neurons, following these stimulations, - an implant control device comprising computer processing means, and suitable for:

. generate electrical stimulation signals for the electrodes,

. receive the corresponding neural responses, and,. record these neural responses in the form of response curves representing the composite action potential measured over time, characterized in that the computer processing means are adapted to perform automatic shape recognition and classify each response curve among a plurality types of curve shapes.

The invention also extends to a method for telemetry of neural response of a cochlear stimulation implant comprising a plurality of implantable electrodes adapted to deliver electrical cochlear stimuli and to collect neural responses corresponding to an electrical action potential. composite generated by cochlear neurons, and induced by these stimulations, process in which:

- electrical stimulation signals are generated for the electrodes,

- the corresponding neural responses are received and recorded in the form of response curves representing the composite action potential measured over time, characterized in that computer processing means are used which are suitable for carrying out automatic shape recognition and for classifying each response curve among a plurality of types of curve shapes.

The invention thus extends to a method of computer processing of neural responses of a cochlear stimulation implant comprising a plurality of implantable electrodes adapted to deliver electrical cochlear stimuli and to collect neural responses corresponding to an action potential. electrical composite generated by the cochlear neurons, induced by these stimulations, and recorded in the form of response curves representing the composite action potential measured over time, a process characterized in that computer processing means are used which are suitable for perform automatic shape recognition and to classify each response curve among a plurality of types of curve shapes. The invention starts from the observation that, in practice, there are a limited number of types of shapes for the neural response curves that can be obtained in all patients.

Among these different types of curve shapes, we can in particular determine those with which the calculation of the amplitude of the neural response is possible, and those for which such a calculation is not possible, and therefore requires to make a new measurement with a configuration of new measurement parameters (acquisition delay, stimulation voltage, amplification gain, etc.). The invention then makes it possible to select, from the response curves, automatically, quickly and objectively, those which are significant, that is to say allow the determination of implant adjustment parameters such as the thresholds T and vs.

The different types of shape of the curves are predetermined, that is to say defined in advance, from saved models and used to initially configure the computer processing means for automatic shape recognition.

Alternatively, the types of shapes can be changed during use if other types appear relevant to the practitioner. For example, when the computer processing means fail to classify a response curve into a known type, the clinician may be asked to carry out this classification himself, either in a known type, or in a new type that it defines with this response curve.

Preferably, even in this variant, at least one set of types of curve shapes is predetermined and is used for the initial configuration of the computer processing means. Several sets can be provided and recorded, for example to allow the clinician to choose one of the sets according to the patient to be treated. For example, a set of types of curve shapes can be provided for young children; another game for adults ... Games of different types of curve shapes can also be provided depending on the nature of the pathology from which the patient is affected.

In an advantageous variant of the invention, the response curves classified as significant can also be used to carry out an automatic calculation of operating parameters of the implant, and in particular of the T-NRT threshold. In another variant, the care of this calculation can be left to the practitioner. In all cases, the procedure is greatly accelerated, which in fact makes it possible to adjust more electrodes of the implant - or even all of the electrodes - in an acceptable duration, and in a more objective manner.

Pattern recognition applied to response curves can be the subject of very many variant embodiments. It is possible, for example, to use, as a means of pattern recognition, an adaptive digital filtering, or a calculation of optimization of mean square errors; or any other means based on a correlation method.

Advantageously and according to the invention, the computer processing means comprise an artificial neural network adapted to perform automatic shape recognition. Such artificial neural network is, by design, learning type and can handle nonlinear problems with high efficiency and high accuracy even when ^• forms recognition are varied, as is the case in practice with neural responses. Artificial neural networks and their functioning are well known in themselves (cf. for example “Neural Network Design.” Hagen MT, Demuth, HB, & Beale, M. (1996). Boston: PWS Publishing; “Neural Networks. A Comprehensive Foundation. ”Haykin, S. (1994). Prentice-Hall: New Jersey). Usually they are used to solve much more complex problems than simply comparing curves.

US-2001/049466 describes an at least partially implantable system of the "dual" type, that is to say comprising both mechanical stimulation and electrical stimulation. This document provides that the system can incorporate telemetry means for the transmission of data between an implanted part of the system and an external unit. Similarly, in an embodiment where the system is fully implantable, after implantation, the parameters such as the patient-specific data can be transmitted transcutaneously to the implant, using software modules which are of preferably dynamic, that is to say adaptive, so as to optimize the functioning of the implant for the patient. As such, a software module can be provided which simulates a "healthy" cochlear amplifier based on an adaptive neural network. The learning of this neural network can be carried out by the wearer of the implant and / or by external assistance. More particularly, this neural network for simulating a "healthy" cochlear amplifier implements the TRA principle so that locally limited areas of the cochlea are stimulated mechanically. This system uses a neural network only to directly optimize the functioning of the implant and dynamically modify its parameters so as to simulate a "healthy" cochlear amplifier. This document therefore does not in any way describe the use of automatic shape recognition to classify curves of neural responses among a plurality of types of shapes of predetermined curves as in the invention.

However, it is one aspect of the invention to have found that the use of such an artificial neural network for the classification of neural responses recorded with the NRT system on a cochlear implant for the adjustment of its functioning is particularly advantageous, simple, reliable and precise and in fact easy and economical to implement. An artificial neural network has the particular advantage of being able to perform effective pattern recognition from shapes of unknown curves and which can be very different, as soon as it has undergone an initial learning phase. An artificial neural network can include several layers of neurons. An artificial neuron is a mathematical object whose functioning mimics the behavior of biological neurons. In principle, each neuron produces a digital response value a from an input vector p according to the formula: a = f (W.p + b) where W is a weighting vector, b is a coefficient of deviation, and f is a transfer function.

A single isolated neuron is only able to solve linear problems. On the other hand, the association of neurons in different layers makes it possible to solve nonlinear and complex classification problems.

The learning phase of the artificial neural network consists of providing numerous examples of known vectors at the input of the network and, simultaneously, the target responses to be provided by the network for each of these examples. Using an appropriate learning algorithm known in itself, the network modifies the W-weighting vectors of each layer of neurons so as to minimize the mean square error given by: e = (η _t - η) ² where η _t is the target value and η is the value of the response provided by the artificial neural network in response to the example provided as input.

Advantageously and according to the invention, the artificial neural network comprises:

- an input vector whose number of scalars depends on the number of measurement points of each response curve,

a first intermediate layer comprising a number of neurons depending on the number of points of the input vector, - a layer of output neurons, the number of neurons of which depends on the number of types of curve shapes learned by the network.

Advantageously and according to the invention, the artificial neural network comprises only two layers of neurons, namely an intermediate layer of single neurons.

Furthermore, advantageously and according to the invention, the types of curve shapes are chosen representative of:

- the absence of a neural response of listed form,

- the presence of positive and / or negative pulses in the response curve. They include at least one type of significant curve shape allowing the evaluation of at least one operating parameter of the implant, and at least one type of non-significant curve shape.

Advantageously and according to the invention, the reliability of the classification made by the neural network is evaluated according to the digital values ηi delivered by the neurons of the output layer.

Advantageously and according to the invention, the computer processing means are adapted to automatically calculate a response amplitude value on each response curve classified into a type of significant curve shape, that is to say corresponding to a real neural response and authorizing such a calculation.

Advantageously and according to the invention, the computer processing means are adapted to: a) automatically determine whether a first response curve obtained for an electrode can be classified into a type of significant curve shape, b) if not, repeat the measurement after automatic modification of measurement parameters then repeat step a); if so, acquire a series of successive response curves by decreasing the electrical stimulation voltage until a response curve classified in a type of non-significant curve forms is obtained, classify each response curve in a type of curve shapes, and calculate automatically a value of at least one parameter relating to electrical stimulation from this series of response curves.

Advantageously and according to the invention, the computer processing means are adapted to automatically calculate a threshold value T-NRT from the response amplitude values calculated automatically.

Furthermore, advantageously and according to the invention, a quality criterion Q of each response curve representative of the regularity of a final portion of the response curve is calculated to determine, according to the value of this quality criterion Q, if the response curve is:

- quality 1: abnormal and not significant, classified in a non-recordable type TNR,

- quality 2: distorted but significant,

- quality 3: clear and significant. Thus, the invention makes it possible, for the first time, to carry out a fully automatic calculation of operating parameters such as the T-NRT threshold.

Advantageously and according to the invention, the quality criterion Q is calculated according to the formula:

NOT

where R (n) corresponds to the ^nth point of the response curve R, ni is the sequence number of the first point of said final portion, N is the total number of points of the response curve R. Other formulas can be used as a variant. The invention also extends to a method implemented in an apparatus according to the invention. Conversely, the invention extends to an apparatus for implementing a method according to the invention.

The invention also relates to an apparatus and a method characterized in combination by all or some of the characteristics mentioned above or below. Other objects, characteristics and advantages of the invention will appear on reading the examples and the description which follow, which refer to the appended figures in which:

FIG. 1 is a schematic view of an apparatus according to the invention,

FIG. 2 illustrates five examples of types of shape of response curves that can be used in an apparatus and a method according to the invention,

FIG. 3 is a diagram illustrating the operation of a neuron,

FIG. 4 is a diagram representing a sigmoid transfer function curve which can be used in an artificial neural network neuron, in an apparatus and a method according to the invention,

FIG. 5 is a diagram illustrating the structure of a neural network which can act as shape recognition means in an apparatus and a method according to the invention,

FIG. 6 represents two examples of forms of response curves Ca, Cb, of the same type, but more or less deformed, the curves Ce and Cd representing the final enlarged portions of the curves Ca and Cb respectively,

FIG. 7 represents four other examples of shapes of response curves for which a practitioner might hesitate in their classification,

FIG. 8 is a functional flow diagram of a fully automatic variant of a method for calculating the T-NRT thresholds according to the invention,

- Figure 9 is a functional flowchart of an exemplary embodiment of a method for classifying response curves according to the invention, - Figure 10 is a diagram comparing a classification performed by a practitioner with that performed by a method and an apparatus according to the invention, - Figures 11 and 12 show examples of various neural response curves obtained in practice.

FIG. 1 represents an apparatus according to the invention which can be used for the clinical measurement of a neural response. This device comprises a cochlear stimulation implant 1 comprising a bundle of intra cochlear electrodes 2. This implant 1 is adapted to be implanted under the skin 17 of the skull of a patient. The intra cochlear electrodes are organized in pairs. That is, in each pair of electrodes 2 in the bundle, one electrode is used to stimulate the auditory nerve, and the neighboring electrode is used to record the response of the auditory nerve to stimulation. Conventionally, such a bundle of intra cochlear electrodes 2 can comprise around twenty electrodes therefore defining around twenty channels. The electrodes allow the measurement and recording of the composite electrical ECAP action potential generated by the auditory nerve following electrical stimulation. As is well known, the cochlear implant 1 comprises a radio frequency antenna 3, a receiver 4, a transmitter 5, a decoder 6, an encoder 7, a current source 8 and an amplifier circuit 9. The antenna 3 communicates through the skin 17 in particular with an external antenna 10 of a transmission / reception unit 11 coupled via a programming control interface 12 PCI ("Programming Control Interface") to a computer unit 13 for processing digital information, which, in the example shown, is a microcomputer. The transmission / reception unit 11 is linked bidirectionally to the computer unit 13, and includes non-volatile storage means incorporating different speech processing methods and different stimulation programs. The computer unit 13 is adapted and programmed to implement the method according to the invention, and in particular to produce automatic pattern recognition means, in accordance for example with the flow diagram of FIG. 9. The other general characteristics of the entire apparatus thus formed, making it possible to carry out neural response telemetry, are known in themselves (cf. in particular US 5758651 or WO 0052963) and can be the subject of numerous variant embodiments. To perform an NRT measurement, the computer unit 13 controls the transmission / reception unit 11 so that the latter transmits a predetermined stimulation signal, for example in the form of a succession of balanced biphasic square slots. This stimulation is emitted on the antenna 10, received through the skin 17 on the antenna 3 of the cochlear implant 1, then transmitted to the receiver 4 then to the decoder 6 and to the current source 8 which supplies an electric current of stimulation to an electrode of the beam 2. The neural response is collected on another electrode close to the beam 2, transmitted to the amplifier 9, then to the encoder 7, and to the transmitter 5 which transmits this response on the antenna 3 received through the skin 17 on the antenna 10 of the transmitting reception unit 11, which retransmits it to the computer unit 13. The computer processing unit is also suitable for being able to adjust the operating parameters of the cochlear implant 1, in particular the levels of electrical stimulation provided by the transmitting / receiving unit 11 to the cochlear implant 1, and determined by values recorded in the transmitting / receiving unit 11.

It should be noted that in practice each neural response, that is to say each response curve, consists of an average of successive measurements carried out with the same electrodes. Each response curve results from the neural response signal (ECAP) sampled over time as indicated in US-5,758,651.

Figures 11 and 12 illustrate examples of response curves obtained in practice. In these figures, the electrode bundle 2 is shown in the lower part. The stimulation electrodes were those bearing the numbers 3, 5, 10, 15 and 20 where a square stimulation wave is represented in the lower part of the electrode bundle 2. The detection electrodes measuring the neural response were the electrodes no. 5, 7, 12, 17 and 22 as shown schematically by arrows in the figures. The top five diagrams in each of Figures 11 and 12 illustrate the neural response curves recorded for different levels of stimulation current, the highest current levels being at the top of each diagram, and the lowest current levels being at the bottom of each diagram. The five diagrams respectively represent the responses obtained to various detection electrodes. The diagrams in the lower part of each figure 11 and 12 illustrate recorded neural response curves by varying another measurement parameter (the interval between excitatory signal and mask signal). In FIG. 11, the response curves obtained are of the type with a single negative peak and with a single positive peak, and are the most common in practice and those sought for the calculation of the implant adjustment parameters. FIG. 12 represents another example of neural response comprising response curves, some of which have two positive peaks (for the detection electrodes No. 12 and 22).

As can be seen in these two figures, the neural response curves can be very varied as regards their shape. The interpretation, classification and exploitation of these response curves by a clinician without using a method and an apparatus according to the invention are therefore long and difficult.

Figure 2 illustrates non-limiting examples of types of response curve shapes that can be chosen. Other types of response curve shapes can be defined and chosen in practice.

In Figure 2, the Ta type corresponds to a clear significant response in which the negative and positive peaks are clearly identifiable. The Tb type corresponds to a neural response in which the negative peak is missing. The Td type corresponds to a neural response in which the positive peak is prominent. The Te type corresponds to a clear significant response in which a negative peak and a positive peak are clearly identifiable and the first point of the response is above the positive peak. The TNR type corresponds to curves which do not represent a neural response, that is to say corresponding only to noise.

The invention makes it possible to recognize the shapes of the response curves obtained (such as those of FIGS. 11 and 12) and to classify them according to predetermined types such as those represented in FIG. 2 by way of example.

The computer unit 13 is adapted to perform automatic shape recognition of the different response curves obtained, and to classify them according to the predetermined types, and this according to any suitable algorithm (adaptive digital filtering, optimization of the mean square error, ...). Preferably, and in accordance with the invention, an artificial neural network is used as an automatic shape recognition means. In this network, each artificial neuron exhibits a behavior as illustrated in FIG. 3. From an input vector p, the neuron provides the output value a given by the function: a = f (W.p + b ), as explained above.

In the examples given below, advantageously and according to the invention, the transfer function can be a log-sigmoid function as shown in FIG. 4. Alternatively, one can choose a linear transfer function, or any other transfer function known appropriate.

The neural network which can be used according to the invention comprises for example two layers 15, 16 of neurons as shown in FIG. 5, namely an intermediate layer 15, and an output layer 16. The input vector 14 comprises a number of scalars dependent on the number of measurement points (at most equal to the latter), that is to say the number of sampling time intervals making it possible to form a response curve.

The various parameters that can be used in such a neural network are for example the following:

- learning rule: back propagation of the error gradient (using for example the MATLAB trainrp function as described in Riedmille, M. and H. Braun, "a direct adaptative method for faster backpropagation learning: the RPROP algorithm" Proceedings of the IEEE International Conference on Neural Networks, 1993).

number of scalars of the input vector 14: corresponding at most to the number of points of each response curve, for example of the order of 30,

number of artificial neurons in the intermediate layer 15: for example equal to the number of scalars of the input vector 12, - number of neurons in the output layer 16: equal to the number of types of response curve shape (equal to 5 in the examples described),

- learning rate: 0.01, - acceptable quadratic error between the target and the network response: 10 ^"6 ,

- transfer function of each neuron: log-sigmoid function of FIG. 4.

To carry out the learning phase, a series of examples of response curves is provided to the neural network. Each example preferably represents an almost perfect neural response associated with a known predetermined type. Before providing the examples of responses to the neural network, all the response values obtained are normalized, that is to say calibrated in the interval [0,1]. Thirty examples for each type of response were used for learning. The output codification for a perfect classification is for example the following:

Ta type: 10000

Type Tb: 01000

Td type: 00100 Te type: 00010

TNR type: 00001

The five values coding each type thus correspond respectively to the five artificial neurons of the output layer 16 of the network.

During the learning phase, thanks to the trainrp function, the artificial neural network modifies the weighting vectors W of the neurons of each layer so as to minimize a quadratic error given by e = (η _t - η) ² where η _t is the target value and η is the value of the network response as defined above. After a sufficient number of examples have been given to the network, the learning phase is completed and the network is ready to process unknown and new response curves. Different learning algorithms have been developed in the state of the art in order to achieve effective learning in artificial neural networks. One of Approaches, known in itself, is based on the back propagation of the error gradient which produces the modification of the weighting vectors of the network using a linear function of the value of the quadratic error of the gradient. The modifications of the weighting vectors are given by the formula: Wi ₊ i - Wi = - oti x gj where Wi is a vector representing the weighting vectors of the different neurons during step i, gi is the quadratic error gradient and oti is the learning rate.

The learning examples are chosen to represent the set of possible neural responses, and in almost perfect forms. In reality, the recorded neural responses are not always clear and perfect. In many cases, they are distorted, which is due for example to the limitations of the recording system (for example to the saturation of the amplification). In addition, when the stimulation level is too low, the neural response can no longer be distinguished from background noise or at least with difficulty. In the case of a distorted recording, an identifiable neural response nevertheless exists and the network should be able to recognize the type of this neural response in the distorted curve obtained, and assess the quality of this response. To do this, advantageously and according to the invention, a quality criterion is defined which makes it possible to take into account the quality of the response obtained. For a neural response, all the values are calibrated in the interval [0,1] so that different responses can be objectively compared. The quality criterion can be chosen, for example, as the average of the distance between two consecutive points of the response thus calibrated, accumulated from, for example, the tenth measurement point towards the end of the response, i.e. from the tenth sampling point. The quality criterion can then, in this example, be calculated as follows:

30

R (n + 1) -R (n)

20 n ₌ 1 fi

where R (n) corresponds to the ^nth point of the response curve R. With this formula, we can consider for example that when Q> 0.05, the response is distorted. Conversely, when Q <0.05, the response is acceptable. FIG. 6 illustrates the application of this rule in two different cases of response curves of the Ta type. The curve Ca in FIG. 6 represents an almost ideal response while the curve Cb represents a distorted response. The curve Ce illustrates the same response as the curve Ca but represented only from the tenth sampling point. Similarly, the curve Cd represents the curve Cb from the tenth sampling point. In the case of the Ce curve, Q is equal to 0.0213, and we see that the response is not distorted. On the contrary, with the curve Cd, Q is equal to 0.0923 and the curve is distorted. Furthermore, if Q> 0.12, it can be assumed that the response is of the TNR type, that is to say abnormal, with no significant identifiable response, and should not be taken into account; whereas if 0.08 <Q <0.12, we can assume that the response is significant, but that the type of response is unknown. In the latter case, the identification of the type of response obtained can be stopped so as to accept manual entry by the practitioner to create a new type of response. Figure 7 illustrates this possibility. In this figure, the curves Ce and Cf resemble the Ta type, but are not exactly identical to the example of the Ta type represented in FIG. 2. The curves Cg and Ch could be classified in the Te type but also in the Ta type. We see from these four examples that the correct classification of a response curve is not always immediate. This is why, in certain cases, the artificial neural network will not be able to clearly identify with certainty the type of the response curve obtained. Thus, the output values given by the artificial neural network can be different from 1. For example, in the cases represented in FIG. 7, the real values given respectively by the five neurons of the output layer 16 of the artificial neural network are those of table 1 below: Table 1

As we can see, in each case the network is able to identify the correct type although the values delivered ηi by the network are not strictly equal to 1. We can assume on this basis that if the result delivered by the neural network artificial is greater than 0.98, the type of response can be considered to be reliably identified. If the value provided by the artificial neural network is between 0.9 and 0.98, the type of response cannot be considered to be reliably identified. And, if the value supplied is less than 0.9, the type of response is unknown, and, if necessary, requires the classification of the response curve into one of the types by the practitioner himself. In the table above, the value η of the response provided by the neural network is defined as being the largest of the values r | i given by the different neurons of the output layer 16: η = Max (ηi). In the example of the curve Ce, this value η is 0.908. In the example of the curve Ch, this value η is equal to 0.7398. If two or more ηi values given by the artificial neural network are greater than 0.5, it is assumed that the type of response is unknown and if all the values η _; are less than 0.5, the type of response is also unknown.

FIG. 8 represents an alternative embodiment of a method according to the invention allowing the fully automatic calculation of the T-NRT thresholds. During the initialization step 40, a patient registration form is created. During the subsequent step 41, the data related to the patient in question are entered. The patient file and the data entered are saved in the database patient 42. In step 43, the correct operation of the various electrodes is checked so as to use only those which are actually in good condition. During the subsequent step 44, a first measurement is carried out from an initial configuration of measurement parameters, on one of the electrodes, and the corresponding response curve is recorded. During the subsequent step 45, an automatic shape recognition and a classification among a predetermined type are carried out in accordance with the method according to the invention.

It is then examined during the subsequent test 46 whether or not the type of response curve obtained corresponds to a significant response. If not, the measurement parameters are modified in step 47 and steps 44, 45 and 46 are repeated. In step 47 for modifying the measurement parameters, a new set of parameters is chosen for example from a plurality of predetermined and prerecorded measurement parameters. The loop of steps 44 to 47 continues as long as no real significant response is detected. When such a significant response is detected during test 46, we then proceed to step 48 of automatic calculation of the T-NRT threshold for the electrode considered. During this automatic calculation step 48, the stimulation voltages are decreased, and for each voltage value, a new response curve is recorded. The shape of this response curve is then recognized and it is classified into a predetermined type according to the invention, until the stimulation voltage is low enough for the type of response curve obtained to correspond to a non-recordable type (no neural response detected). During this same step 48, for each response curve corresponding to a significant response (Ta or Te type), the amplitude of the response is automatically calculated by difference of the voltages between the negative peak and the positive peak. This calculation can be carried out by any known algorithm making it possible to locate the maximums and the minimums on a given response curve. A value of the T-NRT threshold is then calculated by interpolating the calculated response amplitude values. This interpolation leads to a growth line whose extrapolation provides the T-NRT threshold as the point of intersection with the abscissa axis (stimulation voltages).

Test 49 checks whether the previous calculation was carried out on the last active electrode or not. If not, iterates 50 to pass to the subsequent active electrode and all the steps 44 to 48 are repeated for this subsequent electrode. Once all the active electrodes have been analyzed, all of the data is collected and recorded during step 51 in the database 42. As can be seen, the invention makes it possible for the first time to carry out a calculation fully automatic T-NRT threshold. Other parameters than the T-NRT thresholds of all the electrodes could be calculated automatically. We can for example calculate the slope of each growth line, and / or the recovery time of the auditory nerve ... In addition, the determination of the T and C thresholds from the T-NRT threshold can also be carried out in all or part by automatic calculation from these parameters, in particular from the T-NRT threshold.

An example of a method according to the invention for classifying the shape of a response curve among a predetermined type using an artificial neural network is represented by the flowchart in FIG. 9. In this figure, we collect, during the step 21, the neural response. An amplitude calibration process is applied to this neural response, in step 22, so as to normalize the values obtained in the interval [0,1]. Then, in step 23, the quality criterion Q is determined. If Q> 0.12, the response is classified in the type TNR not recordable in step 24. If 0.08 <Q <0, 12, the response is classified, in step 25, in an unknown type to be defined by the practitioner. If 0.05 <Q <0.08, it is considered, during step 26, that the response is distorted and we store (associated with this response) information indicating that it is distorted. If 0 <Q <0.05, it is considered during step 27 that the answer is clear and we store (associated with this response) information indicating that it is clear. Furthermore, after step 23 of determining the quality criterion, the shape of the response curve is analyzed by the artificial neural network during step 28, and the value η is calculated, maximum value from the responses of the neurons of the output layer 16, during step 29. If 0 <η <0.9, we assign an unknown type to the response in step 25. If 0.98 <η <1, we consider, in step 30, that the response is reliable, then we proceed to identify its type during step 31 and recording the type of response associated during step 32 since simultaneously the ^" quality Q ^" criterion makes it possible to determine that the response is distorted or clear following steps 26 or 27. The output of these steps 26, 27 is in fact combined by a logical OR 33 making it possible to validate or inhibit the recording of the type of response during step 32.

If 0.9 <η <0.98, it is determined in step 34 that the response is not reliable, but it is nevertheless possible to identify its type and then to record the response type on closer obtained together with recording that the answer is not reliable.

FIG. 10 represents the results of a study for evaluating the performance of an artificial neural network in the classification of the types of responses. 320 examples of patients were presented on the one hand to a very experienced practitioner, and on the other hand to the artificial neural network as described above.

In a first protocol, the practitioner was asked to classify the curves by examining each curve and estimating the presence or not of a neural response. In this first protocol, the practitioner simply had to indicate the existence or not of a response for each curve.

Out of the 320 examples presented to the practitioner, the latter succeeded in determining the existence of a neural response in 213 curves while the artificial neural network in accordance with the invention considered that 221 curves corresponded to a neural response. The practitioner estimated that his choice was safe with an uncertainty of around 2 curves in 10 examples, ie an error rate of around 20%. The artificial neural network made an error of 8 responses, representing a percentage of correct responses of approximately 96%. In a second protocol, the practitioner was asked to identify among the five types of response curve shapes as shown in Figure 2, the different response curves of 320 examples from of patients. To do this, the practitioner was asked to examine the response curve, then examine the types of shapes corresponding to Figure 2, and then try to identify the type of curve shape for each response curve. . It was pointed out to the practitioner that in case of hesitation on the type of response curve shape, the classification had to be carried out in an unknown type Tu.

The result is that illustrated in FIG. 10. In this figure, the hatched bars represent the values given by the artificial neural network, and the non-hatched bars those given for the practitioner. As can be seen, the number of examples representing the types Tb and Te was not large enough for the recognition percentages to be significant. But in practice, these two types are not clinically exploitable, unlike the types Ta and Te which, in practice, are often encountered. For these types of responses, the percentages are 88.1% and 92.13%. This result is interesting for clinical applications since the time required for a practitioner to carry out such a classification is without comparison much greater than that obtained from the device according to the invention. Thus, to classify the 320 examples, the practitioner put about lb.30 while the device according to the invention provided the results in about 10s. In addition, it should be noted that the practitioner had difficulty providing constant results during his classification given the very large number of curves and the difficulty in carrying out the classification.

It has also been found that the practitioner is not consistent in the selection criteria he applies.

The following table 2 provides the percentage of correct responses given by the artificial neural network according to the invention to achieve this classification, it being assumed that the classification by the practitioner is itself 100% or correct.

Table 2

The types Taj, Tbj, Tdj, Tej are the results obtained on curves considered to be distorted.

Another study consisting in having the growth line measured by clinicians on the one hand, and by a method according to the invention, on the other hand, shows that the dispersion of the results obtained between the clinicians, and the dispersion of the results obtained between the method according to the invention and the clinicians are the same. It is thus demonstrated that the method according to the invention behaves in the same way as a clinician, but remains objective and constant in its selection criteria and its calculations. -

Claims

CLAIMS 1 / - Neural response telemetry device comprising:

- an implant (1) of cochlear stimulation comprising a plurality of implantable electrodes (2) adapted to deliver electrical cochlear stimulations and to collect neural responses corresponding to the composite electrical action potential generated by the cochlear neurons, following these stimulations ,

- an implant control device (11, 12, 13) comprising computer processing means (13), and adapted for:

. generate electrical stimulation signals for the electrodes,

. receive the corresponding neural responses, and,. record these neural responses in the form of response curves representing the composite action potential measured over time, characterized in that the computer processing means (13) are adapted to perform automatic shape recognition and classify each response curve among a plurality of types of curve shapes. 21 - Apparatus according to claim 1, characterized in that the means (13) of computer processing comprise an artificial neural network adapted to carry out the recognition of automatic forrne.

3 / - Apparatus according to claim 2, characterized in that the artificial neural network comprises: - an input vector (14) whose number of scalars depends on the number of measurement points of each response curve,

- a first intermediate layer (15) comprising a number of neurons depending on the input vector,

- a layer (16) of output neurons, the number of neurons of which depends on the number of types of curve shapes. 41 - Apparatus according to one of claims 2 or 3, characterized in that the artificial neural network comprises only two layers (15, 16) of neurons.

5 / - Apparatus according to one of claims 1 to 4, characterized in that the types of curve shapes are representative of:

- the absence of a neural response of listed form,

- the presence of positive and / or negative pulses in the response curve.

6 / - Apparatus according to one of claims 2 to 5, characterized in that it comprises means for evaluating the reliability of the classification made by the neural network according to the digital values r | j delivered by the neurons of the output layer (16).

II - Apparatus according to one of claims 1 to 6, characterized in that the computer processing means (13) are adapted to automatically calculate a response amplitude value on each response curve classified into a type of shape of curves significant.

8 / - Apparatus according to one of claims 1 to 7, characterized in that the computer processing means (13) are suitable for: a) automatically determining whether a first response curve obtained for an electrode can be classified in a type significant curve shapes, b) in the negative, repeat the measurement after automatic modification of measurement parameters then repeat step a); if so, acquire a series of successive response curves by decreasing the electrical stimulation voltage until a response curve classified into a type of shape is obtained. of non-significant curves, classify each response curve into a type of curve shape, and automatically calculate a value of at least one parameter relating to electrical stimulation from this series of response curves.

9 / - Apparatus according to claim 8, characterized in that the means (13) of computer processing are adapted to calculate automatically a threshold value T - NRT from the automatically calculated response amplitude values.

10 / - Apparatus according to one of claims 1 to 9, characterized in that the computer processing means (13) are adapted to calculate a quality criterion Q of each response curve representative of the regularity of a final portion of the response curve to determine, according to the value of this quality criterion Q, whether the response curve is:

- quality 1: abnormal and not significant, classified in a non-recordable type TNR, - quality 2: distorted but significant,

- quality 3: clear and significant.

11 / - Apparatus according to claim 10, characterized in that

NOT

where R (n) corresponds to the ^nth point of the response curve R, ni is the sequence number of the first point of said final portion, N is the total number of points of the response curve R.

12 / - Method for computer processing of neural responses from an implant (1) of cochlear stimulation comprising a plurality of implantable electrodes (2) adapted to deliver electrical cochlear stimuli and to collect neural responses corresponding to an action potential electric composite generated by cochlear neurons, induced by these stimulations, and recorded in the form of response curves representing the composite action potential measured over time, process characterized in that computer processing means (13) are used adapted to perform automatic shape recognition and to classify each response curve among a plurality of types of curve shapes.

13 / - Method according to claim 12, characterized in that means (13) of computer processing are used which comprise a artificial neural network adapted to perform automatic shape recognition.

14 / - Method according to claim 13, characterized in that the artificial neural network comprises: - an input vector (14) whose number of scalars depends on the number of measurement points of each response curve,

- a layer (16) of output neurons, the number of neurons of which depends on the number of types of curve shapes.

15 / - Method according to one of claims 13 or 14, characterized in that the artificial neural network comprises only two layers (15, 16) of neurons.

16 / - Method according to one of claims 12 to 15, characterized in that the types of curve shapes are chosen representative of

- the absence of a neural response of listed form,

- the presence of positive and / or negative pulses in the response curve. 17 / - Method according to one of claims 13 to 16, characterized in that the reliability of the classification made by the neural network is evaluated according to the digital values rji delivered by the neurons of the output layer (16).

18 / - Method according to one of claims 12 to 17, characterized in that one automatically calculates a response amplitude value on each response curve classified into a type of significant curve forms.

19 / - Method according to one of claims 12 to 18, characterized in that: - a first response curve obtained for an electrode is automatically determined can be classified into a type of shape of significant curves, - if not, the measurement is automatically repeated after automatic modification of the measurement parameters, then step a) is repeated; in the affirmative, a series of successive response curves is acquired by decreasing the electrical stimulation voltage until a response curve classified in a type of non-significant shape of curves is obtained, each response curve is classified in a type of curve shape, and a value of at least one parameter relating to electrical stimulation is automatically calculated from this series of response curves.

-20 / - A method according to claim 19, characterized in that a threshold value T - NRT is automatically calculated from the response amplitude values calculated automatically.

21 / - Method according to one of claims 12 to 20, characterized in that a quality criterion Q is calculated for each response curve representative of the regularity of a final portion of the response curve, and it is determined, according to the value of this quality criterion Q, if the response curve is: quality 1: abnormal and not significant, classified in a non-recordable type TNR, quality 2: distorted but significant, quality 3: clear and significant.

22 / - Method according to claim 21, characterized in that N

<l ^{ue: Q =}

where R (n) corresponds to the n ^th point of the response curve R, ni is the sequence number of the first point of said final portion, N is the total number of points of the response curve R.