EP4233069A1

EP4233069A1 - Method for assigning a vertigo patient to a medical specialty

Info

Publication number: EP4233069A1
Application number: EP21800995.9A
Authority: EP
Original assignee: Vertify GmbH
Current assignee: Vertify GmbH
Priority date: 2020-10-20
Filing date: 2021-10-19
Publication date: 2023-08-30
Also published as: WO2022082245A1; US20230389793A1; AT524365A1; CN116438607A

Abstract

The present invention relates to a method for assigning a vertigo patient (SP) to a medical specialty (MF), said method comprising the following steps: - capturing eye movements (AB) of the vertigo patient (SP) in the form of video data (VD), - processing the captured video data (VD) in a neural network (NN), - determining at least one medical specialty (MF) on the basis of the result of the processing in the neural network (NN), - outputting an assignment of the vertigo patient (SP) to the determined at least one medical specialty (MF).

Description

Procedure for assigning a vertigo patient to a medical specialty

The present invention relates to a method for assigning a vertigo patient to a medical specialty, a training method for training a neural network for use in such a method, an assignment system for carrying out such a method, and a corresponding computer program product.

It is well known that patients visit their general practitioner for a wide variety of reasons. After headaches, the second most common reason for seeing a family doctor is dizziness. However, dizziness symptoms can have a variety of causes. A wide variety of specialists are therefore required to diagnose the cause of dizziness symptoms. For example, it is possible that dizziness has a neurological cause. A neurological cause would therefore have to be examined by a neurological specialist. It is also possible that dizziness is due to a problem in the inner ear. In such a case, the specialist needed to diagnose the actual cause would be an ear, nose and throat specialist.

In the case of the known solutions, an assignment and thus a referral to the respective specialist is carried out exclusively on the basis of a time-limited examination by the general practitioner. It can often lead to patients being sent to a number of different specialists who either try to diagnose the cause of the dizziness one after the other or even in parallel. This leads to a high expenditure of time for the specialists as well as for the patient. In addition, this also leads to high costs, since there is often only one cause of the dizziness. This in turn means that in many cases only one of the specialists is actually able to diagnose the cause of the dizziness, while the other specialists consulted cannot make a successful diagnosis. Furthermore, there may be unnecessary use of expensive imaging methods, such as magnetic resonance tomography. The increased costs and the increased effort are either borne by the patient themselves or by society in the form of the respective health insurance. It is the object of the present invention to at least partially eliminate the disadvantages described above. In particular, it is the object of the present invention to achieve assignment to an explicit medical specialty in a cost-effective and simple manner.

The above object is achieved by a method having the features of claim 1, a training method having the features of claim 12, an assignment system having the features of claim 15 and a computer program product having the features of claim 17. Further features and details of the invention result from the dependent claims, the description and the drawings. Features and details that are described in connection with the method according to the invention also apply, of course, in connection with the training method according to the invention, the allocation system according to the invention and the computer program product according to the invention and vice versa, so that the disclosure of the individual aspects of the invention is always referred to reciprocally can be.

According to the invention, the method serves to allocate a dizzy patient to a medical specialty. To do this, such a procedure has the following steps:

Recording of eye movements of the dizzy patient in the form of video data,

processing the captured video data in a neural network,

Determination of at least one medical specialty based on the result of the processing in the neural network,

Outputting an allocation of the dizzy patient to the specific, at least one, medical specialty.

A method according to the invention is therefore used to find out a medical specialty which, with a high degree of probability, is able to diagnose the actual cause of the dizziness in the dizzy patient. In the context of the present invention, a vertigo patient is a patient who suffers from a vertigo symptom. A method according to the invention now makes it possible to provide the assignment of a medical specialty independently or in addition to the experience of the respective general practitioner. The assigned and issued medical specialty is in all probability the specialty in which the cause of the dizziness symptom is still to be diagnosed, so that the allocation to the medical specialty means that the specialist can be consulted directly who is also highly likely to be the cause of the dizziness can diagnose. In the case of multiple referrals based on the state of the art, a possibly reduced sense of responsibility on the part of the individual medical specialists can also be assumed.

In the context of the present invention, a neural network is to be understood as meaning any form of artificial intelligence, for example any form of differentiable programming.

In comparison to the previous solutions, a consultation by a large number of specialists no longer has to take place in parallel and/or one after the other, but only a single specialist or a few specialists can be consulted, which corresponds to the assigned medical specialty. As a result, it is possible to spend significantly less time consulting the specialist with the patient and avoiding unnecessary consultations with other specialists. The reduced expenditure of time also goes hand in hand with a reduced expenditure of costs, since unnecessary specialist visits no longer have to be paid for.

A method according to the invention is based on detecting eye movements of the dizzy patient. This is done with a video sensor, for example in the form of a camera system. A particularly simple and cost-effective embodiment of a method according to the invention, as explained below, can be carried out on a mobile terminal. However, it is of course also possible for special terminals, which have associated video sensors or camera systems, to be installed preferably at a family doctor. In the simplest way, however, a method according to the invention can be used as a software product or as a computer program product, for example in Form of an application, run on a mobile device in the form of a tablet or a mobile phone.

In principle, it is also possible for the eye movements to be recorded separately from the method according to the invention. For example, existing camera systems can record the eye movements in the form of a video file and make them available to the method according to the invention for further processing.

Using the mobile radio device as an example, it will now be explained how a method according to the invention can be carried out. With the help of the mobile device's camera, it is possible to record the eye movements of the dizzy patient in the form of video data. These eye movements can be freely recorded eye movements or guided eye movements. As will be explained later, guided eye movements, which can be recorded on the basis of preferably standardized eye movement tests, are particularly advantageous. For example, movement patterns of the eyes with the smallest movements and/or twitches can enable conclusions to be drawn about the corresponding medical specialties.

Once the vertigo patient's eye movements have been captured in the form of video data, they can be processed using a neural network. In the context of the present invention, a neural network is to be understood as meaning the use of what is known as artificial intelligence. The neural network is a trained neural network, which is provided with corresponding neural network nodes and associated weightings on the basis of a large number of eye movement data labeled in a controlled manner. As a result, the processing of the captured video data in this trained neural network makes it possible for the captured video data to be assigned to a medical specialty with a high level of certainty, in which a diagnosis of the cause of dizziness can also be made in a subsequent consultation with the associated specialist. In other words, the video data is processed in the neural network, with at least one medical specialty being determined by the neural network on the basis of the result of this processing. Finally, in a method according to the invention, this allocation of the dizzy patient to the specific, at least one, medical specialty is output. It is therefore possible that after the recording of the eye movements and the processing in the neural network, an output of one or even several medical specialties is possible directly on the mobile terminal device, which with a high probability allow the cause of the dizziness to be diagnosed following a method according to the invention . The assigned medical specialty means that the dizziness patient then visits the specialist in the assigned medical specialty, where the actual cause of the dizziness can be diagnosed and a suitable therapy can be initiated.

As can be seen on the basis of the above explanation, the allocation to the medical specialty is now at least partially standardized. This is based on the fact that the experience of a large number of experts is stored in the neural network through the corresponding training methods in the form of the weighting of the individual network nodes. The recorded eye movements in the form of the video data, alone or in combination with patent answers explained later, can thus automatically and with a high degree of accuracy assign that medical specialty in which the cause of the dizziness can be diagnosed with a high degree of probability.

It should also be pointed out that a method according to the invention is not used to diagnose a cause, but rather enables a diagnosis to be made following the method according to the invention. While in known procedures a family doctor selected a large number of specialists who had to carry out all diagnostic tests for the cause of dizziness, one or significantly fewer specialists in the area of the assigned medical specialties can now be selected explicitly in accordance with the invention, in order to now be cheaper, faster and above all easier to subsequently diagnose the cause of the dizziness. In this way, a transmission to the wrong medical specialties is avoided or at least the probability is significantly reduced. With regard to the training and the design of the neural network, more detailed explanations will follow later. It can bring advantages if, in a method according to the invention, the video data are anonymized before processing in the neural network, in particular are translated into eye movement parameters. While it is fundamentally possible to process video data directly in a neural network, converting it into anonymous data, for example in the form of eye movement parameters, can bring additional benefits. On the one hand, this makes it possible to provide a compression of the amount of data, so that the anonymized data has a significantly smaller memory size than is the case for the recorded video data. At the same time, this makes it possible to provide anonymization, which then allows these anonymized eye movements to be sent from the mobile device to other devices. This is particularly advantageous when, for example, the neural network is centralized in a network solution, for example in the form of a processing cloud. The processing and use of personal data in a cross-system network is fundamentally based on high security standards. The anonymization step thus allows the anonymized eye movement information to be communicated to a central network without security concerns, while the patient-related data remain on the mobile device. Anonymized eye movement parameters are, for example, directional vectors of the individual eyes, in particular in the form of the viewing direction of the eyes, in the form of a head tilt and/or in the form of a head orientation. However, it is also possible that the eye movement parameters are stored in another form, such as an Euler angle, rotation matrix, or pixel translation. It is also conceivable that further supporting information, such as the head inclination and/or head orientation, can also be recorded as direction vectors, rotation matrix, or quaternions. The anonymity of this data is given by the low sampling rate and the lack of biometric information, such as the interpupillary distance. The anonymization step preferably takes place locally in the respective recording device, so that the separation between patient-related data and anonymized data is made available locally. This translation can be provided, for example, in the form of an algorithm and/or also through the use of one or more neural networks. There are further advantages if, in a method according to the invention, the recorded video data, in particular in anonymous form, is transmitted from a recording device to a processing device in which the processing takes place using the neural network, with the specific medical specialty preferably transmitting back to the recording device and is output at the detection device by means of an output device. As has already been explained in the previous paragraph, there can be advantages in making the processing intelligence in the form of the processing neural network available centrally on a server, for example in a cloud. In particular, if the eye movements are anonymized in the form of the video data, for example by the eye movement parameters explained in the previous paragraph, this anonymized data can be transmitted, preferably wirelessly, to this central cloud server and processed there. If the server is a correspondingly secured variant, non-anonymized data can also be stored and/or processed there. This also makes it possible to present computationally intensive solutions in the form of the neural network centrally on a server, so that the computing capacity does not have to be made available by the mobile end device. The process here is similar to a server-based speech recognition system, for example in a cloud. The eye movements are recorded and, if necessary, pre-processed locally using the detection device, processed centrally using the processing device arranged on the cloud server and sent back for output to the output device, which is in particular designed integrally with the detection device.

Further advantages can be achieved if, in a method according to the invention, during the detection of the eye movements, at least one test video is played and/or instructions are displayed on a display device of a detection device. For example, oculomotor tests such as gaze direction nystagmus, optokinetic nystagmus, gaze tracking movement, fixation suppression can be displayed in the form of image or audio material using a smart device. Such a test video can, for example, have a guide function for a forced or guided eye movement. It is also possible that the test video shows different light-dark sequences (e.g. reproduction of a nystagmus drum with a black-and-white bar pattern). to evoke appropriate eye reactions in the form of eye movements. The individual test videos are to be understood in particular as completed eye movement tests, with a separate eye movement preferably being recorded in the form of video data for each eye movement test. The processing in the subsequent step can, for example, be specific to the respective eye movement test. For example, if a guided eye movement is performed, the recorded eye movements are fed to a first neural network for processing, while in the case of a light-dark variation as an eye movement test, the recorded eye movements are fed to a second, independent, neural network. Here it is easy to see that not just a single, but several different eye movement tests can be combined with a method according to the invention in a simple and cost-effective manner, so that the combined evaluation in correspondingly specifically assigned neural networks, the security of the assignment to the correct medical specialty and so that the advantages of the present invention can be further enhanced. It is also conceivable that additional information is output in addition to the assignment, which information can, for example, provide focal points for the subsequent diagnosis by the specialist.

It also has advantages if, in a method according to the invention, at least one patient's answer to at least one patient's question is recorded in addition to the video data. It is thus possible for one or more patient questions to be displayed on the detection device in a corresponding display section. In the simplest case, this involves querying basic patient information such as age, gender, weight or the like. However, it is also conceivable for information relevant to the dizziness system to be queried, for example in which context dizziness has occurred, since when the dizziness has occurred, whether there was a new prescription for spectacles, or the like. The associated patient responses to the patient questions can be used to enhance different steps of the procedure. For example, they can be used to select different eye movement tests based on patient responses. It is also possible to change weights in processing in the neural network based on the patient's response. In addition, it is still possible to view the processing result of the correlate the neural network with the responses received from the patient and in this way to further increase the certainty of assignment to the correct medical specialty. The patient's questions and the patient's answers preferably remain locally on the recording device and are in particular not transmitted to the neural network of the processing device, or only in an anonymous manner.

Further advantages can be achieved if, in a method according to the invention, the eye movements are recorded for at least two different eye movement tests, with the video data being processed for different eye movement tests in different neural networks. As discussed previously, different eye movement tests can provide different levels of significant information as to which medical specialty is able to diagnose the cause of dizziness. In this embodiment, different eye movements are stored in different video data specifically for each of the eye movement tests carried out and are fed to different neural networks in a correspondingly specific manner. In other words, each neural network outputs its own processing result, which can then be combined to form a common processing result and can be output in the form of a medical specialty assignment. In this case, the individual video data for the eye movement tests preferably remain separate from one another and, in particular, are not mixed with one another.

It can also be advantageous if, in a method according to the invention, a quality control of the video data takes place before and/or during and/or after the detection of the eye movements in the form of video data. Quality control means, for example, the alignment of the video sensor, so it must be checked whether the eyes and the line of sight of the dizzy patient are actually recorded. Ambient parameters such as the lighting situation and the direction of the light can also be monitored. The quality of the recording itself, particularly in the form of shaking of the video data, can also be monitored. In this case, in particular, sensors are used which are already present in the mobile terminal device in the form of the detection device. On the one hand, the quality requirements and the monitoring result can mean that it is necessary to carry out the eye movement test again. Also it is It is possible to use live feedback to inform the user directly while the eye movement test is being carried out that the necessary quality requirements are not currently being met, for example the eyes are not completely within the detection range of the camera, so that the user can already adapt the quality criteria while the test is being carried out will.

Further advantages can result if, in a method according to the invention, the determination of the at least one medical specialty includes a safety factor that includes the accuracy of the assignment of the at least one medical specialty. This is particularly useful when two or more eye movement tests can be evaluated. In combination with this, it becomes possible to output the accuracy in a numerical value in the form of the safety factor, which is specified in particular as a percentage. For example, it can be output that there is a high probability that a diagnosis in the medical field of an ear, nose and throat doctor will be possible. The safety factor, for example in the form of a trust factor, can be 90 percent, for example. In addition, it can be stated that a cause in a neurological area is also fundamentally possible, but with a reduced probability of, for example, 30 percent. In the event that an assessment of the allocation is desired, this allows additional statement information to be given to the dizzy patient or the user of a method according to the invention. This safety factor can also be assigned to individual eye movement tests, so that the validity of the assignment of the medical specialties can be increased even further.

It is also advantageous if, in a method according to the invention, the recorded video data are made available for a manual check. Such a manual check can be carried out, for example, by experts who are able to interpret the eye movements in a manual manner. The method according to the invention is now able to make this recorded video data available to this manual check, in particular in video form. This will make sense in particular if the neural network does not allow assignment to a medical specialty with sufficient certainty. Manual verification results in accuracy in these individual cases is given with a high degree of certainty and can also allow feedback to the neural network, so that a training effect can also be achieved for the individual weighting parameters in the neural network. For example, the method according to the invention can make the captured video data available for the manual check as a function of a limit value for the safety factor mentioned. For example, results of the processing from the neural network can be transmitted to an expert with a safety factor of less than 60 percent, for example, who can then carry out a manual check and report back within the framework of telemedicine. Of course, this is also possible at the direct active request of the user.

There are also advantages if, in a method according to the previous paragraph, the video data is made available for manual verification in an anonymized form, in particular as artificial video data. While it is fundamentally possible to transmit real eye movements in the form of actually recorded video data, anonymous transmission may be desired for the data protection reasons already mentioned. While the neural network can work with mathematical input information in the form of anonymized eye movements, for example in the form of eye movement parameters, experts usually rely on video information for manual verification. It is thus now possible with this embodiment to generate artificial video data in an anonymous manner on the basis of the video data, which transfer the actually recorded eye movements to a virtual and therefore artificial head or artificial eyes. In other words, the expert for the manual check is provided with an eye movement digitized from a real eye movement and thus virtual eye movement on an artificial and therefore virtual head, so that he can carry out a manual check qualitatively and quantitatively, but no personal data of the dizzy patient receives.

It is also advantageous if additional secondary parameters are recorded in a method according to the invention, in particular at least one of the following:

Brightness, lighting situation,

acceleration of a detection device,

Positioning of a detection device.

The above list is a non-exhaustive list. Of course, two or more secondary parameters can also be combined with one another. The use of secondary parameters serves in particular to carry out a quality check in order to check a quality criterion before processing in the neural network during or after the recording of the eye movements. However, this additional information can also be used to make the processing in the neural network even more precise, for example the brightness of the background or environmental parameters such as the ambient temperature when the test is carried out. For example, it is possible that eye movements are worse when looking at a bright background and/or at particularly low ambient temperatures. By recording the corresponding parameters, this can be taken into account in the evaluation. In particular, existing sensors of the detection device are used as corresponding sensors. If the detection device is a mobile terminal, for example, a gyroscope, an acceleration sensor, a lidar, a brightness sensor or similar sensors can be used.

The present invention also relates to a training method for training a neural network for use in a method according to the invention, having the following steps:

Making available a large number of eye movements of vertigo patients in the form of video data, manual labeling of the large number of eye movements,

Training the neural network with the multitude of manually labeled eye movements. In particular, in addition to conspicuous eye movements of dizzy patients, healthy eye movements are also made available for a training method. Manual labeling means that eye movements are manually classified into medically conspicuous eye movements or medically inconspicuous eye movements and a medical specialty is assigned. This manual labeling is carried out by experts who, based on the results of at least one eye movement test carried out, can select the relevant medical specialty in which the cause of dizziness can be diagnosed with a high degree of certainty. A large number of such labeled eye movements are now used to train the neural network in the known way. The training effect leads to the weighting of the corresponding mathematical operation being varied and refined in a single network node of the neural network until the output of the neural network corresponds to the desired result in the form of the labeled eye movement. This requires a large number of eye movement data, which is made available in particular on the basis of existing patient information. However, it is also possible that anonymized data is used here, which is then used as a basis for the training process. Manual labeling also means, for example, the higher-level labeling of an entire group of eye movements that are marked as conspicuous in a higher-level step.

It is also advantageous if, in a training method according to the invention, artificial video data are generated for the provision, which are in particular manipulated manually and/or automatically. This artificial video data can be completely artificially generated. However, it is also conceivable for the artificial video data to be generated on the basis of real video data and then additionally manipulated. If, for example, the eye movement of a dizzy patient is recorded and labeled, this recording can be made anonymous and translated into artificial video data. If not just a single virtual head is used, but instead ten different virtual heads, for example, ten different artificial video data can be generated from a single recording of a dizzy patient, which are then automatically duplicated and made available in larger numbers for training the neural network can be asked. Despite a low labeling effort, a large number of training data can be achieved in a short time.

Further advantages can result if, in a training method according to the invention, the large number of eye movements are at least partially labeled multiple times. This means that the same eye movements of the same dizzy patient are labeled multiple times by different experts in order to avoid unwanted inaccuracy in the neural network. The statements and thus the label results of the individual experts can be weighted differently over the duration of the labeling process and the training process in order to exert different training effects on the neural network according to the different weighting. For example, information such as interaction between different experts and intra-agreement of a single expert could be included in the weighting.

Another object of the present invention is an assignment system for assigning vertigo patients to a medical specialty. Such an allocation system has a detection device for detecting eye movements of the dizzy patient in the form of video data. A processing device for processing the recorded video data in a neural network is also provided. A determination device can be used to determine at least one medical specialty based on the result of the processing in the neural network. In addition, the allocation system has an output device for outputting an allocation of vertigo patients to the specific at least one medical specialty. The detection device, the processing device, the determination device and/or the output device are designed in particular for carrying out a method according to the invention. An allocation system according to the invention thus brings with it the same advantages as have been explained in detail with reference to a method according to the invention.

There can be advantages if, in an allocation system according to the invention, the detection device and the output device are arranged in a local allocation device and at least the processing device is arranged in a central allocation device. Below is the one that has already been explained several times To understand cloud solution, with a mobile device in the form of a z

Tablets, mobile phones, digital glasses systems, digital contact lenses or other digital visual aids are used as a detection device to detect eye movements and then, in particular in an anonymous manner, are transmitted to the cloud and the processing device located there. After the cloud-based processing has been carried out, the processing result is transmitted back to the local allocation device and thus to the mobile device, so that it can be output there to the dizzy patient or the test person.

The subject matter of the present invention is also a computer program product, comprising instructions which cause an allocation system according to the invention to carry out the method steps according to a method according to the invention. A computer program product according to the invention thus brings with it the same advantages as have been explained in detail with reference to an allocation system according to the invention and a method according to the invention.

Further advantages, features and details of the invention result from the following description, in which exemplary embodiments of the invention are described in detail with reference to the drawings. The features mentioned in the claims and in the description can each be essential to the invention individually or in any combination. They show schematically:

1 shows an embodiment of an allocation system according to the invention,

2 shows an embodiment of an allocation system according to the invention,

3 shows an embodiment of an allocation system according to the invention,

4 shows an embodiment of an allocation system according to the invention, 5 shows an embodiment of an allocation system according to the invention,

6 shows an embodiment of an allocation system according to the invention,

7 shows an embodiment of an allocation system according to the invention,

8 shows an embodiment of an allocation system according to the invention,

9 shows an illustration of a training method according to the invention.

FIG. 1 shows an allocation system 10 according to the present invention in its simplest configuration. A separation into a local assignment device 100 and a central assignment device 200 can already be seen here. In principle, however, it would also be conceivable to carry out the entire method in the local assignment device 100 .

A cell phone is shown schematically here as a detection device 20 as the local assignment device 100 of the assignment system 10 . This is able, with the help of a camera, to record the eye movements AB of a dizzy patient SP. This recording leads to recorded eye movements AB in the form of video data VD. This video data VD is then processed in a processing device 30 using a neural network NN, either locally or, as shown in FIG. 1, in a central allocation device 200 . With the aid of the determination device 40, for example in the form of a cloud memory, at least one medical specialty MF is assigned on the basis of the processing result, which is transmitted back to the mobile terminal device as a local assignment device 100. The mobile phone as a detection device 20 is also designed here as an output device 50 and is used to output the assigned medical specialty MF.

A further development of the embodiment of FIG. 1 is shown in FIG. The decisive difference here is that the video data VD are no longer transmitted to the central allocation device 200 in the cloud. Rather, it takes place in local assignment device 100 a translation of the video data VD in eye movement parameters AP, which represent an anonymization of the real video data VD. The transmission to the central assignment device is now anonymous or pseudonymised if the assignment is to be assigned at the end of the method by not transmitting the video data VD but only the eye movement parameters AP. The further course of the processing and the determination as well as the output of the assignment is identical to the description of the embodiment of Figure 1.

FIG. 3 shows a further development of the embodiment from FIG. These patient responses are entered in particular by the doctor or medical attendant accompanying the method or by the patient himself. These patient responses are used either to select different eye movement tests ABT, to select different neural networks NN or to carry out a plausibility check of the assigned medical specialty MF in the output.

FIG. 4 is also a further development of an assignment system 10. Here it is now possible, for example, for an assigned medical specialty MF to determine a safety factor SF with the probability of this being the correct medical specialty MF. If the safety factor in particular falls below a limit value, for example with a safety of less than 60 percent, in this embodiment the central assignment device 200 can allow provision for a manual check by experts EX. A manual selection of a check by an expert EX is of course also conceivable. This can bring advantages when carrying out the method as well as the training method. However, this provision does not take place directly, but in an anonymous manner, in that the anonymized data transmitted to the central allocation device 200 is transferred to a virtual image of a person in the form of artificial video data KVD. This artificial video data KVD can now be checked manually by the experts EX, so that on the one hand a checked output of the assigned medical specialty MF is possible and on the other hand this feedback serves a learning effect of the neural network NN.

FIG. 5 shows a variant of FIG. 4, in which case the verified assignment of the medical specialty MF does not take place via the central assignment device 200, but directly from the experts EX to the local assignment device 100 in the form of telemedicine. In this case, videos are not sent to a medical expert EX, as is already the case with other known telemedical software approaches, but the extracted non-sensitive eye movements AB are inserted into an artificially generated head, so that artificial video data KVD are displayed in the same way as these medical experts EX already are are also accustomed to when examining patients directly, thus increasing the accuracy of assessment.

Figure 6 shows the combination of the embodiment of Figures 4 and 5.

FIG. 7 shows schematically how an eye movement test ABT can run. A light-dark progression/black-white bar pattern, which in particular moves, can thus be shown on the display device 22 of the detection device 10 in the form of a display. Of course, other and more complex representations, such as cartoon characters or the like, are also conceivable within the scope of the present invention. The dizzy patient SP views this video of the eye movement test ABT and during this viewing time his eye movements AB are recorded by the detection device 20 . Due to the video being played, certain eye movements, for example a saccade, are evoked in the viewer by the constant refocusing on a bar from the bar pattern, which provides information on whether the patient's eye movements are healthy or unhealthy. The subsequent course of processing in the neural network NN is again essentially identical in the central allocation device 200 . Here, too, it is again shown how different artificial video data KVD can even be generated, which can then be viewed by one or more experts. This right-hand section of FIG. 7 serves in particular to increase the learning effect or the training effect for the neural network NN. It is also shown that at least one subject matter expert Possibility to manipulate the eye movements so that in particular inconspicuous to conspicuous eye movements are specifically manipulated so that the training set can be supplied with the necessary data.

FIG. 8 shows how different video data VD can be recorded for different eye movement tests ABT and then made anonymous. The middle shows how different eye movement parameters AP are generated for each eye movement test ABT, which can then be processed either in a common neural network NN or in different neural networks NN in the processing device 30 . The combination of these individual test results is now carried out again on the output device 50 in the form of an assignment to one or more medical specialties MF.

FIG. 9 again shows how a training method can be carried out in principle. During a normal run, real video data VD are generated from eye movements AB of the dizzy patient SP. In the manner already explained, these run through the neural network NN and a downstream additional neural network NN for generating artificial video data KVD, which are then made available to the experts EX for manual labeling. This part is used to make the labeled artificial video data KVD available to the neural network NN as training data and in this way to adapt the weighting in the neural network NN for its training.

The upper section of FIG. 9 shows how a large number of training data can be generated from a single video recording. A video recording in the form of video data VD can thus be transmitted to a large number of different virtual heads by a neural network NN, so that a large number of different artificial video data KVD can be generated from a real video recording. It is also conceivable for the artificial video data KVD to be completely generated virtually at the top left, with the eye movements and other properties such as the head orientation or the like being known exactly as a result of the virtual generation. These can be, in particular based on a photorealistic artificial base representation, via a further neural network NN shown in the middle above photo-realistic artificial video data KVD is transmitted at the top right, so that both the virtually generated artificial video data KVD and the photo-realistic artificial video data KVD are then used for training the neural network NN. Even with a relatively small database, this means that a sufficient amount of training data for training the neural network NN is available in the processing device 30 with high accuracy and at high speed.

The above explanation of the embodiments describes the present invention exclusively in the context of examples. It goes without saying that individual features of the embodiments can be freely combined with one another, insofar as this makes technical sense, without departing from the scope of the present invention.

Reference List

10 allocation system

20 detection device

22 display device

30 processing device

40 determination device

50 dispenser

100 local allocation device

200 central allocation device

SP vertigo patient

MF medical specialty

SF safety factor

AB eye movements

ABT eye movement test

AP eye movement parameters

VD video data

KVD artificial video data

PA patient response

PF patient question

NN neural network

EX experts

Claims

22 patent claims

1 . A method for assigning a vertigo patient (SP) to a medical specialty (MF), comprising the steps of:

- Recording of eye movements (AB) of the dizzy patient (SP) in the form of video data (VD),

- processing the captured video data (VD) in a neural network (NN),

- determining at least one medical specialty (MF) based on the result of the processing in the neural network (NN),

- Outputting an assignment of the dizzy patient (SP) to the at least one specific medical specialty (MF).

2. The method according to claim 1, characterized in that the video data (VD) are anonymized before processing in the neural network (NN), in particular in eye movement parameters (AP) are translated.

3. The method according to any one of the preceding claims, characterized in that the recorded video data (VD), in particular in anonymous form, is transmitted from a recording device (20) to a processing device (30), in which the processing by means of the neural network (NN ) takes place, the specific medical specialty (MF) preferably being transmitted back to the detection device (20) and being output at the detection device (20) by means of an output device (50).

4. The method as claimed in one of the preceding claims, characterized in that when the eye movements (AB) are detected, at least one test video is played back on a display device (22) of a detection device (20). Method according to one of the preceding claims, characterized in that at least one patient answer (PA) to at least one patient question (PF) is recorded in addition to the video data (VD). Method according to one of the preceding claims, characterized in that the eye movements (AB) are recorded for at least two different eye movement tests (ABT), the video data (VD) for the different eye movement tests (ABT) being processed in different neural networks (NN). Method according to one of the preceding claims, characterized in that a quality control of the video data (VD) takes place before and/or during and/or after the detection of the eye movements (AB) in the form of video data (VD). Method according to one of the preceding claims, characterized in that the determination of the at least one medical specialty (MF) includes a safety factor (SF) which includes the accuracy of the assignment of the at least one medical specialty (MF). Method according to one of the preceding claims, characterized in that the recorded video data (VD) are made available for manual checking. Method according to Claim 9, characterized in that the video data (VD) are made available for manual checking in anonymous form, in particular as artificial video data (KVD).

Method according to one of the preceding claims, characterized in that additional secondary parameters are recorded, in particular at least one of the following:

- Brightness

- Lighting situation

- Accelerations of a detection device (20)

- Positioning a detection device (20) training method for training a neural network (NN) for use in a method having the features of one of claims 1 to 11, comprising the following steps:

- Making available a large number of eye movements (AB) of vertigo patients (SP) in the form of video data (VD),

- Manual labeling of the multitude of eye movements (AB),

- Train the neural network (NN) with the multitude of manually labeled eye movements (AB). Training method according to Claim 12, characterized in that artificial video data (KVD) are generated for the making available, which are in particular manipulated manually and/or automatically. Training method according to one of Claims 12 or 13, characterized in that the large number of eye movements (AB) are at least partially labeled multiple times. Allocation system (10) for allocation of dizzy patients (SP) to a medical specialty (MF), having a detection device (20) for detecting eye movements (AB) of the dizzy patient (SP) in the form of video data (VD), a processing device (30 ) for processing the captured video data (VD) in a neural network (NN), a determination device (40) for determining at least one 25 medical specialty (MF) based on the result of the processing in the neural network (NN) and an output device (50) for outputting an allocation of the dizzy patient (SP) to the at least one specific medical specialty (MF), wherein the detection device (20 ), the processing device (30), the determination device (40) and/or the output device (50) are designed in particular for carrying out a method having the features of one of claims 1 to 11. Allocation system (10) according to claim 15, characterized in that the detection device (20) and the output device (50) are arranged in a local allocation device (100) and at least the processing device (30) in a central allocation device (200). A computer program product comprising instructions which cause an allocation system (10) having the features of any of claims 15 or 16 to carry out the method steps of any of claims 1 to 11.