Device and method for determining a medical health parameter of a subject using voice analysis, mobile device and computer program

Abstract

Method (100) for determining a health parameter of a subject using voice analysis, comprising the following steps: Providing individual model parameters that indicate a functional relationship between a speech sample of the subject himself or speech characteristics derived from the speech sample and an associated, current health parameter of the subject, Providing (102) a digitized speech sample of the subject, Evaluating (104) the digitized speech sample based on the individual model parameters to obtain measurement information that is within a tolerance range based on an instantaneous value of the subject's health parameter, and Adapting (112) the individual model parameters of the subject by performing a machine learning procedure based on a plurality of speech samples of the subject and on associated invasively measured medical health parameters to obtain adapted individual model parameters, where the current, individual model parameters are replaced by the adapted, individual model parameters (114); and the step (102) of evaluating the speech sample based on the adapted, individual model parameters is carried out (116).

Description

[0001] The present invention relates to a device and a method for the non-invasive determination of an organ function-dependent medical health parameter, e.g., a current blood glucose level, of a subject by means of voice analysis based on a speech sample from the subject. In particular, the present invention relates to a learning voice analysis device, e.g., a learning non-contact diagnostic device, and to a method for operating the learning voice analysis device for the learning non-contact determination of a medical health parameter.

[0002] For self-monitoring of medical health parameters, such as blood sugar levels or glucose concentrations, by a patient at home and for external monitoring in laboratories, hospitals, nursing homes, etc., a variety of procedures and devices that touch and injure (invasive) or irradiate the human body, as well as combined concepts, are known (see, e.g., WO 2011 / 135 562 A1, WO 2014 / 143 452 A1).

[0003] The measurement accuracy achieved so far by control methods and devices varies and is often not specified by the manufacturers. For the subsequent and crucial diagnostic interpretation of the measured values ​​following the recording of the medical health parameter, different classifications, classes, or limit values ​​are typically used, which are adjusted or updated from time to time.

[0004] Most devices and procedures for determining a medical health parameter, particularly blood glucose levels, are technology-oriented rather than user-oriented and involve bodily harm (bloody or invasive). They require a blood sample obtained through injury, i.e., a puncture with a syringe or lancet. These injuries are unpleasant for the individual and always carry the health risk of infection.

[0005] Each blood glucose measurement typically consumes high-quality disposable materials, such as test strips, which are then discarded as waste. Furthermore, it should be noted that the time and expense involved in self-monitoring of blood glucose is considerable and reaches significant economic dimensions (even without considering diabetes treatments) in Germany and worldwide.

[0006] Even well-known, non-invasive methods for blood glucose measurement require at least unpleasant and risky physical contact or even radiation with unknown side effects, and are often neither tested nor established in practice. Furthermore, the measurement inaccuracies, which are crucial for the effectiveness of these methods and often unavoidable, are rarely explicitly stated or readily comprehensible. Moreover, the measurement results are not individualized, meaning they are not tailored to the specific test subject.

[0007] In summary, it can be stated that the measurement results of previously known control procedures for determining a medical health parameter, e.g., the blood sugar level, are neither personally user-oriented nor adapted or individualized to the user.

[0008] US 7,925,508 B1 relates to the detection of extreme hyperglycemia or hypoglycemia based on an automated analysis of speech patterns, whereby a communication device analyzes received speech utterances and then detects a physical target state, such as hyperglycemia or hypoglycemia, of the speaker. The analysis is performed by comparing speech units of the utterance with corresponding baseline speech units received from the same person when that person did not exhibit the physical target state, and based on comparison results from previous steps.

[0009] The scientific publication “TSCHÖPE, Constanze; WOLFF, Matthias: Statistical classifiers for structural health monitoring. In: IEEE Sensors Journal, Vol. 9, No. 11, November 2009, pp. 1567-1576” refers to the determination of statistical classifiers for structural health monitoring.

[0010] The scientific publication “WOLFF, Matthias ; TSCHÖPE, Constanze: Pattern recognition for sensor signals. In: IEEE Conference proceedings, Sensors 2009, 25-28 October 2009, pp. 665-668) refers to pattern recognition for sensor signals.

[0011] US 7 962 342 B1 refers to a dynamic user interface for a temporarily impaired version based on an automatic analysis of language patterns.

[0012] US 7 398 213 B1 relates to a procedure and system for diagnosing a pathological phenomenon using a speech signal.

[0013] US 2004 / 0122790A1 refers to a computer-aided data processing system and a method for automatic learning.

[0014] Based on the existing state of the art, the object underlying the present invention is to create an improved concept for determining a medical health parameter of a subject, which provides both a sufficiently high measurement accuracy for the individual subject and is designed to be as user-friendly as possible.

[0015] This problem is solved by the independent patent claims with method claim 1, apparatus claims 21 and 23 and computer program claim 25.

[0016] Further developments according to the invention are defined in the associated dependent claims.

[0017] A procedure for determining a subject's health parameter using voice analysis comprises the following steps: providing a digitized speech sample of the subject and evaluating the digitized speech sample based on individual model parameters that specify a functional relationship between the speech sample itself or speech features derived from the speech sample and an assigned, instantaneous health parameter, in order to obtain measurement information that is within a tolerance range based on an instantaneous value of the subject's health parameter.

[0018] A device for determining a health parameter of a subject by means of voice analysis comprises a processing unit configured to evaluate a digitized speech sample of the subject based on individual model parameters in order to obtain measurement information that is within a tolerance range based on an instantaneous value of the subject's health parameter, wherein the individual model parameters specify a functional relationship between the speech sample or speech features derived from the speech sample and an associated instantaneous health parameter.

[0019] A mobile device comprises a speech capture device for capturing a speech sample from a subject; the device for determining a health parameter of a subject by means of voice evaluation comprises a processing device designed to evaluate a digitized speech sample of the subject based on individual model parameters in order to obtain measurement information that is based on an instantaneous value of the subject's health parameter within a tolerance range, wherein the individual model parameters specify a functional relationship between the speech sample or speech features derived from the speech sample and an associated instantaneous health parameter; and a display device for displaying an optical, haptic, and / or acoustic signal dependent on the measurement information.

[0020] The core idea of ​​the present invention is to provide a non-invasive method for determining a medical health parameter, such as the current blood glucose level or blood glucose class, individually for each subject. This individual determination of the health parameter is based on a voice analysis of a speech sample from the subject. During the voice analysis, relevant voice characteristics of the subject are extracted from the speech sample. These voice characteristics then serve as the starting point for evaluating the data using individual models or model parameters specific to the subject, employing "intelligent" signal processing.

[0021] Initially, during a so-called learning phase, baseline values ​​for the models or model parameters for voice characteristic analysis can be determined. These models or model parameters are then applied during the measurement phase (core phase) when using the method or device according to the invention. During an adaptation phase, which can also take place in parallel with the use of the method or device according to the invention, the models or model parameters for voice characteristic analysis can then be continuously and individually adapted. The newly adapted models can then be used for subsequent measurements.

[0022] Some organ-function-dependent medical health parameters, such as a subject's current blood glucose level, also affect various organs in the body. Blood glucose is required for essentially all vital functions and thus influences them, including the brain, which controls these functions, and especially the organs that produce and influence sound. Therefore, certain voice characteristics (for example, after specific processing), obtained from a predefined and reproducible speech sample, can be correlated with the current medical health parameter, i.e., the current blood glucose level or blood glucose class, to establish a functional relationship.For this purpose, individual models (model parameters) are provided, which are determined (individually) for each participant and can be individually adapted. These models demonstrate a functional relationship between the speech sample itself or language features derived from the speech sample and the current medical health parameter (e.g., blood glucose level or blood glucose class). The models or model parameters thus serve as the basis for evaluating the voice sample using "intelligent" signal processing.

[0023] The following are some examples (not an exhaustive list) of well-correlated speech features for voice analysis to determine current blood glucose levels. These features are from the third category (statistical and derived features): 1. Spectral roll-off point 0.75, smoothed, 3. LPC coeff. 2. RASTA-PLP filtered spectrum, smoothed, delta, spectral centroid 3. FO-freq., smoothed, spectral centroid 4. FO-freq., smoothed, lin. regression slope 5. Spectral roll-off point 0.75, smoothed, 4. LPC coeff. 6. RASTA-PLP filtered spectrum, smoothed, channel 7, standard derivation of peak distance 7. Spectral roll-off point 0.50, smoothed, 3. LPC coeff. 8. Spectral roll-off point 0.75, smoothed, 2. LPC coeff. 9. FO-freq., smoothed, 2. quadratic regression coeff. 10. FO-freq., smoothed, 1. quadratic regression coeff.

[0024] The personalized models and model parameters are determined for the participant during the learning phase and can be continuously adapted individually during the adaptation phase to achieve the highest possible resulting measurement accuracy. During the learning phase, supervised learning with reference information takes place. The initial generation of the model parameters is performed using machine learning methods from a large database of speech samples from numerous patients with associated invasively measured blood glucose values. However, previously generated speech samples with associated invasively measured blood glucose values, also provided by the participant, can be incorporated.

[0025] Furthermore, the adaptation, which aims to improve and individually adjust the model parameters, can be achieved using machine learning methods by deriving modified model parameters (models) from additional (few) speech samples of the respective subject (patient) with associated invasively measured blood glucose values. These modified or adapted model parameters then replace the currently available model parameters for further measurements or investigations.

[0026] The present invention thus enables non-invasive, highly user-oriented, and individualized acquisition of organ function-dependent medical health parameters, such as the current blood glucose class or current blood glucose level, by evaluating a speech sample from the subject. Since the underlying model parameters are individually adaptable (capable of learning), the measurement accuracy and reliability achieved by the concept according to the invention can be continuously improved for each individual subject.

[0027] Preferred embodiments of the present invention are explained in more detail below with reference to the accompanying drawings. With regard to the figures shown, it should be noted that the functional blocks depicted are to be understood both as elements or features of the device according to the invention and as corresponding process steps of the method according to the invention. The figures show: Fig. 1a a flowchart of the method according to the invention in accordance with an exemplary embodiment; Fig. 1b a block or flow diagram of the concept according to the invention in accordance with an exemplary embodiment; Fig. 2a a flowchart of signal processing with regression analysis according to an exemplary implementation; Fig. 2b a block or flow diagram of (intelligent) signal processing with regression analysis according to an exemplary implementation; Fig.3a a flowchart of signal processing with statistical classification according to an exemplary embodiment; Fig. 3b a block or flow diagram of the (intelligent) signal processing with statistical classification according to an exemplary implementation; Fig. 4a a flowchart of signal processing with linear classification according to an exemplary embodiment; Fig. 4b a block or flow diagram of (intelligent) signal processing with linear classification according to an exemplary embodiment; Fig. 5a a flowchart of signal processing with an artificial neural network according to an exemplary embodiment; Fig. 5b a block or flow diagram of (intelligent) signal processing with an artificial neural network according to an exemplary embodiment; Fig.6 a block or flow diagram of the concept according to the invention during a learning phase for the initial determination of model parameters according to an exemplary embodiment; Fig. 7 a block or flow diagram of the concept according to the invention during an adaptation phase for adapting the individual model parameters according to an exemplary embodiment.

[0028] Before exemplary embodiments of the present invention are explained in detail below with reference to the drawings, it should be noted that identical, functionally equivalent or equivalent elements, objects, functional blocks and / or process steps in the different figures are provided with the same reference numerals, so that the description of these elements, objects, functional blocks and / or process steps shown in different exemplary embodiments is interchangeable or can be applied to one another.

[0029] Fig. Figure 1a shows a flowchart of the inventive method 100 for determining a health parameter of a subject by means of voice characteristic evaluation. Fig. Figure 1b shows a basic block or flow diagram of the concept according to the invention (method and device) for determining a health parameter of a subject by means of a voice analysis.

[0030] Referring to procedure 100 of Fig. 1a, in a (first) step 102, a digitized speech sample (also called a voice sample) of the test subject is provided. The speech sample can be provided, for example, from a voice recording device, from a storage device, etc.

[0031] In a subsequent step 104, the digitized speech sample is evaluated and analyzed based on individual model parameters. These individual model parameters, which are assigned to the subject individually (person-specifically), represent a functional relationship between the speech sample itself or language features derived from the speech sample and an assigned current health parameter. The subject's health parameter is, for example, a medical health parameter dependent on organ function, such as the subject's blood glucose level or a blood glucose class assigned to that level.

[0032] As a result of the evaluation of the digital speech sample (104), measurement information, such as an instantaneous value or a health parameter class, is obtained and made available for display. This measurement information is based on the instantaneous value of the subject's health parameter or indicates this instantaneous value within a tolerance range.

[0033] A confidence interval (also called a tolerance range) is a statistical interval that indicates the precision of estimating the position of a parameter, such as the current health parameter of a subject. The confidence interval is thus the range that, with a certain frequency (the confidence level) of repeated measurement, encompasses the true position of the parameter, i.e., the subject's medical health parameter, in a (theoretical) infinite repetition of the measurement. A confidence level of, for example, 95% might be used, meaning that in this case, (at least) 95% of all confidence intervals calculated based on measured data, i.e., evaluated digitized speech samples, contain the true value for the population under investigation.

[0034] The measurement information can then optionally be reproduced in a further step 106 on a display device as an optical, acoustic and / or haptic output signal.

[0035] The output of the measurement information to the subject or patient can, for example, include a blood glucose class with a reliability indicator (confidence level or confidence interval), or it can display an estimated, instantaneous blood glucose value in mmol / L (mmol*18 = mg / dL) derived from the speech sample, along with a tolerance level. The following five blood glucose categories, for example, "much too low," "too low," "normal," "too high," and "much too high," can be selected for display.

[0036] The above examples of a possible display of measurement information on a display device are merely illustrative; essentially any class and range classifications can be used for displaying the current blood glucose value.

[0037] In the following, a so-called "learning phase" for determining basic values ​​for the model parameters and an "adaptation phase" for the continuous adaptation of the individual model parameters will be discussed, wherein the learning phase and adaptation phase represent optional procedural steps for the inventive method 100 for determining a health parameter of a subject by means of voice evaluation.

[0038] During a (potential) learning phase, prior to step 104 of evaluating the digitized speech sample, step 108 is performed to determine basic model parameters based on a large number of speech samples, e.g., from different subjects in a large database, with associated, invasively measured (medical) health parameters, using a machine learning method. These determined basic model parameters are then used, for example, as the model parameters for step 104 of evaluating the subject's digital speech sample (step 110).

[0039] During the learning phase, "supervised learning" with reference information can be carried out to obtain initial model parameters (basic model parameters) using a machine learning method from a large database of speech samples from a large number of patients or test subjects with associated, invasively measured blood glucose values ​​with tolerance information.

[0040] Supervised learning refers, for example, to an algorithm learning a function or functional relationship from given pairs of input and output values. During the learning process, the correct function value is provided for each input value. The goal of supervised learning is to train the "intelligent signal processing unit" to make associations after several iterations with different inputs and outputs. Automatic classification is a subfield of supervised learning.

[0041] Within the scope of this description, machine learning methods are used by the "intelligent signal processing unit" to automatically obtain model parameters (individual model parameters). The following list presents examples of automatic methods for obtaining the model parameters: Machine learning methods: automatic procedure Linear transformation: Eigenvalue decomposition, singular value decomposition, etc. Feature transformation: Analytical or iterative minimization, etc. Regression: Analytical or iterative minimization, etc. Statistical classifier: EM algorithm (EM = Expectation Maximization) Linear classifier: quadratic optimization, etc. Artificial neural network: Error back propagation, etc.

[0042] Alternatively or additionally to the learning phase, an adaptation phase can also be conducted. During the adaptation phase, the individual model parameters of the subject are "individually adapted" to obtain improved (= adapted) individual model parameters. This involves a machine learning procedure (see above) based on multiple speech samples from the subject and associated invasively measured (determined) medical health parameters. Since, for example, only speech samples from the respective subject are used to adapt the model parameters during the adaptation phase, it is also referred to as individual adaptation.

[0043] The adapted individual model parameters can then replace the current (e.g., non-individual or not yet sufficiently individual and still improvable) model parameters (step 114), whereby the step of evaluating 104 of the digitized speech sample is subsequently carried out based on the new individual model parameters, which are improved or adapted.

[0044] According to the present description, adaptation or individual adaptation is thus understood as an improvement and individual adjustment of the model parameters using machine learning methods (see above) from a few (e.g. a plurality of) speech samples of a subject or patient with associated invasively measured blood glucose values.

[0045] The models or model parameters required for the intelligent signal processing unit, which are used to perform the evaluation of the digital speech sample, represent variables needed for regression or classification, for example, which can be determined or obtained using machine learning methods (see above). The following exemplary list indicates which model parameters can be used in the respective machine learning method: Machine learning methods: Model parameters Linear transformation: Transformation matrix Feature transformation: Function parameters (parametric function) Regression: Regression coefficients Statistical classifier: Mean vectors, covariance matrices, mixture weights, state transition probabilities Linear classifier: Learning constants Artificial neural networks: Net weights.

[0046] The following will now be based on the in Fig. The block diagram or flowchart shown in Figure 1b describes the concept (device and method) according to the invention for determining a health parameter of a subject by means of voice evaluation.

[0047] The device 200 comprises a processing unit 210 with an (intelligent) signal processing unit 212 and stored individual model parameters or models 214, which are individually assigned to the subject and are person-specific. The processing unit 210 is designed to evaluate or analyze a digitized speech sample 216 (also called a speech sample or voice sample) of the subject (patient) based on the individual model parameters 214 in order to obtain measurement information 218 (e.g., classification or instantaneous value). The measurement information 218 corresponds to an instantaneous value of the health parameter, e.g., an instantaneous blood glucose level or an assigned class of instantaneous blood glucose level, of the subject, or indicates this within a tolerance range.

[0048] The individual model parameters 214, which are individually assigned and personalized to the subject, indicate a functional relationship between the speech sample 216 itself or speech characteristics derived from the speech sample 216 and an assigned, current health parameter, or represent such a functional relationship.

[0049] The device 200 for determining a health parameter of a subject by means of voice analysis is specifically designed to perform all steps of the procedure 100 for determining a health parameter by means of voice analysis described above. In particular, the procedure steps described above can also optionally be carried out by the device 200 during the learning phase and during the adaptation phase, e.g., by means of the processing unit 210.

[0050] In Fig.Figure 1b also shows a speech recording device 220 for recording the subject's speech sample. Furthermore, a device 224 for digitizing the subject's speech sample 222 is shown in order to provide the digitized speech sample 216.

[0051] The digitization unit 224 can, for example, be assigned to the speech capture unit 220 or the processing unit 210. Furthermore, the speech sample 222 or the digitized speech sample 216 can be retrieved from a memory (not shown in the figure) by the processing unit 212. Fig. 1b) be provided or transmitted by wired or wireless data transmission from the storage or speech capture device 220 to the processing device 212.

[0052] Furthermore, an optional display device 226 is provided for displaying an optical, haptic, and / or acoustic signal dependent on the measurement information 218. For example, the display device 226 is intended to display a blood glucose reading in the form of a blood glucose class with a reliability indication (confidence or confidence interval) or a determined (estimated) instantaneous blood glucose value in mmol / liter with a tolerance indication.

[0053] The device 200 for determining a health parameter of a subject by means of voice analysis can, for example, be implemented autonomously in a mobile device. Suitable mobile devices include, for example, smartphones, tablets, notebooks, etc., wherein the processing unit 210 can be implemented as a processor of the mobile device in order to execute the inventive method 100 for determining a health parameter of a subject by means of voice analysis, for example, by means of an app. The processing unit 210 can thus be implemented at least partially in hardware and / or software (app) in the mobile device. An application software is defined as, for example, an application software, and a mobile app is defined as application software in the field of mobile operating systems.

[0054] Regarding the text used for the actual speech sample, it is important to ensure that the test subject has access to a measurement set that is as simple and easily reproducible as possible for determining voice characteristics (speech characteristics). As a general rule, the more reproducible the conditions for voice analysis, the better and more accurate the results in the form of measurement information.

[0055] A possible simple measurement sentence could be: "Hello, how are you!" This sentence is, of course, only an example; depending on the language, dialect, etc., correspondingly modified measurement sentences can be used. In this context, it should be emphasized again that the measurement sentence should be as easy and reproducible as possible for the test subject. Therefore, for example, a possible number of words for the measurement sentence is between 3 and 10 or between 4 and 8.

[0056] The following will now be based on the Fig. 2a-b, Fig. 3a-b, Fig. 4a-b and Fig. Sections 5a-b describe various possible, optional implementations of the processing unit 210 with the intelligent signal processing 212. The presented list of different implementations is to be considered exemplary only and not exhaustive.

[0057] The following will now be based on… Fig. 2a-b the concept according to the invention (method 100 and device 200) for determining a health parameter of a subject by means of voice evaluation using a regression analysis is described.

[0058] As shown in the flowchart of Fig. As shown in 2a, the evaluation step 102 (i.e., intelligent signal processing) can be performed using regression analysis.

[0059] First, in step 102a, a feature analysis is performed on the provided, digitized speech sample of the subject to determine or obtain the "primary features" of the digitized speech sample. Then, in step 102b, a feature transformation is performed using the determined primary features and functional parameters (and / or parametric functions) to obtain (ideally) suitable secondary features of the speech sample for regression analysis. Finally, in step 102c, a regression analysis is performed using regression coefficients based on the secondary features to obtain the measurement information, e.g., in the form of an (estimated) current health parameter or blood glucose level of the subject.

[0060] Fig.2b now presents the various functional sections or blocks of the processing unit 210a in the form of a basic block or flow diagram. As in Fig. As shown in Figure 2b, a feature analysis 212a is performed on the provided, digitized speech samples 216 of the test subject in order to obtain the primary features 230 of the digitized speech sample. The primary features 230 of the speech sample are, for example, signal features, spectral features, and / or statistical and dynamic features derived from the signal features and / or spectral features of the digitized speech sample 216.

[0061] The following are some examples of voice characteristics (primary characteristics) (in a non-exhaustive list).

[0062] Signal characteristics can include energy, energy profile, pitch, pitch profile, zero-crossing rate, jitter, shimmer, etc.

[0063] Spectral features can include, for example, formants, a spectral shape, FFT (Fast Fourier Transform), scalograms, Mel, third-octave, octave filter bank spectra or Mel, third-octave, octave filter bank spectrograms, Mel-generalized Cepstrum Coefficients (including Cepstrum, MFCC), line spectral frequencies, line Cepstral quefrencies, modulation spectra, etc.

[0064] Statistical and dynamic features derived from signal and / or spectral features can include, for example, means, moments, quantiles, nth-order differences (delta, delta-delta, ...), context features, etc.

[0065] Subsequently, a feature transformation 212b is performed on the determined primary features 230 using function parameters 214a (and / or parametric functions) to obtain secondary features 232 of the speech sample suitable for a regression analysis 212c. In feature transformation 212b, function parameters 214a (or functions or parametric functions) are applied to the primary features 230 to determine or obtain the secondary features 232 of the speech sample. The secondary features 232 exhibit, for example, a linear or quadratic correlation with the (desired) medical health parameter, such as the current blood glucose level, and / or a normal distribution of the feature values. The secondary features 232 are, for example, normally distributed features of the speech sample that are correlated with the desired medical health parameter and uncorrelated (disjoint) with each other.

[0066] Based on the secondary characteristics 232, a regression calculation 212c is now carried out using coefficients 214b (regression coefficients) to obtain the measurement information 218, e.g. in the form of the determined (estimated) health parameter or blood sugar value.

[0067] The function parameters 214a of the feature transformation 212b and the regression coefficients 214b of the regression calculation 212c can be considered the individual model parameters 214 for evaluating the speech sample. Furthermore, the feature analysis 212a, the feature transformation 212b, and the regression 212c together, for example, constitute the intelligent signal processing 212.

[0068] In a linear regression, for example, the model is specified such that the dependent variable y (y = blood glucose level or blood glucose class) is a linear combination of the parameters β i(= regression coefficients) is, but not necessarily of the independent variable x (x = secondary characteristics). For example, simple linear regression modulates the dependence on one independent variable x: Y = β0 + β1 · x + e. In multiple linear regression, several independent variables or functions of the independent variables are considered. For example, if the term x 2 Adding this to the previous regression, we get: Y = β0 + β1 · x + β2 · x 2 + e. Although the expression on the right-hand side is quadratic in the independent variable x (secondary features), the expression is linear in the parameters β0, β1, and β2. Note that x represents a multitude (a vector) of secondary features, and accordingly, β1 and β2 also represent multitudes (vectors) of regression coefficients. Thus, this is also a linear regression. To determine the model parameters β iFor example, the method of least squares is used.

[0069] The following will now be based on the Fig. 3a-b the concept according to the invention (device and method) for determining a health parameter of a subject by means of voice analysis is described, wherein the step of evaluating the speech sample, i.e. intelligent signal processing, is carried out using a statistical classification.

[0070] At the in Fig. In the flowchart shown in 3a, a feature analysis is first performed in step 102a on the provided, digitized speech sample of the test subject in order to obtain the primary features of the digitized speech sample. Subsequently, in step 102d, a feature transformation of the primary features is performed using a transformation matrix to obtain secondary features of the speech sample.

[0071] Subsequently, in step 102e, a statistical classification is performed using class models or a statistical classifier is applied to determine and provide the measurement information in the form of a class of the current medical health parameter, i.e., for example, the current blood glucose class with confidence or confidence interval, based on the provided secondary characteristics.

[0072] Fig.Figure 3b now presents, in the form of a basic block or flow diagram, the various functional sections or functional blocks of the concept according to the invention (device and method), wherein a statistical classification is carried out during the evaluation of the speech sample, i.e., during intelligent signal processing. Thus, a feature analysis 212a is first carried out with the provided, digitized speech sample 216 of the test subject in order to obtain the primary features 230 of the digitized speech sample (cf. the explanations applicable here as well). Fig. 2b).

[0073] Subsequently, a feature transformation 212d (e.g., a linear transformation) of the primary features 230 is performed using a transformation matrix 214c to obtain the secondary features 234. Through the linear transformation of the primary features 230, the secondary features 234 are dimensionally reduced and uncorrelated (or disjoint) with each other.

[0074] The primary features 230 of the speech sample are, for example, signal features, spectral features and / or statistical and dynamic features derived from the signal features and / or spectral features of the speech sample.

[0075] For example, a linear transformation 212d can be used, such as a principal axis transformation (HAT), a principal component analysis (PCA), a linear discriminant analysis (LDA), an independent component analysis (ICA), etc.

[0076] Finally, using class models 214d, a statistical classification 212e can be carried out or a statistical classifier can be applied to determine and provide the desired measurement information 236 in the form of a class of the current medical health parameter, e.g. a blood glucose class with a confidence interval, based on the secondary characteristics 234 provided.

[0077] For example, model parameters of blood glucose classes are referred to as class models 214d. A Gaussian Mixture Model (GMM) classifier, a Hidden Markov Model (HMM) classifier, etc., can be used as a statistical classifier. The transformation matrix 214c and / or the class models 214d for statistical classification constitute, for example, the individual model parameters 214. Furthermore, the feature analysis 212a, the linear transformation 212d, and the statistical classification 212e constitute, for example, the intelligent signal processing 212.

[0078] The following will now be based on the Fig. 4a-b the concept according to the invention (device and method) for determining a health parameter of a subject by means of voice evaluation is described, wherein the evaluation of the digitized speech sample is carried out using a linear classification.

[0079] Fig.Figure 4a shows a flowchart with the respective steps when the step of evaluating the speech sample is carried out using a linear classification.

[0080] In step 102f, a feature analysis is first performed on the provided, digitized speech sample of the subject to obtain primary features of the digitized speech sample. Furthermore, in step 102g, a linear classification is performed using class models and a linear classifier to determine or obtain the measurement information, e.g., in the form of a class of the medical health parameter, and in particular the blood glucose class with a confidence interval.

[0081] The following will now be based on the in Fig.The principal block or flow diagram shown in Figure 4b describes the respective functional sections or functional blocks of the concept according to the invention (device and method) for determining a health parameter of a subject by means of voice evaluation, wherein the evaluation of the speech sample, i.e. the intelligent signal processing, is carried out using a linear classification.

[0082] First, a feature analysis 212a is performed on the provided digitized speech sample 216 of the subject to obtain the primary features 230 of the digitized speech sample. Subsequently, a linear classification 212f is performed using class models 214d or a linear classifier is applied to the obtained primary features 230 in order to determine and provide the measurement information 236, e.g., in the form of a class of the medical health parameter or as a blood glucose class with a confidence interval.

[0083] The class models 214d of the linear classifier 212f, for example, constitute the individual model parameters. Furthermore, the feature analysis 212a and the linear classification 212f together, for example, constitute the intelligent signal processing 212. A support vector machine (SVM), etc., is used as the linear classifier.

[0084] The following will now be based on the Fig. 5a-b describes the concept according to the invention (device and method) for determining a health parameter of a subject by means of voice evaluation, wherein the evaluation of the speech sample is carried out using artificial neural networks or a convolutional neural network (CNN).

[0085] First, in step 102a, a feature analysis (switch position S1) is performed on the provided, digitized speech sample of the test subject in order to obtain primary features of the digitized speech sample.

[0086] Furthermore, in step 102g, an artificial neural network is applied to the primary features using class models to determine or provide the measurement information, e.g. in the form of a class of the current medical health parameter or the blood glucose class with confidence.

[0087] If the artificial neural network is designed as a convolutional neural network (CNN), step 102a of performing a feature analysis can be omitted (switch position S2).

[0088] The following will now be based on the in Fig.Figure 5b shows the principal block or flow diagram, describing the various functional sections or functional blocks of the concept according to the invention (method and device) for determining a health parameter of a subject by means of voice evaluation, wherein the evaluation of the speech sample is carried out using an artificial neural network or a convolutional neural network (CNN).

[0089] First, a feature analysis 212a can be performed on the provided, digitized speech sample 216 of the subject to obtain the primary features 230 of the digitized speech sample 216. Subsequently, an artificial neural network 212g can be applied to the primary features 230 using class models 214d to determine and provide the measurement information 236, e.g., in the form of a class of the current medical health parameter or the blood glucose class with confidence. If the artificial neural network is configured as a convolutional neural network (CNN), the feature analysis 212a can be omitted, in which case the switch position S2 (= no feature analysis required) is assumed. For the application of other artificial neural networks, the switch position S1 (= feature analysis required) is assumed.

[0090] The class models 214d of the neural network 212g, for example, constitute the individual model parameters, where the class models are weights (network weights) of the neural network. The feature analysis 212a and the artificial neural network 212g can constitute the intelligent signal processing 212 (except when using a CNN - see above). When using a convolutional neural network (CNN), the artificial neural network 212g alone can constitute the intelligent signal processing 212.

[0091] An artificial neural network (ANN) can include a multi-layer perceptron (MLP), a deep neural network (DNN), a recurrent neural network (RNN), a convolutional neural network (CNN), a long short-term memory (LSTM), etc.

[0092] The following will be based on the in Fig.The block diagram or flowchart shown in Figure 6 describes the various functional sections or blocks of learning phase 300 of the concept according to the invention (device and method) for determining a health parameter of a subject by means of voice analysis. The learning phase represents additional, possible process steps for the method according to the invention 100 for determining a health parameter.

[0093] During a potential learning phase, prior to step 104 of evaluating the digitized speech sample, step 108 of determining basic model parameters 214' based on a large number of speech samples, e.g., from different subjects from a large database 310, with associated, invasively measured (medical) health parameters 314, can be performed using a machine learning method 312. These determined basic model parameters 214' are then used, for example, as the model parameters 214 for step 104 of evaluating the digital speech sample of the subject (step 110).

[0094] Furthermore, a speech recording device 220 for recording the subject's speech sample is shown, wherein a device 224 for digitizing the subject's speech sample 222 is shown in order to provide the digitized speech sample 216.

[0095] During the learning phase, "supervised learning" with reference information can be carried out to provide an initial acquisition of model parameters (basic model parameters) using a machine learning method from a large database of speech samples from a large number of patients or test subjects with associated, invasively measured blood glucose values ​​with tolerance information.

[0096] Examples of machine learning methods that can be used include analytical or iterative linear, quadratic or other optimization methods, analytical or iterative maximum likelihood estimation methods (e.g. EM algorithm), boundary minimization and error tracing methods.

[0097] In step 212a, a feature analysis (switch position S1) is first performed on the provided, digitized speech sample of the test subject to obtain primary features of the digitized speech sample. If the artificial neural network is configured as a convolutional neural network (CNN), feature analysis 212a can be omitted, in which case switch position S2 (= no feature analysis required) is assumed. For the application of other artificial neural networks, switch position S1 (= feature analysis required) is assumed.

[0098] The following will be based on the in Fig.The block diagram or flowchart shown in Figure 7 describes the various functional sections or blocks of adaptation phase 400 of the concept according to the invention (device and method) for determining a health parameter of a subject using voice analysis. Adaptation phase 400 represents additional optional process steps for the method according to the invention 100 for determining a health parameter.

[0099] In addition to the learning phase 300, the adaptation phase 400 can also be carried out. During the adaptation phase, the individual model parameters 214 of the subject are "individually adapted" to obtain improved (= adapted) individual model parameters 214'. This involves a machine learning procedure 312 (see above) based on a plurality of speech samples 216 from the subject and on associated invasively measured (determined) medical health parameters. Since, for example, only speech samples from the respective subject are used to adapt the model parameters 214 during the adaptation phase, this is also referred to as individual adaptation.

[0100] Furthermore, a speech recording device 220 for recording the subject's speech sample is shown, wherein a device 224 for digitizing the subject's speech sample 222 is shown in order to provide the digitized speech sample 216.

[0101] Examples of machine learning methods that can be used include analytical or iterative linear, quadratic or other optimization methods, analytical or iterative maximum likelihood estimation methods (e.g. EM algorithm), boundary minimization and error tracing methods.

[0102] In step 212a, a feature analysis (switch position S1) is first performed on the provided, digitized speech sample of the test subject to obtain primary features of the digitized speech sample. If the artificial neural network is configured as a convolutional neural network (CNN), feature analysis 212a can be omitted, in which case switch position S2 (= no feature analysis required) is assumed. For the application of other artificial neural networks, switch position S1 (= feature analysis required) is assumed.

[0103] The adapted individual model parameters can then replace the current individual model parameters (step 114), whereby step 104 of evaluating the digitized speech sample is subsequently carried out based on the new individual model parameters, which are improved or adapted.

[0104] According to the present description, adaptation or individual adaptation is thus understood as an improvement and individual adjustment of the model parameters using machine learning methods (see above) from a few (e.g. a plurality of) speech samples of a subject or patient with associated invasively measured blood glucose values.

[0105] The concept according to the invention for determining a health parameter of a subject by means of voice analysis offers numerous advantages over previous methods for determining a health parameter, and in particular a blood glucose level. The following summarizes the general principles of the concept according to the invention and some key advantages resulting therefrom.

[0106] Blood glucose is required for all vital functions and thus affects these functions, as well as the brain, which controls them, and, for example, the organs that produce and influence sound. The concept according to the invention for determining a health parameter of a subject based on voice analysis can, for example, use sensitive audio sensors / microphones to capture and digitally analyze the subject's acoustic speech signals, whereby abnormalities can be mathematically classified and identified as healthy or ill. Furthermore, stored, pre-trained reference templates (markers or patterns) can be immediately applied for an initial suspected diagnosis, and, if necessary, suitable recommendations for action can be calculated and offered.A device implementing the concept according to the invention for determining a health parameter of a subject by means of voice analysis can, for example, comprise the following functional blocks (functions or devices):. 1. Basic setup of the device and the procedure 2. Person recognition / identification for personalized determination of health parameters (diagnosis) 3. Personalized / Personalized Diagnostic Training and Learning 4. Personalized / Personalized Diagnosis / Application 5. Relearning (calibration) and self-optimization, updating 6. The concept is embedded in data protection and crime prevention functions. 7. The concept utilizes personal data storage and archiving functions. 8. Use of a secure external data and communication interface.

[0107] The concept according to the invention can process technically recorded speech samples from the test subject, both standardized and personalized. In addition to recording blood glucose levels, many other health / disease / vital characteristics and the associated diseases / health can also be correlated, measured, and evaluated using the concept according to the invention.

[0108] The concept according to the invention is essentially applicable to all health sectors and their specialists who deal with the determination of health parameters, such as blood glucose measurement and blood glucose monitoring, for example in (medical) diabetes disease detection, monitoring and prevention.

[0109] Furthermore, the concept according to the invention is also applicable to all areas concerned with healthcare and healthy lifestyles, and which offer services for health detection, monitoring, and prevention. In particular, the concept according to the invention is applicable to, or can be used by, all health-conscious and health-interested persons.

[0110] The concept according to the invention, using digital information and communication technology, thus offers the possibility of flexibly adapting technology to people, their behavior and needs, thereby providing a tailored, personalized, individualized technology and thus avoiding the previously necessary adaptation of people to a rigid, user-unfriendly technology. The concept according to the invention thus enables particularly user-friendly blood glucose meters.

[0111] The concept according to the invention enables the spatially and temporally remote individualized determination of a health parameter of a subject by means of voice analysis, which can be located anywhere in the world (by using, for example, the Internet), whereby user data from the past can also be used, provided that this is stored as digitized recording material and can be accessed.

[0112] The concept according to the invention thus provides the basis for a better quality of life and more user-friendly methods for determining health parameters. The concept according to the invention creates a non-invasive (bloodless), non-contact, non-invasive method for determining a health parameter that adapts to the user's needs and is as minimally burdensome as possible, while simultaneously saving a significant amount of material, time, and costs. The concept according to the invention represents a reliable procedure that also enables long-term monitoring and follow-up measurements. The concept according to the invention can cooperatively alert the user to potential problems (technical, diagnostic) and warn them of emergencies and accidents. Furthermore, the concept according to the invention is flexible, adaptive, and self-learning, and can be flexibly adapted to many different boundary conditions.

[0113] Furthermore, the protection of personal health data and the functions of the concept according to the invention can be safeguarded against criminal cyberattacks. In particular, the concept according to the invention makes it possible to provide a consumer product for "every woman / every man".

[0114] Although some aspects have been described in connection with a device, it is understood that these aspects also constitute a description of the corresponding process, such that a block or component of a device can also be understood as a corresponding process step or as a feature of a process step. Similarly, aspects described in connection with or as a process step also constitute a description of a corresponding block, detail, or feature of a corresponding device. Some or all of the process steps can be performed by (or using) a hardware apparatus, such as a microprocessor, a programmable computer, or an electronic circuit. In some embodiments, some or more of the key process steps can be performed by such an apparatus.

[0115] Depending on specific implementation requirements, embodiments of the invention can be implemented in hardware or in software, or at least partially in hardware or at least partially in software. The implementation can be carried out using a digital storage medium, for example, a floppy disk, a DVD, a Blu-ray disc, a CD, a ROM, a PROM, an EPROM, an EEPROM, a FLASH memory, a hard disk, or another magnetic or optical storage medium, on which electronically readable control signals are stored. These control signals can interact with, or interact with, a programmable computer system in such a way as to execute the respective method. Therefore, the digital storage medium can be computer-readable.

[0116] Some embodiments according to the invention therefore include a data carrier which has electronically readable control signals which are able to interact with a programmable computer system in such a way that one of the methods described herein is carried out.

[0117] In general, embodiments of the present invention can be implemented as a computer program product with a program code, wherein the program code is effective in carrying out one of the methods when the computer program product runs on a computer.

[0118] The program code can also be stored on a machine-readable medium, for example.

[0119] Other embodiments include a computer program for carrying out one of the methods described herein, wherein the computer program is stored on a machine-readable medium. In other words, an embodiment of the method according to the invention is thus a computer program that includes program code for carrying out one of the methods described herein when the computer program is executed on a computer.

[0120] Another embodiment of the methods according to the invention is thus a data carrier (or a digital storage medium or a computer-readable medium) on which the computer program for carrying out one of the methods described herein is recorded. The data carrier or the digital storage medium or the computer-readable medium is typically tangible and / or non-volatile.

[0121] Another embodiment of the method according to the invention is thus a data stream or a sequence of signals that represents the computer program for carrying out one of the methods described herein. The data stream or sequence of signals can be configured, for example, to be transferred via a data communication connection, such as the Internet.

[0122] Another embodiment comprises a processing device, for example a computer or a programmable logic device, which is configured or adapted to perform one of the methods described herein.

[0123] Another embodiment comprises a computer on which the computer program for performing one of the procedures described herein is installed.

[0124] Another embodiment of the invention comprises a device or system designed to transmit a computer program for carrying out at least one of the methods described herein to a receiver. The transmission can be, for example, electronic or optical. The receiver can be, for example, a computer, a mobile device, a storage device, or a similar device. The device or system can, for example, include a file server for transmitting the computer program to the receiver.

[0125] In some embodiments, a programmable logic device (for example, a field-programmable gate array, an FPGA) can be used to perform some or all of the functionalities of the methods described herein. In some embodiments, a field-programmable gate array can interact with a microprocessor to perform one of the methods described herein. Generally, in some embodiments, the methods are performed by any hardware device. This can be general-purpose hardware such as a computer processor (CPU) or method-specific hardware such as an ASIC.

[0126] The embodiments described above merely illustrate the principles of the present invention. It is understood that modifications and variations of the arrangements and details described herein will be obvious to other people skilled in the art. Therefore, it is intended that the invention be limited only by the scope of protection set forth in the following claims and not by the specific details presented herein by way of description and explanation of the embodiments.

Claims

[1] Method (100) for determining a health parameter of a subject by means of voice evaluation, comprising the following steps: Providing individual model parameters that indicate a functional relationship between a speech sample of the subject himself or speech characteristics derived from the speech sample and an associated, current health parameter of the subject, Providing (102) a digitized speech sample of the subject, Evaluating (104) the digitized speech sample based on the individual model parameters to obtain measurement information based on a current value of the subject's health parameter within a tolerance range, and Adapting (112) the individual model parameters of the subject by performing a machine learning process based on a plurality of speech samples of the subject and on associated invasively measured medical health parameters to obtain adapted individual model parameters, where the current, individual model parameters are replaced by the adapted, individual model parameters (114); and the step (102) of evaluating the speech sample is carried out based on the adapted, individual model parameters (116). [2] The method according to claim 1, wherein the health parameter is a medical health parameter (218) of the subject that is dependent on an organ function. [3] Method according to claim 1 or 2, wherein the health parameter (218) is a current blood glucose value of the subject or a blood glucose class associated with the current blood glucose value. [4] Method according to one of the preceding claims, further comprising the following steps: Providing (106) the measurement information on a display device as an optical, acoustic and / or haptic output signal. [5] Method according to one of the preceding claims, wherein the speech features derived from the speech sample are selected from statistical features. [6] Method according to one of the preceding claims, wherein the step of evaluating (102) is carried out using a regression analysis, further comprising the following steps: Carrying out (102a) a feature analysis on the provided digitized speech sample of the subject in order to obtain primary features of the digitized speech sample, Carrying out (102b) a feature transformation with the determined primary features by means of function parameters or parametric functions in order to obtain secondary features of the speech sample suitable for a regression calculation, and Performing (102c) a regression calculation using regression coefficients based on the secondary features to obtain the measurement information. [7] Method according to claim 6, wherein the function parameters of the feature transformation and / or the regression coefficients of the regression calculation form the individual model parameters for evaluating the speech sample. [8] Method according to claim 6 or 7, wherein in the feature transformation the function parameters are applied to the primary features in order to determine secondary features of the speech sample which have a linear or quadratic correlation with the health parameter and / or a normal distribution. [9] Method according to one of claims 6 to 8, wherein the primary features of the speech sample comprise signal features, spectral features and / or statistical and dynamic features derived from the signal features and spectral features of the speech sample. [10] Method according to one of claims 6 to 9, wherein the secondary features are normally distributed, disjoint features of the speech sample correlated with the medical health parameter. [11] Method according to one of claims 1 to 5, wherein the step of evaluating the speech sample is carried out using a statistical classification, comprising the following steps: Carrying out (102a) a feature analysis with the provided digitized speech sample of the subject in order to obtain primary features of the digitized speech sample, performing (102d) a feature transformation of the primary features using a transformation matrix to obtain secondary features, and Performing (102e) a statistical classification using class models to determine the measurement information in the form of a class of the current medical health parameter based on the provided secondary features. [12] The method of claim 11, wherein the transformation matrix for the linear transformation and / or the class models for the statistical classification comprise the individual model parameters. [13] The method according to claim 11 or 12, wherein the primary features of the speech sample comprise signal features, spectral features and / or statistical and dynamic features derived from the signal features and spectral features of the speech sample. [14] Method according to one of claims 11 to 13, wherein the secondary features are dimensionally reduced, uncorrelated and disjoint. [15] A method according to any one of claims 1 to 5, wherein the step of evaluating the speech sample is performed using a linear classification, further comprising the following steps: Carrying out (102a) a feature analysis with the provided digitized speech sample of the subject in order to obtain primary features of the digitized speech sample, and Performing (102f) a linear classification with the primary features using class models to obtain the measurement information. [16] The method of claim 15, wherein the class models of the linear classification form the individual model parameters. [17] Method according to one of claims 1 to 5, wherein the step of evaluating the speech sample is carried out using artificial neural networks, further comprising the following steps: Carrying out (102a) a feature analysis on the provided digitized speech sample of the subject in order to obtain primary features of the digitized speech sample, and Applying (102g) an artificial neural network to the primary features using class models to determine the measurement information. [18] The method of claim 17, wherein the class models of the artificial neural network are network weights of the neural network and form the individual model parameters. [19] Method according to one of claims 1 to 5, wherein the step of evaluating the speech sample is carried out using a convolutional neural network (CNN), comprising the following steps: Applying the convolutional neural network to the provided digitized speech sample using class models to obtain the measurement information. [20] The method of claim 19, wherein the class models of the convolutional neural network are network weights of the neural network and form the individual model parameters. [21] Device (200) for determining a health parameter (218) of a subject by means of voice evaluation, having the following features: a processing device (210) which is designed to evaluate a digitized speech sample (216) of the subject based on individual model parameters (214), to obtain measurement information (218) based on an instantaneous value of the subject's health parameter within a tolerance range, where the individual model parameters indicate a functional relationship between the speech sample or speech characteristics derived from the speech sample and an associated, current health parameter, and wherein the processing device (210) is designed to carry out the method (100) for determining a health parameter of a subject by means of voice evaluation according to one of claims 1 to 20. [22] The apparatus of claim 21, wherein the processing means is implemented at least partially in hardware and / or software in a mobile device. [23] Mobile device with the following features: a speech recording device (220) for recording a speech sample of a test subject, the device (200) according to one of claims 21 to 22, and a display device (226) for displaying an optical, haptic and / or acoustic signal dependent on the measurement information. [24] Mobile device according to claim 23, which is designed as a smartphone, tablet or notebook. [25] Computer program with a program code for carrying out the method according to one of claims 1 to 20, when the computer program runs on a computer or a processor.

Images