Parkinson's disease prediction device and method

Abstract

To provide a Parkinson disease prediction apparatus and method for predicting Parkinson disease of a speaker through speech analysis of the speaker.SOLUTION: The method includes the steps of: extracting a syntactic combination from audio data including a speaker's speech result; verifying accuracy of a Parkinson disease prediction model by changing conditions for preprocessing the audio data and the syntactic combination; determining, as audio data for Parkinson disease prediction, a syntactic combination that ranks high in a result of verifying the accuracy of the Parkinson disease prediction model; and inputting, to the Parkinson disease prediction model, a speaker's speech result corresponding to the audio data for the Parkinson disease prediction and obtaining a Parkinson disease prediction result for the speaker as an output of the Parkinson disease prediction model.SELECTED DRAWING: Figure 34

Description

[Technical Field]

[0001] The present invention relates to an apparatus and method for predicting Parkinson's disease. [Background technology]

[0002] Parkinson's disease is a neurodegenerative disease that causes impaired movement due to destruction of the dopamine nervous system in the substantia nigra of the brain, located in the center of the brainstem.

[0003] It is known that the majority of Parkinson's disease patients have speech problems, which may mean that the speech patterns of patients who may have Parkinson's disease and the sentences they speak can indirectly indicate whether or not they have the disease.

[0004] The above-mentioned background art is technical information that the inventor possessed for the purpose of deriving the present invention or that he acquired in the process of deriving the present invention, and is not necessarily publicly known art that was made public to the general public prior to the filing of the present invention.

[0005] This invention is a technology developed through the "Bizcall-based, optimized early dementia screening technology commercialization for the elderly" project, project number SU220002, of the Seoul Industrial Development Agency's 2022 Growth Stage Scale-Up Technology Commercialization Support Project. Summary of the Invention [Problem to be solved by the invention]

[0006] An object of the present invention is to provide an apparatus and method for predicting Parkinson's disease in a speaker through speech analysis of the speaker.

[0007] The problems to be solved by the present invention are not limited to those described above, and other problems and advantages of the present invention not mentioned above can be understood from the following description and will be more clearly understood in the embodiments of the present invention. Furthermore, it will be understood that the problems to be solved and advantages of the present invention can be realized by the means and combinations thereof set forth in the claims. [Means for solving the problem]

[0008] The Parkinson's disease prediction method of this embodiment is a Parkinson's disease prediction method executed by a processor of a Parkinson's disease prediction device, and includes the steps of extracting combinations of syntax from speech data including a speaker's speech results, verifying the accuracy of the Parkinson's disease prediction model by changing the conditions for preprocessing the speech data and the combinations of syntax, determining the highest ranked combinations of syntax from the results of verifying the accuracy of the Parkinson's disease prediction model as speech data for Parkinson's disease prediction, and inputting the speaker's speech results corresponding to the speech data for Parkinson's disease prediction into the Parkinson's disease prediction model, and obtaining a Parkinson's disease prediction result for the speaker at the output of the Parkinson's disease prediction model.

[0009] The Parkinson's disease prediction device of this embodiment includes a processor and a memory operably connected to the processor for storing at least one code executed by the processor. The memory can store code that, when executed via the processor, causes the processor to extract syntactic combinations from speech data including a speaker's speech results, change the conditions for preprocessing the speech data and the syntactic combinations to verify the accuracy of a Parkinson's disease prediction model, determine the syntactic combinations that rank highest based on the results of verifying the accuracy of the Parkinson's disease prediction model as speech data for Parkinson's disease prediction, input the speaker's speech results corresponding to the speech data for Parkinson's disease prediction into the Parkinson's disease prediction model, and obtain a Parkinson's disease prediction result for the speaker from the output of the Parkinson's disease prediction model.

[0010] In addition, other methods for implementing the present invention, other systems, and computer-readable recording media storing computer programs for carrying out the methods can also be provided.

[0011] Other aspects, features, and advantages beyond those described above will become apparent from the following drawings, claims, and detailed description of the invention. [Effects of the Invention]

[0012] According to the present invention, it is possible to provide a Parkinson's disease prediction device and method for predicting Parkinson's disease in a speaker by analyzing the speaker's voice.

[0013] The effects of the present invention are not limited to those mentioned above, and other effects not mentioned will be clearly understood by those skilled in the art from the following description. [Brief explanation of the drawings]

[0014] [Figure 1] FIG. 1 is a block diagram showing an outline of the configuration of a Parkinson's disease prediction device according to this embodiment. [Figure 2] FIG. 2 is a diagram illustrating an example of a language regulation showing the number and order of Hangul consonants and vowels according to the present embodiment. [Figure 3] FIG. 3 is a diagram illustrating an example of a combination of syntax extracted by the Parkinson's disease prediction device according to this embodiment. [Figure 4] FIG. 4 is a diagram illustrating an example of a combination of syntax extracted by the Parkinson's disease prediction device according to this embodiment. [Figure 5] FIG. 5 is a diagram illustrating an example of a combination of syntax extracted by the Parkinson's disease prediction device according to this embodiment. [Figure 6] FIG. 6 is an exemplary diagram illustrating accuracy verification of the Parkinson's disease prediction device according to this embodiment. [Figure 7]FIG. 7 is an exemplary diagram illustrating accuracy verification of the Parkinson's disease prediction device according to this embodiment. [Figure 8] FIG. 8 is an exemplary diagram illustrating accuracy verification of the Parkinson's disease prediction device according to this embodiment. [Figure 9] FIG. 9 is an exemplary diagram illustrating accuracy verification of the Parkinson's disease prediction device according to this embodiment. [Figure 10] FIG. 10 is an exemplary diagram illustrating accuracy verification of the Parkinson's disease prediction device according to this embodiment. [Figure 11] FIG. 11 is an exemplary diagram illustrating accuracy verification of the Parkinson's disease prediction device according to this embodiment. [Figure 12] FIG. 12 is an exemplary diagram illustrating accuracy verification of the Parkinson's disease prediction device according to this embodiment. [Figure 13] FIG. 13 is an exemplary diagram illustrating accuracy verification of the Parkinson's disease prediction device according to this embodiment. [Figure 14] FIG. 14 is an exemplary diagram illustrating accuracy verification of the Parkinson's disease prediction device according to this embodiment. [Figure 15] FIG. 15 is an exemplary diagram illustrating accuracy verification of the Parkinson's disease prediction device according to this embodiment. [Figure 16] FIG. 16 is an exemplary diagram illustrating accuracy verification of the Parkinson's disease prediction device according to this embodiment. [Figure 17] FIG. 17 is an exemplary diagram illustrating accuracy verification of the Parkinson's disease prediction device according to this embodiment. [Figure 18] FIG. 18 is an exemplary diagram illustrating accuracy verification of the Parkinson's disease prediction device according to this embodiment. [Figure 19] FIG. 19 is an exemplary diagram illustrating accuracy verification of the Parkinson's disease prediction device according to this embodiment. [Figure 20] FIG. 20 is an exemplary diagram illustrating accuracy verification of the Parkinson's disease prediction device according to this embodiment. [Figure 21] FIG. 21 is an exemplary diagram illustrating accuracy verification of the Parkinson's disease prediction device according to this embodiment. [Figure 22]FIG. 22 is an exemplary diagram illustrating accuracy verification of the Parkinson's disease prediction device according to this embodiment. [Figure 23] FIG. 23 is an exemplary diagram illustrating accuracy verification of the Parkinson's disease prediction device according to this embodiment. [Figure 24] FIG. 24 is an exemplary diagram illustrating accuracy verification of the Parkinson's disease prediction device according to this embodiment. [Figure 25] FIG. 25 is an exemplary diagram illustrating accuracy verification of the Parkinson's disease prediction device according to this embodiment. [Figure 26] FIG. 26 is an exemplary diagram illustrating accuracy verification of the Parkinson's disease prediction device according to this embodiment. [Figure 27] FIG. 27 is an exemplary diagram illustrating accuracy verification of the Parkinson's disease prediction device according to this embodiment. [Figure 28] FIG. 28 is an exemplary diagram illustrating accuracy verification of the Parkinson's disease prediction device according to this embodiment. [Figure 29] FIG. 29 is an exemplary diagram illustrating accuracy verification of the Parkinson's disease prediction device according to this embodiment. [Figure 30] FIG. 30 is an exemplary diagram illustrating accuracy verification of the Parkinson's disease prediction device according to this embodiment. [Figure 31] FIG. 31 is an exemplary diagram illustrating accuracy verification of the Parkinson's disease prediction device according to this embodiment. [Figure 32] FIG. 32 is an exemplary diagram illustrating accuracy verification of the Parkinson's disease prediction device according to this embodiment. [Figure 33] FIG. 33 is a block diagram showing an outline of the configuration of a Parkinson's disease prediction device according to another embodiment. [Figure 34] FIG. 34 is a flowchart illustrating the Parkinson's disease prediction method according to this embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0015] The advantages and features of the present invention, as well as methods for achieving them, will become apparent from the detailed description of the embodiments accompanied by the accompanying drawings. However, the present invention is not limited to the embodiments presented below, and can be implemented in various different forms, and it should be understood that the present invention includes all modifications, equivalents, and alternatives falling within the spirit and technical scope of the present invention. The embodiments presented below are provided to fully disclose the present invention and to fully convey the scope of the invention to those skilled in the art to which the present invention pertains. In describing the present invention, if a detailed description of related publicly known technology is considered to obscure the gist of the present invention, such detailed description will be omitted.

[0016] The terms used in this application are merely used to describe particular embodiments and are not intended to limit the present invention. The singular includes the plural unless the context clearly dictates otherwise. In this application, terms such as "comprise" or "have" are intended to specify the presence of features, numbers, steps, operations, components, parts, or combinations thereof described herein, and should be understood not to preclude the presence or additional possibility of one or more other features, numbers, steps, operations, components, parts, or combinations thereof. Terms such as "first" and "second" can be used to describe various components, but the components should not be limited by these terms. These terms are used solely to distinguish one component from another.

[0017] Furthermore, in this application, a "module" may be a hardware component such as a processor or circuitry, and / or a software component executed by a hardware component such as a processor.

[0018] Furthermore, in the present application documents, square brackets ([ ]) may contain the English notation of the syllable, syntactic, or syntactic combination written before the square brackets, and the English notation may represent the Korean pronunciation of the syllable, syntactic, or syntactic combination. Here, a syllable may refer to the smallest unit of speech that is most easily intuitional when spoken. In this embodiment, a syllable may include, for example, [ga], [ha], etc. Furthermore, a syntactic may be a unit of segmentation as the smallest component of a sentence. In this embodiment, a syntactic may include, for example, "[aeiou]", "[nanana]", "[rarara]", etc. Furthermore, a syntactic combination may include the result of combining one or more of the above-mentioned syntactic combinations. In this embodiment, a syntactic combination may include, for example, "[gagagananana]", "[aeiougagagananana]", "[nananasasasa]", etc.

[0019] Hereinafter, an embodiment of the present invention will be described in detail with reference to the accompanying drawings. In the description with reference to the accompanying drawings, the same or corresponding components will be given the same drawing numbers, and duplicate descriptions will be omitted.

[0020] In the following embodiments, terms such as first and second are not used in a limiting sense but are used to distinguish one component from another.

[0021] In the following embodiments, singular expressions include plural expressions unless the context clearly indicates otherwise.

[0022] In the following embodiments, terms such as "comprise" or "have" mean the presence of the features or components described in this specification, but do not preclude the possibility that one or more other features or components may be added.

[0023] Certain process sequences may be performed out of the order described when an embodiment is otherwise operable, for example, two processes described as successive may be performed substantially simultaneously or may be performed in the reverse order from that described.

[0024] Figure 1 is a block diagram shown to explain the general configuration of the Parkinson's disease prediction device of this embodiment, Figure 2 is an example diagram of grammatical rules indicating the number and order of Hangul consonants and vowels of this embodiment, Figures 3 to 5 are example diagrams of combinations of syntax extracted by the Parkinson's disease prediction device of this embodiment, and Figures 6 to 32 are example diagrams for explaining accuracy verification of the Parkinson's disease prediction device of this embodiment.

[0025] The Parkinson's disease prediction device 100 according to the present embodiment may exist independently in the form of a server, or the Parkinson's disease prediction function provided by the Parkinson's disease prediction device 100 may be implemented in the form of an application and installed in a user terminal (not shown). The user terminal may access the Parkinson's disease prediction application and / or Parkinson's disease prediction site provided by the Parkinson's disease prediction device 100 to receive the Parkinson's disease prediction service.

[0026] 1 to 32, the Parkinson's disease prediction device 100 may include a collection unit 110, a generation unit 120, an extraction unit 130, a verification unit 140, a determination unit 150, an acquisition unit 160, and a control unit 170.

[0027] The collection unit 110 can collect voice data including the results of a speaker's speech. In this embodiment, the voice data can include a voice recording file of a pre-set material being read aloud. Here, the pre-set material can include a document, file, etc. including syllables, sentences, or combinations of sentences, which will be described later. Furthermore, the voice data can include data in which one of the syllables, sentences, or combinations of sentences, which will be described later, is spoken for a pre-set time (e.g., 1 to 2 seconds). Furthermore, the voice data can include data in which the syllables, sentences, or combinations of sentences, which will be described later, are spoken in a high pitch and data in which the syllables, sentences, or combinations of sentences, which will be described later, are spoken in a low pitch.

[0028] In this embodiment, the audio data and the audio signal can be described in combination with the same meaning.

[0029] The generation unit 120 may generate a Parkinson's disease prediction model. In this embodiment, the Parkinson's disease prediction model can be generated by training a deep neural network model that has been pre-trained to predict a speaker's Parkinson's disease using voice data including the speaker's speech. Here, the deep neural network model may be a model that receives voice data including the speaker's speech as input and is trained in a supervised learning manner using training data labeled with any of "normal," "Parkinson's disease," "multiple system atrophy," and "cerebellar atrophy," which are included in Parkinson's disease-related diseases.

[0030] The generation unit 120 may train the initially set deep neural network model in a supervised learning manner using the labeled training data. Here, the initially set deep neural network model is an initial model designed to be configured as a Parkinson's disease prediction model, and parameter values ​​are set to arbitrary initial values. The initially set deep neural network model is trained using the above training data and the parameter values ​​are optimized, thereby completing a Parkinson's disease prediction model that can accurately predict Parkinson's disease in a speaker.

[0031] The extraction unit 130 can extract combinations of syntaxes from speech data including the results of a speaker's speech. The combinations of syntaxes in this embodiment can include combinations of multiple different first syntaxes, combinations of multiple different second syntaxes, and combinations of multiple different third syntaxes.

[0032] The combination of the plurality of different first syntaxes extracted by the extraction unit 130 from the speech data may be made up of at least one first syllable that combines a predetermined consonant and a predetermined vowel.

[0033] A predetermined consonant included in a plurality of different first syntactic combinations can be combined with the above-mentioned vowels in a consonant order according to grammatical rules. Figure 2 shows a consonant order 201 according to grammatical rules. Referring to Figure 2, the consonant order 201 can be "[giyeok], [nieun], [digeut], [rieul], [mieum], [bieup], [shiot], [ieung], [jieut], [chieut], [kiuek], [tieut], [pieup], [hieut]."

[0034] The predetermined vowel included in the combination of the plurality of different first syntaxes may be determined as any single vowel. In this embodiment, the predetermined vowel may be [a], which is the first vowel in the vowel order 202 disclosed in the grammar rule shown in FIG. 2.

[0035] The extraction unit 130 generates a first syllable from the speech data by combining the nth consonant to the (n+k)th consonant according to the consonant order with a predetermined vowel, and repeats the first syllable a preset number of times (for example, three times), thereby extracting it as a combination of the first syntax. Here, n includes a natural number, and k includes 0 and a natural number.

[0036] In one embodiment, when n=1 and k=0, the first syntactic combination may be "[gagaga]". In this embodiment, "[gagaga]" may also be named a syntactic in accordance with the definition above. In another embodiment, when n=1 and k=3, the first syntactic combination may be "[gagaganananadadadararara]".

[0037] 3 shows combinations of first syntaxes extracted from speech data by the extraction unit 130. Referring to FIG. 3, combinations of first syntaxes can include "[gagaga]" to "[gagaga...hahaha]".

[0038] The combinations of the multiple different second syntaxes extracted by the extraction unit 130 from the speech data may be configured by combining one or more of the combinations of a basic vowel set, a second syllable, and a first syntax.

[0039] In this embodiment, the basic vowel set can include "[aeiou]". In this embodiment, the second syllable can be generated by combining a predetermined diphthong and a predetermined vowel. In this embodiment, the predetermined diphthong included in the second syllable can be [ssan bieup], which is the third diphthong in the diphthong order 203 according to the grammar rules disclosed in FIG. 2. Furthermore, the predetermined vowel included in the second syllable can be [wa], which is the fourth diphthong in the diphthong order 204 according to the grammar rules disclosed in FIG. 2. Thus, the second syllable can be [ppawa]. In this embodiment, the second syllable can be repeated a predetermined number of times (e.g., three times) so as to be included in the second syntactic combination.

[0040] 4 shows combinations of the second syntax extracted from the speech data by the extraction unit 130. Referring to FIG. 4, the combinations of the second syntax can include "[aeiou]" to "[aeiou...ppawappawappawa]".

[0041] The combination of the plurality of different third syntaxes extracted by the extraction unit 130 from the speech data may include a combination of a third syllable and any of the syntaxes included in the combination of the plurality of different second syntaxes.

[0042] In this embodiment, the third syllable can be generated by combining any consonant and any vowel. In this embodiment, any consonant included in the third syllable can be specified as "[nieun], [digeut], [shiot], [ieung], or [hieut] in the consonant order 201 according to the grammar rules shown in FIG. 2. In addition, any vowel included in the third syllable may be [a], which is the first vowel in the vowel order 202 according to the grammar rules shown in FIG. 2. From this, the third syllable can be specified as "[na]", "[da]", "[sa]", "[a]", or "[ha]". In this embodiment, the third syllable can be repeated a predetermined number of times (e.g., three times) so as to be included in the third syntactic combination.

[0043] In this embodiment, the third syntax combination can be specified as "[aeiougagaga...mamama]", which is a combination of any syntax included in the second syntax combination.

[0044] In this embodiment, the combinations of multiple different third syntaxes are different from the combinations of multiple different second syntaxes and the combinations of multiple different third syntaxes, and can be identified as combinations of syntaxes that rank in the top 1 to 4 in terms of accuracy of the Parkinson's disease prediction model after enhancing the speech data.

[0045] 5 shows a plurality of different combinations of third syntaxes extracted from the speech data by the extraction unit 130. Referring to Fig. 5, the plurality of different combinations of third syntaxes can be identified as "[nananasasasa]", "[dadadaaaa]", and "[aaahahaha]", which are combinations of a third syllable repeated a preset number of times, and "[aeiougagaga...mamam]", which is included in a plurality of different combinations of second syntaxes.

[0046] The verification unit 140 can verify the accuracy of the Parkinson's disease prediction model by changing the combination of conditions and syntax for preprocessing the speech data.

[0047] In this embodiment, the conditions for preprocessing the audio data may include conditions for performing one or more of acoustic preprocessing, syntax, data padding, statistical properties, data splitting, label imbalance imputation, scaling, outlier removal, and floating-point formatting.

[0048] The acoustic preprocessing can be classified into first acoustic preprocessing and second acoustic preprocessing. The control unit 170 can select either the first acoustic preprocessing or the second acoustic preprocessing and output the selected result to the verification unit 140. The verification unit 140 can verify the accuracy of the Parkinson's disease prediction model by executing either the selected first acoustic preprocessing or the selected second acoustic preprocessing.

[0049] In this embodiment, the first acoustic pre-processing may include channel equalization and sampling rate equalization. The channel equalization may mean unifying the number of channels for recording audio data to one (monaural). The sampling rate equalization may mean unifying the sampling rate to a certain number of times per second (e.g., 16,000 times). Here, the sampling rate may mean the number of samples per time when converting an analog audio signal to a digital audio signal.

[0050] In this embodiment, the second audio pre-processing may include channel equalization, sampling rate equalization, volume normalization, high-pass filtering, DC offset removal, and noise reduction (stationary, 40%). The channel equalization and sampling rate equalization processes are similar to those described above and will not be described here. The volume of audio data may vary depending on the speaker's voice volume or the speaker's distance from the recording device (not shown). Volume normalization may refer to normalizing the overall audio data according to the average volume of the audio data to make the volume of audio data recorded in different ways relatively uniform, as described above. High-pass filtering may refer to passing only audio signals with frequencies higher than a cutoff frequency (e.g., 90 Hz). DC offset removal may refer to zeroing out the DC offset. Here, the DC offset may represent the average strength of a waveform. A DC offset may not be zero, which indicates that a DC offset has occurred. A DC offset may occur when audio is not recorded properly or when there is a malfunction in the recording device. Noise reduction processing can mean reducing noise to obtain a clean signal. Here, stationary can mean applying a noise threshold to the entire interval of the audio signal, and 40% can mean reducing the amplitude of the audio signal by 40% when noise is detected.

[0051] The syntax can be classified into the above-mentioned combinations of multiple different first syntaxes, combinations of multiple different second syntaxes, and combinations of multiple different third syntaxes. The control unit 170 can select one or more of the combinations of multiple different first syntaxes, combinations of multiple different second syntaxes, and combinations of multiple different third syntaxes, and output them to the verification unit 140. The verification unit 140 can execute one or more of the selected combinations of multiple different first syntaxes, combinations of multiple different second syntaxes, and combinations of multiple different third syntaxes to verify the accuracy of the Parkinson's disease prediction model. For convenience of explanation, the syntax in this embodiment can be excluded from the conditions for preprocessing the speech data.

[0052] Data padding may refer to adding specific values ​​to empty intervals before and after data to equalize the size of the data. In this embodiment, data padding may include audio data padding, which may refer to adding specific values ​​to silent intervals before and after audio data to equalize the size of the audio data. Types of data padding may include edge, repeat, tile, and zero. Edge may fill the preceding silent interval with the start value of the data and the following silent interval with the end value of the data. Repeat may fill the silent interval with a repeated individual value (e.g., 1). Tile may fill the silent interval with a repeated whole value (e.g., 1234). Zero may fill the silent interval with 0. The control unit 170 may select one or more of the types of data padding and output the selected type to the verification unit 140. The verification unit 140 may execute one or more of the selected types of data padding to verify the accuracy of the Parkinson's disease prediction model.

[0053] The statistical characteristics may include at least one of gender, volume, size (root mean square), maximum amplitude, duration, and average syntactic duration per speaker, which can be extracted from the speech data. The control unit 170 may select one or more of the statistical characteristics and output them to the verification unit 140. The verification unit 140 may execute one or more of the selected statistical characteristics to verify the accuracy of the Parkinson's disease prediction model.

[0054] Data division may include dividing the entire dataset used in the Parkinson's disease prediction model at a certain ratio. In this embodiment, the entire dataset can be divided into a training set, a validation set, and a test set. For example, the training set, validation set, and test set can be divided in a ratio of 64:16:20. The control unit 170 can determine the division ratio of the entire dataset and output it to the verification unit 140. The verification unit 140 can divide the speech data at the determined division ratio (64:16:20) of the entire dataset to verify the accuracy of the Parkinson's disease prediction model.

[0055] Compensating for data imbalance may mean amplifying speech data to compensate for the imbalance. To compensate for data imbalance, one or more of the following algorithms can be used: the synthetic minority over sampling technique (SMOTE) algorithm, the synthetic minority over sampling technique of nominal and continuous (SMOTE-NC) algorithm, and a weight-balancing algorithm. The SMOTE algorithm can amplify data of classes present at low ratios by newly generating them using a k-NN algorithm. The SMOTE-NC algorithm can be used for data containing a mixture of categorical data and continuous data. The class-weighting algorithm can adjust the learning of the Parkinson's disease prediction model by assigning a larger loss weight to data with a lower class ratio. The control unit 170 can select one or more of the SMOTE algorithm, the SMOTE-NC algorithm, and the class-weighting algorithm and output them to the verification unit 140. The verification unit 140 can execute the selected one or more of the SMOTE algorithm, the SMOTE-NC algorithm, and the class-weighting algorithm to verify the accuracy of the Parkinson's disease prediction model.

[0056] Scaling may refer to adjusting the range of data values. Data is scaled because if the data values ​​are too large or too small, they may converge to 0 or diverge infinitely during the model learning process. In this embodiment, scaling may be classified into normalization or standardization scaling. Normalization may refer to scaling such that feature values ​​lie between 0 and 1. Standardization may refer to scaling such that the mean of feature values ​​is 0 and the variance of feature values ​​is 1. The control unit 170 may select normalization or standardization from the data scaling and output the selected data scaling to the verification unit 140. The verification unit 140 may verify the accuracy of the Parkinson's disease prediction model using normalization or standardization from the selected data scaling.

[0057] Outlier removal can be performed to improve the accuracy of the model during training. In this embodiment, an isolation forest algorithm can be used as the algorithm for removing outliers. The control unit 170 can select whether to remove the outliers or not, and output the selection to the verification unit 140. The verification unit 140 can verify the accuracy of the Parkinson's disease prediction model by performing the selected outlier removal or no outlier removal.

[0058] Floating-point formats are data representation methods, and in this embodiment, can be classified into float32 and float64. float32 represents data in 32 bits, and float64 represents data in 64 bits. float32 can represent data with half the capacity of float64, but there is a possibility of data loss. float64 can represent data in more detail with less loss. The control unit 170 can select whether the data representation is float32 or float64, and output the selected data to the verification unit 140. The verification unit 140 can represent the speech data in the selected float32 or float64 format to verify the accuracy of the Parkinson's disease prediction model.

[0059] In some embodiments, the preprocessing of the audio data may further include audio feature extraction. In this embodiment, methods for extracting audio features can be classified into short-time Fourier transform chromagram (STFT), constant-Q transform chromagram (CQT), chroma energy normalized statistics (CENS), Mel-scaled spectrogram, mel-frequency cepstral coefficients (MFCC), and tempogram. STFT divides an audio signal into intervals of a certain length and applies a Fourier transform to each interval to extract a spectrum over time. CQT converts an audio signal into the frequency domain to extract vector values ​​representing the pitch frequency distribution. CENS extracts normalized chroma vector values ​​using short-term harmonics of the audio signal. Mel-scaled spectrograms can extract features by analyzing Mel-scaled spectrograms. MFCC samples audio data at predetermined time intervals and then analyzes the spectrum to extract vector values ​​of characteristic sounds. TEMPOGRAM can extract mid-level features that represent local tempo characteristics of an audio signal. In this embodiment, the control unit 170 can set one or more of the above-mentioned methods for extracting audio features to be executed by default and output the results to the verification unit 140. The verification unit 140 can execute one or more of the above-mentioned methods for extracting audio features by default when verifying the accuracy of the Parkinson's disease prediction model. In an optional embodiment, the control unit 170 can select one or more of the above-mentioned methods for extracting audio features and output them to the verification unit 140. The verification unit 140 can execute one or more of the selected methods for extracting audio features to verify the accuracy of the Parkinson's disease prediction model.

[0060] In an optional embodiment, the conditions for preprocessing the speech data may further include syllable processing. In this embodiment, syllable processing can be classified into mean processing and concatenation processing. Mean processing can include processing for averaging syllables spoken in a high pitch and syllables spoken in a low pitch. Concatenation processing can include processing for concatenating a syllable spoken in a high pitch with a syllable spoken in a low pitch. For example, when concatenating "[gagagananan]," which is one of a plurality of different first syntax combinations, the result may be the concatenation of a high-pitched "[gagaga]" and a low-pitched "[nanana]." In this embodiment, the control unit 170 can set one or more of the mean processing and the concatenation processing to be performed by default and output the results to the verification unit 140. When verifying the accuracy of the Parkinson's disease prediction model, the verification unit 140 can perform one or more of the mean processing and the concatenation processing by default. In any embodiment, the control unit 170 can select one or more of the mean processing and the concat processing and output them to the verification unit 140. The verification unit 140 can execute the selected one or more of the mean processing and the concat processing to verify the accuracy of the Parkinson's disease prediction model.

[0061] In this embodiment, the verification unit 140 may perform first to seventh verifications to verify the accuracy of the Parkinson's disease prediction model.

[0062] The verification unit 140 can perform a first verification that verifies the accuracy of the Parkinson's disease prediction model based on a first preprocessing condition that preprocesses the speech data and a combination of a plurality of different first syntaxes.

[0063] 6 is a table showing first conditions for preprocessing speech data for a first validation run. Referring to FIG. 6, the first preprocessing conditions may include, among the above-described conditions for preprocessing speech data, performing second acoustic preprocessing, performing data padding including edges, repeats, tiles, and zeros, performing statistical characteristics including gender, volume, size, maximum amplitude, length, and average syntactic length by speaker, performing data splitting in which the training set, validation set, and test set are divided into 64:16:20, performing data imbalance imputation using the SMOTE-NC algorithm, performing scaling including standardization, performing outlier removal using isolation forest, and performing floating-point formatting including float64.

[0064] The verification unit 140 can perform a second verification to verify the accuracy of the Parkinson's disease prediction model based on a second preprocessing condition for preprocessing the speech data and a combination of a plurality of different second syntaxes.

[0065] 9 is a table showing second conditions for preprocessing speech data for second verification. Referring to FIG. 9, the second preprocessing conditions may include, among the above-described conditions for preprocessing speech data, performing first acoustic preprocessing, performing data padding including edges, repeats, tiles, and zeros, performing statistical characteristics including gender, volume, size, maximum amplitude, length, and average syntactic length for each speaker, performing data splitting in which the training set, validation set, and test set are divided into 64:16:20, performing data imbalance imputation including the SMOTE-NC algorithm, performing scaling including standardization, performing outlier removal including isolation forest, and performing floating-point formatting including float64. In this embodiment, the second verification may have a different combination of acoustic preprocessing and syntax compared to the first verification.

[0066] The verification unit 140 can perform a third preprocessing condition for preprocessing the speech data and a third verification for verifying the accuracy of the Parkinson's disease prediction model based on a combination of a plurality of different third syntaxes.

[0067] 12 is a table showing third conditions for preprocessing speech data for third verification. Referring to FIG. 12, the third preprocessing conditions may include the following conditions for preprocessing speech data: performing first acoustic preprocessing; performing data padding including edges, repeats, tiles, and zeros; performing statistical characteristics including gender, volume, size, maximum amplitude, length, and average syntactic length for each speaker; performing data splitting in which the training set, validation set, and test set are divided into 64:16:20; performing data imbalance imputation using the SMOTE-NC algorithm; performing scaling including standardization; performing outlier removal using isolation forest; and performing floating-point formatting including float64. In this embodiment, the third verification may have different syntactic combinations compared to the second verification.

[0068] The verification unit 140 can perform a fourth preprocessing condition for preprocessing the speech data and a fourth verification for verifying the accuracy of the Parkinson's disease prediction model based on a combination of a plurality of different third syntaxes.

[0069] 15 is a table showing fourth conditions for preprocessing speech data for the fourth validation run. Referring to FIG. 15, the fourth preprocessing conditions may include the following conditions for preprocessing speech data: performing first acoustic preprocessing; performing data padding including edges, repeats, tiles, and zeros; performing statistical characteristics including gender, volume, size, maximum amplitude, length, and average syntactic length by speaker; performing data splitting in which the training set, validation set, and test set are divided into 64:16:20; performing data imbalance imputation using the SMOTE-NC algorithm; performing scaling including standardization; performing outlier removal using isolation forest; and performing floating-point formatting including float64. In this embodiment, the fourth validation run may differ in acoustic preprocessing from the third validation run.

[0070] The verification unit 140 can perform a fifth preprocessing condition for preprocessing the speech data and a fifth verification for verifying the accuracy of the Parkinson's disease prediction model based on a combination of a plurality of different third syntaxes.

[0071] FIG. 18 is a table showing fifth conditions for preprocessing speech data for the fifth validation run. Referring to FIG. 18, the fifth preprocessing conditions may include the following conditions for preprocessing speech data: performing first acoustic preprocessing; performing data padding including edges, repeats, tiles, and zeros; performing statistical characteristics including gender, volume, size, maximum amplitude, length, and average syntactic length by speaker; performing data splitting in which the training set, validation set, and test set are divided into 64:16:20; performing data imbalance compensation including the SMOTE algorithm and class weighting algorithm; performing scaling including standardization; performing no outlier removal; and performing floating-point formatting including float64. In this embodiment, the fifth validation run may differ from the fourth validation run in terms of acoustic preprocessing, statistical characteristics, data imbalance compensation, and outlier removal.

[0072] The verification unit 140 can perform a sixth verification to verify the accuracy of the Parkinson's disease prediction model based on a sixth preprocessing condition for preprocessing the speech data, a combination of a plurality of different first syntaxes, and a combination of a plurality of different second syntaxes.

[0073] 21 is a table showing sixth conditions for preprocessing speech data for sixth verification. Referring to FIG. 21, the sixth preprocessing conditions may include the following conditions for preprocessing speech data: performing first acoustic preprocessing; padding data including edges; evaluating statistical characteristics including gender, volume, size, maximum amplitude, length, and average syntactic length for each speaker; splitting the data into a training set, validation set, and test set in a ratio of 64:16:20; correcting data imbalances using the SMOTE-NC algorithm; scaling including standardization; removing outliers using isolation forest; and using a floating-point format including float64. In this embodiment, the sixth verification may differ from the fifth verification in terms of syntactic combination, data padding, statistical characteristics, correcting data imbalances, and removing outliers.

[0074] The verification unit 140 can perform a seventh verification to verify the accuracy of the Parkinson's disease prediction model based on a seventh preprocessing condition for preprocessing the speech data, a combination of a plurality of different first syntaxes, and a combination of a plurality of different second syntaxes.

[0075] 24 is a table showing seventh conditions for preprocessing speech data for the seventh validation run. Referring to FIG. 24, the seventh preprocessing conditions may include, among the above-described conditions for preprocessing speech data, performing second acoustic preprocessing, performing data padding including edges, evaluating statistical characteristics including gender, volume, size, maximum amplitude, length, and average syntactic length for each speaker, performing data splitting in which the training set, validation set, and test set are divided into 64:16:20, performing data imbalance imputation including the SMOTE-NC algorithm, performing scaling including standardization, performing outlier removal including isolation forest, and performing floating-point formatting including float64. In this embodiment, the seventh validation run may differ in acoustic preprocessing compared to the sixth validation run.

[0076] The determination unit 150 can determine the combination of syntaxes ranked highly in the accuracy verification result of the Parkinson's disease prediction model output from the verification unit 140 as the speech data for Parkinson's disease prediction.

[0077] The determination unit 150 can determine, as the speech data for predicting Parkinson's disease, the combination of first syntaxes ranked highest from the result of the first verification output from the verification unit 140. Here, the combination of first syntaxes ranked highest may be a first syllable generated by combining the first to eleventh consonants with a predetermined vowel according to the consonant order, repeated a preset number of times.

[0078] 7 is a diagram showing the accuracy of different first syntax combinations ranked from top 1 to top 10 according to the result of the first verification. Referring to FIG. 7, the determination unit 150 can determine [gagagananana...kakaka], which is one of the first syntax combinations ranked highest (top 1), as speech data for predicting Parkinson's disease.

[0079] Figure 8 is a diagram showing the results of comparing the average accuracy of elements included in the first preprocessing condition according to the results of the first verification. Referring to Figure 8, 801 shows the results of comparing the average accuracy of different combinations of the first syntax. 802 shows the results of comparing the average accuracy for speech features by speech feature extraction method. 803 shows the results of comparing the average accuracy by syllable processing. 804 shows the results of comparing the average accuracy by data padding type.

[0080] The determination unit 150 can determine the highest ranked basic vowel set from the second verification result output from the verification unit 140 as the voice data for Parkinson's disease prediction.

[0081] 10 is a diagram showing the accuracy of different second syntax combinations ranked from top 1 to top 10 according to the results of the second verification. Referring to FIG. 10, the determination unit 150 can determine the basic vowel set, i.e., [aeiou], which is one of the second syntax combinations ranked highest (top 1), as the speech data for Parkinson's disease prediction.

[0082] FIG. 11 is a diagram showing the results of comparing the average accuracy of elements included in the second preprocessing conditions according to the results of the second verification. Referring to FIG. 11, 1101 shows the results of comparing the average accuracy of different combinations of the second syntax. 1102 shows the results of comparing the average accuracy for speech features by speech feature extraction method. 1103 shows the results of comparing the average accuracy by syllable processing. 1104 shows the results of comparing the average accuracy by data padding type.

[0083] The determination unit 150 can determine the highest ranked combination of third syntax as the speech data for predicting Parkinson's disease from the result of the third verification output from the verification unit 140. Here, the highest ranked combination of third syntax may be a combination of different first syllables generated by combining the third consonant and the eighth consonant with a predetermined vowel according to the consonant order, repeated a predetermined number of times.

[0084] 13 is a diagram showing the accuracy of different third syntax combinations ranked from top 1 to top 10 according to the results of the third verification. Referring to FIG. 13, the determination unit 150 can determine [dadadaaaa], which is one of the third syntax combinations ranked highest (top 1), as the speech data for Parkinson's disease prediction. In this embodiment, [dadadaaaa] ranked highest (top 1) may be one of the third syntax combinations in which the Mel-scaled spectrogram extraction method is performed in the speech feature extraction included in the third preprocessing condition, the mean processing is performed in the syllable processing, and zero is performed as the data padding type.

[0085] Fig. 14 is a diagram showing the results of comparing the average accuracy of elements included in the third preprocessing condition according to the results of the third verification. Referring to Fig. 14, 1401 shows the results of comparing the average accuracy of different combinations of the third syntax. 1402 shows the results of comparing the average accuracy for speech features by speech feature extraction method. 1403 shows the results of comparing the average accuracy by syllable processing. 1404 shows the results of comparing the average accuracy by data padding type.

[0086] The determination unit 150 can determine, as the speech data for predicting Parkinson's disease, the combination of the third syntax ranked highest from the result of the fourth verification output from the verification unit 140. Here, the combination of the third syntax ranked highest may be a combination of different third syllables generated by combining the eighth and fourteenth consonants according to the consonant order with a predetermined vowel, repeated a preset number of times.

[0087] 16 is a diagram showing the accuracy of different third syntax combinations ranked from top 1 to top 10 according to the result of the fourth verification. Referring to FIG. 16, the determination unit 150 can determine [aaahahaha], which is one of the third syntax combinations ranked highest (top 1), as the speech data for predicting Parkinson's disease. In this embodiment, [aaahahaha] ranked highest (top 1) may be one of the third syntax combinations for which the Mel-scaled spectrogram extraction method is performed in the speech feature extraction included in the fourth preprocessing condition, mean processing is performed in the syllable processing, and edge is performed as the type of data padding.

[0088] Fig. 17 is a diagram showing the results of comparing the average accuracy of elements included in the fourth preprocessing condition according to the results of the fourth verification. Referring to Fig. 17, 1701 shows the results of comparing the average accuracy of different combinations of the third syntax. 1702 shows the results of comparing the average accuracy for speech features by speech feature extraction method. 1703 shows the results of comparing the average accuracy by syllable processing. 1704 shows the results of comparing the average accuracy by data padding type.

[0089] The determination unit 150 can determine, as the speech data for predicting Parkinson's disease, the combination of the third syntax ranked highest from the result of the fifth verification output from the verification unit 140. Here, the combination of the third syntax ranked highest may be a combination of different third syllables generated by combining the second and seventh consonants according to the consonant order with a predetermined vowel, repeated a preset number of times.

[0090] 19 is a diagram showing the accuracy of different third syntax combinations ranked from top 1 to top 10 according to the results of the fifth verification. Referring to FIG. 19, the determination unit 150 can determine [nananasasasa], which is one of the third syntax combinations ranked highest (top 1), as the speech data for predicting Parkinson's disease. In this embodiment, [nananasasasa] ranked highest (top 1) may be one of the third syntax combinations for which the Mel-scaled spectrogram extraction method is performed in the speech feature extraction included in the fifth preprocessing condition, concat processing is performed in the syllable processing, and edge is performed as the data padding type.

[0091] Figure 20 is a diagram showing the results of comparing the average accuracy of elements included in the fifth preprocessing condition according to the results of the fifth verification. Referring to Figure 20, 2001 shows the results of comparing the average accuracy of different combinations of the third syntax. 2002 shows the results of comparing the average accuracy for speech features by speech feature extraction method. 2003 shows the results of comparing the average accuracy by syllable processing. 2004 shows the results of comparing the average accuracy by data padding type.

[0092] The determination unit 150 can determine, as the speech data for predicting Parkinson's disease, the combination of first syntaxes ranked highest from the result of the sixth verification output from the verification unit 140. Here, the combination of first syntaxes ranked highest may be a combination of a plurality of different first syllables generated by combining the first to ninth consonants according to the consonant order with a predetermined vowel, repeated a preset number of times.

[0093] Furthermore, the determination unit 150 can determine the top two ranked combinations of first syntaxes as the speech data for predicting Parkinson's disease. Here, the top two ranked combinations of first syntaxes may be a preset number of repetitions of a plurality of different first syllables generated by combining the first to seventh consonants according to the consonant order with a predetermined vowel.

[0094] Furthermore, the determination unit 150 can determine the top three ranked combinations of first syntaxes as the speech data for predicting Parkinson's disease. Here, the top three ranked combinations of first syntaxes may be a preset number of repetitions of a plurality of different first syllables generated by combining the first to sixth consonants according to the consonant order with a predetermined vowel.

[0095] Furthermore, the determination unit 150 can determine the top four ranked combinations of first syntaxes as the speech data for predicting Parkinson's disease. Here, the top four ranked combinations of first syntaxes may be a preset number of repetitions of a plurality of different first syllables generated by combining the first to fifth consonants according to the consonant order with a predetermined vowel.

[0096] 22 is a diagram showing the accuracy of different combinations of first syntaxes and different combinations of second syntaxes, ranked from top 1 to top 10, according to the result of the sixth verification. Referring to FIG. 22, the determination unit 150 can determine that [gagagananana...jajaja], which is one of the first syntax combinations ranked highest (top 1), is speech data for predicting Parkinson's disease.

[0097] Furthermore, the determination unit 150 can determine that [gagagananana...sasasa], which is one of the first syntax combinations ranked in the top two places, is speech data for predicting Parkinson's disease.

[0098] Furthermore, the determination unit 150 can determine that [gagagananana...bababa], which is one of the combinations of the first syntax ranked in the top three, is the speech data for predicting Parkinson's disease.

[0099] Furthermore, the determination unit 150 can determine that [gagagananana...mamama], which is one of the first syntax combinations ranked in the top four, is the speech data for predicting Parkinson's disease.

[0100] FIG. 23 shows the results of comparing the average accuracy for different combinations of first syntaxes and different combinations of second syntaxes according to the results of the sixth verification.

[0101] The determination unit 150 can determine, as the speech data for predicting Parkinson's disease, the combination of first syntaxes ranked highest from the result of the seventh verification output from the verification unit 140. Here, the combination of first syntaxes ranked highest may be a combination of a plurality of different first syllables generated by combining the first to ninth consonants according to the consonant order with a predetermined vowel, repeated a preset number of times.

[0102] Furthermore, the determination unit 150 can determine the top two ranked combinations of first syntaxes as the speech data for predicting Parkinson's disease. Here, the top two ranked combinations of first syntaxes may be a preset number of repetitions of a plurality of different first syllables generated by combining the first to seventh consonants according to the consonant order with a predetermined vowel.

[0103] 25 is a diagram showing the accuracy of different combinations of first syntaxes and different combinations of second syntaxes, ranked from top 1 to top 10, according to the result of the seventh verification. Referring to FIG. 25, the determination unit 150 can determine that [gagagananana...jajaja], which is one of the first syntax combinations ranked highest (top 1), is speech data for predicting Parkinson's disease.

[0104] Furthermore, the determination unit 150 can determine that [gagagananana...sasasa], which is one of the first syntax combinations ranked in the top two places, is speech data for predicting Parkinson's disease.

[0105] FIG. 26 shows the results of comparing the average accuracy for different combinations of first syntaxes and different combinations of second syntaxes according to the results of the seventh test.

[0106] The acquisition unit 160 inputs the speech results of the speaker corresponding to the voice data for Parkinson's disease prediction determined by the determination unit 150 into a Parkinson's disease prediction model, and can acquire the Parkinson's disease prediction result for the speaker from the output of the Parkinson's disease prediction model.

[0107] The control unit 170 can control the overall operation of the Parkinson's disease prediction device 100. The control unit 170 can include any type of device capable of processing data, such as a processor. Here, the term "processor" may refer to a data processing device implemented in hardware, having a physically structured circuit for executing functions expressed by code or instructions included in a program. Examples of such data processing devices implemented in hardware include a microprocessor, a central processing unit (CPU), a processor core, a multiprocessor, an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), etc., but the scope of the present invention is not limited thereto.

[0108] In an optional embodiment, the generation unit 120 may generate a gender classification model. The gender classification model in this embodiment may be generated by training a deep neural network model that has been pre-trained to classify the gender of a speaker using audio data including the speaker's speech results. Here, the deep neural network model may be a model that is trained in a supervised learning manner using training data that uses audio data including the speaker's speech results as input and labels the speaker's gender.

[0109] The generation unit 120 may train the initially set deep neural network model in a supervised learning manner using the labeled training data. Here, the initially set deep neural network model is an initial model designed to be configured as a gender classification model, and parameter values ​​are set to arbitrary initial values. The initially set deep neural network model is trained using the above training data and the parameter values ​​are optimized, thereby completing the gender classification model that can accurately classify the gender of a speaker.

[0110] The verification unit 140 can verify the accuracy of the gender classification model based on a combination of a plurality of different first syntaxes, a combination of a plurality of different second syntaxes, and a combination of a plurality of different third syntaxes. The verification unit 140 can apply the above-mentioned conditions for preprocessing the speech data when verifying the accuracy of the gender classification model.

[0111] The determination unit 150 can determine the highest ranked combination of third syntaxes as the speech data for gender classification based on the result of verifying the accuracy of the gender classification model. Here, the highest ranked combination of third syntaxes may be a combination of different first syllables generated by combining the third consonant and the eighth consonant with a predetermined vowel according to the consonant order, repeated a predetermined number of times.

[0112] 27 is a diagram showing the accuracy rankings of different combinations of first syntaxes, different combinations of second syntaxes, and different combinations of third syntaxes, from top 1 to top 10, according to the results of accuracy verification of the gender classification model. Referring to FIG. 27, the determination unit 150 can determine [dadadaaaa], which is one of the third syntax combinations ranked highest (top 1), as the speech data for gender classification.

[0113] Figure 28 shows the results of comparing the average accuracy for different combinations of first syntaxes and different combinations of second syntaxes according to the results of the accuracy verification of the gender classification model.

[0114] The acquiring unit 160 inputs the speech result of the speaker corresponding to the voice data determined by the determining unit 150 into the gender classification model, and can acquire the gender classification result for the speaker from the output of the gender classification model.

[0115] In an optional embodiment, the generation unit 120 may generate a syntactic performance comparison model. In this embodiment, the syntactic performance comparison model may be generated by training a deep neural network model that is pre-trained to output performance comparison results for each syntax included in the speech data using speech data including the speech results of a speaker. Here, the deep neural network model may be a model that is trained in a supervised learning manner using training data that uses speech data including the speech results of a speaker as input and that labels the performance comparison results for each syntax.

[0116] The generation unit 120 may train the initially set deep neural network model in a supervised learning manner using the labeled training data. Here, the initially set deep neural network model is an initial model designed to be configured for syntax performance comparison, and parameter values ​​are set to arbitrary initial values. The initially set deep neural network model is trained using the above training data and the parameter values ​​are optimized, so that it can be completed as a syntax performance comparison model that can accurately compare the performance of each syntax.

[0117] The verification unit 140 can verify the accuracy of the syntax performance comparison model for Parkinson's disease prediction based on a predetermined syntax extracted from speech data including the speaker's speech results. The verification unit 140 can apply the above-mentioned conditions for preprocessing the speech data when verifying the accuracy of the syntax performance comparison model.

[0118] Based on the results of verifying the accuracy of the performance comparison model of syntax for Parkinson's disease prediction, the determination unit 150 can determine the top-ranked syntax as the speech data for the performance comparison of syntax for Parkinson's disease prediction.

[0119] Based on the result of verifying the accuracy of the performance comparison model of syntaxes for Parkinson's disease prediction, the determination unit 150 can determine the first syntax ranked first highest as the speech data for the performance comparison of syntaxes for Parkinson's disease prediction. Here, the first syntax ranked first highest may be a first syllable generated by combining the fourth consonant according to the consonant order and a predetermined vowel, repeated a predetermined number of times.

[0120] Furthermore, the determination unit 150 may determine the second highest ranked first syntax as speech data for performance comparison of syntaxes for predicting Parkinson's disease. Here, the second highest ranked first syntax may be a first syllable generated by combining an eighth consonant according to the consonant order and a predetermined vowel, repeated a preset number of times.

[0121] Furthermore, the determination unit 150 can determine the third highest ranked basic vowel set as speech data for performance comparison of syntax for Parkinson's disease prediction.

[0122] Furthermore, the determination unit 150 may determine the first syntax ranked in the fourth highest position as speech data for performance comparison of syntaxes for predicting Parkinson's disease. Here, the first syntax ranked in the fourth highest position may be a first syllable generated by combining an eleventh consonant according to the consonant order and a predetermined vowel, repeated a preset number of times.

[0123] Furthermore, the determination unit 150 can determine the first syntax ranked in the fifth highest rank as speech data for performance comparison of syntaxes for predicting Parkinson's disease. Here, the first syntax ranked in the fifth highest rank may be a first syllable generated by combining the second consonant according to the consonant order and a predetermined vowel, repeated a preset number of times.

[0124] Furthermore, the determination unit 150 may determine the first syntax ranked in the sixth highest rank as speech data for performance comparison of syntaxes for predicting Parkinson's disease. Here, the first syntax ranked in the sixth highest rank may be a first syllable generated by combining a sixth consonant according to the consonant order and a predetermined vowel, repeated a preset number of times.

[0125] Furthermore, the determination unit 150 can determine the first syntax ranked in the seventh highest rank as speech data for performance comparison of syntaxes for predicting Parkinson's disease. Here, the first syntax ranked in the seventh highest rank may be a first syllable generated by combining the twelfth consonant according to the consonant order and a predetermined vowel, repeated a preset number of times.

[0126] Furthermore, the determination unit 150 can determine the first syntax ranked in the eighth highest ranking as speech data for performance comparison of syntaxes for predicting Parkinson's disease. Here, the first syntax ranked in the eighth highest ranking may be a first syllable generated by combining the 14th consonant according to the consonant order and a predetermined vowel, repeated a preset number of times.

[0127] Furthermore, the determination unit 150 can determine the first syntax ranked in the 9th highest as speech data for comparing the performance of syntaxes for predicting Parkinson's disease. Here, the first syntax ranked in the 9th highest is a combination of a first syllable generated by combining the 13th consonant according to the consonant order with a predetermined vowel, a first syllable generated by combining the 12th consonant with a predetermined vowel, and a first syllable generated by combining the 11th syllable with a predetermined vowel.

[0128] Furthermore, the determination unit 150 can determine the first syntax ranked in the 10th highest ranking as speech data for performance comparison of syntaxes for predicting Parkinson's disease. Here, the first syntax ranked in the 10th highest ranking may be a first syllable generated by combining the fifth consonant according to the consonant order and a predetermined vowel, repeated a preset number of times.

[0129] FIG. 29 is a diagram showing the accuracy of a plurality of different first syntaxes, ranked from top 1 to top 10, according to the results of accuracy verification of the syntax performance comparison model. Referring to FIG. 29, the determination unit 150 may determine [rarara], one of the first syntaxes ranked first, as speech data for syntax performance comparison. The determination unit 150 may determine [aaa], one of the first syntaxes ranked second, as speech data for syntax performance comparison. The determination unit 150 may determine [aeiou], a basic vowel set ranked third, as speech data for syntax performance comparison. The determination unit 150 may determine [kakaka], one of the first syntaxes ranked fourth, as speech data for syntax performance comparison. The determination unit 150 may determine [nanana], one of the first syntaxes ranked fifth, as speech data for syntax performance comparison. The determination unit 150 can determine [bababa], one of the first syntaxes ranked sixth, as the speech data for the syntax performance comparison. The determination unit 150 can determine [tatata], one of the first syntaxes ranked seventh, as the speech data for the syntax performance comparison. The determination unit 150 can determine [hahaha], one of the first syntaxes ranked eighth, as the speech data for the syntax performance comparison. The determination unit 150 can determine [pataka], one of the first syntaxes ranked ninth, as the speech data for the syntax performance comparison. The determination unit 150 can determine [mamama], one of the first syntaxes ranked tenth, as the speech data for the syntax performance comparison.

[0130] Figure 30 shows the results of comparing the average accuracy for different primary syntaxes and basic vowel sets according to the results of the accuracy verification of the performance comparison model for syntax.

[0131] The acquisition unit 160 inputs the speech results of the speaker corresponding to the audio data determined by the determination unit 150 into a syntax performance comparison model, and can acquire the syntax performance comparison results for predicting Parkinson's disease for the speaker from the output of the syntax performance comparison model.

[0132] In an optional embodiment, the generation unit 120 may generate a first Parkinson's disease accuracy classification model. The first Parkinson's disease accuracy classification model in this embodiment may be generated by training a deep neural network model pre-trained to classify the accuracy of a speaker's Parkinson's disease using audio data including the speaker's speech. Here, the deep neural network model may be a model trained in a supervised learning manner using audio data including the speaker's speech and training data labeled with one or more of the first and second classes and a third class. The first class in this embodiment may include the disease with the highest accuracy among Parkinson's disease-related diseases: normal, Parkinson's disease, multiple system atrophy, and cerebellar atrophy. The second class may include the disease with the top two accuracy among Parkinson's disease-related diseases: normal, Parkinson's disease, multiple system atrophy, and cerebellar atrophy. The third class may include the remaining Parkinson's disease-related diseases excluding the first or second class.

[0133] The generation unit 120 may train the initially set deep neural network model in a supervised learning manner using the labeled training data. Here, the initially set deep neural network model is an initial model designed to be configured as a first Parkinson's disease accuracy classification model, and parameter values ​​are set to arbitrary initial values. The initially set deep neural network model is trained using the above training data and the parameter values ​​are optimized, thereby completing the first Parkinson's disease accuracy classification model that can accurately classify the accuracy of Parkinson's disease.

[0134] The verification unit 140 can verify the accuracy of the first Parkinson's disease accuracy classification model based on a combination of a plurality of different first syntaxes, a combination of a plurality of different second syntaxes, and a combination of a plurality of different third syntaxes. The verification unit 140 can apply the above-mentioned conditions for preprocessing the speech data when verifying the accuracy of the first Parkinson's disease accuracy classification model.

[0135] The determination unit 150 can determine the combination of sentences ranked highly from the accuracy verification results of the first Parkinson's disease accuracy classification model output from the verification unit 140 as speech data for classifying the accuracy for Parkinson's disease.

[0136] The determination unit 150 can determine the highest ranked combination of third syntax as the speech data for classifying the accuracy of Parkinson's disease from the accuracy verification result of the first Parkinson's disease accuracy classification model output from the verification unit 140. Here, the highest ranked combination of third syntax can correspond to a third class including the remaining Parkinson's disease-related diseases excluding cerebellar atrophy and cerebellar atrophy as the first class. Furthermore, the highest ranked combination of third syntax may be a predetermined number of repetitions of different third syllables generated by combining the third and eighth consonants according to the consonant order with a predetermined vowel.

[0137] Furthermore, the determination unit 150 can determine the top two ranked combinations of third syntaxes as the speech data for classifying the accuracy of Parkinson's disease. Here, the top two ranked combinations of third syntaxes can correspond to a third class including the remaining Parkinson's disease-related disorders excluding normal and normal as the first class. Furthermore, the top two ranked combinations of third syntaxes may be a preset number of repetitions of different third syllables generated by combining the eighth and fourteenth consonants according to the consonant order with a predetermined vowel.

[0138] Furthermore, the determination unit 150 can determine the top three ranked combinations of third syntaxes as the speech data for classifying the accuracy of Parkinson's disease. Here, the top three ranked combinations of third syntaxes can correspond to a third class including the remaining Parkinson's disease-related diseases excluding multiple system atrophy and multiple system atrophy as the first class. Furthermore, the top three ranked combinations of third syntaxes may be a predetermined number of repetitions of different third syllables generated by combining the second and seventh consonants according to the consonant order with a predetermined vowel.

[0139] Furthermore, the determination unit 150 can determine the top four ranked combinations of third syntaxes as the speech data for classifying the accuracy of Parkinson's disease. Here, the top four ranked combinations of third syntaxes can correspond to a third class including Parkinson's disease as the first class and the remaining Parkinson's disease-related diseases excluding Parkinson's disease. Furthermore, the top four ranked combinations of third syntaxes may be a predetermined number of repetitions of different third syllables generated by combining the eighth and fourteenth consonants according to the consonant order with a predetermined vowel.

[0140] Figure 31 is a diagram showing the accuracy rankings from top 1 to top 10 for different combinations of first syntaxes, different combinations of second syntaxes, and different combinations of third syntaxes, based on the results of accuracy verification of the first Parkinson's disease accuracy classification model.

[0141] 31, the determination unit 150 may determine [dadadaaaa], which is one of the first ranked third syntactic combinations, as the speech data for the first Parkinson's disease accuracy classification. Here, the first ranked [dadadaaaa] may be the syntactic combination that can most accurately distinguish between cerebellar atrophy (D) as the first class and the remaining Parkinson's disease-related disorders (X) excluding cerebellar atrophy as the third class, namely, normal, Parkinson's disease, and cerebellar atrophy.

[0142] Furthermore, the determination unit 150 may determine [aaahahaha], which is one of the third syntax combinations ranked second, as the speech data for the first Parkinson's disease accuracy classification. Here, [aaahahaha] ranked second may be the syntax combination that can distinguish, with the highest accuracy, between normal (A) as the first class and Parkinson's disease, multiple system atrophy, and cerebellar atrophy, which are the remaining Parkinson's disease-related diseases (X) excluding normal as the third class.

[0143] Furthermore, the determination unit 150 can determine [nananasasasa], which is one of the third ranked third syntactic combinations, as the speech data for the first Parkinson's disease accuracy classification. Here, [nananasasasa], which is ranked third, can be the syntactic combination that can distinguish, with the highest accuracy, multiple system atrophy (C) as the first class from the remaining Parkinson's disease-related diseases (X) excluding multiple system atrophy as the third class, namely normal, Parkinson's disease, and cerebellar atrophy.

[0144] Furthermore, the determination unit 150 may determine [aaahahaha], which is one of the third syntactic combinations ranked fourth, as the speech data for the first Parkinson's disease accuracy classification. Here, [aaahahaha], which is ranked third, may be the syntactic combination that can distinguish, with the highest accuracy, Parkinson's disease (B) as the first class from the remaining Parkinson's disease-related diseases (X) excluding Parkinson's disease as the third class, namely normal, multiple system atrophy, and cerebellar atrophy.

[0145] The acquisition unit 160 inputs the speech result of the speaker corresponding to the voice data determined by the determination unit 150 into the first Parkinson's disease accuracy classification model, and can acquire the first Parkinson's disease accuracy classification result for the speaker from the output of the Parkinson's disease accuracy classification model.

[0146] In an optional embodiment, the generation unit 120 may generate a second Parkinson's disease accuracy classification model. The second Parkinson's disease accuracy classification model in this embodiment may be generated by training a deep neural network model that has been pre-trained to classify the accuracy of a speaker's Parkinson's disease using audio data including the speaker's speech. Here, the deep neural network model may be a model trained in a supervised learning manner using audio data including the speaker's speech and training data labeled with two or more of the fourth, fifth, and sixth classes. The fourth class in this embodiment may include the diseases with the highest accuracy among normal, Parkinson's disease, multiple system atrophy, and cerebellar atrophy, which are included in Parkinson's disease-related diseases. The fifth class may include the diseases with higher accuracy among the remaining three Parkinson's disease-related diseases excluding the fourth class. The sixth class may include the diseases with the highest accuracy among the remaining two Parkinson's disease-related diseases excluding the fourth and fifth classes.

[0147] The generation unit 120 may train the initially set deep neural network model in a supervised learning manner using the labeled training data. Here, the initially set deep neural network model is an initial model designed to be configured as a second Parkinson's disease accuracy classification model, and parameter values ​​are set to arbitrary initial values. The initially set deep neural network model is trained using the above training data and the parameter values ​​are optimized, thereby completing the second Parkinson's disease accuracy classification model that can accurately classify the accuracy of Parkinson's disease.

[0148] The verification unit 140 can verify the accuracy of the second Parkinson's disease accuracy classification model based on a combination of a plurality of different first syntaxes, a combination of a plurality of different second syntaxes, and a combination of a plurality of different third syntaxes. The verification unit 140 can apply the above-mentioned conditions for preprocessing the speech data when verifying the accuracy of the second Parkinson's disease accuracy classification model.

[0149] The determination unit 150 can determine the combination of sentences ranked highly based on the result of verifying the accuracy of the second Parkinson's disease accuracy classification model as speech data for which the accuracy is classified as Parkinson's disease.

[0150] Based on the result of verifying the accuracy of the second Parkinson's disease accuracy classification model, the determination unit 150 can determine the first highest ranked combination of first syntax as the speech data for classifying the accuracy for Parkinson's disease. Here, the first highest ranked combination of first syntax can correspond to normal as the fourth class and cerebellar atrophy as the fifth class. Furthermore, the first highest ranked combination of first syntax may be a first syllable generated by combining the first to seventh consonants according to the consonant order with a predetermined vowel, repeated a predetermined number of times.

[0151] Furthermore, the determination unit 150 can determine the combination of first syntaxes ranked second highest as the speech data for classifying the accuracy of Parkinson's disease. Here, the combination of first syntaxes ranked second highest can correspond to normal as the fourth class and multiple system atrophy as the fifth class. Furthermore, the combination of first syntaxes ranked second highest may be a first syllable generated by combining the first to seventh consonants according to the consonant order with a predetermined vowel, repeated a predetermined number of times.

[0152] Furthermore, the determination unit 150 can determine the combination of first syntaxes ranked third highest as speech data for classifying the accuracy of Parkinson's disease. Here, the combination of first syntaxes ranked third highest can correspond to multiple system atrophy as a fourth class and cerebellar atrophy as a fifth class. Furthermore, the combination of first syntaxes ranked third highest may be a first syllable generated by combining the first to seventh consonants according to the consonant order with a predetermined vowel, repeated a predetermined number of times.

[0153] Furthermore, the determination unit 150 can determine the combination of first syntaxes ranked in the fourth highest ranking as speech data for classifying the accuracy of Parkinson's disease. Here, the combination of first syntaxes ranked in the fourth highest ranking can correspond to Parkinson's disease as a fourth class and cerebellar atrophy as a fifth class. Furthermore, the combination of first syntaxes ranked in the fourth highest ranking may be a first syllable generated by combining the first to seventh consonants according to the consonant order with a predetermined vowel, repeated a predetermined number of times.

[0154] Furthermore, the determination unit 150 can determine the combination of first syntaxes ranked in the fifth highest as the speech data for classifying the accuracy of Parkinson's disease. Here, the combination of first syntaxes ranked in the fifth highest can correspond to normal as the fourth class, multiple system atrophy as the fifth class, and cerebellar atrophy as the sixth class. Furthermore, the combination of first syntaxes ranked in the fifth highest may be a first syllable generated by combining the first to seventh consonants according to the consonant order with a predetermined vowel, repeated a predetermined number of times.

[0155] Furthermore, the determination unit 150 can determine the combination of first syntaxes ranked in the sixth highest ranking as the speech data for classifying the accuracy of Parkinson's disease. Here, the combination of first syntaxes ranked in the sixth highest ranking can correspond to normal as the fourth class and Parkinson's disease as the fifth class. Furthermore, the combination of first syntaxes ranked in the sixth highest ranking may be a first syllable generated by combining the first to ninth consonants with a predetermined vowel according to the consonant order, repeated a predetermined number of times.

[0156] Furthermore, the determination unit 150 can determine the combination of first syntaxes ranked in the seventh highest ranking as the speech data for classifying the accuracy of Parkinson's disease. Here, the combination of first syntaxes ranked in the seventh highest ranking can correspond to normal as the fourth class, Parkinson's disease as the fifth class, and cerebellar atrophy as the sixth class. Furthermore, the combination of first syntaxes ranked in the seventh highest ranking may be a first syllable generated by combining the first to seventh consonants according to the consonant order with a predetermined vowel, repeated a predetermined number of times.

[0157] Furthermore, the determination unit 150 can determine the combination of first syntaxes ranked in the eighth highest ranking as the speech data for classifying the accuracy for Parkinson's disease. Here, the combination of first syntaxes ranked in the eighth highest ranking can correspond to Parkinson's disease as the fourth class and multiple system atrophy as the fifth class. Furthermore, the combination of first syntaxes ranked in the eighth highest ranking may be a first syllable generated by combining the first to ninth consonants in accordance with the consonant order with a predetermined vowel, the first syllable being repeated a preset number of times.

[0158] Furthermore, the determination unit 150 can determine a combination of first syntaxes ranked in the 9th highest as speech data for classifying accuracy for Parkinson's disease. Here, the combination of first syntaxes ranked in the 9th highest can correspond to normal as the 4th class, Parkinson's disease as the 5th class, and multiple system atrophy as the 6th class. Furthermore, the combination of first syntaxes ranked in the 9th highest may be generated by combining the first to seventh consonants according to the consonant order with a predetermined vowel to generate a first syllable, and repeating the first syllable a preset number of times.

[0159] Furthermore, the determination unit 150 can determine the combination of first syntaxes ranked in the 10th highest as the speech data for classifying the accuracy of Parkinson's disease. Here, the combination of syntaxes ranked in the 10th highest may correspond to Parkinson's disease as the fourth class, multiple system atrophy as the fifth class, and cerebellar atrophy as the sixth class. Furthermore, the combination of syntaxes ranked in the 10th highest may be a first syllable generated by combining the first to seventh consonants in accordance with the consonant order with a predetermined vowel, repeated a predetermined number of times.

[0160] Figure 32 is a diagram showing the accuracy rankings from top 1 to top 10 for different combinations of first syntaxes, different combinations of second syntaxes, and different combinations of third syntaxes, based on the results of accuracy verification of the second Parkinson's disease accuracy classification model.

[0161] 32, the determination unit 150 may determine that [gagagananana...sasasa], which is one of the first ranked syntactic combinations, is the speech data for the second Parkinson's disease accuracy classification. Here, the first ranked [gagagananana...sasasa] may be the syntactic combination that can distinguish between normal (A) as the fourth class and cerebellar atrophy (D) as the fifth class with the highest accuracy.

[0162] Furthermore, the determination unit 150 may determine that [gagagananana...sasasa], which is one of the combinations of the first syntax ranked second, is the speech data for the second Parkinson's disease accuracy classification. Here, [gagagananana...sasasa] ranked second may be the combination of syntax that can distinguish between normal (A) as the fourth class and multiple system atrophy (C) as the fifth class with the highest accuracy.

[0163] Furthermore, the determination unit 150 can determine that [gagagananana...sasasa], which is one of the first syntactic combinations ranked third, is the speech data for the second Parkinson's disease accuracy classification. Here, [gagagananana...sasasa] ranked third can be the syntactic combination that can distinguish, with the highest accuracy, between multiple system atrophy (C) as the fourth class and cerebellar atrophy (D) as the fifth class.

[0164] Furthermore, the determination unit 150 can determine that [gagagananana...sasasa], which is one of the first syntax combinations ranked fourth, is the speech data for the second Parkinson's disease accuracy classification. Here, [gagagananana...sasasa] ranked fourth can be the syntax combination that can distinguish Parkinson's disease (B) as the fourth class from cerebellar atrophy (D) as the fifth class with the highest accuracy.

[0165] Furthermore, the determination unit 150 can determine that [gagagananana...sasasa], which is one of the first syntactic combinations ranked fifth, is the speech data for the second Parkinson's disease accuracy classification. Here, [gagagananana...sasasa] ranked fifth can be the syntactic combination that can distinguish, with the highest accuracy, normal (A) as the fourth class, multiple system atrophy (C) as the fifth class, and cerebellar atrophy (D) as the sixth class.

[0166] Furthermore, the determination unit 150 can determine that [gagagananana...jajaja], which is one of the first syntax combinations ranked sixth, is the speech data for the second Parkinson's disease accuracy classification. Here, [gagagananana...jajaja] ranked sixth can be the syntax combination that can distinguish between normal (A) as the fourth class and Parkinson's disease (B) as the fifth class with the highest accuracy.

[0167] Furthermore, the determination unit 150 can determine that [gagagananana...sasasa], which is one of the first syntactic combinations ranked seventh, is the speech data for the second Parkinson's disease accuracy classification. Here, [gagagananana...sasasa] ranked seventh can be the syntactic combination that can distinguish, with the highest accuracy, normal (A) as the fourth class, Parkinson's disease (B) as the fifth class, and cerebellar atrophy (D) as the sixth class.

[0168] Furthermore, the determination unit 150 can determine that [gagagananana...jajaja], which is one of the first syntactic combinations ranked eighth, is the speech data for the second Parkinson's disease accuracy classification. Here, [gagagananana...jajaja] ranked eighth can be the syntactic combination that can distinguish Parkinson's disease (B) as the fourth class from multiple system atrophy (C) as the fifth class with the highest accuracy.

[0169] Furthermore, the determination unit 150 can determine that [gagagananana...sasasa], which is one of the first syntactic combinations ranked ninth, is the speech data for the second Parkinson's disease accuracy classification. Here, [gagagananana...sasasa] ranked ninth can be the syntactic combination that can distinguish, with the highest accuracy, between normal (A) as the fourth class, Parkinson's disease (B) as the fifth class, and multiple system atrophy (C) as the sixth class.

[0170] Furthermore, the determination unit 150 can determine that [gagagananana...sasasa], which is one of the first syntax combinations ranked 10th, is the speech data for the second Parkinson's disease accuracy classification. Here, [gagagananana...sasasa] ranked 10th can be the syntax combination that can distinguish Parkinson's disease (B) as the fourth class, multiple system atrophy (C) as the fifth class, and cerebellar atrophy (D) as the sixth class with the highest accuracy.

[0171] The acquisition unit 160 inputs the speech result of the speaker corresponding to the voice data determined by the determination unit 150 into the second Parkinson's disease accuracy classification model, and can acquire the Parkinson's disease accuracy classification result for the speaker from the output of the second Parkinson's disease accuracy classification model.

[0172] Figure 33 is a block diagram showing an outline of the configuration of a Parkinson's disease prediction device according to another embodiment. In the following description, parts that overlap with the description of Figures 1 to 32 will be omitted. With reference to Figure 33, a Parkinson's disease prediction device 100 according to another embodiment can include a processor 180 and a memory 190.

[0173] In this embodiment, the processor 180 is capable of processing the functions performed by the collection unit 110, the generation unit 120, the extraction unit 130, the verification unit 140, the determination unit 150, the acquisition unit 160, and the control unit 170 disclosed in FIG.

[0174] Such a processor 180 can control the overall operation of the Parkinson's disease prediction device 100. Here, the term "processor" can refer to, for example, a data processing device built into hardware having a circuit physically structured to execute functions expressed by codes or instructions included in a program. Examples of such data processing devices built into hardware include processing devices such as a microprocessor, a central processing unit, a processor core, a multiprocessor, an ASIC, and an FPGA, but the scope of the present invention is not limited thereto.

[0175] The memory 190 is operatively coupled to the processor 180 and is capable of storing at least one code associated with operations to be performed by the processor 180 .

[0176] Furthermore, memory 190 may temporarily or permanently store data processed by processor 180. Memory 190 may include a magnetic recording medium or flash memory, although the scope of the present invention is not limited thereto. Such memory 190 may include internal and / or external memory, and may include storage devices such as volatile memory such as DRAM, SRAM, or SDRAM; non-volatile memory such as OTPROM, PROM, EPROM, EEPROM, mask ROM, flash ROM, NAND flash memory, or NOR flash memory; flash drives such as SSDs, CF cards, SD cards, Micro-SD cards, Mini-SD cards, xD cards, or Memory Sticks; or HDDs.

[0177] Figure 34 is a flowchart for explaining the Parkinson's disease prediction method according to this embodiment. In the following explanation, parts that overlap with the explanations of Figures 1 to 33 will be omitted. The Parkinson's disease prediction method according to this embodiment will be explained on the assumption that the Parkinson's disease prediction device 100 executes the method using the processor 180 with the help of peripheral components.

[0178] Referring to FIG. 34, in step S3410, processor 180 can extract syntactic combinations from speech data including the results of a speaker's utterance.

[0179] The processor 180 can extract syntactic combinations including a plurality of different first syntactic combinations, a plurality of different second syntactic combinations, or a plurality of different third syntactic combinations. The plurality of different first syntactic combinations can be composed of at least one first syllable combining a predetermined consonant and a predetermined vowel from the audio data. The predetermined consonant included in the plurality of different first syntactic combinations is combined with a predetermined vowel in a consonant order according to a grammar rule, and the predetermined vowel included in the plurality of different first syntactic combinations can be determined as any single vowel. From the audio data, a first syllable is generated by combining an nth consonant to an (n+kth)th consonant according to the consonant order with a predetermined vowel, and the first syllable is repeated a predetermined number of times and extracted as a first syntactic combination, where n includes a natural number and k may include 0 and a natural number. The plurality of different second syntax combinations may be configured by combining a second syllable that combines a basic vowel set, a predetermined diphone, and a predetermined vowel, with one or more of the first syntax combinations. The plurality of different third syntax combinations may include at least one third syllable that combines any consonant and any vowel, with any of the syntax combinations included in the plurality of different second syntax combinations.

[0180] In this embodiment, processor 180 can generate a Parkinson's disease prediction model before extracting syntactic combinations. The Parkinson's disease prediction model in this embodiment can be generated by training a deep neural network model that has been pre-trained to predict Parkinson's disease for a speaker using audio data including the speaker's speech results. Here, the deep neural network model may be a model that uses audio data including the speaker's firing results as input and is trained using a supervised learning method using training data labeled with one of normal, Parkinson's disease, multiple system atrophy, and cerebellar atrophy, which are included in Parkinson's disease-related disorders.

[0181] In step S3420, processor 180 may vary the combination of conditions and syntax for preprocessing the speech data to verify the accuracy of the Parkinson's disease prediction model.

[0182] The processor 180 can execute a first preprocessing condition for preprocessing speech data and a first validation condition for validating the accuracy of the Parkinson's disease prediction model based on a combination of multiple different first syntaxes. In this embodiment, the first preprocessing condition can include conditions for executing a second acoustic preprocessing, performing data padding including edges, repeats, tiles, and zeros, evaluating statistical characteristics including gender, volume, size, maximum amplitude, length, and average syntax length by speaker, splitting the data into a training set, validation set, and test set in a ratio of 64:16:20, performing data imbalance imputation including the SMOTE-NC algorithm, scaling including standardization, removing outliers including isolation forest, and using floating-point formatting including float64.

[0183] Processor 180 can execute second preprocessing conditions for preprocessing speech data and second validation for validating the accuracy of the Parkinson's disease prediction model based on a combination of multiple different second syntaxes. In this embodiment, the second preprocessing conditions can include conditions for executing first acoustic preprocessing, performing data padding including edges, repeats, tiles, and zeros, performing statistical characteristics including gender, volume, size, maximum amplitude, length, and average syntax length by speaker, performing data splitting in which the training set, validation set, and test set are divided into 64:16:20, performing data imbalance imputation including the SMOTE-NC algorithm, performing scaling including standardization, performing outlier removal including isolation forest, and performing floating-point formatting including float64.

[0184] The processor 180 can perform a third preprocessing condition for preprocessing the speech data and a third validation condition for validating the accuracy of the Parkinson's disease prediction model based on a combination of multiple different third syntaxes. In this embodiment, the third preprocessing condition can include conditions for performing a first acoustic preprocessing, performing data padding including edges, repeats, tiles, and zeros, performing statistical characteristics including gender, volume, size, maximum amplitude, length, and average syntax length by speaker, performing data splitting in which the training set, validation set, and test set are divided into 64:16:20, performing data imbalance imputation including the SMOTE-NC algorithm, performing scaling including standardization, performing outlier removal including isolation forest, and performing floating-point formatting including float64.

[0185] Processor 180 can perform a fourth preprocessing condition for preprocessing speech data and a fourth validation condition for validating the accuracy of the Parkinson's disease prediction model based on a combination of multiple different fourth syntaxes. In this embodiment, the fourth preprocessing condition can include conditions for performing a second acoustic preprocessing, performing data padding including edges, repeats, tiles, and zeros, performing statistical characteristics including gender, volume, size, maximum amplitude, length, and average syntax length by speaker, performing data splitting in which the training set, validation set, and test set are divided into 64:16:20, performing data imbalance imputation including the SMOTE-NC algorithm, performing scaling including standardization, performing outlier removal including isolation forest, and performing floating-point formatting including float64.

[0186] The processor 180 can execute a fifth preprocessing condition for preprocessing the speech data and a fifth validation condition for validating the accuracy of the Parkinson's disease prediction model based on a combination of multiple different third syntaxes. In this embodiment, the fifth preprocessing condition can include conditions for executing a first acoustic preprocessing, performing data padding including edges, repeats, tiles, and zeros, performing statistical characteristics including gender, volume, size, maximum amplitude, length, and average syntax length by speaker, performing data splitting in which the training set, validation set, and test set are divided into 64:16:20, performing data imbalance imputation including the SMOTE algorithm and the class weighting algorithm, performing scaling including standardization, performing no outlier removal, and using floating-point formats including float64.

[0187] The processor 180 can execute a sixth preprocessing condition for preprocessing speech data, and a sixth validation condition for validating the accuracy of the Parkinson's disease prediction model based on a combination of multiple different first syntaxes and a combination of multiple different second syntaxes. In this embodiment, the sixth preprocessing condition can include conditions for executing a first acoustic preprocessing, performing data padding including edge padding, performing statistical characteristics including gender, volume, size, maximum amplitude, length, and average syntax length by speaker, performing data splitting in which the training set, validation set, and test set are divided into 64:16:20, performing data imbalance imputation including the SMOTE-NC algorithm, performing scaling including standardization, performing outlier removal including isolation forest, and performing floating-point formatting including float64.

[0188] Processor 180 can execute seventh preprocessing conditions for preprocessing speech data, and seventh validation conditions for validating the accuracy of the Parkinson's disease prediction model based on a combination of multiple different first syntaxes and a combination of multiple different second syntaxes. In this embodiment, the seventh preprocessing conditions can include conditions for executing second acoustic preprocessing, performing data padding including edge padding, performing statistical characteristics including gender, volume, size, maximum amplitude, length, and average syntax length by speaker, performing data splitting in which the training set, validation set, and test set are divided into 64:16:20, performing data imbalance imputation including the SMOTE-NC algorithm, performing scaling including standardization, performing outlier removal including isolation forest, and performing floating-point formatting including float64.

[0189] In step S3430, processor 180 can determine the top-ranked combinations of syntax as speech data for Parkinson's disease prediction based on the results of verifying the accuracy of the Parkinson's disease prediction model.

[0190] The processor 180 can determine the highest ranked combination of first syntaxes based on the result of the first verification of the accuracy of the Parkinson's disease prediction model as the speech data for Parkinson's disease prediction. Here, the highest ranked combination of first syntaxes may be a first syllable generated by combining the first to eleventh consonants according to the consonant order with a predetermined vowel, repeated a preset number of times.

[0191] The processor 180 can determine the top-ranked basic vowel set from the results of the second verification of the accuracy of the Parkinson's disease prediction model as the speech data for Parkinson's disease prediction.

[0192] The processor 180 can determine the highest ranked combination of third syntaxes based on the result of the third verification of the accuracy of the Parkinson's disease prediction model as the speech data for Parkinson's disease prediction. Here, the highest ranked combination of third syntaxes may be a preset number of repetitions of different third syllables generated by combining the third consonant and the eighth consonant with a predetermined vowel according to the consonant order.

[0193] Processor 180 can determine the highest ranked combination of third syntaxes from the result of the fourth verification of the accuracy of the Parkinson's disease prediction model as the speech data for Parkinson's disease prediction. Here, the highest ranked combination of third syntaxes may be a preset number of repetitions of different third syllables generated by combining the eighth and fourteenth consonants according to the consonant order with a predetermined vowel.

[0194] Processor 180 can determine the highest ranked combination of third syntaxes from the result of the fifth verification of the accuracy of the Parkinson's disease prediction model as the speech data for Parkinson's disease prediction. Here, the highest ranked combination of third syntaxes may be a preset number of repetitions of different third syllables generated by combining the second and seventh consonants according to the consonant order with a predetermined vowel.

[0195] The processor 180 can determine the highest ranked combination of first syntaxes based on the results of the sixth verification of the accuracy of the Parkinson's disease prediction model as the speech data for Parkinson's disease prediction. Here, the highest ranked combination of first syntaxes may be a combination of a plurality of different first syllables generated by combining the first to ninth consonants according to the consonant order with a predetermined vowel, repeated a predetermined number of times. The processor 180 can determine the top two ranked combinations of first syntaxes as the speech data for Parkinson's disease prediction. Here, the top two ranked combinations of first syntaxes may be a combination of a plurality of different first syllables generated by combining the first to seventh consonants according to the consonant order with a predetermined vowel, repeated a predetermined number of times. The processor 180 can determine the top three ranked combinations of first syntaxes as the speech data for Parkinson's disease prediction. Here, the first syntax combinations ranked in the top three may be a predetermined number of times repeating a plurality of different first syllables generated by combining the first to sixth consonants according to the consonant order with a predetermined vowel. Processor 180 can determine the first syntax combinations ranked in the top four as the speech data for predicting Parkinson's disease. Here, the first syntax combinations ranked in the top four may be a predetermined number of times repeating a plurality of different first syllables generated by combining the first to fifth consonants according to the consonant order with a predetermined vowel.

[0196] The processor 180 can determine the highest ranked combination of first syntaxes based on the results of the seventh verification of the accuracy of the Parkinson's disease prediction model as the speech data for Parkinson's disease prediction. Here, the highest ranked combination of first syntaxes may be a combination of a plurality of different first syllables generated by combining the first to ninth consonants according to the consonant order with a predetermined vowel, repeated a predetermined number of times. The processor 180 can determine the top two ranked combinations of first syntaxes as the speech data for Parkinson's disease prediction. Here, the top two ranked combinations of first syntaxes may be a combination of a plurality of different first syllables generated by combining the first to seventh consonants according to the consonant order with a predetermined vowel, repeated a predetermined number of times.

[0197] In step S3440, the processor 180 inputs the speaker's speech results corresponding to the voice data for Parkinson's disease prediction into the Parkinson's disease prediction model, and obtains the Parkinson's disease prediction result for the speaker at the output of the Parkinson's disease prediction model.

[0198] In an optional embodiment, processor 180 can generate a gender classification model. The gender classification model in this embodiment can be generated by training a deep neural network model that has been pre-trained to classify the gender of a speaker using audio data including the speaker's speech. Here, the deep neural network model may be a model trained in a supervised learning manner using training data that uses audio data including the speaker's speech and labels the speaker's gender. Processor 180 can verify the accuracy of the gender classification model based on a plurality of different combinations of first syntaxes, a plurality of different combinations of second syntaxes, and a plurality of different combinations of third syntaxes. Based on the results of verifying the accuracy of the gender classification model, processor 180 can determine the highest-ranked combination of third syntaxes as the audio data for gender classification. Here, the highest-ranked combination of third syntaxes may be a predetermined number of repetitions of different third syllables generated by combining the third and eighth consonants with a predetermined vowel according to the consonant order. The processor 180 can input the speech result of the speaker corresponding to the determined voice data into a gender classification model, and obtain a gender classification result for the speaker at the output of the gender classification model.

[0199] In an optional embodiment, processor 180 may generate a syntax performance comparison model. In this embodiment, the syntax performance comparison model may be generated by training a deep neural network model that is pre-trained to output performance comparison results for each syntax included in the speech data using speech data including the speaker's speech results. Here, the deep neural network model may be a model that is trained in a supervised learning manner using training data that uses speech data including the speaker's speech results and labels the performance comparison results for each syntax. Processor 180 may verify the accuracy of the syntax performance comparison model for Parkinson's disease prediction based on a predetermined syntax extracted from the speech data including the speaker's speech results. Processor 180 may determine a highly ranked syntax from the result of verifying the accuracy of the syntax performance comparison model for Parkinson's disease prediction as speech data for use in syntax performance comparison for Parkinson's disease prediction. Processor 180 may determine a first syntax that is ranked first from the result of verifying the accuracy of the syntax performance comparison model for Parkinson's disease prediction as speech data for use in syntax performance comparison for Parkinson's disease prediction. Here, the first syntax ranked first highest may be a first syllable generated by combining the fourth consonant according to the consonant order and a predetermined vowel, repeated a predetermined number of times. Processor 180 can determine the first syntax ranked second highest as speech data for performance comparison of syntaxes for Parkinson's disease prediction. Here, the first syntax ranked second highest may be a first syllable generated by combining the eighth consonant according to the consonant order and a predetermined vowel, repeated a predetermined number of times. Processor 180 can determine the basic vowel set ranked third highest as speech data for performance comparison of syntaxes for Parkinson's disease prediction. Processor 180 can determine the first syntax ranked fourth highest as speech data for performance comparison of syntaxes for Parkinson's disease prediction.Here, the first syntax ranked in the fourth highest ranking may be a first syllable generated by combining the eleventh consonant according to the consonant order and a predetermined vowel, repeated a predetermined number of times. Processor 180 can determine the first syntax ranked in the fifth highest ranking as speech data for performance comparison of syntaxes for Parkinson's disease prediction. Here, the first syntax ranked in the fifth highest ranking may be a first syllable generated by combining the second consonant according to the consonant order and a predetermined vowel, repeated a predetermined number of times. Processor 180 can determine the first syntax ranked in the sixth highest ranking as speech data for performance comparison of syntaxes for Parkinson's disease prediction. Here, the first syntax ranked in the sixth highest ranking may be a first syllable generated by combining the sixth consonant according to the consonant order and a predetermined vowel, repeated a predetermined number of times. Processor 180 can determine the first syntax ranked in the seventh highest ranking as speech data for performance comparison of syntaxes for Parkinson's disease prediction. Here, the first syntax ranked 7th highest may be a first syllable generated by combining the 12th consonant according to the consonant order and a predetermined vowel, repeated a predetermined number of times. The processor 180 can determine the first syntax ranked 8th highest as speech data for performance comparison of syntaxes for Parkinson's disease prediction. Here, the first syntax ranked 8th highest may be a first syllable generated by combining the 14th consonant according to the consonant order and a predetermined vowel, repeated a predetermined number of times. The processor 180 can determine the first syntax ranked 9th highest as speech data for performance comparison of syntaxes for Parkinson's disease prediction. Here, the first syntax ranked 9th highest may be a combination of a first syllable generated by combining the 13th consonant according to the consonant order with a predetermined vowel, a first syllable generated by combining the 12th consonant with a predetermined vowel, and a first syllable generated by combining the 11th syllable with a predetermined vowel. The processor 180 may determine the tenth highest ranked first syntax as the speech data for comparing the performance of syntaxes for predicting Parkinson's disease.Here, the first syntax ranked in the 10th place may be a first syllable generated by combining the fifth consonant according to the consonant order and a predetermined vowel, repeated a predetermined number of times. Processor 180 inputs the speech result of the speaker corresponding to the determined voice data into a syntax performance comparison model, and can obtain the syntax performance comparison result for predicting Parkinson's disease for the speaker from the output of the syntax performance comparison model.

[0200] In any embodiment, the processor 180 can generate a first Parkinson's disease accuracy classification model. The first Parkinson's disease accuracy classification model in this embodiment can be generated by training a deep neural network model pre-trained to classify the accuracy of a speaker's Parkinson's disease using audio data including the speaker's speech. Here, the deep neural network model may be a model trained in a supervised learning manner using audio data including the speaker's speech and training data labeled with one or more of the first and second classes and a third class. The first class in this embodiment may include the disease with the highest accuracy among Parkinson's disease-related diseases: normal, Parkinson's disease, multiple system atrophy, and cerebellar atrophy. The second class may include the disease with the top two accuracy among Parkinson's disease-related diseases: normal, Parkinson's disease, multiple system atrophy, and cerebellar atrophy. The third class may include the remaining Parkinson's disease-related diseases excluding the first or second class. Processor 180 can verify the accuracy of the first Parkinson's disease accuracy classification model based on a plurality of different first syntax combinations, a plurality of different second syntax combinations, and a plurality of different third syntax combinations. Processor 180 can determine the highest-ranked syntax combinations based on the results of verifying the accuracy of the first Parkinson's disease accuracy classification model as speech data for classifying the accuracy of Parkinson's disease. Processor 180 can determine the highest-ranked third syntax combination based on the results of verifying the accuracy of the first Parkinson's disease accuracy classification model as speech data for classifying the accuracy of Parkinson's disease. Here, the highest-ranked third syntax combination may correspond to a third class including the remaining Parkinson's disease-related disorders excluding cerebellar atrophy and cerebellar atrophy as the first class. Furthermore, the highest-ranked third syntax combination may be a predetermined number of repetitions of different third syllables generated by combining the third and eighth consonants according to the consonant order with a predetermined vowel.The processor 180 can determine the top two ranked combinations of third syntax as the speech data for classifying the accuracy of Parkinson's disease. Here, the top two ranked combinations of third syntax may correspond to a third class including the remaining Parkinson's disease-related disorders excluding the first class, normal, and normal. The top two ranked combinations of third syntax may be a predetermined number of repetitions of different third syllables generated by combining the eighth and fourteenth consonants according to the consonant order with a predetermined vowel. The processor 180 can determine the top three ranked combinations of third syntax as the speech data for classifying the accuracy of Parkinson's disease. Here, the top three ranked combinations of third syntax may correspond to a third class including the remaining Parkinson's disease-related disorders excluding the first class, multiple system atrophy, and multiple system atrophy. The top three ranked third syntax combinations may be a combination of a different third syllable generated by combining the second and seventh consonants according to the consonant order with a predetermined vowel, repeated a predetermined number of times. The processor 180 can determine the top four ranked third syntax combinations as speech data for classifying the accuracy of Parkinson's disease. The top four ranked third syntax combinations may correspond to a third class including Parkinson's disease as the first class and Parkinson's disease-related disorders other than Parkinson's disease. The top four ranked third syntax combinations may be a combination of a different third syllable generated by combining the eighth and fourteenth consonants according to the consonant order with a predetermined vowel, repeated a predetermined number of times. The processor 180 can input the speech results of the speaker corresponding to the determined speech data into the first Parkinson's disease accuracy classification model and obtain the first Parkinson's disease accuracy classification result for the speaker from the output of the Parkinson's disease accuracy classification model.

[0201] In any embodiment, the processor 180 can generate a second Parkinson's disease accuracy classification model. The second Parkinson's disease accuracy classification model in this embodiment can be generated by training a deep neural network model that has been pre-trained to classify the accuracy of a speaker's Parkinson's disease using audio data including the speaker's speech. Here, the deep neural network model may be a model trained in a supervised learning manner using audio data including the speaker's speech and training data labeled with two or more of the fourth, fifth, and sixth classes. The fourth class in this embodiment may include the diseases with the highest accuracy among normal, Parkinson's disease, multiple system atrophy, and cerebellar atrophy, which are included in Parkinson's disease-related diseases. The fifth class may include the diseases with higher accuracy among the remaining three Parkinson's disease-related diseases excluding the fourth class. The sixth class may include the diseases with the highest accuracy among the remaining two Parkinson's disease-related diseases excluding the fourth and fifth classes. Processor 180 can verify the accuracy of the second Parkinson's disease accuracy classification model based on a plurality of different first syntax combinations, a plurality of different second syntax combinations, and a plurality of different third syntax combinations. Processor 180 can determine the highest-ranked syntax combinations based on the results of verifying the accuracy of the second Parkinson's disease accuracy classification model as speech data for classifying the accuracy of Parkinson's disease. Based on the results of verifying the accuracy of the second Parkinson's disease accuracy classification model, processor 180 can determine the first highest-ranked first syntax combination as speech data for classifying the accuracy of Parkinson's disease. Here, the first highest-ranked first syntax combination may correspond to normal as the fourth class and cerebellar atrophy as the fifth class. Furthermore, the first highest-ranked first syntax combination may be a first syllable generated by combining the first to seventh consonants with a predetermined vowel according to the consonant order, repeated a predetermined number of times. The processor 180 may determine the second highest ranked combination of the first syntax as the speech data for classifying the accuracy for Parkinson's disease.Here, the combination of first syntax ranked second highest may correspond to normal as the fourth class and multiple system atrophy as the fifth class. Furthermore, the combination of first syntax ranked second highest may be a combination of the first consonant from the first to seventh consonants in the consonant order with a predetermined vowel, the first syllable being repeated a predetermined number of times. The processor 180 can determine the combination of first syntax ranked third highest as the speech data for classifying the accuracy of Parkinson's disease. Here, the combination of first syntax ranked third highest may correspond to multiple system atrophy as the fourth class and cerebellar atrophy as the fifth class. Furthermore, the combination of first syntax ranked third highest may correspond to the first syllable from the first consonant from the seventh consonant in the consonant order with a predetermined vowel, the first syllable being repeated a predetermined number of times. The processor 180 can determine the combination of first syntax ranked fourth highest as the speech data for classifying the accuracy of Parkinson's disease. Here, the first syntax combination ranked fourth may correspond to Parkinson's disease as the fourth class and cerebellar atrophy as the fifth class. Furthermore, the first syntax combination ranked fourth may be a combination of the first consonant to the seventh consonant in the consonant order with a predetermined vowel, the first syllable being repeated a predetermined number of times. The processor 180 may determine the first syntax combination ranked fifth as the speech data for classifying the accuracy of Parkinson's disease. Here, the first syntax combination ranked fifth may correspond to normal as the fourth class, multiple system atrophy as the fifth class, and cerebellar atrophy as the sixth class. Furthermore, the first syntax combination ranked fifth may be a combination of the first consonant to the seventh consonant in the consonant order with a predetermined vowel, the first syllable being repeated a predetermined number of times. The processor 180 may determine the sixth highest ranked first syntax combination as the speech data classifying accuracy for Parkinson's disease.Here, the combination of first syntax ranked in the sixth highest ranking may correspond to normal as the fourth class and Parkinson's disease as the fifth class. Alternatively, the combination of first syntax ranked in the sixth highest ranking may be a combination of the first consonant from the ninth consonant in the consonant order with a predetermined vowel, the combination generating a first syllable repeated a predetermined number of times. The processor 180 may determine the combination of first syntax ranked in the seventh highest ranking as speech data for classifying the accuracy of Parkinson's disease. Here, the combination of first syntax ranked in the seventh highest ranking may correspond to normal as the fourth class, Parkinson's disease as the fifth class, and cerebellar atrophy as the sixth class. Alternatively, the combination of first syntax ranked in the seventh highest ranking may be a combination of the first consonant from the 1st consonant to the 7th consonant in the consonant order with a predetermined vowel, the combination generating a first syllable repeated a predetermined number of times. The processor 180 can determine the combination of first syntaxes ranked in the eighth highest ranking as speech data for classifying the accuracy of Parkinson's disease. Here, the combination of first syntaxes ranked in the eighth highest ranking can correspond to Parkinson's disease as the fourth class and multiple system atrophy as the fifth class. The combination of first syntaxes ranked in the eighth highest ranking can also be a first syllable generated by combining the first to ninth consonants with a predetermined vowel according to the consonant order, repeated a predetermined number of times. The processor 180 can determine the combination of first syntaxes ranked in the ninth highest ranking as speech data for classifying the accuracy of Parkinson's disease. Here, the combination of first syntaxes ranked in the ninth highest ranking can correspond to normal as the fourth class, Parkinson's disease as the fifth class, and multiple system atrophy as the sixth class. The ninth highest ranked first syntax combination may be a first syllable generated by combining the first to seventh consonants with a predetermined vowel according to the consonant order, repeated a predetermined number of times. Processor 180 may determine the tenth highest ranked first syntax combination as the speech data for which accuracy is to be classified for Parkinson's disease.Here, the first syntax combinations ranked in the 10th top ranking may correspond to Parkinson's disease as the fourth class, multiple system atrophy as the fifth class, and cerebellar atrophy as the sixth class. Furthermore, the first syntax combinations ranked in the 10th top ranking may be a first syllable generated by combining the first to seventh consonants with a predetermined vowel according to the consonant order, repeated a predetermined number of times. Processor 180 inputs the speech result of the speaker corresponding to the speech data determined by determination unit 150 into the second Parkinson's disease accuracy classification model, and can obtain the Parkinson's disease accuracy classification result for the speaker from the output of the second Parkinson's disease accuracy classification model.

[0202] The above-described embodiments of the present invention may be implemented in the form of a computer program executable on a computer via various components, and such a computer program may be recorded on a computer-readable medium, which may include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CD-ROMs and DVDs, magneto-optical media such as floptical disks, and hardware devices specially configured to store and execute program instructions, such as ROM, RAM, flash memory, etc.

[0203] On the other hand, the computer programs may be those specially designed and constructed for the purposes of the present invention, or they may be those well known and available to those skilled in the computer software art. Examples of computer programs include not only machine code, such as produced by a compiler, but also high-level language code that can be executed by a computer using an interpreter, etc.

[0204] The use of the term "said" and similar indicators in the present specification (particularly in the claims) can correspond to both the singular and the plural. Furthermore, when a range is stated in the present invention, it is considered to include inventions to which individual values ​​belonging to said range are applied (unless otherwise specified), and is the same as if each individual value constituting said range were stated in the detailed description of the invention.

[0205] Unless explicitly stated or opposed to the order of steps constituting a method according to the present invention, the steps may be performed in any suitable order. The order in which the steps are described is not necessarily intended to limit the present invention. The use of all examples or exemplary terms (e.g., etc.) in the present invention is merely for the purpose of explaining the present invention in detail, and the scope of the present invention is not limited by such examples or exemplary terms unless otherwise limited by the claims. Furthermore, those skilled in the art will recognize that various modifications, combinations, and variations can be made according to design conditions and factors within the scope of the appended claims or their equivalents.

[0206] Therefore, the spirit of the present invention should not be limited to the above-described embodiments, and all scopes equivalent to or modified equivalently from the scope of the claims below, as well as the scope of the spirit of the present invention, will be considered to belong to the scope of the present invention.

Claims

1. 1. A Parkinson's disease prediction method executed by a processor of a Parkinson's disease prediction device, comprising: generating a Parkinson's disease prediction model; A step of extracting a combination of sentences from speech data including a speech result of a speaker; a step of verifying the accuracy of the Parkinson's disease prediction model by changing the conditions for preprocessing the speech data and the combination of the syntax; determining a combination of sentences ranked highly based on the results of verifying the accuracy of the Parkinson's disease prediction model as speech data for predicting Parkinson's disease; inputting a speech result of a speaker corresponding to the speech data for Parkinson's disease prediction into the Parkinson's disease prediction model, and obtaining a Parkinson's disease prediction result of the speaker as an output of the Parkinson's disease prediction model; The condition for preprocessing the audio data includes a condition for performing one or more of acoustic preprocessing, data padding, and floating-point formatting; The Parkinson's disease prediction model comprises: The prediction is generated by training a deep neural network model, which has been pre-trained to predict Parkinson's disease for a speaker, using speech data including the speech results of the speaker; The deep neural network model The model is trained using a supervised learning method with training data that uses speech data containing the speaker's speech as input and labels the model with one of the following Parkinson's disease-related disorders: normal, Parkinson's disease, multiple system atrophy, and cerebellar atrophy. Parkinson's disease prediction methods.

2. The above syntax combinations are: a plurality of different first syntactic combinations each consisting of at least one first syllable combining a predetermined consonant and a predetermined vowel; a combination of a basic vowel set, a second syllable that combines a predetermined diphone and a predetermined vowel, and a plurality of different second constructions that combine at least one of the combinations of the first constructions; at least one third syllable consisting of a combination of any consonant and any vowel; and a combination of a plurality of different third syntaxes, including any combination of syntaxes including the combination of the plurality of different second syntaxes; The method for predicting Parkinson's disease according to claim 1.

3. the predetermined consonants included in the plurality of different first syntactic combinations are combined with the predetermined vowels in a consonant order that conforms to a language regulation; The predetermined vowel included in the combination of the plurality of different first syntaxes is determined as any single vowel. The method for predicting Parkinson's disease according to claim 2.

4. The step of extracting syntactic combinations from the data comprises: a step of generating the first syllable from the speech data by combining the nth consonant to the (n+k)th consonant according to the consonant order with the predetermined vowel, and repeating the first syllable a preset number of times to extract it as a combination of the first syntax, The n includes a natural number, and the k includes 0 and a natural number. The method for predicting Parkinson's disease according to claim 3 .

5. The step of verifying accuracy includes: and verifying accuracy of the Parkinson's disease prediction model based on second preprocessing conditions for preprocessing the speech data and combinations of the plurality of different second syntaxes; The second pretreatment conditions are: a condition for performing first acoustic pre-processing with a uniform channel for recording the audio data and a uniform sampling rate for the audio data; The method for predicting Parkinson's disease according to claim 4.

6. The step of verifying accuracy includes: and verifying accuracy of the Parkinson's disease prediction model based on a seventh preprocessing condition for preprocessing the speech data, the combinations of the plurality of different first syntaxes, and the combinations of the plurality of different second syntaxes; The seventh pretreatment condition is: a second acoustic pre-processing step of unifying channels for recording the audio data, unifying sampling rates of the audio data, normalizing the audio data according to an average volume, filtering the audio data in a preset band, removing DC offset from the audio data, and removing noise from the audio data; and a condition for using edge among types of data padding that satisfies a silent section specific value of the audio data and performs a size equalization process on the audio data. The method for predicting Parkinson's disease according to claim 4.

7. A computer-readable recording medium storing a computer program for executing the method according to any one of claims 1 to 6 using a computer.

8. It is a Parkinson's disease prediction device, a processor; a memory operatively coupled to the processor and storing at least one code to be executed by the processor; The memory, when executed by the processor, Generate a Parkinson's disease prediction model, The processor extracts combinations of sentences from speech data containing the results of a speaker's speech; Verifying the accuracy of the Parkinson's disease prediction model by changing the conditions for preprocessing the speech data and the combination of the syntax; determining, based on the results of verifying the accuracy of the Parkinson's disease prediction model, combinations of sentences ranked highly as speech data for predicting Parkinson's disease; inputting a speech result of a speaker corresponding to the speech data for Parkinson's disease prediction into the Parkinson's disease prediction model, and storing a code for causing the speaker to obtain a Parkinson's disease prediction result of the speaker from the output of the Parkinson's disease prediction model; The condition for preprocessing the audio data includes a condition for performing one or more of acoustic preprocessing, data padding, and floating-point formatting; The memory may include: When generating the Parkinson's disease prediction model, storing a code generated by training a pre-trained deep neural network model to predict Parkinson's disease for a speaker using audio data including the speaker's speech results; The deep neural network model A Parkinson's disease prediction device that is a model trained using a supervised learning method with training data that uses as input voice data including the speaker's speech results and is labeled with one of the following Parkinson's disease-related diseases: normal, Parkinson's disease, multiple system atrophy, and cerebellar atrophy.

Images