Causal Voice Digital Biomarker Analysis Method and Device for Parkinson's Disease

Abstract

The present invention discloses a causal voice digital biomarker analysis method and device for Parkinson's disease, belonging to the field of intelligent medical technology, including: collecting voice data of Parkinson's disease patients and healthy people and performing data preprocessing; extracting features from the voice data to obtain a number of voice features; using a voice iterative causal discovery algorithm to establish a preliminary causal graph between the voice features and Parkinson's disease, introducing a trade-off factor, and obtaining an updated causal graph according to the conditional independence test results between the trade-off factor and the voice features; finding the equivalence classes between the voice features in the updated causal graph, and obtaining a final causal graph according to the equivalence classes; during actual detection, judging whether the voice comes from a Parkinson's disease patient according to the similarity between the voice features in the final causal graph and the voice features in the input voice. The present invention provides more clues and basis for the detection of early Parkinson's disease by constructing a causal graph between voice data and Parkinson's disease.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent medicine, and particularly relates to a causal speech digital biomarker analysis method and device for Parkinson's disease. Background Art

[0002] The importance of early diagnosis of Parkinson's disease cannot be ignored. First of all, early diagnosis can reduce the pain of patients, because Parkinson's disease is often difficult to detect in the early stage of onset. If the disease develops to the late stage, it may cause various complications, such as cognitive impairment, depression, etc., seriously affecting the daily life and working ability of patients. Secondly, early diagnosis can improve the treatment effect, because the selection and implementation of treatment methods need to accurately grasp the development stage and characteristic manifestations of the disease, and early diagnosis provides an important reference for treatment. In addition, early diagnosis can also reduce the waste of medical resources and avoid unnecessary diagnosis and treatment work.

[0003] Common early diagnosis means in modern medicine include medical imaging technology, laboratory testing technology, molecular diagnosis technology, etc., which are methods involving sophisticated instruments or precise detections. Although these methods have high detection accuracy, the equipment cost is expensive, and the participation of professional physicians or testing personnel is required, resulting in relatively high human and material costs.

[0004] In response to the above problems, the patent document with the publication number CN118072946A discloses a device, an electronic device and a storage medium for evaluating the location of brain iron deposition in Parkinson's disease, including: a processor; and a memory storing computer instructions for evaluating the location of brain iron deposition in Parkinson's disease. When the computer instructions are executed by the processor, the device is caused to perform the following operations: obtaining target data related to evaluating the order of brain iron deposition in Parkinson's disease patients and normal subjects, where the target data includes multiple indicators, and the multiple indicators include clinical indicators and quantitative susceptibility mapping (QSM) indicators of the target brain region; constructing a causal relationship graph representing the causal relationship between the clinical indicators and the target brain region based on the target data; determining the deposition order of brain iron in the target brain region based on the relationship links in the causal relationship graph; and evaluating the location of brain iron deposition in different stages of Parkinson's disease according to the deposition order of brain iron in the target brain region.

[0005] This invention evaluates the location of brain iron deposition in different stages of Parkinson's disease by obtaining clinical indicators and quantitative susceptibility mapping (QSM) indicators related to Parkinson's disease and constructing a causal relationship graph. However, on the one hand, the data set including clinical indicators and quantitative susceptibility mapping indicators collected by this invention has obvious complexity. On the other hand, due to the analysis of brain iron deposition requiring a certain period of time, this solution is not very suitable for early detection.

[0006] The patent document with the publication number CN117877534A discloses a Parkinson's disease atlas classification method based on sample augmentation, including: converting the original speech signal into an original spectrogram; establishing a deep convolutional adversarial network and training the deep convolutional adversarial network; inputting the original spectrogram into the trained deep convolutional adversarial network to generate augmented samples; evaluating the augmented samples using evaluation metrics and selecting spectrograms with higher quality; merging the original spectrogram with the selected spectrograms and inputting them into the ConvNeXt model for atlas classification.

[0007] The invention uses a deep convolutional adversarial network (DCGAN) and a self-attention mechanism to generate augmented spectrogram samples to solve the problem of insufficient data in Parkinson's disease detection. However, due to the complexity of Parkinson's disease, the data generated by DCGAN is difficult to comprehensively reflect the features contained in real speech signals, resulting in insufficient accuracy in Parkinson's disease detection. Summary of the Invention

[0008] The purpose of the present invention is to provide a causal speech digital biomarker analysis method and device for Parkinson's disease, and design a speech causal discovery algorithm for describing the causal relationship between speech features and Parkinson's disease, so that early signs of Parkinson's disease can be discovered based on the speech features contained in speech data.

[0009] To achieve the above invention purpose, the technical solutions provided by the present invention are as follows:

[0010] In the first aspect, a causal speech digital biomarker analysis method for Parkinson's disease provided by an embodiment of the present invention includes the following steps:

[0011] Step 1: Use a speech acquisition system to collect speech data of Parkinson's disease patients and healthy people and perform data preprocessing to obtain speech digital biomarkers;

[0012] Step 2: Perform empirical mode decomposition on the speech digital biomarkers to obtain a number of intrinsic mode functions, and use a self-attention mechanism to extract a number of speech features from the number of intrinsic mode functions;

[0013] Step 3: Use a speech iterative causal discovery algorithm to establish a preliminary causal graph between speech features and Parkinson's disease, introduce a trade-off factor for adjusting the correlation degree between speech features, and update the preliminary causal graph according to the conditional independence test results between the trade-off factor and speech features to obtain an updated causal graph;

[0014] Step 4: Find the equivalence classes between speech features in the updated causal graph, restore the updated causal graph to a final causal graph according to the equivalence classes, and perform credibility verification;

[0015] Step 5: During actual detection, based on the similarity between the voice features in the final causal graph and the voice features in the input voice signal, when the similarity is higher than a preset threshold, it is determined whether the voice signal is from a Parkinson's disease patient.

[0016] Further, in Step 1, the voice acquisition system includes an antenna module for signal acquisition, a radio frequency module for signal amplification, filtering, and down-conversion, and a baseband processing module for data interface module configuration.

[0017] Further, in Step 1, the data preprocessing includes:

[0018] Noise reduction of the voice digital biomarker based on spectral subtraction to eliminate background noise;

[0019] Normalization of the volume and spectrum of the voice digital biomarker to normalize the voice data;

[0020] Time alignment of voice digital biomarkers of different lengths to ensure the comparability of the timing between the voice features extracted from the voice digital biomarkers.

[0021] Further, in Step 2, the intrinsic mode function contains the frequency information and timing characteristics of the voice digital biomarker. The intrinsic mode function containing multiple voice features is extracted from the voice digital biomarker through empirical mode decomposition, which facilitates subsequent screening of voice features strongly correlated with Parkinson's disease from multiple voice features by using the self-attention mechanism and ensures the completeness of the voice features.

[0022] Further, in Step 2, the extracted voice features include: audio signal waveform, spectral features, Mel frequency cepstral coefficients, pitch and speech rate, voice energy, zero-crossing rate, time-domain features, and voice segmentation features; and all the extracted voice features are abstracted into timing nodes. When abstracting the voice features into timing nodes and constructing the causal graph, the previous timing node points to the next timing node to obtain a preliminary causal graph.

[0023] Further, in Step 3, the preliminary causal graph is updated according to the trade-off factor and the conditional independence test results between the voice features to obtain an updated causal graph. Specifically:

[0024] In each iteration, the conditional independence test results are weighted by the trade-off factor to obtain an updated edge set, which is expressed by the formula:

[0025] E updated =E initial +α×CI Test Results

[0026] Among them, E updated represents the updated edge set, and E initial represents the edge set of the initial causal graph. α represents the trade-off factor, and CI Test Results represents the conditional independence test results between temporal nodes;

[0027] Reconstruct the causal graph between temporal nodes according to the updated edge set to obtain the updated causal graph.

[0028] Furthermore, in step 4, the updated causal graph is restored to the final causal graph according to the equivalence class, specifically:

[0029] Perform a conditional independence test on the temporal nodes in the updated causal graph to find the equivalence classes in the temporal nodes;

[0030] According to the equivalence class, update the edge set of the updated causal graph to obtain the final edge set, and draw the final causal graph according to the final edge set and all temporal nodes.

[0031] Preferably, according to expert experience, the credibility of the final causal graph is verified, specifically:

[0032] Count the total number of connections between temporal nodes in the final causal graph, denoted as Total Relationships;

[0033] According to expert experience, judge the accuracy of the connections between temporal nodes, and count the number of correct connections, denoted as Correct Relationships;

[0034] The ratio of the number of correct connections to the total number of connections is used as the credibility of the final causal graph, which is expressed by the formula:

[0035]

[0036] Among them, Confidence represents the credibility.

[0037] In a second aspect, to achieve the above invention objective, an embodiment of the present invention further provides a causal voice digital biomarker analysis device for Parkinson's disease, including a data acquisition module, a feature extraction module, a preliminary causal graph construction module, a final causal graph restoration module, and an actual application module;

[0038] The data acquisition module is used to collect voice data of Parkinson's disease patients and healthy people using a voice acquisition system and perform data preprocessing to obtain voice digital biomarkers;

[0039] The feature extraction module is used to perform empirical mode decomposition on the speech digital biomarker to obtain a number of intrinsic mode functions, and a self-attention mechanism is used to extract a number of speech features from the number of intrinsic mode functions;

[0040] The preliminary causal graph construction module is used to use the speech iterative causal discovery algorithm to establish a preliminary causal graph between the speech features and Parkinson's disease, introduce a trade-off factor for adjusting the correlation degree between the speech features, and update the preliminary causal graph according to the conditional independence test results between the trade-off factor and the speech features to obtain an updated causal graph;

[0041] The final causal graph restoration module is used to find the equivalence classes between the speech features in the updated causal graph, and restore the updated causal graph to the final causal graph according to the equivalence classes;

[0042] The actual application module is used for actual detection. According to the similarity between the speech features in the final causal graph and the speech features in the input speech signal, when the similarity is higher than a preset threshold, it is judged whether the speech signal is from a Parkinson's disease patient.

[0043] In a third aspect, to achieve the above-mentioned invention purpose, an embodiment of the present invention also provides a causal speech digital biomarker analysis device for Parkinson's disease, including a memory and a processor. The memory is used to store a computer program, and the processor is used to implement the causal speech digital biomarker analysis method for Parkinson's disease provided by the embodiment of the first aspect of the present invention when executing the computer program.

[0044] In a fourth aspect, to achieve the above-mentioned invention purpose, an embodiment of the present invention also provides a computer-readable storage medium. A computer program is stored on the storage medium, and when the computer program is used by a computer, it implements the causal speech digital biomarker analysis method for Parkinson's disease provided by the embodiment of the first aspect of the present invention.

[0045] The beneficial effects of the present invention are as follows:

[0046] (1) By collecting speech data, the present invention discovers potential early signs of Parkinson's disease based on the speech features in the speech data. The speech data collection is simple, non-invasive, and fast, and does not require complex equipment and long waiting time, making it easy to popularize.

[0047] (2) The present invention uses empirical mode decomposition to decompose a series of intrinsic mode functions containing speech frequency and time series features from speech data, and then uses a self-attention mechanism to screen out speech features with strong correlation with Parkinson's disease from the intrinsic mode functions, ensuring the accuracy and completeness of the speech features extracted from the speech data.

[0048] (3) In view of the causal relationship between the extracted speech features and the signs of Parkinson's disease, the present invention designs a speech iterative causal discovery algorithm specifically for speech data analysis, and constructs a causal graph between speech features and Parkinson's disease based on the speech iterative causal discovery algorithm. Through the causal graph, the corresponding relationship from speech features to Parkinson's disease can be clearly revealed. Moreover, since speech data often contains features related to the early stage of Parkinson's disease, therefore, the causal graph between speech features and the signs of Parkinson's disease constructed by the present invention helps to detect early Parkinson's disease earlier.

[0049] (4) The present invention constructs a complete causal speech digital biomarker analysis device including speech signal acquisition, feature decomposition and extraction, and causal graph construction, which can provide standardized early detection of Parkinson's disease, has strong scalability, can be compatible with different medical devices, and is easy to promote. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a flowchart of a causal speech digital biomarker analysis method for Parkinson's disease provided by an embodiment of the present invention.

[0051] Figure 2 is an effect diagram of empirical mode decomposition of the first speech data provided by an embodiment of the present invention.

[0052] Figure 3 is a preliminary causal graph provided by an embodiment of the present invention.

[0053] Figure 4 is an updated causal graph provided by an embodiment of the present invention.

[0054] Figure 5 is a schematic diagram of equivalence classes provided by an embodiment of the present invention.

[0055] Figure 6 is a final causal graph provided by an embodiment of the present invention.

[0056] Figure 7 is a causal graph after real-time update of speech data provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0057] To make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and do not limit the protection scope of the present invention.

[0058] As Figure 1 shown, the embodiment provides a causal speech digital biomarker analysis method for Parkinson's disease, including the following steps:

[0059] S110. Use a voice acquisition system to collect voice data of Parkinson's disease patients and healthy people, and perform data preprocessing to obtain voice digital biomarkers.

[0060] In this embodiment, to better collect the voice data of two groups of people, namely Parkinson's disease patients and healthy people, as the first voice digital biomarker, a voice acquisition system for voice data collection is designed. By setting the two groups of people to be comparable in demographic characteristics, and to ensure the diversity of the first voice digital biomarker, the voice data of the two groups of people includes different voice environments and accents, so as to ensure that subtle voice changes related to Parkinson's disease can be recognized during subsequent feature extraction.

[0061] Under a unified environment, use the voice acquisition system designed by the present invention to collect voice data. At the same time, the voice acquisition system performs data preprocessing on the collected first voice digital biomarker to reduce noise interference and improve data quality. Specifically as follows:

[0062] The voice acquisition system designed by the present invention includes an antenna module, a radio frequency module, and a baseband signal processing module. Among them, the antenna module is used to collect the voice data of the two groups of people, convert the electromagnetic wave signal into an electrical signal, and obtain the first voice digital biomarker. The radio frequency module is used to perform signal amplification, filtering, and down-conversion processing on the first voice digital biomarker, and convert the medium and high-frequency first voice digital biomarker into a baseband signal. The baseband signal processing module is used to configure the data interface and store the first voice digital biomarker in the DDR cache to ensure the real-time accessibility of the voice data.

[0063] Perform data preprocessing on the voice digital biomarker in the DDR cache. Denoise the voice digital biomarker based on spectral subtraction to eliminate background noise; perform standardization of volume and spectrum on the denoised first voice digital biomarker to achieve voice data normalization. In addition, it is also necessary to perform time alignment on voice digital biomarkers of different lengths to ensure the comparability of the timings between the subsequent extracted voice features.

[0064] Out of privacy and ethical considerations, the voice acquisition system designed by the present invention is also required to follow relevant privacy and ethical guidelines to protect the privacy rights and interests of the subjects.

[0065] S120. Perform empirical mode decomposition on the voice digital biomarker to obtain a number of intrinsic mode functions, and use the self-attention mechanism to extract a number of voice features from the number of intrinsic mode functions.

[0066] Empirical mode decomposition (EMD) is a signal processing technique that can decompose a complex time series signal (here it is a voice signal) into a series of intrinsic mode functions (IMFs), and each IMF represents a component of a different frequency.

[0067] By decomposing the voice digital biomarker into IMFs, the subtle fluctuations and changes in the signal can be better captured, which helps to analyze different frequency characteristics in the voice digital biomarker, including pitch, rhythm, etc., so as to extract voice features from the voice digital biomarker as completely as possible. For example, the irregular fluctuations in pitch or the discontinuity in rhythm may be signs in the voice data of Parkinson's disease patients.

[0068] The process of using EMD to decompose the voice digital biomarker is as follows:

[0069] EMD decomposes the time series data into a series of intrinsic mode functions (IMFs) and a residual trend term. For the voice signal s(t), the decomposition of EMD can be expressed as:

[0070]

[0071] where IMF i (t) is the i-th intrinsic mode function, representing different frequency components of the voice signal, and r(t) is the residual trend, representing the overall trend of the signal.

[0072] As Figure 2 shown, Figure 2 The figure shows the effect of decomposing the voice data of a group of late-stage Parkinson's disease patients into 5 intrinsic mode functions (IMFs) using empirical mode decomposition (EMD). Each IMF i (t) captures specific frequency components in the voice signal, which is very important for understanding features such as the rhythm and pitch changes of the voice. In Parkinson's disease patients, these IMFs may reveal unique change patterns, such as irregular fluctuations in pitch or discontinuity in rhythm.

[0073] Convert the IMFs corresponding to each segment of the voice digital biomarker into query (Q), key (K), and value (V) matrices, and use the self-attention mechanism to extract key features from the first voice digital biomarker. The conversion process is expressed by the formula:

[0074] Q = IMF W Q , K = IMF W K , V = IMF W V

[0075] where W Q , W K , W V are weight matrices used to convert the IMFs into the corresponding Q, K, and V matrices.

[0076] The self-attention mechanism calculates the similarity between the query and the key, and is used to screen out the speech features with high relevance to Parkinson's disease from multiple speech features obtained by empirical mode decomposition. This is achieved through the following formula:

[0077]

[0078] Among them, Attention(Q, K, V) represents applying the self-attention mechanism to Q, K, and V. QK T calculates the dot product between the query and the key to measure the similarity; is the scaling factor used to stabilize the gradient; the softmax function ensures that the sum of all attention scores is 1.

[0079] The present invention adopts empirical mode decomposition combined with the self-attention mechanism. The key speech features extracted from the speech digital biomarkers include:

[0080] ① Audio signal waveform: The model assigns greater attention weights to significant fluctuations or important audio events to focus on the key audio changes in the speech.

[0081] ② Spectral features: The self-attention mechanism highlights the spectral segments related to the speech content, such as formants in the speech, to better capture the spectral information of the speech.

[0082] ③ Mel-frequency cepstral coefficients (MFCC): The model focuses on the changes in specific MFCC coefficients, which may be related to the speaker's speech features or the phonetic attributes of the speech content.

[0083] ④ Pitch and speech rate: The self-attention mechanism adjusts its attention to more comprehensively capture the pitch and speech rate changes in the speech, especially when expressing emotions or emphasizing.

[0084] ⑤ Speech energy: The model assigns more attention to the high-energy segments in the speech or the parts where the speech intensity changes significantly.

[0085] ⑥ Zero-crossing rate: The self-attention mechanism emphasizes the rapid changes or noise events in the speech to capture the zero-crossing rate features in the speech.

[0086] ⑦ Time-domain features: The model focuses on specific time periods in the speech, such as the beginning, middle, or end parts of the speech, to better understand the time-domain structure of the speech.

[0087] ⑧ Speech segmentation features: The self-attention mechanism adjusts between different segments of the speech to capture possible speech transitions or speech events.

[0088] S130. Use the voice iterative causal discovery algorithm to establish a preliminary causal graph between voice features and Parkinson's disease. Introduce a trade-off factor for adjusting the correlation degree between voice features. Update the preliminary causal graph according to the conditional independence test results between the trade-off factor and voice features to obtain the updated causal graph.

[0089] Abstract the voice features collected in S120 into time-series nodes to form an initial node set V. These time-series nodes represent different voice features, such as pitch, speech rate, pronunciation clarity, etc. By introducing an initial edge set E initial .

[0090] The present invention innovatively proposes a voice iterative causal discovery algorithm (Voice Iterative Causal Discovery, VICD) for early Parkinson's disease detection. Use the voice iterative causal discovery algorithm to construct a preliminary causal graph between voice features and Parkinson's disease as Figure 3 shown. The process of constructing the initial causal graph is as follows:

[0091] VICD first constructs a partially connected graph and connects all pairs of time-series nodes with directed edges. As long as k>0, a causal direction that conforms to the time series can be obtained. As shown, the constructed initial causal graph is described by the formula: Figure 2 shown.

[0092] G initial =(V, E initial )

[0093] where E initial is the initial edge set. This initial causal graph provides a basis for subsequent causal analysis, facilitating the subsequent algorithm to accurately capture potential causal relationships related to Parkinson's disease from voice data.

[0094] Furthermore, conduct a conditional independence test on the possible causal relationships between time-series nodes. When the conditional independence is satisfied between two time-series nodes, retain the connection and direction of the time-series nodes in the initial causal graph. When the conditional independence is not satisfied, remove the unnecessary edges. Based on the assumption of time consistency, further remove the homologous edges.

[0095] During the conditional independence test, in this embodiment, the instantaneous conditional independence test (Momentary Conditional Independence Tests, MCI) is specifically used for conditional independence testing:

[0096] Detect time-series node and time-series node When considering conditional independence, for k > 0, perform sequential detection. The detection condition is: That is, determine whether the following independence exists:

[0097]

[0098] Among them, represents the parent nodes of. During the iteration process, gradually refine the initial causal graph by considering conditional independence tests between speech features. This process aims to identify potential causal relationships existing between speech features to more accurately distinguish Parkinson's disease patients from healthy people.

[0099] To more flexibly adapt to the diversity of data, introduce a trade-off factor α for adjusting the association degree between speech features. The introduction of this parameter enables the algorithm to dynamically adjust the causal graph in subsequent iterations to better reflect the potential relationships between speech features.

[0100] By combining the results of conditional independence tests, dynamically adjust the initial edge set E initial , to obtain the updated edge set E updated . The association parameter α here acts as a trade-off factor to ensure that the algorithm remains sensitive to speech digital biomarkers when refining the causal graph. The calculation formula for the updated edge set E updated is:

[0101] E updated = E initial + α × CI Test Results

[0102] Among them, CI Test Results represents the results of conditional independence tests.

[0103] According to the updated edge set E updated , combined with all temporal nodes, draw the updated causal graph, as shown in Figure 4 . By designing the trade-off factor as needed, the VICD algorithm can dynamically adjust the causal graph according to the characteristics of speech digital biomarkers. For example, when there is noise in the speech features, set the trade-off factor to a smaller value to accurately reflect the potential relationships between speech features. In this way, innovation and originality are demonstrated in iterative causal discovery, ensuring the efficiency and credibility of early screening for Parkinson's disease.

[0104] S140, find the equivalence classes between speech features in the updated causal graph, and restore the updated causal graph to the final causal graph according to the equivalence classes.

[0105] The equivalence class is the Markov equivalence class, which means that multiple structures can describe the same conditional probability (Bayesian probability),Figure 5 There are three equivalence classes. Among them, for (a), we have:

[0106] P(X)P(Z|X)P(Y|Z) = P(Z)P(X|Z)P(Y|Z)

[0107] For (b), we have

[0108] P(Y)P(Z|Y)P(X|Z) = P(Z)P(X|Z)P(Y|Z)

[0109] For (c), we have

[0110] P(Z)P(X|Z)P(Y|Z)

[0111] Among them, P(·) represents conditional probability, and X, Y, and Z respectively represent arbitrary nodes. As a schematic illustration of the equivalence class, it is different from the aforementioned timing nodes. For example Figure 5 In the cases of the three equivalence classes shown, these directed acyclic graphs can describe the same conditional independence information, and the main difference lies in the direction of the connected edges. Therefore, according to the idea of the equivalence class, the present invention finds the equivalence classes of all timing nodes in the updated causal graph, that is, finds a suitable edge set, and then draws the final causal graph.

[0112] Specifically, the VICD algorithm performs conditional independence tests on the timing nodes in the updated causal graph, randomly permutes and combines all the timing nodes to obtain several equivalence classes, and finds the unique equivalence class as the final causal graph by judging the conditional independence among all the timing nodes in the equivalence class, as Figure 6 shown. Through the idea of the equivalence class, the problems of confounding factors and selection bias that may exist in the updated causal graph are effectively solved, and the reliability of the final causal graph is improved.

[0113] In addition, the method of the present invention also sets up a cloud server to provide data storage and transmission, so as to have the ability to present real-time voice features, ensuring that the conditional independence and causal relationship among the timing nodes are correctly presented after any iteration, as Figure 7 shown, which is a causal graph obtained after iteration of updated voice data. And the whole process and results are stored in the cloud server, which can ensure the real-time nature and accessibility of the data. This design of presenting real-time voice features makes the method of the present invention applicable to the construction and analysis of causal graphs of dynamically changing voice data.

[0114] Finally, verify the accuracy and usability of the final causal graph:

[0115] Statistical the total number of connected edges among the timing nodes in the final causal graph, denoted as Total Relationships;

[0116] According to expert experience, judge the accuracy of the edges between timing nodes, count the number of correct edges, and denote it as Correct Relationships;

[0117] Use the ratio of the number of correct edges to the total number of edges as the credibility of the final causal graph, which is expressed by the formula:

[0118]

[0119] Among them, Confidence represents the credibility.

[0120] S150. During actual detection, according to the final causal graph and the speech features in the input speech signal, judge whether the speech signal is from a Parkinson's disease patient.

[0121] During actual detection, collect and preprocess the user's speech data through a speech acquisition system, perform speech feature extraction, and compare the distribution and size of the multiple speech features obtained by extraction with the distribution and size of the corresponding timing nodes in the final causal graph. When the similarity of multiple speech features to the timing nodes is higher than the preset threshold, it can be determined that this segment of speech data is from a Parkinson's disease patient.

[0122] Based on the same inventive concept, the embodiment of the present invention also provides a causal speech digital biomarker analysis device for Parkinson's disease, including a data acquisition module, a feature extraction module, a preliminary causal graph construction module, a final causal graph restoration module, and an actual application module;

[0123] Among them, the data acquisition module is used to collect the speech data of Parkinson's disease patients and healthy people through a speech acquisition system and perform data preprocessing to obtain speech digital biomarkers;

[0124] The feature extraction module is used to perform empirical mode decomposition on the speech digital biomarkers to obtain a number of intrinsic mode functions, and use a self-attention mechanism to extract a number of speech features from the number of intrinsic mode functions;

[0125] The preliminary causal graph construction module is used to use a speech iterative causal discovery algorithm to establish a preliminary causal graph between speech features and Parkinson's disease, introduce a trade-off factor for adjusting the correlation degree between speech features, and update the preliminary causal graph according to the trade-off factor and the conditional independence test results between speech features to obtain an updated causal graph;

[0126] The final causal graph restoration module is used to find the equivalence class between speech features in the updated causal graph, and restore the updated causal graph to the final causal graph according to the equivalence class;

[0127] When the actual application module is used for actual detection, according to the similarity between the voice features in the final causal graph and the voice features in the input voice signal, when the similarity is higher than a preset threshold, it is determined whether the voice signal is from a Parkinson's disease patient.

[0128] For the causal voice digital biomarker analysis device for Parkinson's disease provided by the embodiments of the present invention, since it basically corresponds to the method embodiments, the relevant parts can be referred to the partial descriptions of the method embodiments. The device embodiments described above are only illustrative. The units described as separated components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of the present invention. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0129] Based on the same inventive concept, the embodiments also provide a causal voice digital biomarker analysis device for Parkinson's disease, including a memory and a processor. Among them, the memory is used to store a computer program, and the processor is used to implement the above-mentioned causal voice digital biomarker analysis method for Parkinson's disease when executing the computer program.

[0130] The causal voice digital biomarker analysis device for Parkinson's disease proposed by the embodiments of the present invention can be a device such as a computer. The device embodiments can be implemented through software, or through hardware or a combination of software and hardware. Taking software implementation as an example, it is formed by the processor of any device with data processing capabilities reading the corresponding computer program instructions in the non-volatile memory into the memory for operation. From the hardware level, in addition to the processor, memory, network interface, and non-volatile memory, the causal voice digital biomarker analysis device provided by the embodiments of the present invention usually includes other hardware according to the actual functions of the device with data processing capabilities, which will not be elaborated here.

[0131] Based on the same inventive concept, the embodiments also provide a computer-readable storage medium, on which a computer program is stored. When the computer program is used by a computer, it implements the above-mentioned causal voice digital biomarker analysis method for Parkinson's disease.

[0132] The computer-readable storage medium may be an internal storage unit of any device with data processing capabilities described in any of the foregoing embodiments, such as a hard disk or a memory. Further, the computer-readable storage medium may include both an internal storage unit of any device with data processing capabilities and an external storage device. The computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capabilities, and may also be used to temporarily store data that has been output or is to be output.

[0133] It should be noted that the causal speech digital biomarker analysis device for Parkinson's disease, the causal speech digital biomarker analysis device for Parkinson's disease, and the computer-readable storage medium provided in the foregoing embodiments all belong to the same concept as the embodiments of the causal speech digital biomarker analysis method for Parkinson's disease. For the specific implementation process, please refer to the embodiments of the causal speech digital biomarker analysis method for Parkinson's disease, which will not be elaborated here.

[0134] The above is only a preferred embodiment of the present invention and does not impose any formal restrictions on the present invention. Although the implementation process of the present invention has been described in detail above, those familiar with the art can still modify the technical solutions described in the foregoing examples or make equivalent replacements for some of the technical features. Any modifications, equivalent replacements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. A causal speech digital biomarker analysis method for Parkinson's disease, characterized in that It includes the following steps: Step 1: Use a voice acquisition system to collect voice data of Parkinson's disease patients and healthy people and perform data preprocessing to obtain voice digital biomarkers; Step 2: Perform empirical mode decomposition on the voice digital biomarkers to obtain a number of intrinsic mode functions, and use a self-attention mechanism to extract a number of voice features from the number of intrinsic mode functions; Step 3: Use a voice iterative causal discovery algorithm to establish a preliminary causal graph between voice features and Parkinson's disease, introduce a trade-off factor for adjusting the correlation degree between voice features, and update the preliminary causal graph according to the conditional independence test results between the trade-off factor and voice features to obtain an updated causal graph; Step 4: Find the equivalence classes between voice features in the updated causal graph, restore the updated causal graph to a final causal graph according to the equivalence classes, and perform credibility verification; Step 5: During actual detection, according to the similarity between the voice features in the final causal graph and the voice features in the input voice signal, determine whether the voice signal is from a Parkinson's disease patient when the similarity is higher than a preset threshold.

2. The causal voice digital biomarker analysis method for Parkinson's disease according to claim 1, wherein In Step 1, the voice acquisition system includes an antenna module, a radio frequency module, and a baseband processing module.

3. The causal voice digital biomarker analysis method for Parkinson's disease according to claim 1, characterized in that In Step 1, the data preprocessing includes: Noise reduction of the voice digital biomarkers based on spectral subtraction; Standardization of the volume and spectrum of the voice digital biomarkers; Time alignment of voice digital biomarkers of different lengths.

4. The causal speech digital biomarker analysis method for Parkinson's disease according to claim 1, wherein In Step 2, the intrinsic mode functions contain the frequency information and time series characteristics of the voice digital biomarkers.

5. The causal speech digital biomarker analysis method for Parkinson's disease according to claim 1, wherein In Step 2, the extracted voice features include: audio signal waveform, spectral features, Mel frequency cepstral coefficients, pitch and speech rate, voice energy, zero-crossing rate, time domain features, and voice segmentation features; and all the extracted voice features are abstracted into time series nodes.

6. The causal speech digital biomarker analysis method for Parkinson's disease according to claim 5, wherein In Step 3, the updating of the preliminary causal graph according to the conditional independence test results between the trade-off factor and voice features to obtain an updated causal graph is specifically: In each iteration, use the trade-off factor to weight the conditional independence test results to obtain an updated edge set, which is expressed by the formula: E updated = E initial + α × CITestResults Among them, E updated represents the updated edge set, and E initial represents the edge set of the initial causal graph, α represents the trade-off factor, and CITestResults represents the conditional independence test results between temporal nodes; Reconstruct the causal graph between time series nodes according to the updated edge set to obtain an updated causal graph.

7. The causal voice digital biomarker analysis method for Parkinson's disease according to claim 6, wherein In Step 4, restoring the updated causal graph to a final causal graph according to the equivalence classes is specifically: Perform a conditional independence test on the time series nodes in the updated causal graph to find the equivalence classes in the time series nodes; According to the equivalence classes, update the edge set of the updated causal graph to obtain a final edge set, and draw a final causal graph according to the final edge set and all the time series nodes.

8. A causal voice digital biomarker analysis device for Parkinson's disease, characterized in that, It includes a data acquisition module, a feature extraction module, a preliminary causal graph construction module, a final causal graph restoration module, and an actual application module; The data acquisition module is used to collect voice data of Parkinson's disease patients and healthy people by using a voice acquisition system and perform data preprocessing to obtain voice digital biomarkers; The feature extraction module is used to perform empirical mode decomposition on the voice digital biomarker to obtain a number of intrinsic mode functions, and uses a self-attention mechanism to extract a number of voice features from the number of intrinsic mode functions; The preliminary causal graph construction module is used to establish a preliminary causal graph between voice features and Parkinson's disease by using the voice iterative causal discovery algorithm, introduce a trade-off factor for adjusting the correlation degree between voice features, and update the preliminary causal graph according to the trade-off factor and the conditional independence test result between voice features to obtain an updated causal graph; The final causal graph restoration module is used to find the equivalent classes between voice features in the updated causal graph, restore the updated causal graph to the final causal graph according to the equivalent classes, and perform credibility verification; When actually detecting, the actual application module is used to judge whether the voice signal comes from a Parkinson's disease patient according to the similarity between the voice features in the final causal graph and the voice features in the input voice signal when the similarity is higher than a preset threshold.

9. A causal speech digital biomarker analysis device for Parkinson's disease, comprising a memory and a processor, wherein the memory is used for storing computer programs, characterized in that, When executing the computer program, the processor is used to implement the causal voice digital biomarker analysis method for Parkinson's disease according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When using a computer, the computer program is used to implement the causal voice digital biomarker analysis method for Parkinson's disease according to any one of claims 1-7.

Images