System and method for detecting cognitive decline using speech analysis

Abstract

A system and method for detecting cognitive decline in a subject using a classification system for detecting cognitive decline in a subject based on speech samples.SOLUTION: The classification system is trained using speech data corresponding to audio recordings of speech from normal and cognitively impaired patients to generate an ensemble classifier that includes a plurality of component classifiers and an ensemble module. Each of the plurality of component classifiers is a machine learning classifier configured to generate a component output that identifies the sample data as corresponding to a normal patient or a cognitive patient. The machine learning classifier is generated based on a subset of the available features. The ensemble module receives the component outputs from all of the component classifiers and generates an ensemble output that identifies the sample data as corresponding to normal or cognitively impaired patients based on the component outputs.SELECTED DRAWING: None

Description

[Technical Field]

[0001] (Priority Claim) This application is a part of the "System and Method" filed on April 15, 2019. for Predicting Cognitive Decline" Priority is claimed to U.S. Provisional Patent Application No. 62 / 834,170, and the entire contents of this U.S. Provisional Patent Application are hereby incorporated by reference. The contents of which are incorporated herein by reference. [Background technology]

[0002] Mild cognitive impairment (MCI) is a slight but noticeable impairment of cognitive abilities, including memory and thinking skills. The changes caused by MCI are felt daily. People with MCI who do not meet the diagnostic criteria for dementia may have symptoms that are not severe enough to affect their daily lives. However, in individuals with MCI, Alzheimer's disease (AD) ) or other forms of dementia. Early intervention can improve outcomes. This can provide better prospects for

[0003] Episodic memory is the memory of an event or "episode." Contains a directional (newly encountered information) or retrograde (past events) component. Episodic memory decline occurs earliest in patients with preclinical / prodromal AD and persists throughout the course of the disease. Assessment of verbal episodic memory decline in MCI may indicate early cognitive changes. and as a screening tool for timely detection of early / presymptomatic AD and initiation of treatment. It can be used. Summary of the Invention [Means for solving the problem]

[0004] An exemplary embodiment of the present invention relates to a method for detecting cognitive decline in a subject. The method includes measuring speech of the subject in response to a first set of instructions delivered to the subject. obtaining subject baseline speech data corresponding to a plurality of audio recordings; Further audio recordings of the subject's speech in response to a second set of instructions delivered to and acquiring corresponding subject-based speech data. A step of extracting multiple features from the line speech data and the subject test speech data. and normalizing the subject test speech data using the subject baseline speech data. and generating subject test data by using the trained algorithm. The method further includes analyzing the subject test data using the sample classifier. The ensemble classifier includes multiple component classifiers and an ensemble module. Each of the multiple component classifiers classifies the subject test data into normal patients or patients with cognitive decline. Each component is configured to generate a corresponding component output that identifies it as a The subject classifier is configured to analyze a subset of features selected from the plurality of features. The ensemble module receives the component outputs from the component classifiers. Based on the component outputs, the subject test data is then assigned to correspond to normal or cognitively impaired patients. The multiple component classification is configured to generate an ensemble output that identifies the components. The device measures the speech antecedents of normal and cognitively impaired groups in response to the first set of instructions. Training baseline speech data corresponding to the audio recording, as well as responses to a second set of instructions Training tests corresponding to previous audio recordings of speech from groups of normal and cognitively impaired patients. It is trained using speech data.

[0005] An apparatus for detecting cognitive decline in a subject is provided, the apparatus generating an audio output. and an audio output facility configured to receive the audio signal and accommodate recording of the audio signal. a voice input facility configured to generate data, and a display; The device includes a processor and a non-transitory memory device including an instruction set executable by the processor. and a computer-readable storage medium for transmitting the first set of instructions to the subject. directing the audio output facility to audibly provide a plurality of audio signals in response to the first set of instructions; Subject baseline speech data corresponding to multiple audio recordings of the subject's speech from a voice input facility and audibly provide a second set of instructions to the subject. and directing the audio output facility to further record the subject's speech in response to the second set of instructions. The subject test speech data corresponding to the voice recording is received from the voice input equipment, and the subject base Extract multiple features from the line speech data and the subject test speech data, By normalizing the subject test speech data using the baseline speech data, Generate subject test data using the trained ensemble classifier. Analyzes the data and generates an output indicating whether the subject is likely to have cognitive decline and operates to direct the display to provide a visual representation of the output to the user. The apparatus may include a memory configured to store the trained ensemble classifier. The trained ensemble classifier comprises a plurality of component classifiers and and an ensemble module. Each of the plurality of component classifiers is a subject test. Generate component outputs that identify the data as corresponding to a normal patient or a patient with cognitive decline. Each component classifier is configured to select a subset of features from a plurality of features. The ensemble module is configured to analyze a subset of the component classifications. receiving component outputs from the device and generating subject test data based on the component outputs; The method is configured to generate an ensemble output that identifies the patients as corresponding to normal or cognitively impaired patients. The plurality of component classifiers are configured to classify normal and cognitively impaired responses to the first set of instructions. A training baseline speech data set corresponding to the previous audio recordings of the speech of the lower patient group was used. and speech patterns of a group of normal and cognitively impaired patients in response to a second set of commands. It is trained using training test speech data corresponding to previous speech recordings.

[0006] In another exemplary embodiment, a computer-implemented method for training a classification system is provided. The classification system assesses cognitive decline in a subject by evaluating the subject's speech. The method is configured to detect the presence of leukemia cells in a sample based on the normal and cognitively impaired patients. Obtain training baseline speech data and training test speech data from the loop The training baseline speech data includes normal and non-normal speech in response to a first set of instructions. The training test speech data corresponded to audio recordings of speeches from a group of patients with cognitive decline. , audio recordings of the speech of a group of normal and cognitively impaired patients in response to the second set of instructions. Furthermore, the method includes: (i) training baseline speech data; and (ii) ) extracting a plurality of features from the training test speech data. An ensemble classifier is generated that includes the component classifiers and the ensemble module. Each of the plurality of component classifiers further comprises generating a normalized version of the sample data. and configured to generate component outputs that identify the patient as corresponding to the cognitive impairment patient. Each component classifier classifies a subset of features selected from multiple features. The ensemble module is configured to analyze the components from the component classifiers. and receives the component output and classifies the sample data into normal or The method is configured to generate an ensemble output that identifies the cognitive impairment as corresponding to a patient with cognitive decline. The method normalizes the training test speech data using the training baseline speech data. generating a training dataset by generating a training data set; and further training an ensemble classifier using the resulting data.

[0007] A system for training a classification system is also provided. The classification system comprises: The system is configured to detect cognitive decline based on speech samples from the subject. The training baseline speech data and training a database configured to store test speech data for training; Line speech data were collected from the normal and cognitively impaired groups in response to the first set of instructions. and the training test speech data corresponds to an audio recording of the speech in response to a second set of instructions. The system responds to audio recordings of speech from a group of normal and cognitively impaired patients. The computing device further comprises a computing device operably connected to communicate with the database. a processor and a non-transitory computer containing a set of instructions executable by the processor. and a computer-readable storage medium. The instruction set includes training baseline speech data and and training test speech data from the database, (i) a training baseline speech data set; (ii) extracting features from the speech data and (iii) the training test speech data, and Generate an ensemble classifier comprising component classifiers and an ensemble module. Each of the plurality of component classifiers can operate as follows: configured to generate component outputs that are identified as corresponding to a normal patient or a patient with cognitive decline. Each component classifier is constructed using a subset of features selected from the plurality of features. The ensemble module is configured to analyze the components from the component classifiers. It receives the component output and converts the sample data into normal patient data based on the component output. or to generate an ensemble output that identifies the plurality of subjects as corresponding to a cognitive decline patient. The instruction set uses the training baseline speech data to train the test speech data. Generate a training dataset by normalizing The system is further operable to train an ensemble classifier. The system further comprises a memory configured to store the ensemble classifier.

[0008] These and other aspects of the invention are described in the following claims, including the drawings and accompanying claims. This will be apparent to those skilled in the art after reading this detailed description. [Brief explanation of the drawings]

[0009] [Figure 1] 1 illustrates a system for training a classification system for detecting cognitive decline based on speech samples of subjects according to an exemplary embodiment of the present application. [Figure 2] 1 illustrates a method for training a classification system for detecting cognitive decline based on speech samples of subjects according to an exemplary embodiment of the present application. [Figure 3] 1 illustrates an ensemble classifier comprising multiple component classifiers and an ensemble module according to an exemplary embodiment of the present application. [Figure 4] 4 illustrates a method for independently selecting a subset of speech features for each component classifier of the example ensemble classifier of FIG. 3. [Figure 5] 1 illustrates an apparatus for detecting cognitive decline based on speech samples of a subject according to an exemplary embodiment of the present application. [Figure 6] 1 illustrates a method for detecting cognitive decline based on a speech sample of a subject according to an exemplary embodiment of the present application. [Figure 7] 1 illustrates an exemplary system including an ensemble classifier for detecting cognitive decline based on speech samples of subjects according to Example I of the present application. [Figure 8] 1 shows data corresponding to the average number of words recalled across different steps in an exemplary embodiment of the Rey Auditory Verbal Learning Test (RAVLT) test according to Example I of the present application. DETAILED DESCRIPTION OF THE INVENTION

[0010] This application describes a method for testing a Word List Recall (WLR) test using a computer-implemented method. Data corresponding to speech samples of subjects or patients obtained from neuropsychological tests such as and for detecting and / or predicting cognitive decline, particularly MCI, by analyzing the data. Neuropsychological testing involves screening a subject's cognitive abilities, such as memory. Furthermore, this application is based on the WLR speech samples of the subjects. Based on this, cognitive decline (e.g., MCI) can be detected, and / or cognitive decline or dementia (e.g., For example, a system for training a classification system configured to predict the onset of AD. This includes systems and methods.

[0011] Figure 1 shows a method for detecting and / or assessing cognitive decline, particularly MCI, based on speech samples from subjects. An exemplary embodiment of a system 100 for training a classification system for predicting The system 100 collects data corresponding to audio recordings of previously conducted WLR tests. The system includes a database 110 for storing various types of data, including , database 110, as further described below, contains training baseline speech Data 112 and training test speech data 1, each acquired under different experimental conditions. 14. The database 110 includes one or more non-transitory computers. The program may be stored on a computer-readable storage medium.

[0012] The database 110 may perform some or all of the data stored in the database 110. The computing device 120 may provide the data stored in the database 110 to the computing device 120, or the computing device 120 may use the data stored in the database 110. operatively connected to the computing device 120 to allow retrieval of some or all of the data. As shown in FIG. 1, the database 110 may be connected to a communication network 140. (e.g., the Internet, wide area networks, local area networks, cellular network, etc.) to the computing device 120. It is also possible to connect the database 110 directly to the computing device 120 via a wired connection. In this embodiment, the computing device 120 includes a processor 122 and a computer data-accessible medium 124 and a computing device 120 to receive data and / or instructions; and / or input / output for transmitting data and / or instructions from the computing device 120. and a power supply 126. The processor 122 may include, for example, one or more microprocessors. may be included and stored in a computer-accessible medium 124 (e.g., a memory storage device). The computer-accessible medium 124 may include, for example, It may be a non-transitory computer-accessible medium having executable instructions therein. The system 100 generates and trains an ensemble classifier 300. A memory storage device provided separately from the computer-accessible medium 124 for storing The computing device 120 may further include a memory storage device 130. The memory storage device 130 is part of the computing device 120. or may be external to computing device 120 and operably connected to computing device 120. The memory storage device 130 may be configured to detect and / or predict cognitive decline in a subject. In another embodiment, the processor may be connected to a separate computing device (not shown) for The ensemble classifier 300 is another method for detecting and / or predicting cognitive decline in a subject. The data may be stored in a separate memory storage device (not shown) connected to the individual computing device (not shown). good.

[0013] FIG. 2 shows a method for detecting and / or assessing cognitive decline in a subject based on speech samples of the subject. 2 illustrates an exemplary embodiment of a method 200 for training a classification system for prediction. In particular, the method 200 generates speech samples (e.g., WLR speech samples) The samples were analyzed to determine whether they were from normal patients or patients with cognitive decline (e.g., MCI patients). Generate and train an ensemble classifier 300 to determine whether a sample is correlated. The WLR speech sample is used to assess verbal episodic memory, for example, using the RAVLT. The data used corresponds to previously recorded audio files from various WLR tests, e.g. The data can be obtained from the database 110 that can store the data. is an examiner-administered test that can be used to measure verbal episodic memory. This is a tool based on word lists that correlate with cognitive ability, such as learning speed, short-term and slow-term learning. Extended verbal memory, recall ability after interference, recognition memory, and learning patterns (serial position effect) This can be used to detect and / or generate scores relating to any verbal memory.

[0014] In step 202, the computing device 110 selects normal and cognitive A training base corresponding to multiple sets of prior speech recordings of a group of speech deafness patients. Receive multiple sets of line speech data 112. Each set of data 112 is used for listening to a word list and for playing and speaking the same word list. Corresponding speech of a group of normal and cognitively impaired patients responding to the same set of instructions For example, the training baseline speech data 112 may include a first word list, a second word list, a third word list, a fourth word list, a fifth word list, a sixth word list, a sixth word list, a seventh word list, a seventh word list, a eighth ... and normal and normal responses in response to a first set of instructions regarding the reproduction and speech of the first word list immediately thereafter. In one particular embodiment, the training base corresponds to the speech of a group of people with cognitive decline. The baseline speech data 112 is a learning test for a WLR test, such as the RAVLT test. The data may include data corresponding to audio recordings of groups of normal and cognitively impaired patients from Step 202 identifies feature discriminant signals contained within the training test speech data 114. and / or averages that can be compared with the training test speech data 114 to enhance To establish baseline characteristics, any of the training baseline speech data 112 A suitable number of sets of training baseline speech The data 112 may include learning phase data from any suitable WLR test. Baseline speech data 112, particularly learning phase data, is used for normal and cognitively impaired Training test speech data to generate a quantitative representation of cognitive load for a group of patients. 4 serves as a baseline characteristic against which the For training, at least three sets of baseline speech data 112, at least five sets of baseline speech data 112, One set, or at least ten sets, may be used. In

[0043] , five sets of training baseline speech data 112 are used.

[0015] Additionally, the computing device 110 retrieves a set of training test speech data from the database 110. The training test speech data set 114 is further described below. As can be seen, the method 200 generates training data for training the ensemble classifier 300. To generate the dataset 330 and / or to generate the ensemble classifier 300 The training test speech data 114 is used for training baseline speech data. 112 responded to at least one different set of instructions for reproduction and speech Corresponding to previous audio recordings of speech from groups of normal and cognitively impaired patients. The instruction set involves having the patient listen to a different word list and then immediately repeating the different word list. As another example, different sets of instructions can be For the patient, the same word list was used for the training baseline speech data 112. Listen to the same word list, but after the distractor task and / or for a period of time (e.g., at least or at least about 10 minutes, or at least about 20 minutes, or at least about 30 minutes After a delay of approximately 10 minutes (or approximately 30 minutes), the child can be asked to play and speak the word list. In one particular embodiment, the training test speech data 112 is generated using, for example, RAVLT. Normal and post-disturbance tests and / or time delay tests of WLR tests such as The data may include data corresponding to audio recordings of speech from a group of patients with cognitive decline.

[0016] In one particular embodiment, cognitive decline in a subject is assessed based on speech samples of the subject. The method 200 for training a classification system for detecting and / or predicting based on normal Use clinical test data obtained from patients and MCI patients as part of the RAVLT trial It can be reproduced through the different steps of an exemplary embodiment of the RAVLT test. The data corresponding to the average number of words found is shown in Figure 8. As shown in Figure 8, The VLT test consists of several different tests (e.g., 802) that are part of the learning phase of the RAVLT test. For example, tests I to V (811 to 815) may be included. The computing device 110 uses the RAVLT test as training baseline speech data 112. Speech data can be acquired as part of the training phase 802 of FIG. 8 shows that the RAVLT test is a different type of regeneration test, namely, the disturbance test B 822m after disturbance test The processing unit 110 includes a stage 824 and a 20-minute delayed playback 826. The training test speech data 114 for the step 202 is Speech data collected at any one of them can be retrieved.

[0017] In step 204, the computing device 110 receives the speech from the database 110. a set of training baseline speech data 112; and (ii) a plurality of speech samples from each of the sets 114 of training and testing speech data. Specifically, the computing device 110 analyzes and extracts the background features from each of the data sets. The computing device 110 extracts a set of speech features of the same type from the speech data. Speech features from the dataset are collected as utterances, such as mean and standard deviation data values ​​corresponding to acoustic characteristics. Extracting acoustic features of any suitable type for analyzing audio recordings of spoken speech. For example, speech features can be speech characteristics listed and defined in Table 1 below. In some embodiments, the recording may include one or more exemplary acoustic characteristics. In this case, speech features are the sum of these exemplary acoustic features across frames of the speech recording. It may include all or a subset of the mean and / or standard deviation data values ​​corresponding to the characteristic. do.

[0018] [Table 1]

[0019] The computing device 120 may extract any suitable number of speech features from each speech data set. As the number of speech features increases, the system and method of the present application It can improve predictive and / or analytical performance, but may be computationally intensive. Therefore, to balance predictive and / or analytical performance with computational efficiency, an appropriate number In some embodiments, the computing device 120 At least 24, at least 30, at least As many as 50, or at least 100, different speech features can be extracted. In an embodiment, the computing device 120 selects 5 to 150 speech sequences from each speech data set. Extract speech features, 10-100 speech features, 12-50 speech features It is possible.

[0020] In step 206, the computing device 120 receives the training data from the database 110. A set of test speech data 114 and the speech data obtained in the previous step (step 204) Based on the relevant speech features of the training test speech data 114, the answer As shown in FIG. 3, the ensemble classifier 300 is It comprises a plurality of component classifiers 310 and an ensemble module 320. Each of the component classifiers 310 classifies the sample data into samples of normal patients or cognitively impaired patients. machine learning configured to generate component outputs that identify the data as corresponding to the The component classifier 310 is a classifier that uses data obtained from normal and MCI patients. When trained using the data, each of the component classifiers 310 classifies the sample data Generate component outputs that identify the sample data as corresponding to normal or MCI patients. More specifically, a component classifier is a machine learning classifier configured to generate a Each of the components 310 is a support vector machine (SVM). Each of the classifiers 310 may be implemented using, for example, a logistic regression module, a fast gradient boosting module, or a module, a random forest module, or a naive Bayes module. It is contemplated that a supervised learning module such as the component classifier 310 may be used. Each of the speech patterns is generated by the computing device 120 and is used to compare the speech patterns of the training test speech data 114. Analyze a downsampled subset of the track features. For 0, the subset of speech features is shown in Figure 4 and described further below. 400. The ensemble classifier 300 may be any suitable It can include a number N of component classifiers 310. For example, an ensemble classifier 3 00 is at least 10, at least 20, at least 30, or at least 50 In one particular embodiment, the component classifier 310 may include The sample classifier 300 includes 30 component classifiers 310 .

[0021] The ensemble module 320 extracts components from all the component classifiers 310. and receives the component output and classifies the sample data as normal or MC based on the component output. The system is configured to generate an ensemble output 340 that identifies the patient as corresponding to the patient. The sample module 320 samples the components provided by each component classifier 310. Any suitable method for determining the ensemble output 340 based on the component outputs. For example, the ensemble module 320 can use Bagging, which takes into account the component outputs from each classifier 310 with equal weighting. Alternatively, an aggregating method can be used. Alternatively, an ensemble module 320 is a component, such as adaptive boosting or gradient boosting. Other methods can be used in which the output signals are given different weights.

[0022] FIG. 4 shows how the subset of speech features for each component classifier 310 can be independently 4 illustrates an exemplary embodiment of a method 400 for selecting a target. Step 206 selects the desired speech features to generate each component classifier 310. Each time a downsampled subset of σ is selected, the method 400 is repeated. , the computing device 120 calculates the speech features for generating the first component classifier 311. 4. Use method 400 to select the desired downsampled subset of The speech data 114 is analyzed to generate the following component classifiers 312: Repeat the analysis to select another downsampled subset of peach features and perform the calculation The device 120 selects a downsampled subset of the speech features to generate the Nth component. The subsequent component classifiers 313 to 317 are then used to generate the component classifier 317. The analysis is repeated to generate 7. The method 400 generates By selecting different downsampled subsets of speech features for It models only a subset of all available speech features, but overall Component classifiers 310 that provide sampling across a larger number of speech features This allows the computing device 120 to generate an ensemble classifier 300 having the following features: This structure of the sample classifier 300 allows each component classifier 310 to By limiting the analysis to only a subset of the speech features selected for the component classifier, This allows for prediction and classification that incorporates a larger number of speech features while providing improved computational efficiency. and / or provide modules with improved analytical performance.

[0023] In step 402, the computing device 120 analyzes the training test speech data 114. to obtain a subsample of the training test speech data 114. As such, this subsample is a component generated for a smaller number of speech features. To identify the desired parameters and features for the classifier 310, the computing device 120 The subsamples are used for the first set of training test speech data 114 from normal patients. 1 number of samples and a second number of training test speech data 114 from cognitive decline patients. The first number of samples includes training test speech data 1 from normal patients. The second number of samples may be randomly selected from the entries in the 14. The speech samples may be randomly selected from the entries in the training test speech data 114 from the patient. In step 402, two different classifications of patients (i.e., normal patients and cognitively impaired patients) are selected. The second number of samples was compared to the first number of samples to provide a substantial balance between the number of patients. At least 80% of the samples, at least 90% of the first number of samples, or at least 90% of the first number of samples Preferably, the subsample is at least 95% of the first number of samples. The ratio of the second number to the sample is balanced to be 1:1. By subsampling, the remainder of the method 400 is performed to separate two classes: normal patients and This allows for evenly balanced data across patients with cognitive decline. , the training test speech data 114 may contain more normal subjects than cognitively impaired subjects. and therefore may result in unbalanced data between the two classifications. Balanced subsampling is done towards the larger category, i.e., normal patients. Biased classifiers (leading to false negatives in the classifier, which are not recognized as corresponding to patients with cognitive decline) (which may result in missing patients who should be identified) This helps to address any imbalance in classification that may exist in the data set. The classifier 300 may be selected using substantially equal or equal proportions of normal and cognitively impaired patients. Several individual component classifiers for the subset of speech features selected. 0, thus resulting in a balanced classifier, which can be used to compare different speech features. It samples discontinuously and therefore can learn from the entire training dataset. do.

[0024] In step 404, the computing device 120 extracts the sub-samples of the training test speech data 114. Analyzes all of the speech features of a sample and ranks them based on predetermined criteria Speech features are observed in a subsample of the training and testing speech data 114. Any suitable statistical method for identifying the features that most significantly contribute to the feature discrimination signal as determined by the feature discriminant signal. Specifically, speech features can be ranked based on a training test. Based on the feature importance of each speech feature for a subsample of the speech data 114 Specifically, the component classifier 310 can be an SVM. If so, the computing device 120 may extract the speech features from the previous step (step 402). Differences between normal and cognitively impaired patients (e.g., can be ranked based on their importance in specifying a decision boundary (e.g., More specifically, each speech feature is calculated by dividing the speech feature into subsamples of the training test speech data 114. Ranking by corresponding coefficients to the sample-based SVM hyperplane (e.g., decision boundary) The speech features with smaller coefficients to the SVM hyperplane are closer to the SVM decision boundary. are considered relatively insignificant in the specification of boundaries and therefore less important in features; The rank is lower.

[0025] In step 406, the computing device 120 calculates the value of ... The speech features generated by the Select a subset of speech features based on (select only the top x speech features) The selected subset of x speech features is calculated by the computing device 120 as follows: Generate a component classifier 310 to analyze the downsampled x features. The x highest ranked speech features are selected to generate a Thus, the computing device 120 extracts the feature-discriminating signals contained in the training test speech data 114. A component classifier 310 that models the features that contribute most significantly to the signal is The computational cost of the speech classifier 310 is reduced by the number of downsampled speech features, x. The most highly ranked arbitrary suitable number x is generated while keeping the features low. For example, step 406 may select the highest ranked feature. At least 10, at least 20, or at least 30 features may be selected. In one particular embodiment, the top 20 ranked features are: For each component classifier 310 is selected.

[0026] Returning to method 200, in step 208, computing device 120 repeats the previous step ( training data for training the ensemble classifier 300 generated from step 206 Specifically, the computing device 120 generates a training baseline speech data set 330. The speech data 112, especially the training baseline speech data 112, are used to generate the correct Generate a training data set 330 containing the normalized training test speech data 114. In one embodiment, the training baseline data 112 is obtained from a WLR test (e.g., RA The training dataset 330 includes the learning phase data from the training test (VLT test). By normalizing the test speech data114 with the training phase data from the WLR test, The training datasets obtained from this embodiment are generated from normal and cognitively impaired patients. Specifically, the computing device 120 provides a quantitative value corresponding to the cognitive load of the group. , the average of the corresponding features across multiple sets of training baseline speech data 112 The training dataset 33 includes each feature of the training test speech data 114 normalized by More specifically, each feature of the training dataset 330 is generated by From the features of the speech data 114, multiple sets of training baseline speech data 112 are generated. The training baseline can be obtained by subtracting the mean value of the features over Speech features of the training and testing speech data 114 are calculated from multiple sets of speech data 112. Normalizing , enhances the feature-discriminative signals contained in the training test speech data 114. and the predictions and / or the predictions of the ensemble classifier 300 trained on such normalized data. Or the analytical performance can be improved.

[0027] In step 210, the computing device 120 executes the training generated by step 208. The ensemble classifier 300 is trained using the data set 330 to generate the trained ensemble classifier. Each of the component classifiers 310 generates a sample classifier 516. The training data corresponds to the subset of features selected to be modeled by the The dataset is trained on a downsampled portion of the dataset 330. To train the component classifier 310, a unique downsampling of the training dataset 330 is performed. In other words, a training part is provided for training the component classifiers 311. The downsampled portion of the dataset 330 is used to classify the other component classifiers 312- This is different from the downsampled portion of the training dataset 330 used to train 317. The trained ensemble classifier 516 comprises the component classifiers 310. Each of the component classifiers 310 applies a selected subset of the speech features to the training data. For each feature generated from training the component classifier 310 with the set 330, The trained ensemble classifier 516 may be used to generate a classifier with a weighting coefficient value. and detecting and / or detecting cognitive decline, in particular MCI, in a subject. or loading new patient speech data onto a user device for analysis to predict It is possible.

[0028] FIG. 5 illustrates a method for detecting and / or assessing cognitive decline in a subject based on speech samples of the subject. 5 shows an exemplary apparatus 500 for predicting. The apparatus 500 uses a trained ensemble The trained ensemble classifier 516 is used to generate the ensemble classifier 516, which is then trained according to the method 20 described above. The device 500 receives an audio signal and records the audio signal. For example, the voice input device 506 is provided to generate data corresponding to the voice input device 506. 506 may include a microphone for capturing the subject's vocal speech. Additionally, the device 500 includes an audio output facility 508 for generating audio output. For example, the audio output facility 508 can be used to give the subject instructions for the WLR test. The device 500 may be equipped with a speaker to provide a visual signal to the user. The device 500 further comprises a display 512 for generating a visual output. receive data or instructions to and / or send data or instructions from device 500 The device 500 may further include an input / output device 510 for receiving data. The processor 502 further comprises a computer-accessible medium 504. is operably connected to an audio output facility 508 and a display 512 to output the sound of the device 500. The processor 502 also controls the audio and visual output. operably connected to a device that corresponds to an audio recording captured by the audio input facility 506. The processor 502 may be, for example, one or more microprocessors. and stored on a computer-accessible medium 504 (e.g., a memory storage device). The computer-accessible medium 504 may, for example, include The device 500 may be a non-transitory computer-accessible medium containing executable instructions. is a computer-accessible medium for storing the trained ensemble classifier 516. The system may further include a memory storage device 514 provided separately from the body 504. The storage device 514 may be part of the device 500 or may be external to the device 500. , may be operatively connected to the device 500.

[0029] FIG. 6 shows a subject's speech based on a speech sample according to an exemplary embodiment of the present application. 6 illustrates an exemplary method 600 for detecting and / or predicting cognitive decline in a subject. Specifically, the method 600 obtains a WLR speech sample from a subject, and the sample is compared with a normal subject. The correlation will be determined whether the sample is more strongly correlated with the normal or cognitively impaired patients. In step 602, a first set of instructions is given to the subject. The clinician operating the device 500 or the audio output facility 506 may The processor 502 may provide a plurality of audio inputs from the audio input facility 506. and a subject vector corresponding to a plurality of audio recordings of the subject's speech in response to the first set of instructions. The subject also receives baseline speech data. to provide a second set of instructions, and the processor 502 directs the audio output facility 506 to audibly providing a second set of instructions, and the processor 502 audibly providing a second set of instructions to the test subject. The first voice data corresponding to the subject test speech data is received from the voice input facility 506. The instruction set follows the steps described above to generate the training baseline speech data 112. The second set of instructions corresponds to the instructions used in step 202, and the second set of instructions corresponds to the training test speech data. This may correspond to the instructions used in step 202 to generate data 114.

[0030] In step 604, processor 502 performs the steps described above with respect to step 204. In a similar manner, the subject's baseline speech data and the subject's test speech data In step 606, the processor 502 analyzes and extracts a plurality of features from the In a manner similar to that described above for the training data set in step 208, Each feature of the subject-test speech data is compared with the corresponding feature of the subject-baseline speech data. Similar to step 208, the normalized test data is generated. In one embodiment, baseline speech data is obtained from a WLR test administered to a subject. The subject test data includes learning phase data from a clinical trial (e.g., the RAVLT trial). By normalizing the subject test speech data with the training phase data from the WLR test, The subject test data obtained from this embodiment is generated based on the cognitive load of the subject. In step 608, the processor 502 performs a trained The trained ensemble classifier 516 is used to analyze the subject test data. Each of the component classifiers 310 of the ensemble classifier 516 processes the subject test data A component that analyzes and identifies subject test data as corresponding to normal patients or cognitively impaired patients. The ensemble module of the trained ensemble classifier 516 generates the component output. 320 receives the component output from the component classifier 310 and classifies the component Analyzing the test output to identify the subject test data as corresponding to normal patients or cognitively impaired patients. Further, the processor 502 generates an ensemble output. Based on the ensemble output generated from the analysis of the subject test data by the classifier 516 , whether the subject is at high risk of neurodegeneration and / or whether the subject is experiencing cognitive decline. It can also generate an output that indicates whether the ensemble is likely to If the output identifies the subject test data as corresponding to a cognitive decline patient, High risk of dementia and / or high likelihood that the subject will have cognitive decline In contrast, the output is that the ensemble output is a subset of the subject test data. that the subject is not at increased risk of neurodegeneration when identified as corresponding to a and / or may indicate that the subject is not likely to have cognitive decline. At step 610, the processor 502 configures the display to provide a visual display indicative of the output. Conducts play 512.

[0031] generated from the analysis of subject test data by a trained ensemble classifier 516 The ensemble output and / or the output generated by the device 500 may be used to assess risk of neurodegeneration. and / or impaired verbal episodic memory, including but not limited to: Slower learning speed, decreased short-term verbal memory, decreased delayed verbal memory, and decreased recall ability after interfering stimuli To screen and identify patients with cognitive decline, including cognitive impairment, cognitive decline, and cognitive impairment The subject test data can be recognized by the ensemble output and provided to the clinician. Subjects identified as corresponding to cognitive impairment patients are those with a neurodegenerative disorder such as AD or another type of dementia. Therefore, the risk of developing the disease may be higher. The resulting ensemble output and / or the output generated by the device 500 may be in a dangerous state. Patients in this condition may be further analyzed to confirm the ensemble output and / or the output from method 600. This study can assist clinicians in directing further cognitive testing to identify the underlying cause of cognitive impairment. The claimed devices and methods provide a more accurate diagnosis of neurodegenerative diseases, dementia, or AD than traditional diagnostic methods. earlier identification of patients at risk for HIV and improved care for these patients and / or allow treatment to begin at an earlier stage of cognitive decline. Finally, the ensemble output indicates that the subject test data corresponds to a cognitive decline patient, and and / or wherein the output generated by the device 500 is a measure of risk of neurodegeneration in the subject. processor 5 if the individual indicates that the individual has a high risk of cognitive decline and / or is likely to have cognitive decline. 02 generates instructions for administration of the treatment to the subject. The instructions may direct the administration of therapy automatically without the need for intervening user intervention. do.

[0032] Suitable treatments include treatments for improving cognitive ability and / or ameliorating cognitive decline. In one example, the treatment may be, for example, a digital treatment (e.g., This may include non-pharmacological treatments to improve the subject's cognitive abilities, such as brain training modules. In some embodiments, the ensemble output is generated based on the subject test data. and / or the output generated by the device 500 corresponds to a patient with a low blood pressure. that individuals are at increased risk of neurodegeneration and / or that subjects are likely to suffer from cognitive decline If the brain shows high levels of cognitive impairment, digital therapy can be automatically initiated. The process involves administering brain training exercises to improve cognitive performance in subjects. The brain training module contains a set of instructions executable by the brain training exerciser. The system displays user commands to the subject and receives input from the user in response to such commands. For this purpose, it may be connected to a user interface.

[0033] In another example, the treatment comprises one or more therapeutic agents to prevent and / or ameliorate the progression of neurodegeneration. Treatment may involve the administration of medicinally active substances. In particular, treatment may involve preventing the progression of dementia. and / or ameliorate, or more specifically prevent and / or ameliorate the progression of AD The administration of one or more pharmaceutically active substances to one or more of the pharmaceutically active substances Suitable medicinally active substances include those that target the beta-amyloid protein in the brain or Prevents or reduces tau protein aggregation, improves the resilience of brain synapses or cells, and one or more of these to regulate the expression of the oE4 gene, regulate neuroinflammation-related pathways, etc. In one example, the medicinally active substance is an antipsychotic. These may include antidepressants, acetylcholinesterase inhibitors, and the like.

[0034] Those skilled in the art will appreciate that the exemplary embodiments described herein may be implemented as separate software modules. any number of ways, such as as a module, or as a combination of hardware and software. For example, the exemplary method may be implemented in a non-transitory storage medium. When stored and compiled, it may be used on one or more processor cores or separate processors. A program or programs containing lines of code that can be executed by a processor. A system according to one embodiment may include multiple processor cores and the above exemplary and an instruction set that executes on the plurality of processor cores to perform the method. The processor core or a separate processor may be implemented in any suitable electronic device, e.g. a device capable of communicating with an on-board processing unit within the device or at least a portion of the device external processing devices, such as mobile computing devices, smartphones, may be incorporated into a computing tablet, computing device, etc. or may communicate with it. [Example]

[0035] Example I In Example I, a total of 106 subjects were included, 84 of whom were normal and 22 of whom had MCI. For patients, an exemplary word list recall test was administered (similar to the RAVLT test) It was performed manually by the examiner and also by a computerized device. Then, the first command is given to listen to the first word list and immediately play and pronounce the first word list. A set of instructions was provided. The first word list consisted of the following words: drum, helmet, Includes music, coffee, school, parents, machines, gardens, radio, farmers, noses, sailors, colors, houses, and rivers. Speech was recorded for each patient, and baseline speech was recorded corresponding to the speech recording. A set of matched data was generated. The presentation of the first word list and its immediate playback were recorded in the patient group. A total of five iterations were performed across the entire group, generating five different baseline speech datasets. did.

[0036] The patients then completed the interference task. Specifically, each patient was asked to Listen to the list and then provide a second set of instructions to play and speak a second list of words. The second word list consisted of the following words: desk, ranger, bird, shoe, and strike. Includes a tape, mountain, glasses, towel, cloud, boat, lamb, bell, pencil, church, and fish. and recording speech data corresponding to the speech recording for the obstruction test. After the distractor task, the patient was asked to play the first word list. Speech was recorded for each patient and the speech was recorded for the post-disturbance test. A test speech data set corresponding to the patient's recording was generated. After a 20-minute delay, the patient The subjects were again asked to play and pronounce the first word list. Record the speech and collect test speech data corresponding to the speech recording for the delayed recall test. A set was generated.

[0037] For example, across speech frames of an audio recording of a speech (as described above in Table 1) ) An exemplary set of speech features, such as mean and standard deviation values ​​of exemplary acoustic features, is used as a basis Five different sets of line speech data, test speech data for interference tests the test speech data set for the post-interference test, and the test speech data set for the delayed recall test. The speech data sets were extracted from each of the test speech data sets. 700 is shown in Figure 7. As shown in Figure 7, the ensemble classifier of Example I 700 includes 30 component classifiers 701 to 730. Each of 701 to 730 is a training data set correlated with control patients and a training data set correlated with MCI patients. The top scores are determined based on a subsample of the training data balanced against the training data. Support vectors generated based on a downsampled subset of 20 features Specifically, the subsamples were 20 samples corresponding to control patients. and 20 entries corresponding to MCI patients.

[0038] As shown in FIG. 7, the ensemble classifier 700 classifies each of the features of the test speech data. We normalized each feature by the mean of the corresponding feature across the five sets of baseline speech data. Train and validate by 10-fold cross-validation using a set of data 760 containing As shown in FIG. 7, the data 760 is randomly distributed and divided into equal-sized blocks. 0 splits. Of these 10 splits, data from 9 splits are The remaining splits are used as the training data 770 for the sample classifier 700. Another partition is used as the validation data 780, and the remaining partition is used as the training data. The training data 770 is used, and each division is used once as validation data 780. Training and validation are repeated until a satisfactory result is obtained. Example I uses a 10-fold cross-validation method. The ensemble classifier 700 is constructed using k-fold cross-validation, where k is any suitable positive integer. It is thought that it can be verified using

[0039] For each of the interference test, post-interference test, and delayed recall test, tests were performed across each of the 10 sections. The means of certain performance measures generated based on the evidence data are shown in Table 2 below. The post-interference test showed the highest difference between MCI patients and normal subjects, as shown by the performance scale of 2. It has another signal, then a jamming test, then a delayed playback test.

[0040] [Table 2]

[0041] The specific embodiments disclosed herein are intended to illustrate certain aspects of the present invention. The invention described and claimed herein is intended to be encompassed by these embodiments. Any equivalent embodiments are intended to be within the scope of the present invention. Indeed, various embodiments of the present invention in addition to those shown and described herein may be practiced. Modifications will be apparent to those skilled in the art from the above description. Such modifications also include the accompanying It is intended that the invention be within the scope of the claims. All publications cited herein are incorporated by reference. No. 6,299,499, filed on Dec. 1, 2003, the entire contents of which are incorporated herein by reference.

Claims

1. 1. A method for detecting cognitive decline in a subject, comprising: a plurality of speeches of the subject in response to a first set of instructions delivered to the subject; obtaining subject baseline speech data corresponding to the audio recording of Further recording of the subject's speech in response to a second set of instructions delivered to the subject. acquiring subject test speech data corresponding to the audio recording; A plurality of speech data are extracted from the subject baseline speech data and the subject test speech data. Extracting features; The subject's baseline speech data is used to correct the subject's test speech data. generating subject test data by normalizing; analyzing the subject test data using a trained ensemble classifier; Including, The trained ensemble classifier Multiple Component Classifiers and Ensemble Modules Including, Each of the plurality of component classifiers classifies the subject test data into normal patients or configured to generate component outputs that identify the component outputs as corresponding to a cognitive decline patient, each a component classifier for analyzing a subset of features selected from the plurality of features; It is structured as follows: The ensemble module extracts the component classes from the component classifiers. and receiving the component outputs and, based on the component outputs, or a cognitive decline patient, The plurality of component classifiers classify normal and undetected components in response to the first set of instructions. Training baseline speech corresponding to prior audio recordings of speech from a group of cognitively impaired patients. and the group of normal and cognitively impaired patients who responded to the second set of instructions. is trained using training test speech data corresponding to a previous audio recording of the speech of ,method.

2. The subset of features of the trained ensemble classifier may be The multiplexing is performed on a sub-sample of the training test speech data based on predetermined criteria. The subsamples are then ranked based on the number of features in the training test samples corresponding to normal patients. a first number of samples of speech data and a first number of samples of the training test speech data corresponding to a cognitive decline patient; a second number of samples of the data; and selecting the subset of features from the plurality of features based on a predetermined ranking threshold; To do a step of: The method of claim 1 , wherein each of the two is independently selected.

3. wherein the second number of samples is at least 80% of the first number of samples. The method according to item 2.

4. wherein the ratio of the first number of samples to the second number of samples is 1:

1. The method according to claim 3.

5. The method of claim 2 , wherein the predetermined criterion is feature importance.

6. The multiple component classifiers are trained using training baseline speech data. Uses a training dataset generated by normalizing the test speech data. The method of claim 1 , wherein the training is performed using a

7. Detecting mild cognitive impairment (MCI) in said subjects and comparing the results of normal and cognitively impaired patients The method of claim 1 , wherein the group consists of normal subjects and MCI subjects.

8. Each of the plurality of component classifiers is based on the training baseline speech data and and a plurality of weighted feature coefficients trained using the training test speech data. Item 1. The method according to item 1.

9. The method of claim 1 , wherein each of the plurality of component classifiers is a machine learning classifier. method.

10. Each of the plurality of component classifiers is a support vector machine (SVM).

10. The method according to claim 9.

11. If the ensemble output identifies the subject test data as corresponding to a cognitive decline patient, If the subject is found to be suffering from cognitive decline, an output is displayed indicating that the subject is likely to be suffering from cognitive decline. If the sample output identifies the subject test data as corresponding to a normal patient, displaying an output indicating that the individual is unlikely to have cognitive decline; The method of claim 1 further comprising:

12. If the ensemble output identifies the subject test data as corresponding to a cognitive decline patient, If so, the subject is determined to have a verbal episodic memory decline based on the ensemble output. Generate output that indicates The method of claim 1 further comprising:

13. The first set of instructions corresponds to a word list recall test, and the second set of instructions corresponds to a time 13. The method of claim 12, corresponding to an inter-delayed word list recall test.

14. If the ensemble output identifies the subject test data as corresponding to a cognitive decline patient, In such cases, treatment to improve cognitive ability or reverse cognitive decline is provided. The method of claim 1 further comprising:

15. If the ensemble output identifies the subject test data as corresponding to a cognitive decline patient, In such cases, administering a substance that is effective as a medicine to prevent or improve the progression of dementia. The method of claim 1 further comprising:

16. If the ensemble output identifies the subject test data as corresponding to a cognitive decline patient, In this case, the aggregation of beta-amyloid protein or tau protein in the brain of the subject is administering medicinally effective substances to prevent or alleviate The method of claim 1 further comprising:

17. 1. An apparatus for detecting cognitive decline in a subject, comprising: an audio output facility configured to generate an audio output; configured to receive an audio signal and generate data corresponding to a recording of said audio signal; voice input equipment, The display and a processor, and The audio output is adapted to audibly provide a first set of instructions to the subject multiple times. Commanding the power facilities, corresponding to a plurality of audio recordings of the subject's speech in response to the first set of instructions. receiving subject baseline speech data from said speech input facility; said audio output facility to audibly provide a second set of instructions to said subject; He directed a subject's speech corresponding to a further audio recording of the subject's speech in response to a second set of instructions; receiving subject test speech data from said voice input facility; The subject's baseline speech data and the subject's test speech data are extracted from a plurality of Extract the features of The subject's baseline speech data is used to generate subject test speech data. generating subject test data by normalizing; Analyzing the subject test data using a trained ensemble classifier to identify the subject generating an output indicating whether the subject is likely to have cognitive decline; directing the display to provide a visual representation of the output to a user; a set of instructions executable by said processor to operate as follows: a non-transitory computer-readable storage medium having a memory configured to store the trained ensemble classifier; Equipped with The trained ensemble classifier Multiple Component Classifiers and Ensemble Modules Equipped with Each of the plurality of component classifiers classifies the subject test data into normal patients or configured to generate component outputs that identify the component outputs as corresponding to a cognitive decline patient, each a component classifier for analyzing a subset of features selected from the plurality of features; It is structured as follows: The ensemble module extracts the component classes from the component classifiers. and receiving the component outputs and, based on the component outputs, or a cognitive decline patient, The plurality of component classifiers may be configured to classify normal and cognitive classes in response to the first set of instructions. Training baseline speech corresponding to previous recordings of speech from a group of hypotensive patients data, and the group of normal and cognitively impaired patients who responded to the second set of instructions. It is trained using training test speech data corresponding to the preceding audio recording of the speech. A device.

18. The subclassification of features for each of the plurality of component classifiers from the plurality of features. The set includes: The multiplexing is performed on a sub-sample of the training test speech data based on predetermined criteria. The subsamples are then ranked based on the number of features in the training test samples corresponding to normal patients. a first number of samples of speech data and training test speech data corresponding to cognitive decline patients; a second number of samples; and selecting the subset of features from the plurality of features based on a predetermined ranking threshold; To do a step of: The device of claim 17 , wherein the plurality of electrodes are independently selected.

19. configured to detect cognitive decline in a subject based on a speech sample of the subject; 1. A computer-implemented method for training a classification system comprising: From the group of normal and cognitively impaired patients, normal and cognitively impaired patients responded to the first set of instructions. training baseline speech data corresponding to audio recordings of the speech of said group of patients; and sounds of the speech of said group of normal and cognitively impaired patients in response to a second set of instructions. obtaining training test speech data corresponding to the voice recording; generating an ensemble classifier, the classifier comprising: (i) a training baseline speech; (ii) extracting a plurality of features from the training test speech data; 、 a plurality of component classifiers and an ensemble module; Each of the plurality of component classifiers classifies the sample data as normal or cognitively impaired. configured to generate component outputs that identify the component as corresponding to the patient, a component classifier configured to analyze a subset of features selected from the plurality of features; It has been made The ensemble module extracts the component classifiers from the component classifiers. and receiving an output from said component, and classifying said sample data as a normal patient or and generating an ensemble output that identifies the ensemble output as corresponding to a cognitive decline patient. generating an ensemble classifier; The training baseline speech data is used to correct the training test speech data. generating a training dataset by normalizing the training the ensemble classifier using the training dataset; A computer-implemented method comprising:

20. The subset of features may be: The multiplexing is performed on a sub-sample of the training test speech data based on predetermined criteria. The subsamples are then ranked based on the number of features in the training test samples corresponding to normal patients. a first number of samples of speech data and training test speech data corresponding to cognitive decline patients; a second number of samples; and selecting the subset of features from the plurality of features based on a predetermined ranking threshold; To do and generating the plurality of feature classes for each of the plurality of component classifiers by a step including:

20. The computer-implemented method of claim 19, wherein the characteristics are independently selected from the characteristics.

21. configured to detect cognitive decline in a subject based on a speech sample of the subject; 1. A system for training a classification system comprising: Training baseline speech data and training data from groups of normal and cognitively impaired patients configured to store test speech data, and is a set of audio of the speech of the group of normal and cognitively impaired patients in response to a first set of instructions. corresponding to the recording, the training test speech data includes normal and a database corresponding to audio recordings of speech of said group of cognitive decline patients; a processor operatively connected to communicate with the database; and a non-transitory computer-readable storage medium containing a set of instructions executable by the The instruction set is: The training baseline speech data and the training test speech data are received from the database, generating an ensemble classifier, (i) using the training baseline speeds; (ii) extracting a plurality of features from the training test speech data; a plurality of component classifiers and an ensemble module; Each of the plurality of component classifiers classifies the sample data as a normal patient or a cognitively impaired patient. and each component configured to generate a component output that identifies the component as corresponding to the patient. a component classifier configured to analyze a subset of features selected from the plurality of features; It is composed of The ensemble module extracts the component classes from the component classifiers. and receiving the component outputs and analysing the sample data based on the component outputs. is configured to generate an ensemble output that is identified as corresponding to a patient with cognitive decline. 、 Generate an ensemble classifier, The training baseline speech data is used to generate the training test speech data. Generate a training dataset by normalizing it; Training the ensemble classifier using the training dataset. a computing device operable to: a memory configured to store the trained ensemble classifier; 1. A system for training a classification system, comprising:

22. The instruction set comprises: The multiplexing is performed on a sub-sample of the training test speech data based on predetermined criteria. The subsamples are then ranked based on the number of features in the training test samples corresponding to normal patients. a first number of samples of speech data and training test speech data corresponding to cognitive decline patients; a second number of samples; and selecting the subset of features from the plurality of features based on a predetermined ranking threshold; To do and generating a classification result from the plurality of features into each of the plurality of component classifiers by a step including the steps of: and further operable to independently select said subset of features for each of the plurality of features.

22. A system for training a classification system according to clause 21.

Images