Device and related components for constructing a classification model for identifying different stages of heart failure

Abstract

The present invention discloses an apparatus and related components for constructing a classification model for identifying different stages of heart failure. The apparatus includes a sample processing unit for converting collected analog voice signals into digital voice signals, preprocessing the digital voice signals, and extracting features from the preprocessed digital voice signals to obtain multiple categories of voice feature samples; a model training unit for constructing a classification model for identifying different stages of heart failure, and using the multiple categories of voice feature samples to train and optimize the classification model to obtain an optimal classification model. This optimal classification model can more accurately identify different stages of heart failure.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a device for constructing a classification model for identifying different stages of heart failure based on sound features, a computer-readable storage medium, and a computer device. Background Art

[0002] Heart failure is a complex clinical syndrome caused by abnormal changes in cardiac structure and / or function due to various reasons. It is a serious and terminal stage of various common cardiovascular diseases. Currently, according to the development and progression of heart failure, it is divided into four stages: heart failure risk stage (stage A), pre-heart failure stage (stage B), heart failure stage (stage C), and end-stage heart failure stage (stage D). The purpose of staging heart failure is to achieve early detection, early diagnosis, and early intervention, especially for the timely identification and treatment of patients at risk of heart failure and those in the pre-heart failure stage. Early intervention is of great significance in delaying ventricular remodeling and heart failure progression, protecting cardiac function, improving quality of life, and reducing rehospitalization rates.

[0003] In existing technologies, the identification of heart failure stages usually relies on medical history, physical examination, laboratory tests, cardiac imaging tests and functional tests. These tests are often performed only when patients seek medical treatment due to symptoms, which is not conducive to early identification. In recent years, the development of hemodynamic or lung water content monitoring through implantable devices, such as CardioMEMS, MultiSENSE, ReDS and other sensor devices, as well as evaluation through HeartLogic multi-sensor index and alarm algorithm, can achieve early warning of patients' heart failure decompensation events. However, the equipment of the above methods is expensive and invasive, requiring the implantation of sensors or the installation of pacemakers. They are only suitable for a small number of patients with severe or refractory heart failure and are not suitable for screening patients in the risk period and pre-heart failure stage of heart failure. On the basis of strengthening standardized diagnosis and treatment and patient education, developing non-invasive, convenient and universal monitoring and early warning methods, identifying patients with different stages of heart failure, and strengthening home monitoring and early warning are the key to chronic disease management of heart failure, reducing rehospitalization rates and lowering mortality. There are also technologies in the existing technology that train and learn certain parameters to identify classification models for certain diseases, but the data used by these classification models also rely on a lot of examination data, and the recognition accuracy is not high. Summary of the Invention

[0004] The embodiments of the present invention aim to provide an apparatus for constructing a classification model based on sound features that can accurately identify different stages of heart failure, a computer-readable storage medium, and a computer device.

[0005] In a first aspect, an embodiment of the present invention provides a device for constructing a classification model for identifying different stages of heart failure based on sound features, comprising:

[0006] A sample processing unit, configured to convert the collected voice analog signal into a voice digital signal, preprocess the voice digital signal, and extract features from the preprocessed voice digital signal to obtain multiple types of voice feature samples;

[0007] A model training unit is used to construct a classification model for identifying the stages of heart failure, and use the multi-class speech feature samples to train and optimize the classification model to obtain an optimal classification model, wherein the classification model for identifying heart failure stage A and heart failure stage B is an AdaBoost classification model based on the original variables; the classification model for identifying heart failure stage B and stage C is an AdaBoost classification model based on Lasso dimensionality reduction; and the classification model for identifying heart failure stages A, B, and C is an AdaBoost classification model based on Lasso dimensionality reduction.

[0008] In the second aspect, an embodiment of the present invention further provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the following method when executing the computer program: converting the collected voice analog signal into a voice digital signal, preprocessing the voice digital signal, and extracting features from the preprocessed voice digital signal to obtain multiple categories of voice feature samples; constructing a classification model for identifying the stages of heart failure, and training and optimizing the classification model using the multiple categories of voice feature samples to obtain an optimal classification model, wherein the classification model for identifying stage A of heart failure and stage B of heart failure is an AdaBoost classification model based on original variables; the classification model for identifying stage B and stage C of heart failure is an AdaBoost classification model based on Lasso dimensionality reduction; and the classification model for identifying stage A, stage B, and stage C of heart failure is an AdaBoost classification model based on Lasso dimensionality reduction.

[0009] In a third aspect, an embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the following method is implemented: converting the collected voice analog signal into a voice digital signal, preprocessing the voice digital signal, and extracting features from the preprocessed voice digital signal to obtain multiple categories of voice feature samples; constructing a classification model for identifying the stages of heart failure, and using the multiple categories of voice feature samples to train and optimize the classification model to obtain an optimal classification model, wherein the classification model for identifying stage A of heart failure and stage B of heart failure is an AdaBoost classification model based on the original variables; the classification model for identifying stage B and stage C of heart failure is an AdaBoost classification model based on Lasso dimensionality reduction; and the classification model for identifying stage A, stage B and stage C of heart failure is an AdaBoost classification model based on Lasso dimensionality reduction.

[0010] The embodiment of the present invention is based on the sound features that can reflect different stages, so as to achieve the purpose of constructing a classification model for identifying the corresponding stages for different stages, thereby improving the accuracy of model recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0012] Figure 1 A schematic diagram of the structure of a device for building a classification model for identifying different stages of heart failure based on sound features provided by an embodiment of the present invention;

[0013] Figure 2 This is a diagram showing the effects of different heart failure stages on jitter provided by an embodiment of the present invention;

[0014] Figure 3 This is a diagram showing the effects of different heart failure stages on Shimmer provided by an embodiment of the present invention;

[0015] Figure 4 This is a diagram showing the effect of different heart failure stages on harmonic difference provided by an embodiment of the present invention;

[0016] Figure 5 This is a diagram showing the effects of different heart failure stages on HNR provided by an embodiment of the present invention;

[0017] Figure 6This is a diagram showing the effect of different heart failure stages on Alpha Ratio provided by an embodiment of the present invention;

[0018] Figure 7 This is a diagram showing the effect of different heart failure stages on voiced / unvoiced duration provided by an embodiment of the present invention;

[0019] Figure 8 This is a diagram showing the effects of Loudness on different heart failure stages provided by an embodiment of the present invention;

[0020] Figure 9 This is a diagram showing the impact of the Hammarberg Index on different heart failure stages provided by an embodiment of the present invention;

[0021] Figure 10 This is a diagram showing the effects of spectral slope on different heart failure stages provided by an embodiment of the present invention;

[0022] Figure 11 A cross-correlation coefficient diagram of glottal cycles in different heart failure stages provided by an embodiment of the present invention;

[0023] Figure 12 A nonlinear analysis diagram of voice characteristics for different heart failure stages provided by an embodiment of the present invention;

[0024] Figure 13 A graph showing changes in voice acoustic characteristics based on cepstrum for different heart failure stages provided by an embodiment of the present invention;

[0025] Figure 14 The sample-level ROC curve of the optimal model (i.e., AdaBoost of the original variables) provided by the embodiment of the present invention;

[0026] Figure 15 The sample-level ROC curve of the optimal model (i.e., AdaBoost with Lasso dimensionality reduction) provided by the embodiment of the present invention is shown;

[0027] Figure 16 This is a sample-level ROC curve diagram of the optimal model (i.e., AdaBoost using Lasso dimensionality reduction) provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0028] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0029] It will be understood that when used in this specification and the appended claims, the terms “comprises” and “comprising” indicate the presence of described features, integers, steps, operations, elements and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof.

[0030] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the present invention. As used in the specification and appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms unless the context clearly indicates otherwise.

[0031] It should be further understood that the term "and / or" used in the present description and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

[0032] See below Figure 1 , Figure 1 This is a schematic diagram of the structure of a device for building a classification model for identifying different stages of heart failure based on sound features provided by an embodiment of the present invention. The device includes a processing unit 101 and a model training unit 201;

[0033] The sample processing unit 101 is used to convert the collected voice analog signal into a voice digital signal, preprocess the voice digital signal, and extract features from the preprocessed voice digital signal to obtain multiple types of voice feature samples.

[0034] In this unit, sound is an analog signal that needs to be converted into a digital signal before it can be processed by a computer. The sample processing unit includes: a first conversion unit for sampling, a second conversion unit for quantization, and a third conversion unit for encoding;

[0035] The first conversion unit is specifically used to convert the time-continuous voice analog signal into a time-discrete and amplitude-continuous signal sample according to a predetermined sampling period;

[0036] The sampling period is the time interval between two adjacent sampling points, and the sampling frequency is the reciprocal of the sampling period. For example, a sampling frequency of 8kHz means that 8,000 samples are collected in 1 second. Therefore, the higher the sampling frequency, the higher the sound restoration degree and the more realistic the sound.

[0037] The second conversion unit is specifically used to convert each signal sample with continuous amplitude values ​​(analog quantity) into a discrete value (digital quantity), and express it in binary to obtain digital data;

[0038] Quantized signal samples are usually represented in binary. Sampling accuracy refers to the number of binary bits occupied by each signal sample, which can reflect the quality of the sound. For example, common CDs use a 16-bit sampling depth, which can represent 65535 (2^16) different values. DVDs use a 24-bit sampling depth, and most telephone devices use an 8-bit sampling depth.

[0039] The third conversion unit is specifically used to convert the digital data into a binary code stream to obtain a voice digital signal.

[0040] Coding is the process of converting sampled and quantized digital data into a binary code stream for computer storage, processing, and transmission. Pulse code modulation can achieve the highest fidelity level. The data transmission rate of sound can be calculated by the sampling frequency and accuracy: data transmission rate (bps) = sampling frequency * accuracy * number of channels. At the same time, the data volume of the sound signal can be calculated: data volume (byte) = data transmission rate * duration / 8.

[0041] In this embodiment, in order to improve the quality of the voice digital signal and retain more voice information, the voice digital signal needs to be preprocessed before the voice digital signal is analyzed, that is, the sample processing unit also includes: a synchronization unit, an endpoint detection unit, an emphasis unit, a framing and windowing unit.

[0042] The synchronization unit is used to synchronize multiple voice digital signals to a unified sampling rate by using a downsampling method;

[0043] Different speech acquisition devices may have different sampling frequencies, durations, and locations, resulting in varying speech amplitudes. Therefore, the original speech needs to be downsampled before signal analysis. For example, if the original signal has a sampling frequency of 22050 Hz, this embodiment downsamples the signal to 16000 Hz, based on the Nyquist sampling theorem, to reduce computational complexity, improve signal processing efficiency, and preserve the key speech components.

[0044] The endpoint detection unit is used to perform endpoint detection on the voice digital signal after the unified sampling rate, and distinguish the voice area and the non-voice area;

[0045] Endpoint detection, also known as voice activity detection, aims to distinguish speech from non-speech areas. Specifically, it aims to accurately detect the start and end points of speech within noisy speech, remove silence and noise, and identify the truly effective speech content. Common methods include the dual-threshold method based on short-term energy and short-term zero-crossing rate. Because voiced speech has higher energy than unvoiced speech, and the zero-crossing rate of unvoiced speech is higher than that of silence, endpoint detection can be achieved by first using short-term energy to detect voiced speech and then using the zero-crossing rate to extract unvoiced speech.

[0046] The emphasis unit is used to emphasize the high-frequency part of the voice region and increase the high-frequency resolution of the voice;

[0047] Since the average power spectrum of the voice signal is affected by glottal excitation and oral-nasal radiation, the power spectrum decreases with the increase of frequency, and the energy of the voice is mainly concentrated in the low-frequency part. Therefore, pre-emphasis processing is required. The purpose is to emphasize the high-frequency part of the voice, remove the influence of lip radiation, and increase the high-frequency resolution of the voice, so that the signal spectrum can use the same signal-to-noise ratio to obtain the spectrum in the entire frequency band from low frequency to high frequency, which is convenient for spectrum analysis or vocal tract parameter analysis. Generally, pre-emphasis is achieved through a high-pass digital filter with a transfer function, that is, H(z) = 1 - az -1 , where a is the pre-emphasis coefficient, and the range is 0.9 < a < 1.0. Generally, a = 0.97 is taken.

[0048] The framing and windowing unit is used to frame and window the emphasized voice region to obtain multiple voice signal segments.

[0049] The voice signal is a non-stationary continuous analog signal with time-varying characteristics. However, within a short time range (generally within 10 - 30 ms), its characteristics remain basically unchanged, that is, relatively stable. Therefore, the voice signal has short-term stationarity, which means that any analysis and processing of the voice signal must be "short-term analysis", segment the voice signal to analyze its characteristic parameters, and each segment is called a "frame", and the frame length is generally 10 - 30 ms. In order to make the frames transition smoothly and maintain continuity between frames, the overlapping segmentation method can be used for framing. The overlapping part between the previous frame and the next frame is called the frame shift, and the ratio of the frame shift to the frame length is generally 0 - 0.5. In this embodiment, the frame length is taken as 25 ms and the frame shift is taken as 10 ms.

[0050] Due to the short-term stationarity of the voice signal, in addition to framing the signal, windowing processing is also required. The purpose is to emphasize the voice waveform near the sampled samples and weaken the rest of the waveform. Commonly used window functions include rectangular windows, Hamming windows, etc. The rectangular window has a high spectral resolution, but the adjacent harmonic interference is relatively serious, losing high-frequency components and resulting in the loss of waveform details, while the Hamming window is the opposite.

[0051] In this embodiment, the sample processing unit further includes: an extraction unit and a merging unit;

[0052] Among them, the extraction unit is used to extract multi-dimensional first voice feature samples using the openSMILE open source toolkit and extract multi-dimensional second voice feature samples using python;

[0053] Among them, the extraction unit is used to merge the multi-dimensional first voice feature samples and multi-dimensional second voice feature samples to obtain multi-class voice feature samples.

[0054] It should be noted that the first speech feature sample and the second speech feature sample can be extracted according to actual needs. The multi-category speech feature samples extracted in this embodiment have a total of 100 dimensions. The first speech feature sample uses the eGeMAPS feature set, which is an extended feature set of GeMAPS. This feature set is an 88-dimensional manual feature extracted by the openSMILE open source toolkit, which includes 18 low-level descriptors (LLDs) and adds 5 spectral features (MFCC1-4 and spectral flux) and 2 frequency-related features (i.e., the bandwidth of the second resonance peak and the third resonance peak) on the basis of GeMAPS, including frequency, energy / amplitude-related features and spectral features. In addition, a 12-dimensional second speech feature sample was extracted using python.

[0055] In this embodiment, the first speech feature sample includes: a pitch feature, a frequency perturbation feature, a formant feature, an amplitude perturbation feature, a loudness feature, a harmonic-to-noise ratio feature, a harmonic difference feature, an alpha ratio feature, a Hammarberg coefficient feature, a spectral slope feature, a Mel-frequency cepstral coefficient feature, a spectral flow feature, a loudness peak ratio feature, a continuous sound region and a silent region feature, and an equivalent sound level feature. Detailed information about the first speech feature sample is shown in Table 1.

[0056] Table 1

[0057]

[0058]

[0059] In Table 1, the pitch feature is the fundamental frequency of vocal cord vibration, which represents the number of times the vocal cord vibrates per second, and the feature description is log F0, calculated on the semitone frequency scale, starting from 27.5 Hz; the frequency perturbation feature is the frequency change between adjacent cycles of the sound wave, and the feature description is the deviation within a single continuous fundamental tone cycle; the resonance peak feature is the center frequency and bandwidth of the first, second and third resonance peaks, and the energy ratio of the first three resonance peaks to the fundamental tone; the amplitude perturbation feature is the feature description: reflecting the amplitude change between adjacent cycles of the sound wave; the loudness feature is the size of the sound; the harmonic-to-noise ratio feature is the feature description: the proportion of periodically repeated harmonic components in the sound wave of the steady-state vowel; the harmonic difference feature is the feature description: the energy ratio of the first fundamental tone harmonic H1 to the second fundamental tone harmonic H2 or the energy ratio of the first fundamental tone harmonic H1 to the third resonance peak H3; the α ratio feature is the feature description: 50-1000H The energy sum of z is divided by the energy sum of 1-5kHz; Hammarberg coefficient feature, feature description: that is, the strongest energy peak of 0-2kHz is divided by the strongest energy peak of 2-5kHz; spectral slope feature, that is, feature description: the linear regression slope of the logarithmic power spectrum in the range of 0-500Hz and 500-1500Hz; Mel-cephalometric coefficient feature, that is, feature description: Mel-cephalometric coefficients 1-4; spectral flow feature, that is, feature description: the spectral difference between two adjacent frames; loudness peak ratio feature, that is, feature description: the number of loudness peaks per second; continuous sound area and silent area feature, that is, feature description: the duration of continuous voiced sound (F0>0) and the duration of unvoiced sound (F0=0); equivalent sound level feature, that is, the equivalent sound level refers to the average value of the A sound level according to energy over a certain period of time.

[0060] In this embodiment, the second speech feature samples include: a glottal noise excitation ratio feature, a vocal cord excitation ratio feature, a cycle density entropy feature, a detrended fluctuation analysis feature, a sample entropy feature, and a multi-scale entropy feature. Detailed information about the second speech feature samples is shown in Table 2.

[0061] Table 2

[0062]

[0063]

[0064] In Table 2, the glottal noise excitation ratio feature (GNE) and the vocal fold vibration excitation ratio feature (VFER) are both used to quantify the energy proportion of the normal speech signal in which each frequency band is simultaneously excited by the glottal pulse and the noise signal in which each frequency band is disorderly excited by the chaotic noise (usually caused by incomplete closure of the glottis). For example, to calculate GNE parameters, for the original speech signal with a sampling frequency of 44.1kHz, first downsample the signal to 10KHz; then use the inverse filtering method to find the opening and closing time points of each glottis, and thus find its opening and closing time series; then for each time series, use a filter with a bandwidth of 500Hz to filter out the frequency bands of 0-500Hz, 500-1000Hz, 1000-1500Hz, etc. up to 11.5KHz; for each frequency band, use the first five frequency bands of the low frequency band (1Hz-2.5KHz) as the signal, and the remaining high frequency band (2.5KHz-11.5KHz) as the noise, and calculate the SEO and TKEO energy values ​​of the signal and noise respectively; finally, the signal-to-noise ratio (SNR) and the noise-to-signal ratio (NSR) are calculated based on the calculated energy values. The calculation of VFER parameters follows a similar process to GNE, except that the signal downsampling process is omitted. In the second step of GNE, the DYSPA algorithm is used instead of inverse filtering to determine the glottal opening and closing sequences. In other words, GNE and VFER first detect glottal pulses within a given time window using inverse filtering (for GNE) or DYSPA (for VFER). The original sound is then divided into two components: noise above 2.5 kHz and energy signal below 2.5 kHz. Combining the concepts of SEO and TKEO, the energy values ​​of the signal in different frequency bands are calculated to obtain the signal-to-noise ratio (SNR) measured by the empirical mode decomposition excitation ratio (EMDER-ER). After calculating the values ​​for each time window, the average and standard values ​​of GNE and VFER are calculated. The following parameters can be calculated: GNE_SEO_SNR, GNE_TKEO_SNR, GNE_mean, GNE_std, VFER_SEO_SNR, VFER_TKEO_SNR, VFER_mean, VFER_std.

[0065] The Cycle Density Entropy (RPDE) is a method used in the fields of dynamical systems, random processes, and time series analysis to determine the periodicity or repetitiveness of a signal. Its value ranges from 0 to 1. For quasi-periodic signals, its value is 0, while for uniform white noise, its value is close to 1. Its calculation method is as follows:

[0066] The first step is to transform the time series X n =[x n,x n+r ,x n+2r ,…,x n+(M-1)r According to Taken's embedding theory, the system is projected into a phase space; here, M is the embedding dimension, and T is the embedding delay, all of which are obtained through a parameter optimization algorithm. The second step is to draw an M-dimensional region with a radius of ∘ at each point Xn in the phase space. The time difference between each arrival and departure of the time series is recorded, and this time difference is plotted as a histogram. Finally, the histogram is normalized to obtain the regression density function P(T).

[0067] By the formula

[0068]

[0069] Where T max The maximum delay embedded in the phase space can be used to obtain the value of the cycle density entropy.

[0070] Detrended fluctuation analysis (DFA) is a scaling exponent calculation method used to eliminate the influence of trend terms in time series on fluctuation analysis. It is used to analyze the long-range correlation of speech signals, that is, to determine whether the noise terms in the time series have positive or negative autocorrelation. One of its advantages is that it can effectively filter out trend components of various orders in the sequence and detect long-range correlations containing noise and superimposed with polynomial trend signals. It is suitable for long-range power-law correlation analysis of non-stationary time series.

[0071] The implementation method is as follows:

[0072] 1. First, for the sequence x(t), calculate its cumulative deviation y(t)

[0073]

[0074] Where: is the mean value of the sequence x(t).

[0075] Here, we first filter out the mean value of the time series. Since cyclic or fluctuating components may exist in general time series, a time series may also have random components, and filtering out these components of the series will be very helpful.

[0076] 2. Reconstruct the sequence by dividing Y(t) into m non-overlapping intervals of length s, where m = [n / s] (rounded to an integer). Since the sequence length is not always an integer multiple of the increment s, a small amount of data at the end of the sequence may not be used. Therefore, perform the same operation on the reversed order of the sequence to obtain a total of 2N intervals of equal length.

[0077] 3. For each interval v, use the least squares method to perform a first-order linear fitting on the S data points contained in each interval respectively.

[0078]

[0079] 4. Calculate the mean square error after detrending for each interval (here, calculate using separate formulas for the forward and reverse orders):

[0080]

[0081] 5. Calculate the mean value of all equal-length intervals and take the square root to obtain the DFA fluctuation function:

[0082]

[0083] 6. If the runoff time series {x(t)} has long-range power-law correlation, then F(s) and s satisfy the following power-law relationship:

[0084] ln(F(s)) ~ hln(s)

[0085] Taking the logarithm of both sides of the above equation gives:

[0086] F(s) ~ s h

[0087] In the scatter plot in the double logarithmic coordinate (ln(s), ln(F(s))), use the least squares method to fit the data points, where the slope of the straight-line part is the Hurst exponent.

[0088] Relationship between the Hurst exponent and correlation

[0089] [[ID=..38]](1) When 0.5 < h < 1, it indicates that the time series has long-range correlation and shows a state of continuously increasing trend, that is, in a certain time period, it is an increasing (decreasing) trend, and in the next time period, it will also be an increasing (decreasing) trend, and the closer h is to 1, the stronger the correlation.

[0090] (2) When h = 0.5, it indicates that the time series is uncorrelated and is an independent random process, that is, the current state does not affect the future state.

[0091] (3) When 0 < h < 0.5, it indicates that the runoff time series only has negative correlation and shows an anti-persistent state, that is, if the time series is an increasing (decreasing) trend in a certain time period, then in the next time period, it will be a decreasing (increasing) trend.

[0092] Empirical Decomposition Mode Ratio (EMD-ER): For the original speech signal with a sampling frequency of 44.1kHz, it can be decomposed into a finite number of Intrinsic Mode Functions (IMFs). The decomposed IMF components contain local characteristic signals of the original signal at different time scales. The first decomposed eigenmode function is a high-frequency noise signal, and the subsequent decomposed eigenmode function is the actual useful signal. According to the energy operator formula And Shannon entropy, we can calculate the SEO, TKEO and Shannon entropy of each IMF. When calculating the signal-to-noise ratio, the first four IMFs are used as noise signals, and the formula is (u is the SEO, TKEO and Shannon entropy of each IMF, D is the number of IMFs obtained by decomposition) can be used to obtain parameters related to the signal-to-noise ratio. When calculating the signal-to-noise ratio, first take the logarithm of each IMF, take the first two IMFs as noise signals, and then calculate the SEO, TKEO and Shannon entropy of each IMF. The formula is: (u is the value of SEO, TKEO and Shannon entropy of each IMF, and D is the number of IMFs obtained by decomposition) parameters related to the noise-to-signal ratio can be obtained.

[0093] Sample entropy feature (SampEn) is an improved method for measuring the complexity of time series based on approximate entropy;

[0094] Multi-scale entropy feature (MSEn) extends the sample entropy to multiple time scales and calculates the complexity of signals at different time scales.

[0095] The glottis quotient (GQ) measures the stability of vocal cord vibration by calculating the mean and standard deviation of a speech signal. First, the DYSPA algorithm is used to identify the glottal opening and closing points. A speech signal is divided into several glottal opening and closing segments. The mean and standard deviation of each segment is then calculated.

[0096] MFCCs are cepstral parameters extracted from the Mel-scale frequency domain. The Mel-scale describes the nonlinear characteristics of the human ear. From left to right, the 39-dimensional MFCCs include a logarithmic energy and 12 cepstral parameters, a difference operation, and a difference-of-difference operation.

[0097] The model training unit 201 is used to construct a classification model for identifying the stages of heart failure, and use the multi-class speech feature samples to train and optimize the classification model to obtain the optimal classification model, wherein the classification model for identifying heart failure stage A and heart failure stage B is an AdaBoost classification model based on the original variables; the classification model for identifying heart failure stage B and stage C is an AdaBoost classification model based on Lasso dimensionality reduction; and the classification model for identifying heart failure stages A, B and C is an AdaBoost classification model based on Lasso dimensionality reduction.

[0098] In one embodiment, a binary classification model and a dimensionality reduction method are combined to optimize the classifier algorithm. In this embodiment, six classification models and two dimensionality reduction methods are selected, namely, support vector machine classification model (SVM), decision tree classification model (DT), adaptive boosting classification model (Ada Boost), least absolute shrinkage and selection operator classification model (LASSO), ridge regression classification model (Ridge regression), elastic network classification model (Elastic Net), principal component analysis (PCA) dimensionality reduction, and LASSO dimensionality reduction.

[0099] The final model is shown in Table 3, which includes three models: ① classification model based on original variables; ② classification model based on principal components after PCA dimensionality reduction; ③ classification model based on LASSO feature selection and feature variables.

[0100] Table 3

[0101]

[0102] The apparatus of the embodiment of the present invention further includes: an evaluation unit, configured to perform model evaluation on the classification model according to a holdout method.

[0103] When building a model, we should try to avoid "overfitting," where the classifier treats some characteristics of the training samples as general properties that all potential samples will have, resulting in a decrease in generalization performance, manifested as the final model performing well on the training set but poorly on the test set. We should also avoid "underfitting," where the general properties of the training samples have not yet been learned, resulting in poor performance on both the training and test sets. "Overfitting" cannot be avoided, but can only be alleviated, while "underfitting" can be overcome by increasing the number of features, increasing model complexity, and reducing the regularization coefficient. Ultimately, we hope that the model will have a good fit to the training dataset (i.e., training error) and a good fit to unknown datasets (i.e., test sets) (i.e., generalization ability).

[0104] Model building typically requires dividing sample data into a training set and a test set, which are mutually exclusive. Model evaluation requires a test set to test the learner's ability to discriminate against new samples, approximating the generalization error with the "test error" on the test set. Common methods include holdout, cross-validation, and bootstrapping.

[0105] Holdout method: Directly divide the dataset into two mutually exclusive sets, one set as the training set and the other as the test set. The division of the training set and the test set should maintain the consistency of the data distribution as much as possible to avoid the impact of additional bias introduced by the data division process on the final result. Because the estimation results obtained by a single use of the holdout method are often unreliable, it is generally necessary to use several random divisions and repeat the experimental evaluation and take the average value as the evaluation result of the holdout method. Generally, 2 / 3 to 4 / 5 of the samples are used for training, and the rest are used for testing. The leave-one-out method is adopted in this embodiment, which is one of the methods of the holdout method.

[0106] The device of the embodiment of the present invention also includes: a performance measurement unit, which is used to perform performance measurement on the classification model according to predetermined indicators, wherein the predetermined indicators include error rate and accuracy, precision rate and recall rate, F1 value, specificity, sensitivity, ROC curve and AUC, and unweighted average recall rate.

[0107] Evaluating the generalization performance of a classifier requires a metric to measure the model's generalization ability, known as a performance metric. For binary classification problems, a confusion matrix can be formed based on the combination of the true category and the classifier's predicted category. This matrix includes four categories: true positive (TP), true negative (TN), false positive (FP), and false negative (FN). Table 4 shows the confusion matrix.

[0108] Table 4

[0109]

[0110] The predetermined indicators are:

[0111] Accuracy (ACC): Also known as precision, this refers to the ratio of correctly classified samples to the total number of samples. It is applicable to both binary and multi-classification tasks and can be used to determine the overall accuracy rate. However, in cases of imbalanced samples, accuracy becomes ineffective and cannot be used as a metric for measuring results. Therefore, other indicators are needed to supplement the accuracy.

[0112] ACC=(TP+TN) / (TP+FN+FP+TN)

[0113] Error rate (ERR): refers to the ratio of all misclassified samples to the total number of samples, ERR = 1-ACC.

[0114] ERR=(FN+FP) / (TP+FN+FP+TN)

[0115] Precision (P): Also known as accuracy, it mainly refers to the prediction results and refers to the proportion of all samples predicted to be positive to the actual positive samples.

[0116] P=TP / (TP+FP)

[0117] Recall rate (R): also known as recall rate or sensitivity (SEN), mainly for the original sample, refers to the proportion of samples predicted to be positive among the actual positive samples.

[0118] R=TP / (TP+FN)

[0119] Specificity: (SPE): refers to the proportion of samples that are predicted to be negative among those that are actually negative.

[0120] SPE=TN / (FP+TN)

[0121] F1 Score: It is the harmonic mean of precision and recall, ranging from 0 to 1. It is often used in statistics to measure the accuracy of binary classification (or multi-task binary classification) models.

[0122] F1=2*P*R / (P+R)

[0123] ROC curve and AUC: The full name of the ROC curve is "Receiver Operating Characteristic Curve". The ordinate of the ROC curve is the true positive rate (i.e. sensitivity), and the abscissa is the false positive rate (1-sensitivity). The coordinate points are obtained at different thresholds and connected. The closer the ROC curve is to the diagonal line, the lower the accuracy of the model. Assuming that the ROC curve of classifier A can "enclose" classifier B, it can be said that classifier A has better classification performance. However, when the ROC curves of the two classifiers intersect, it is difficult to judge the performance of the two. In this case, the area under the ROC curve can be used for measurement, namely AUC (Area Under ROC Curve). Since the ROC curve is generally above the straight line y=x, the value is generally 0.5 to 1. The larger the AUC value (area), the better the classifier performance.

[0124] Unweighted Average Recall (UAR): If the classifier labels are unevenly distributed, traditional evaluation metrics (such as ACC, P, R, F1, etc.) will lead to overly optimistic results for the class with a large number of samples. In this case, UAR can be used as a performance metric to avoid overfitting the proposed classifier method to a certain class. Specific embodiments

[0126] Patient data: From April 2021 to December 2022, a total of 101 patients were included. According to the stage of heart failure, they were divided into group A (stage A, n=35), group B (stage B, n=26), and group C (stage C, n=40). At the same time, 29 volunteers without heart failure were included in group N (n=29). There were no significant differences in gender, body mass index (BMI), systolic blood pressure, hemoglobin value, creatinine, low-density lipoprotein cholesterol (LDL-C), history of coronary heart disease, history of hypertension, history of diabetes, history of smoking, history of drinking, and history of dyslipidemia among the three groups with different stages of heart failure. The differences in age, creatinine, troponin, N-type brain natriuretic peptide precursor, left ventricular ejection fraction, and left ventricular internal diameter among the three groups with different stages of heart failure were statistically significant (P<0.05). The specific clinical data are shown in Table 5:

[0127] Table 5

[0128] Period A (n=35) Phase B (n=26) Phase C (n=40) P Gender (%, M) 28(68％) 28(97％) 32(80％) 0.014 age 46±12 51±13 57±12 ＜0.001 BMI 25.7±3.5 26.9±3.8 25.3±4.8 0.246 Systolic blood pressure (mmHg) 139±20 141±24 121±18 ＜0.001 Hemoglobin 145±15 146±11 141±28 0.518 Troponin (ng / ml) 0.005(0.003) 0.011(0.013) 0.031(0.045) ＜0.001 NT-ProBNP (pg / ml) 31.0(45.4) 59.0(168.4) 1391.0(1754.0) ＜0.001 LDL-c (mmol / L) 2.90±1.09 2.65±1.12 2.54±1.10 0.326 Left ventricular ejection fraction (%) 67±5 66±5 41±10 ＜0.001 Left ventricular internal diameter (diastolic mm) 45±3 47±4 61±10 ＜0.001 History of coronary heart disease (%) 15(37％) 14(48％) 23(57％) 0.168 History of hypertension (%) 24(59％) 24(83％) 19(48％) 0.011 History of diabetes (%) 9(22％) 8(28％) 11(28％) 0.809 Smoking history (%) 7(17％) 11(38％) 16(40％) 0.053 History of dyslipidemia (%) 13(32％) 12(41％) 13(33％) 0.664

[0129] In this example, a total of 130 patients who met the inclusion criteria were included, with a total of 4055 voice samples and a total effective duration of 2.216 hours. Among them, 63 patients (%) were aged 30-50 years, and 67 patients (%) were aged >50 years. The heart failure groups were as follows: Group A (35 patients (%), containing 1085 voice samples, with an effective duration of 0.574 hours); Group B (26 patients (%), containing 849 voice samples, with an effective duration of 0.462 hours); Group C (40 patients (%), containing 1231 voice samples, with an effective duration of 0.715 hours); Group N (29 patients (%), containing 890 voice samples, with an effective duration of 0.465 hours); and the control group (18 patients (%), containing 890 voice samples, with an effective duration of 0.465 hours).

[0130] The speech features used in this embodiment use the eGeMAPS feature set, an extension of GeMAPS. This feature set is an 88-dimensional handcrafted feature set extracted by the openSMILE open-source toolkit. It includes 18 low-level descriptors (LLDs). Furthermore, based on GeMAPS, five spectral features (MFCCs 1-4 and spectral flux) and two frequency-related features (the bandwidths of the second and third formants) are added. These features include frequency, energy / amplitude-related features, and spectral features. In addition, 12 additional features were extracted using Python: GNE_SEO_SNR, GNE_TKEO_SNR, GNE_mean, GNE_std, VFER_SEO_SNR, VFER_TKEO_SNR, VFER_mean, VFER_std, cycle-by-cycle density entropy (RPDE), detrended fluctuation analysis (DFA), sample entropy (SampEn), and multiscale entropy (MSEn). Therefore, the feature parameters used in this study total 100 dimensions. By analyzing the 100-dimensional speech features of patients with different heart failure stages, it was found that the speech features reflecting voice roughness and breath were significantly affected by different heart failure stages.

[0131] As for voice roughness, it reflects the ability to control the glottis and vocal cords and the degree of hoarseness. The main indicators include jitter, shimmer, harmonic difference, HNR, Alpha Ratio, etc.

[0132] Jitter represents the deviation within a single continuous pitch cycle, reflecting the sound quality characteristics of the voice rhythm. The statistical parameters are mean and standard deviation. There are two features in total, and there are significant differences in different heart failure stages. The experimental results are as follows Figure 2 shown.

[0133] Shimmer represents the difference in peak amplitude between adjacent pitch cycles and also reflects the sound quality characteristics of voice rhythm. The statistics are mean and standard deviation. There are two features in total, and both heart failure types have a significant impact. The experimental results are as follows Figure 3 shown.

[0134] Harmonic difference represents H1-H2: the energy ratio of the first fundamental harmonic H1 to the second fundamental harmonic H2. The statistics are mean and standard deviation. There are 2 features in total, and both heart failure types have a significant impact. H1-A3: +The energy ratio of the first fundamental harmonic H1 to the third resonance peak A3. The statistics are mean and standard deviation. There are 2 features in total, and both heart failure types have a significant impact. The experimental results are as follows Figure 4 shown.

[0135] HNR stands for harmonic noise ratio, which is the proportion of periodically repeated harmonic components in the sound waves of steady-state vowels. The statistics are mean and standard deviation; there are two features in total, and the type of heart failure has a significant impact. The experimental results are as follows Figure 5 shown.

[0136] Alpha Ratio represents the power sum of 50-1000Hz divided by the power sum of 1-5kHz. The statistics are the mean and standard deviation of the voiced area and the mean of the unvoiced area. There are three features in total, and the type of heart failure has a significant impact. The experimental results are as follows. Figure 6 shown.

[0137] Vocal breathiness, which reflects the rhythm and intensity of the voice, includes indicators such as loudness, voiced / unvoiced duration, Hammarberg Index, and spectral slope. There are significant differences in vocal breathiness across different heart failure stages.

[0138] The voiced / unvoiced duration represents the duration of continuous voiced sounds (F0>0), and the statistics are the average length and standard deviation, with a total of 2 features. The duration of unvoiced sounds (F0=0), the statistics are the average length and standard deviation, with a total of 2 features. The number of voiced sound areas per second, with a total of 1 feature. The type of heart failure has a significant impact on all 5 features. The experimental results are as follows Figure 7 shown.

[0139] Loudness represents the statistical parameters of mean, standard deviation, 20 / 50 / 80 percentiles, 20-80 percentile range, and mean and standard deviation of the slope of rising / falling speech signals; there are 10 features in total, and different heart failure stages have significant effects. The experimental results are as follows Figure 8 shown.

[0140] The Hammarberg Index represents the strongest energy peak between 0 and 2 kHz divided by the strongest energy peak between 2 and 5 kHz.

[0141] The statistics are the mean and standard deviation of the voiced area and the mean of the unvoiced area, a total of 3 features, and the type of heart failure has a significant impact. The experimental results are as follows Figure 9 shown.

[0142] Spectral Slope represents the linear regression slope (attenuation rate) of the logarithmic power spectrum in the range of 0-500Hz and 500-1500Hz. It is the slope of the oblique spectral envelope. The larger the slope, the greater the attenuation of the signal outside the frequency division slope. The statistics are the mean and standard deviation of the voiced area and the mean of the unvoiced area. There are 6 features in total, and the type of heart failure has a significant impact on 5 features. The experimental results are as follows: Figure 10 shown.

[0143] The cross-correlation coefficient of each glottal cycle or the ratio of energy above 2.5KHz to energy below 2.5KHz. There are 4 characteristics in total, and the type of heart failure has a significant impact. The experimental results are as follows Figure 11 shown.

[0144] For nonlinear analysis, nonlinear parameters are more suitable for describing the intrinsic characteristics of acoustic signals with poor periodicity, and many heart failure speech often exhibits poor periodicity. Entropy reflects the disordered distribution of speech information in the frequency domain, while sample entropy can reflect the complexity of heart sound signals in the time domain. This study first extracted nonlinear parameters such as recurrence period density entropy (RPDE), detrended fluctuation analysis (DFA), sample entropy, and scale entropy to describe the characteristics of periodic, aperiodic, and chaotic speech signals.

[0145] Cycle Density Entropy (RPDE), Detrending Fluctuation Analysis (DFA), Sample Entropy (SampEn), and Multi-Scale Entropy (MSEn) all reflect the roughness of the sound. There are four features in total, and each has a significant impact on the type of heart failure. The experimental results are as follows: Figure 12 shown.

[0146] Mel-scale cepstral coefficients are cepstral parameters extracted from the frequency domain using the Mel scale, which describes the nonlinear characteristics of the human ear. Cepstral-based acoustic feature parameters can effectively mitigate analysis inaccuracies caused by irregularities in the fundamental frequency.

[0147] Mel cepstral coefficients 1-4. Statistics are the mean and standard deviation of the whole and voiced segments; there are 16 features in total, and the type of heart failure has a significant impact on 15 features. The experimental results are as follows Figure 13 shown.

[0148] Among them, it should be noted that Figure 2-13 The horizontal axis represents the grouping in the heart failure staging study, and the vertical axis represents the corresponding speech feature parameters, such as Figure 2 In the figure, two speech feature parameters, jitterLocal_sma3nz_ameam and jitterLocal_sma3nz_stddevNorm, are extracted for Jitter. The number of corresponding speech feature parameters extracted is consistent with that in Table 1.

[0149] In this example, based on the original 100-dimensional eGeMAPs features of the voices of patients with different heart failure stages, we further conducted a binary classification experiment, using PCA and LASSO dimensionality reduction methods, respectively. We compared six different classifiers, including support vector machine (SVM), decision tree (DT), adaptive boosting (Ada Boost), least absolute shrinkage and selection operator (LASSO), ridge regression, and elastic net. We observed the performance of different classifiers in identifying patients with different heart failure stages and identified the optimal classification model.

[0150] 1. A, B binary classification experiment (leave one out)

[0151] A binary classification recognition model was built using the heart failure stage A and stage B population as dependent variables. Based on the classification model with the original variables substituted into it, the Ada Boost classifier model was the best among the original 100-dimensional speech feature classification models, with an accuracy (Accuracy) of 0.869, a precision (P) of 0.846, a recall (R) of 0.846, and an F1 score of 0.846. After PCA dimensionality reduction, the Ada Boost classification model was the best, and the accuracy of model training using the training set was 0.738; after LASSO dimensionality reduction, the Ada Boost classification model was the best, and the accuracy of model training using the training set was 0.770, both of which were lower than the binary classification recognition model based on the original features, which may be related to the feature loss after dimensionality reduction. The optimal model, that is, the original AdaBoost, has a sample-level ROC curve as shown below. Figure 14 The results show that speech features can identify patients in stages A and B of heart failure, suggesting that speech features can be used to preliminarily screen patients at risk for heart failure and those in the early stages of heart failure with target organ damage. The binary classification accuracy (mean, standard deviation) based on the original 100-dimensional eGeMAPs features is shown in Table 6.

[0152] Table 6

[0153]

[0154] A binary classification recognition model was built with the heart failure stage A and stage B population as the dependent variable, and the classification model was based on the original variable substitution.

[0155] The AdaBoost confusion matrix is ​​shown in Table 7 and Table 8:

[0156] Table 7

[0157]

[0158]

[0159] Table 8

[0160]

[0161] The results of the classification model using principal component substitution based on PCA dimensionality reduction are shown in Table 9:

[0162] Table 9

[0163]

[0164] The AdaBoost confusion matrix is ​​shown in Table 10 and Table 11:

[0165] Table 10

[0166]

[0167] Table 11

[0168]

[0169] The results of the classification model using LASSO-based dimensionality reduction and substituting the feature variables are shown in Table 12:

[0170] Table 12

[0171]

[0172]

[0173] The AdaBoost confusion matrix is ​​shown in Table 13 and Table 14:

[0174] Table 13

[0175]

[0176] Table 14

[0177]

[0178] After LASSO dimensionality reduction, the Ada Boost classification model is the best. The accuracy of model training using the training set is 0.770, which is lower than that of the binary classification recognition model based on the original features. This may be related to the feature loss after dimensionality reduction. The importance of the original 100-dimensional AdaBoost features (where the importance of 2 dimensions is 0) is shown in Table 15.

[0179] Table 15

[0180]

[0181]

[0182] Summary of the results of the A and B binary classification experiments (leave one out)

[0183] The performance evaluation indicators of the original classification model in the test set are shown in Table 16:

[0184] Table 16

[0185]

[0186] The performance evaluation indicators of the "PCA dimensionality reduction + classification model" in the test set are shown in Table 17:

[0187] Table 17

[0188]

[0189] The performance evaluation indicators of "LASSO feature selection + classification model" in the test set are shown in Table 18:

[0190] Table 18

[0191]

[0192] The sample-level ROC curve of the optimal model (original AdaBoost) (AUC = 0.793) is as follows Figure 14 As shown in the figure, the horizontal and vertical axes represent the false positive rate and the true positive rate respectively, and the Receiver Operating Characteristic is the receiver operating curve.

[0193] 2. B, C binary classification results (leave one out)

[0194] A binary classification recognition model was built with the heart failure stage B and stage C population as the dependent variable. Based on the classification model with the original variables substituted, the Ada Boost classifier model was the best in the original 100-dimensional speech feature classification model, with an accuracy (Accuracy) of 0.788, a precision (Pricision, P) of 0.771, a recall (Recall, R) of 0.925, and an F1 score of 0.841. Through PCA dimensionality reduction, the classification models of Ada Boost and SVM were similar, and the accuracy of model training using the training set was 0.773; the accuracy of the Elastic classification model training was 0.742, both of which were lower than the binary classification recognition model based on the original features. After LASSO dimensionality reduction, the Ada Boost classification model was the best, and the accuracy of model training using the training set was 0.803, which was higher than the accuracy of the binary classification recognition model based on the original features. Figure 15As shown in the figure, the optimal model, AdaBoost with Lasso dimensionality reduction, achieved a sample-level area under the receiver operating characteristic (AUC) of 0.819. These results demonstrate that speech features can identify individuals with stage B and stage C heart failure, suggesting that speech features can be used to identify patients with pre-stage heart failure (with target organ damage) and those who have experienced symptomatic heart failure.

[0195] The binary classification accuracy (mean, standard deviation) based on the original 100-dimensional eGeMAPs features is shown in Table 19:

[0196] Table 19

[0197]

[0198] The AdaBoost confusion matrix is ​​shown in Table 20 and Table 21:

[0199] Table 20

[0200]

[0201] Table 21

[0202]

[0203] Based on PCA dimensionality reduction, the classification model using principal component substitution is shown in Table 22:

[0204] Table 22

[0205]

[0206] The AdaBoost confusion matrix is ​​shown in Table 23 and Table 24:

[0207] Table 23

[0208]

[0209] Table 24

[0210]

[0211] The SVM confusion matrix is ​​shown in Table 25 and Table 26:

[0212] Table 25

[0213]

[0214] Table 26

[0215]

[0216] Based on LASSO dimensionality reduction, the classification model in which the feature variables are substituted is shown in Table 27:

[0217] Table 27

[0218]

[0219] The AdaBoost confusion matrix is ​​shown in Table 28 and Table 29:

[0220] Table 28

[0221]

[0222] Table 29

[0223]

[0224] The features with LASSO regularization coefficients not equal to 0 are 66 in total, and 34 of them are 0, as shown in Table 30:

[0225] Table 30

[0226]

[0227]

[0228]

[0229] B, C binary classification experiment (leave one out method) results summary

[0230] The performance evaluation indicators of the original classification model in the test set are shown in Table 31:

[0231] Table 31

[0232]

[0233] The performance evaluation indicators of the "PCA dimensionality reduction + classification model" in the test set are shown in Table 32:

[0234] Table 32

[0235]

[0236] The performance evaluation indicators of the "LASSO feature selection + classification model" in the test set are shown in Table 33:

[0237] Table 33

[0238]

[0239] The sample-level ROC curve of the optimal model (original AdaBoost) (AUC = 0.793) is as follows Figure 15 As shown in the figure, the horizontal and vertical axes represent the false positive rate and the true positive rate respectively, and the Receiver Operating Characteristic is the receiver operating curve.

[0240] 3. AB, C binary classification experiment (leave one out)

[0241] A binary classification recognition model was built using patients with heart failure in stages A and B (AB) and C as dependent variables. Based on the classification model with the original variables substituted, the Ada Boost classifier model was the best in the original 100-dimensional speech feature classification model, with an accuracy (Accuracy) of 0.802, a precision (P) of 0.857, a recall (R) of 0.600, and an F1 score of 0.706. After PCA dimensionality reduction, the SVM classification model was the best, but the accuracy of the model training using the training set dropped to 0.723; after LASSO dimensionality reduction, the Ada Boost classification model was the best, and the accuracy of the model training using the training set was 0.812, which was higher than the accuracy of the binary classification recognition model based on the original features. Figure 16 As shown, the optimal model, AdaBoost with Lasso dimensionality reduction, achieved a sample-level area under the receiver operating characteristic (AUC) of 0.731. These results demonstrate that speech features can distinguish patients with pre-stage heart failure (stage B) at risk for heart failure (stage A) and with target organ damage from patients with stage C who have experienced symptomatic heart failure. This indicates that speech features can help identify patients who have previously experienced symptomatic heart failure. Table 34 shows the binary classification accuracy (mean, standard deviation) based on the original 100-dimensional eGeMAPs features.

[0242] Table 34

[0243]

[0244] The AdaBoost confusion matrix is ​​shown in Table 35 and Table 36:

[0245] Table 35

[0246]

[0247] Table 36

[0248]

[0249] Based on PCA dimensionality reduction, the classification model using principal component substitution is shown in Table 37:

[0250] Table 37

[0251]

[0252] The SVM confusion matrix is ​​shown in Table 38 and Table 39:

[0253] Table 38

[0254]

[0255] Table 39

[0256]

[0257] The results of dimensionality reduction based on LASSO are shown in Table 40:

[0258] Table 40

[0259]

[0260] The AdaBoost confusion matrix is ​​shown in Table 41 and Table 42:

[0261] Table 41

[0262]

[0263] Table 42

[0264]

[0265] The features with LASSO regularization coefficients not equal to 0 have a total of 56 dimensions, and 44 of them are 0, as shown in Table 43:

[0266] Table 43

[0267]

[0268]

[0269] AB, C binary classification results (leave one out)

[0270] The performance evaluation indicators of the original classification model in the test set are shown in Table 44:

[0271] Table 44

[0272]

[0273]

[0274] The performance evaluation indicators of the "PCA dimensionality reduction + classification model" in the test set are shown in Table 45:

[0275] Table 45

[0276]

[0277] The performance evaluation indicators of the "LASSO feature selection + classification model" in the test set are shown in Table 46:

[0278] Table 46

[0279]

[0280] The sample-level ROC curve of the optimal model (original AdaBoost) (AUC = 0.793) is as follows Figure 16 As shown in the figure, the horizontal and vertical axes represent the false positive rate and the true positive rate respectively, and the Receiver Operating Characteristic is the receiver operating curve.

[0281] The above experiments have shown that (1) patients with different heart failure stages have different voice characteristics. Based on the eGeMAPS feature set and a total of 100-dimensional features extracted by python, the main indicators reflecting the roughness of the voice of patients with different heart failure stages, such as jitter, shimmer, harmonic difference, HNR, alpha ratio, etc., are different between different heart failure stages. (2) The main indicators reflecting the breathiness of the voice, including loudness, voiced / unvoiced duration, Hammarberg Index, spectral slope, etc., also have significant differences between different heart failure stages. Among these sound indicators, the basic characteristics of the voice, including frequency, energy / amplitude related characteristics, nonlinearity and other characteristics, are different between different heart failure stages. (3) The contribution of speech features in different stages is not exactly the same. (4) Based on the original 100-dimensional eGeMAPs features of the voices of patients with heart failure at different stages, a classification model was constructed using a binary classification method. The performance of the original variables, PCA dimensionality reduction, LASSO dimensionality reduction, and different classifiers in identifying patients with heart failure at different stages was compared, and the classification model was optimized. The optimal model for identifying patients with heart failure at stage A and stage B was the AdaBoost classification method based on the original variables, with an ROC curve AUC of 0.793; the optimal model for identifying patients with heart failure at stage B and stage C was the AdaBoost classification method based on Lasso dimensionality reduction, with an ROC curve AUC of 0.819; the optimal model for distinguishing between stage AB and stage C was the AdaBoost classification method based on Lasso dimensionality reduction, with an ROC curve AUC of 0.731.

[0282] The present invention also provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the following method when executing the computer program: converting the collected voice analog signal into a voice digital signal, preprocessing the voice digital signal, and extracting features from the preprocessed voice digital signal to obtain multiple categories of voice feature samples; constructing a classification model for identifying heart failure stages, and training and optimizing the classification model using the multiple categories of voice feature samples to obtain an optimal classification model, wherein the classification model for identifying heart failure stage A and heart failure stage B is an AdaBoost classification model based on original variables; the classification model for identifying heart failure stage B and stage C is an AdaBoost classification model based on Lasso dimensionality reduction; and the classification model for identifying heart failure stages A, B, and C is an AdaBoost classification model based on Lasso dimensionality reduction.

[0283] The present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the following method is implemented: converting the collected voice analog signal into a voice digital signal, preprocessing the voice digital signal, and extracting features from the preprocessed voice digital signal to obtain multiple categories of voice feature samples; constructing a classification model for identifying heart failure stages, training and optimizing the classification model using the multiple categories of voice feature samples to obtain an optimal classification model, wherein the classification model for identifying heart failure stage A and heart failure stage B is an AdaBoost classification model based on original variables; the classification model for identifying heart failure stage B and stage C is an AdaBoost classification model based on Lasso dimensionality reduction; and the classification model for identifying heart failure stages A, B and C is an AdaBoost classification model based on Lasso dimensionality reduction.

[0284] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same and similar parts between the various embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description. It should be pointed out that for ordinary technicians in this technical field, without departing from the principle of the present invention, several improvements and modifications can be made to the present invention, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

[0285] It should also be noted that, in this specification, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus comprising the element.

Claims

1. A device for constructing a classification model for identifying different stages of heart failure based on sound features, characterized in that: include: A sample processing unit, configured to convert the collected voice analog signal into a voice digital signal, preprocess the voice digital signal, and extract features from the preprocessed voice digital signal to obtain multiple types of voice feature samples; A model training unit is used to construct a classification model for identifying heart failure stages, and to train and optimize the classification model using the multi-class speech feature samples to obtain an optimal classification model, wherein the classification model for identifying heart failure stage A and heart failure stage B is an AdaBoost classification model based on the original variables; the classification model for identifying heart failure stage B and stage C is an AdaBoost classification model based on Lasso dimensionality reduction; and the classification model for identifying heart failure stages A, B, and C is an AdaBoost classification model based on Lasso dimensionality reduction; An evaluation unit, configured to perform model evaluation on the classification model according to a holdout method; The sample processing unit includes: An extraction unit, configured to extract a multi-dimensional first speech feature sample using the openSMILE open source toolkit, and to extract a multi-dimensional second speech feature sample using Python; a merging unit, configured to merge the multi-dimensional first speech feature sample and the multi-dimensional second speech feature sample to obtain multi-category speech feature samples; The first speech feature samples include: pitch feature, frequency perturbation feature, formant feature, amplitude perturbation feature, loudness feature, harmonic-to-noise ratio feature, harmonic difference feature, alpha ratio feature, Hammarberg coefficient feature, spectral slope feature, Mel-cephalogram coefficient feature, spectral flow feature, loudness peak ratio feature, continuous sound area and silent area feature, equivalent sound level feature; the second speech feature samples include: glottal noise excitation ratio feature, vocal cord excitation ratio feature, cycle density entropy feature, detrending fluctuation analysis feature, sample entropy feature, and multi-scale entropy feature.

2. The device for constructing a classification model for identifying different stages of heart failure based on sound features according to claim 1, characterized in that: The sample processing unit includes: A first conversion unit is used to convert the time-continuous voice analog signal into a time-discrete and amplitude-continuous signal sample according to a predetermined sampling period; A second conversion unit is used to convert each signal sample with continuous amplitude values ​​into a discrete value and express it in binary to obtain digital data; The third conversion unit is used to convert the digital data into a binary code stream to obtain a voice digital signal.

3. The device for constructing a classification model for identifying different stages of heart failure based on sound features according to claim 1, characterized in that: The sample processing unit further includes: A synchronization unit, configured to synchronize multiple voice digital signals to a uniform sampling rate by using a downsampling method; An endpoint detection unit is used to perform endpoint detection on the voice digital signal after the unified sampling rate, and distinguish between the voice area and the non-voice area; an emphasis unit, configured to emphasize the high frequency portion of the speech region to increase the high frequency resolution of the speech; The framing and windowing unit is used to perform framing and windowing on the emphasized speech area to obtain multiple speech signal segments.

4. The device for constructing a classification model for identifying different stages of heart failure based on sound features according to claim 1, characterized in that: Also includes: A performance measurement unit is used to measure the performance of the classification model according to predetermined indicators, wherein the predetermined indicators include error rate and accuracy, precision and recall, F1 value, specificity, sensitivity, ROC curve and AUC, and unweighted average recall.

5. A computer device, characterized in that: The invention comprises a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the following method is implemented: converting a collected voice analog signal into a voice digital signal, preprocessing the voice digital signal, and extracting features from the preprocessed voice digital signal to obtain multiple types of voice feature samples; Constructing a classification model for identifying heart failure stages, training and optimizing the classification model using the multi-class speech feature samples to obtain an optimal classification model, wherein the classification model for identifying heart failure stage A and heart failure stage B is an AdaBoost classification model based on original variables; the classification model for identifying heart failure stages B and C is an AdaBoost classification model based on Lasso dimensionality reduction; and the classification model for identifying heart failure stages A, B, and C is an AdaBoost classification model based on Lasso dimensionality reduction; and performing model evaluation on the classification models according to a holdout method; The converting the collected voice analog signal into a voice digital signal, preprocessing the voice digital signal, and extracting features from the preprocessed voice digital signal to obtain a multi-class voice feature sample comprises: extracting a multi-dimensional first voice feature sample using an openSMILE open source toolkit, and extracting a multi-dimensional second voice feature sample using Python; and merging the multi-dimensional first voice feature sample and the multi-dimensional second voice feature sample to obtain a multi-class voice feature sample; The first speech feature samples include: pitch feature, frequency perturbation feature, formant feature, amplitude perturbation feature, loudness feature, harmonic-to-noise ratio feature, harmonic difference feature, alpha ratio feature, Hammarberg coefficient feature, spectral slope feature, Mel-cephalogram coefficient feature, spectral flow feature, loudness peak ratio feature, continuous sound area and silent area feature, equivalent sound level feature; the second speech feature samples include: glottal noise excitation ratio feature, vocal cord excitation ratio feature, cycle density entropy feature, detrending fluctuation analysis feature, sample entropy feature, and multi-scale entropy feature.

6. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, the computer program implements the following method: converting the collected voice analog signal into a voice digital signal, preprocessing the voice digital signal, and extracting features from the preprocessed voice digital signal to obtain multiple types of voice feature samples; Constructing a classification model for identifying heart failure stages, training and optimizing the classification model using the multi-class speech feature samples to obtain an optimal classification model, wherein the classification model for identifying heart failure stage A and heart failure stage B is an AdaBoost classification model based on original variables; the classification model for identifying heart failure stages B and C is an AdaBoost classification model based on Lasso dimensionality reduction; and the classification model for identifying heart failure stages A, B, and C is an AdaBoost classification model based on Lasso dimensionality reduction; and performing model evaluation on the classification models according to a holdout method; The converting the collected voice analog signal into a voice digital signal, preprocessing the voice digital signal, and extracting features from the preprocessed voice digital signal to obtain a multi-class voice feature sample comprises: extracting a multi-dimensional first voice feature sample using an openSMILE open source toolkit, and extracting a multi-dimensional second voice feature sample using Python; and merging the multi-dimensional first voice feature sample and the multi-dimensional second voice feature sample to obtain a multi-class voice feature sample; The first speech feature samples include: pitch feature, frequency perturbation feature, formant feature, amplitude perturbation feature, loudness feature, harmonic-to-noise ratio feature, harmonic difference feature, alpha ratio feature, Hammarberg coefficient feature, spectral slope feature, Mel-cephalogram coefficient feature, spectral flow feature, loudness peak ratio feature, continuous sound area and silent area feature, equivalent sound level feature; the second speech feature samples include: glottal noise excitation ratio feature, vocal cord excitation ratio feature, cycle density entropy feature, detrending fluctuation analysis feature, sample entropy feature, and multi-scale entropy feature.

Images