Ballistocardiogram atrial fibrillation auxiliary diagnosis system based on signal transformation and random convolution kernel

Abstract

The invention belongs to the technical field of medical instruments, and particularly relates to a ballistocardiogram atrial fibrillation auxiliary diagnosis system based on signal transformation and a random convolution kernel. The system comprises a signal preprocessing module, a signal conversion module, a feature extraction module and a classification module. According to the method, firstly, BCG (ballistocardiogram) signals are subjected to normalization processing, noise is filtered out, then, autocorrelation transformation and energy signal transformation are carried out on each section of signals to achieve data enhancement, and convolution processing is carried out on the autocorrelation signals and the energy signals through 3000 random convolution kernels; dimensionality reduction is carried out on each sequence obtained through convolution by using maximum value pooling and positive value proportion pooling, and finally automatic diagnosis of atrial fibrillation and sinus rhythm is realized through a logistic regression classifier. According to the method, the problem of difficulty in automatic identification of atrial fibrillation caused by BCG waveform variation and individual difference can be effectively solved, and a solid technical support is provided for a long-term heart health monitoring system in a daily life scene.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical instruments, and particularly relates to a ballistocardiogram atrial fibrillation auxiliary diagnosis system based on signal transformation and random convolution kernels. Background Art

[0002] Atrial fibrillation (AF) is the arrhythmia disease with the highest morbidity and mortality globally. The clinical manifestations of AF patients vary, ranging from asymptomatic to severe palpitations, dyspnea, chest pain, and even syncope. Moreover, with the aging of the population and the increase in chronic diseases, the incidence of AF shows an upward trend year by year. Many AF patients may have no obvious symptoms, so AF is often called the "invisible killer". It is estimated that about 15% of AF patients are still not diagnosed in time, which constitutes a huge potential problem and brings significant medical and economic burdens to society and families. Therefore, AF monitoring based on non-contact sensing technology has become a new research hotspot.

[0003] Currently, traditional AF detection still relies on surface electrocardiogram (ECG). However, traditional ECG detection usually needs to be carried out in a hospital or clinic by professionals and requires attaching electrodes to the patient's body surface. This limitation affects its application and promotion in the fields of long-term monitoring and home health management. The BCG sensor does not need to directly contact the skin and can be integrated into a mattress, seat cushion, or wearable device. The signal acquisition process does not interfere with the user and can continuously record cardiac activities during sleep or daily activities, which is particularly suitable for capturing intermittent AF. Currently, the AF detection algorithms based on BCG mainly include traditional machine learning and deep learning based on feature extraction. Among them, the traditional machine learning methods based on feature extraction often need to rely on the experience of researchers to manually design features, which is subjective. Deep learning often brings an increase in model complexity, not only increasing the computational cost but also possibly introducing unnecessary error sources. Currently, AF detection based on BCG still faces many challenges. First, the datasets used in the research are relatively small. Second, individual differences, as an important feature of BCG signals, cannot be ignored in their impact on BCG signals. Finally, periodic features, as the key indicators for distinguishing atrial fibrillation and sinus rhythm, have not been given enough attention in some previous studies. In summary, it is necessary to design an atrial fibrillation auxiliary diagnosis algorithm suitable for BCG signals to provide strong support for the early screening and management of cardiovascular diseases. Summary of the Invention

[0004] In view of the above requirements for AF detection algorithms and the technical deficiencies of existing BCG research, the present invention provides a BCG atrial fibrillation auxiliary diagnosis system based on signal transformation and random convolution kernels to solve the problem of difficult atrial fibrillation recognition caused by problems such as BCG waveform diversity and individual differences, avoid manual feature extraction, overcome the complex calculations of deep models, and utilize the periodic rhythm of sinus rhythm as an important feature for distinguishing sinus rhythm and atrial fibrillation.

[0005] The atrial fibrillation assisted diagnosis system of ballistocardiogram based on signal transformation and random convolution kernels provided by the present invention realizes the signal transformation of BCG through two data augmentation techniques of autocorrelation transformation and energy signal transformation, combines random convolution kernels to extract signal features, and uses two methods of maximum value and positive value ratio for feature selection, and uses a logistic regression model to realize the discrimination between atrial fibrillation and sinus rhythm; specifically includes: a signal preprocessing module, a signal transformation module, a feature extraction module and a classification module, wherein:

[0006] The signal preprocessing module is used to preprocess the collected BCG signal; specifically, it includes segmenting the collected BCG into 30-second segments, then normalizing the signal to the same range using maximum-minimum normalization, and then selecting the mother wavelet Daubechie 6 to decompose the original BCG signal, filtering out low-frequency signals irrelevant to heartbeat information, and selecting the 0.6-15Hz frequency band part related to heartbeat for signal reconstruction to obtain a filtered signal.

[0007] The signal transformation module is used to perform autocorrelation transformation and energy signal transformation on the BCG signal obtained by the signal preprocessing module to obtain its autocorrelation signal and energy signal respectively; wherein:

[0008] The acquisition method of the autocorrelation signal transformation is: obtaining the autocorrelation signal of the BCG signal through autocorrelation transformation to reveal the regularity and periodicity of cardiac activity;

[0009] The acquisition method of the energy signal transformation is: first squaring the BCG signal to obtain an unsmoothed energy signal, and then using a time-domain mean filter for smoothing to obtain the final energy signal;

[0010] Input the above-mentioned autocorrelation signal and energy signal into the feature extraction module respectively;

[0011] The feature extraction module includes two parts: random convolution kernels and pooling, and is used to extract features from the autocorrelation signal and energy signal. Specifically, first convolve each autocorrelation signal or energy signal with k convolution kernels to obtain k convolution sequences, and then use maximum pooling and positive value ratio pooling to reduce the dimension of each convolution sequence, so as to extract two features of maximum value and positive value ratio in each convolution sequence. Therefore, k maximum values and k positive value ratios, a total of 2k features, can be extracted from each autocorrelation signal or energy signal;

[0012] The classification module uses logistic regression as a classifier and introduces L2 regularization and cross-entropy loss function;

[0013] Furthermore:

[0014] In the signal transformation module, two parallel BCG signals, namely the autocorrelation signal and the energy signal, are obtained through signal transformation. The specific method is as follows:

[0015] (1) Perform autocorrelation transformation on the preprocessed BCG signal to obtain its autocorrelation signal x acf , and the specific method is as follows:

[0016] According to formula (1), perform autocorrelation transformation on the preprocessed discrete BCG signal x wave [n]:

[0017]

[0018] where and σ represent the mean value and standard deviation of the BCG time series x wave respectively, and N is the length of the BCG time series x wave ;

[0019] (2) Convert the preprocessed BCG signal into its energy signal E[n]. The specific method is as follows:

[0020] ① According to formula (2), perform energy signal transformation on the preprocessed discrete BCG signal x wave [n] to obtain the unsmoothed BCG energy signal;

[0021]

[0022] where W[k] is a rectangular sliding window with a window length of W1;

[0023] ② According to formula (3), use a mean filter with a window length of W2 to smooth E1[n] to obtain the final energy signal E[n];

[0024]

[0025] (3) Input the obtained autocorrelation signal and energy signal into the subsequent feature extraction module respectively.

[0026] In the feature extraction module, the random convolution kernel and double pooling method are applied to extract features from the autocorrelation signal and the energy signal. The specific process is as follows:

[0027] (1) Randomly generate k random convolution kernels, each with a kernel length of l kernel , a weight of ω, a bias of b, a dilation coefficient of d, and a padding of p, a total of 5 parameters;

[0028] (2) Perform random convolution processing on the autocorrelation signal and energy signal obtained by the signal conversion module. According to formula (4), each autocorrelation signal or energy signal is convolved with the k random convolution kernels obtained in step (1):

[0029]

[0030] Where X represents the autocorrelation signal x acf And the energy signal E[n], * represents convolution;

[0031] (3) Use two pooling methods, maximum value and positive value ratio, to reduce the dimension of the signal obtained after convolving the BCG signal. The pooling size is the length of the signal obtained after convolution. Among them, the maximum value pooling selects the maximum value in the convolved sequence as the feature, and the positive value ratio selects the ratio of the number of positive elements to the total number of elements in the convolved sequence as the feature according to formula (5).

[0032]

[0033] Where M is the total number of elements in the selected convolved sequence.

[0034] Each parameter is randomly selected according to the following rules, and a total of k groups are selected to generate k random convolution kernels for subsequent convolution:

[0035] Kernel length (l kernel ); Randomly selected from (3, 6, 12, 24);

[0036] Weight (ω): Where Represents a normal distribution with a mean of 0 and a variance of 1;

[0037] Bias (b): Where Represents a uniform distribution with a mean of (-1, 1);

[0038] Dilation coefficient (d): Where Represents rounding down; Represents a uniform distribution of (0, A), l input Represents the length of the input signal;

[0039] Padding (p): Or p = 0.

[0040] In the present invention, the classification module uses logistic regression as the output layer, introduces L2 regularization to effectively prevent the model from overfitting, and evaluates the pros and cons with the cross-entropy loss function. Description of the Drawings

[0041] Figure 1System block diagram of an embodiment of the present invention.

[0042] Figure 2 Measured atrial fibrillation BCG signals and sinus rhythm BCG signals of an embodiment of the present invention and the autocorrelation signals obtained by their conversion. In the figure, (a), (b), (c), and (d) are the original signals and their autocorrelation signals of 4 different subjects respectively.

[0043] Figure 3 Measured atrial fibrillation BCG signals and sinus rhythm BCG signals of an embodiment of the present invention and the energy signals obtained by their conversion. In the figure, (a), (b), (c), and (d) are the original signals and their energy signals of 4 different subjects respectively.

[0044] Figure 4 Feature extraction module and classification module of an embodiment of the present invention. Detailed implementation manners

[0045] The following further describes specific embodiments of the present invention with reference to the accompanying drawings. The following embodiments are used to illustrate the present invention, but not to limit the scope of the rights of the present invention.

[0046] In this embodiment, an atrial fibrillation assisted diagnosis system for ballistocardiogram based on signal transformation and random convolution kernels, as Figure 1 shown, includes the following steps:

[0047] Step 1: The signal preprocessing module preprocesses the signal to obtain a pure BCG signal.

[0048] Step 1.1: Collect BCG signals of atrial fibrillation subjects and sinus rhythm subjects during nighttime sleep through a BCG acquisition system;

[0049] Step 1.2: Manually extract atrial fibrillation and sinus rhythm signals in the BCG signal;

[0050] Step 1.3: Divide all signals into non-overlapping data segments with a length of 30 seconds;

[0051] Step 1.4: Use maximum-minimum normalization for the segmented data to normalize all data to the range of [0, 1];

[0052] Step 1.5: Apply wavelet transform to filter out noise. Use Daubechie 6 as the mother wavelet to perform 7-layer wavelet decomposition on the BCG signal to obtain 7 detail components (D1 - D7) and 1 approximation component (A7), and use detail components D3 to D6 to reconstruct the BCG signal. The reconstructed signal is denoted as x wave .

[0053] Step 2: The signal transformation module performs signal transformation on the preprocessed signal to transform it into an autocorrelation signal and an energy signal, asFigure 2 and Figure 3 as shown

[0054] Step 2.1: Transform the preprocessed BCG signal into its autocorrelation signal x acf , specifically as follows:

[0055] According to formula (1), perform autocorrelation transformation on the preprocessed discrete BCG signal x wave [n], and σ represent the mean and standard deviation of the BCG time series x wave respectively, N is the length of the BCG time series x wave . Since the BCG signal sampling rate is 125 Hz and the duration of each BCG signal segment is 30 seconds, the value of N is set to 3750 here;

[0056] Step 2.2: Convert the preprocessed BCG signal into its energy signal E[n], and the specific method is as follows:

[0057] ① According to formula (2), perform energy signal transformation on the preprocessed discrete BCG signal x wave [n] to obtain the unsmoothed BCG energy signal. Among them, W[k] is a rectangular sliding window, and its window length is W1, and W1 is set to 30;

[0058] ② According to formula (3), use a mean filter with a window length of W2 to smooth E1[n] to obtain the final energy signal E[n]. Here, W2 is set to 30;

[0059]

[0060] Step 3: The feature extraction module extracts features for subsequent classification, as Figure 4 shown

[0061] Step 3.1: Randomly generate 3000 random convolution kernels, and the kernel length of each convolution kernel is l kernel , with weight ω, bias b, dilation coefficient d, and padding p, a total of 5 parameters; each parameter is randomly selected according to the following rules, and a total of 3000 groups are selected. The 3000 randomly generated convolution kernels are used for subsequent convolution:

[0062] l kernel , the kernel length, is randomly selected from (3, 6, 12, 24);

[0063] ω, the weight, where represents a normal distribution with a mean of 0 and a variance of 1;

[0064] b, the bias; Among them, represents the mean distribution of (-1, 1);

[0065] d, the expansion coefficient; Among them, represents rounding down; represents the mean distribution of (0, A), l input represents the length of the input signal;

[0066] p, padding; or p = 0.

[0067] Step 3.2: Perform random convolution processing on the autocorrelation signal and energy signal obtained by the signal conversion module. According to formula (4), each autocorrelation signal or energy signal is convolved with the 3000 random convolution kernels obtained in Step 3.1. Among them, X represents the autocorrelation signal x acf and the energy signal E[n], and * represents convolution;

[0068] Step 3.3: Use two pooling methods, maximum pooling and positive value ratio pooling, to reduce the dimension of the signal obtained after convolving the BCG signal. The pooling size is the length of the signal obtained after convolution. Among them, maximum pooling selects the maximum value in the convolved sequence as the feature, and positive value ratio pooling selects the ratio of the number of positive elements to the total number of elements in the convolved sequence as the feature according to formula (5). The maximum value feature and the positive value ratio feature are combined for subsequent classification.

[0069]

[0070] where M is the total number of elements in the selected convolved sequence.

[0071] Step 4: Use logistic regression as the output layer, introduce L2 regularization to effectively prevent the model from overfitting, and evaluate the quality with the cross-entropy loss function. The output is to determine whether the sample is atrial fibrillation or sinus rhythm.

[0072] According to this embodiment, cross-patient five-fold cross-validation is performed on 22,251 non-overlapping 30-second BCG signals of 50 subjects, and the obtained classification accuracy is 96.1% ± 1.9%. The precision rate for atrial fibrillation recognition reaches 96.5%, the recall rate reaches 97.3%, and the F1 score reaches 96.9%.

[0073] The above is the preferred embodiment of the present invention, but the protection scope of the present invention is not limited to this example only. Any other examples made by those skilled in the art within the technical scope disclosed by the present invention through simple replacement and substitution should be covered within the protection scope of the present invention. The protection scope of the present invention is subject to the protection scope defined by the claims.

Claims

1. An atrial fibrillation assisted diagnosis system for ballistocardiogram based on signal transformation and random convolution kernels, characterized in that It includes a signal preprocessing module, a signal transformation module, a feature extraction module, and a classification module, where: The signal preprocessing module is used to preprocess the collected BCG signal; specifically, it includes segmenting the collected BCG into 30 - second segments, then normalizing the signal to the same range using max - min normalization, and then selecting the mother wavelet Daubechie 6 to decompose the original BCG signal, filtering out the low - frequency signals unrelated to the heartbeat information, and selecting the 0.6 - 15 Hz frequency band part related to the heartbeat for signal reconstruction to obtain the filtered signal; The signal transformation module is used to perform autocorrelation transformation and energy signal transformation on the BCG signal obtained by the signal preprocessing module, respectively obtaining its autocorrelation signal and energy signal, where: The method for obtaining the autocorrelation signal transformation is: obtaining the autocorrelation signal of the BCG signal through autocorrelation transformation to reveal the regularity and periodicity of cardiac activity; The method for obtaining the energy signal transformation is: first squaring the BCG signal to obtain the unsmoothed energy signal, and then using a time - domain mean filter for smoothing to obtain the final energy signal; Input the above - mentioned autocorrelation signal and energy signal into the feature extraction module respectively; The feature extraction module includes two parts: a random convolution kernel and pooling, and is used to extract features from the autocorrelation signal and energy signal; first, convolve each autocorrelation signal or energy signal with k convolution kernels to obtain k convolution sequences, and then use max - pooling and positive - value ratio pooling to reduce the dimension of each convolution sequence, so as to extract two features: the maximum value and the positive - value ratio in each convolution sequence; thus, k maximum values and k positive - value ratios, a total of 2k features, are extracted from each autocorrelation signal or energy signal; The classification module uses logistic regression as the classifier and introduces L2 regularization and cross - entropy loss function.

2. The ballistocardiogram atrial fibrillation auxiliary diagnosis system according to claim 1, wherein In the signal transformation module, the signal transformation obtains two parallel BCG signals - the autocorrelation signal and the energy signal. The specific method is: (1) Perform autocorrelation transformation on the preprocessed BCG signal to obtain its autocorrelation signal x acf , and the calculation formula is as follows: Among them, and σ respectively represent the mean value and standard deviation of the BCG time series x wave , N is the length of the BCG time series x wave ; (2) Convert the preprocessed BCG signal into its energy signal E[n]. The specific method is as follows: ①Perform an energy signal transformation on the preprocessed discrete BCG signal x wave [n] to obtain the unsmoothed BCG energy signal; the calculation formula is as follows: where W[k] is a rectangular sliding window with a window length of W1; ② Smooth E1[n] using a mean filter with a window length of W2 to obtain the final energy signal E[n]. The calculation formula is: (3) Input the obtained autocorrelation signal and energy signal into the subsequent feature extraction module respectively.

3. The ballistocardiogram atrial fibrillation assisted diagnosis system according to claim 1, characterized in that In the feature extraction module, a random convolution kernel and a dual - pooling method are applied to extract features from the autocorrelation signal and energy signal. The specific process is: (1) Randomly generate k random convolutional kernels, each with a kernel length of l kernel , with weight ω, bias b, dilation coefficient d, and padding p; (2) Perform random convolution processing on the autocorrelation signal and energy signal obtained by the signal transformation module, that is, convolve each autocorrelation signal or energy signal with the k random convolution kernels obtained in step (1): The calculation formula is: where X represents the autocorrelation signal x acf and the energy signal E[n], and * represents convolution; (3)Dimensionality reduction is performed on the signal obtained after convolving the BCG signal using two pooling methods: maximum value pooling and positive value ratio pooling. The pooling size is the length of the signal obtained after convolution. Among them, the maximum value pooling selects the maximum value in the convolved sequence as the feature, and the positive value ratio selects the ratio of the number of positive elements to the total number of elements in the convolved sequence as the feature according to formula (5); Where M is the total number of elements selected in the convolved sequence.

4. The ballistocardiogram atrial fibrillation assisted diagnosis system according to claim 3, wherein: l kernel : randomly selected from (3, 6, 12, 24); ω: represents a normal distribution with a mean of 0 and a variance of 1; b: Represents the mean distribution of (-1, 1); d: Denotes rounding down; Denotes the mean distribution of (0, A), l input Denotes the length of the input signal; p: or p = 0.

5. The ballistocardiogram atrial fibrillation assisted diagnosis system according to claim 1, wherein The classification module uses logistic regression as the output layer, introduces L2 regularization to effectively prevent model overfitting, and evaluates the advantages and disadvantages using the cross-entropy loss function.

Images