Millimeter wave radar heart failure diagnosis system based on cnn and lstm sequence regression hybrid model

Abstract

The application discloses a millimeter wave radar heart failure diagnosis system based on a CNN and LSTM ordinal regression hybrid model, and comprises the following modules: a radar signal processing module, which is used for acquiring radar signals and forming a slow time-distance complex matrix; a vital sign signal extraction module, which is used for obtaining a vital sign time sequence signal from the complex matrix; a heartbeat waveform separation module, which is used for generating a filtering interval according to a heart rate and extracting a heartbeat vibration waveform; a deep feature learning module, which is used for constructing a feature extraction network for the heartbeat vibration waveform and outputting a continuous score prediction; and an ordinal regression diagnosis module, which is used for establishing a learnable threshold based on the continuous score prediction and generating an ejection function and diastolic function grading result. The application realizes accurate modeling of non-contact heart failure grading diagnosis of the millimeter wave radar, and improves the convenience and accuracy of heart function evaluation.

Description

Technical Field

[0001] This invention relates to the field of signal processing technology, and in particular to a millimeter-wave radar heart failure diagnostic system based on a hybrid model of CNN and LSTM ordinal regression. Background Technology

[0002] Heart failure is an end-stage syndrome caused by impaired cardiac filling or ejection function, leading to decreased cardiac output and pulmonary / systemic congestion. Early, continuous monitoring can provide early warning before irreversible damage occurs.

[0003] Existing home self-testing methods are highly subjective and have low specificity; biomarkers and imaging rely on equipment and have poor continuity; wearable devices have limited parameters and poor compliance; and invasive monitoring carries high risks and costs.

[0004] FMCW millimeter-wave radar boasts advantages such as all-weather operation, non-contact operation, and privacy protection. It can acquire respiratory and cardiac signals through transmit-echo micro-displacement demodulation, extracting cardiac vibration patterns for heart failure diagnosis. To this end, a "millimeter-wave radar heart failure diagnostic system based on a hybrid model of CNN and LSTM ordinal regression" is proposed. This system uses joint modeling of deep features and temporal dependencies to classify and assess cardiac vibration patterns. The radar is designed for bedside, ward, and home applications, requiring no wearing or patch, thus improving compliance. The system employs multi-domain features and robust preprocessing to address posture and occlusion interference. Combined with ordinal regression corresponding to NYHA / risk stratification, it reduces false alarms and missed alarms. Furthermore, it supports long-term trend tracking and individualized threshold settings, providing objective evidence for follow-up and remote management. Summary of the Invention

[0005] One objective of this invention is to propose a millimeter-wave radar heart failure diagnostic system based on a hybrid model of CNN and LSTM ordinal regression. This invention constructs a vital sign extraction algorithm, a heartbeat separation algorithm, a deep feature learning structure, and an ordinal regression grading mechanism to obtain the grading of cardiac systolic and diastolic functions from millimeter-wave radar vibration signals. It has the advantages of non-contact measurement, precise modeling, high grading accuracy, and strong clinical scalability.

[0006] A millimeter-wave radar heart failure diagnostic system based on a hybrid model of CNN and LSTM ordinal regression according to an embodiment of the present invention includes:

[0007] The radar signal processing module acquires radar signals, performs mixing, sampling and mapping, and constructs a slow-time-range complex matrix.

[0008] The vital signs signal extraction module is used to select the slow-time complex sequence of the target corresponding range cell from the slow-time-range complex matrix, perform arctangent deharmonic phase unwinding processing, and generate vital signs time series signals.

[0009] The heartbeat waveform separation module is used to perform heart rate calculation on the vital sign time sequence signal. It extracts the high-frequency components of the heartbeat based on bandpass filtering, generates the heart rate value using the envelope autocorrelation method, and performs passband filtering and multi-band bandpass filtering based on the heart rate value to extract the heartbeat vibration waveform.

[0010] The deep feature learning module receives heartbeat vibration waveforms, extracts short-term local features through a convolutional neural network structure, extracts long-term dependencies through a long short-term memory network structure, and concatenates them into the output of the deep learning network to form a continuous score prediction.

[0011] The ordinal regression diagnostic module receives continuous score predictions, divides the output space of continuous feature representations through learnable thresholds, and generates systolic function grading results and diastolic function grading results.

[0012] Optionally, modules can be integrated using the following methods:

[0013] S1. Acquire radar signals, perform mixing, sampling and mapping, and construct a slow time-range complex matrix;

[0014] S2. Select the slow-time complex sequence of the target corresponding distance cell from the slow-time-range complex matrix, perform arctangent demodulation and phase dewinding processing to form a vital signs time series signal;

[0015] S3. Perform heart rate calculation on the vital signs time series signal, extract the high-frequency components of the heartbeat based on bandpass filtering, and use the envelope autocorrelation method to search for the peak position to generate the heart rate value;

[0016] S4. Perform passband filtering based on the heart rate value to filter out the respiratory fundamental frequency energy in the vital signs time sequence signal and obtain the preliminary filtered output;

[0017] S5. Based on the heart rate value, construct several frequency band filtering intervals, and sequentially perform multi-band bandpass filtering on the preliminary filtering output to extract the heartbeat vibration waveform;

[0018] S6. Convolutional neural network structure and long short-term memory network structure are constructed based on heartbeat vibration waveform input. The sequence extracts short-term local features through the convolutional neural network structure and extracts long-term dependencies through the long short-term memory network structure. They are then concatenated into the output of the deep learning network to form continuous score prediction.

[0019] S7. Input the continuous score prediction into the ordinal regression output structure, divide the output space of continuous feature representation according to the medical rating and through the learnable threshold, and generate the systolic function classification result and diastolic function classification result as the heart failure diagnosis output.

[0020] Optionally, S1 specifically includes:

[0021] S11. Transmit radar signals to the target area and receive echo signals, perform frequency mixing on the echo signals, and generate intermediate frequency signals.

[0022] S12. Perform sampling processing on the intermediate frequency signal and perform distance dimension transformation on the sampled data frame by frame. The sampled data of each frame forms several distance units in the distance dimension.

[0023] S13. Arrange the distance dimension transformation results of several consecutively acquired frames in chronological order. The first dimension of the matrix consists of distance units, and the second dimension consists of consecutively acquired frames, thus constructing a slow-time-distance complex matrix.

[0024] Optionally, the heart rate calculation process in S3 specifically includes:

[0025] S31. Construct a bandpass filter by inputting the vital signs time series signal into the bandpass filter to obtain the first filtered sequence;

[0026] S32. Perform envelope extraction processing on the first filtered sequence and calculate autocorrelation to generate an autocorrelation sequence;

[0027] S33. Search for the target peak in the autocorrelation sequence, obtain the position and amplitude of the target peak, and calculate the candidate heart rate value based on the position of the target peak.

[0028] S34. Set the heart rate zone and peak amplitude threshold. When the candidate heart rate value is within the heart rate zone and the amplitude of the target peak is greater than the peak amplitude threshold, the candidate heart rate value is determined as the heart rate value and output to S4. When the candidate heart rate value is not within the heart rate zone or the amplitude of the target peak is not greater than the peak amplitude threshold, return to S1 to re-execute the radar signal acquisition process.

[0029] Optionally, the passband filtering process in S4 specifically includes:

[0030] S41. Read the vital signs timing signal and heart rate value, organize the vital signs timing signal into a single-column sequence according to the sampling order, and input the heart rate value into the processing flow for passband parameter setting.

[0031] S42. Determine the start and end frequencies of the heart rate-related frequency band based on the heart rate value, and use the start and end frequencies as passband frequency band configuration parameters to generate a set of frequency parameters for passband filter construction.

[0032] S43. Establish a passband filter based on the set of frequency parameters, configure the coefficients, order and calculation structure of the passband filter, input the vital signs time sequence signal into the passband filter point by point and perform traversal operations according to the calculation order of the filter to form the first filtering sequence.

[0033] S44. The first filtered sequence is rearranged into a continuous sequence according to the sampling time order, which is used as the initial filtered output.

[0034] Optionally, S5 specifically includes:

[0035] S51. Read the preliminary filter output and heart rate value, use the heart rate value as a frequency construction parameter to input into the multi-band configuration process, and calculate the heart rate base frequency according to the heart rate value.

[0036] S52. Generate several frequency band intervals based on the heartbeat fundamental frequency, take the integer multiple of the heartbeat fundamental frequency as the center frequency of each frequency band interval, construct a frequency band set consisting of several frequency bands, and match the frequency band set with the sampling parameters of the preliminary filter output;

[0037] S53. Establish several bandpass filters in sequence according to the frequency band set. Configure the filter coefficients, set the order and initialize the operation structure for each bandpass filter. Input the preliminary filter output to each bandpass filter and calculate the output point by point according to the operation order of the bandpass filters to generate the filter sequence corresponding to each frequency band.

[0038] S54. Traverse the filter sequences of each frequency band in frequency band order, align all filter sequences by time index and perform point-by-point superposition processing to form a heartbeat vibration waveform.

[0039] Optionally, S6 specifically includes:

[0040] S61. Read the heartbeat vibration waveform, divide the heartbeat vibration waveform into sequence segments according to a preset length, perform amplitude normalization processing on the sequence segments, and copy the processed sequence segments into two copies, and form a dataset with the corresponding ejection fraction and diastolic function; the diastolic function index corresponds to the ratio of the early diastolic blood flow velocity of the mitral valve to the early diastolic motion velocity of the tissue Doppler valve annulus;

[0041] S62. Input the two datasets into two sets of deep learning network structures with the same architecture. Both sets of deep learning network structures are composed of convolutional neural network structure, long short-term memory network structure and fully connected structure in the same order, and complete the structure initialization according to the same convolution configuration, pooling configuration, number of long short-term memory network layers and fully connected layer settings.

[0042] S63. In the two sets of deep learning network structures, the input sequences of each are sequentially input into the convolutional neural network structure to perform layer-by-layer convolution and pooling operations to generate local feature sequences. The local feature sequences are then input into the long short-term memory network structure to perform time-series calculations to form the feature vectors corresponding to the network.

[0043] S64. In the two sets of deep learning network structures, the feature vectors are input into their respective fully connected structures to perform matrix calculations, generating continuous score predictions for the corresponding ejection fraction and continuous score predictions for the corresponding diastolic function index, respectively.

[0044] S65. The two sets of continuous fractional predictions are used as the output of ejection function prediction and diastolic function prediction, respectively.

[0045] Optionally, S7 specifically includes:

[0046] S71. Read the continuous score prediction of ejection function and the continuous score prediction of diastolic function, and determine the number of ordinal regression thresholds based on the number of medical rating levels.

[0047] S72. Configure a corresponding number of learnable thresholds for the ordinal regression output structure according to the order of medical rating levels, and initialize the learnable thresholds.

[0048] S73. Input the continuous score prediction of the ejection function into the ordinal regression output structure, perform difference operation on the continuous score prediction and each learnable threshold in sequence, calculate the cumulative probability of the result of each difference operation and generate a probability sequence, and perform step judgment on the probability sequence in a preset order to form the ordinal prediction output of the ejection function.

[0049] S74. Input the continuous score prediction of diastolic function into the ordinal regression output structure, perform difference operation on the continuous score prediction and each learnable threshold in sequence, calculate the cumulative probability of the result of each difference operation and generate a probability sequence, and perform step judgment on the probability sequence in a preset order to form the ordinal prediction output of diastolic function.

[0050] S75. Use the ordinal prediction output of ejection function and the ordinal prediction output of diastolic function as the diagnostic results of heart failure.

[0051] Optionally, the deep learning network update process specifically includes:

[0052] Random sampling algorithm is applied to the training set to form disordered training data, and the network parameters are initialized according to the learning rate and Adam optimizer. The cross-entropy loss function is used as the loss calculation method.

[0053] The training and validation sets are input into the model in batches to perform forward computation, loss calculation and parameter update. When the change in the loss value in consecutive iterations is lower than the preset threshold, the convergence condition is met and training is stopped.

[0054] The beneficial effects of this invention are:

[0055] (1) By constructing a vital sign extraction process and a heartbeat separation algorithm, the heartbeat vibration component can be accurately extracted using a multi-band bandpass filtering mechanism. Clear and stable cardiac mechanical activity waveforms can be obtained under complex vibration backgrounds, thereby improving the reliability of subsequent analysis results.

[0056] (2) By replicating the same heartbeat vibration sequence as dual inputs and inputting them into a deep learning network with the same structure, independent modeling of two medical indicators, ejection fraction and diastolic function, can be achieved, effectively characterizing the differences between the systolic and diastolic phases of the heart and enhancing the model's ability to distinguish different stages of heart failure.

[0057] (3) By introducing a learnable threshold through the ordinal regression output structure and configuring an ordinal judgment mechanism based on the number of medical ratings, the network can output heart failure grading results with a grade relationship, realize the unified modeling of continuous score prediction and ordinal grade prediction for clinical indicators, and improve the accuracy and consistency of heart failure diagnosis at the grading level. Attached Figure Description

[0058] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0059] Figure 1 This is a flowchart of the millimeter-wave radar heart failure diagnosis system based on a hybrid model of CNN and LSTM ordinal regression proposed in this invention.

[0060] Figure 2 This is a radar echo signal image of a stationary human target in the millimeter-wave radar heart failure diagnosis system based on a hybrid model of CNN and LSTM ordinal regression proposed in this invention.

[0061] Figure 3 This is a vital sign signal map carrying respiratory and heartbeat information of the millimeter-wave radar heart failure diagnosis system based on a hybrid model of CNN and LSTM ordinal regression proposed in this invention.

[0062] Figure 4 This is a waveform of the extracted heartbeat signal from the millimeter-wave radar heart failure diagnosis system based on a hybrid model of CNN and LSTM ordinal regression proposed in this invention.

[0063] Figure 5 This is a comparison of the heartbeat waveforms of a cardiac health monitor and a heart failure patient in the millimeter-wave radar heart failure diagnostic system based on a hybrid model of CNN and LSTM ordinal regression proposed in this invention.

[0064] Figure 6 This is a structural diagram of the heart failure diagnosis model of the millimeter-wave radar heart failure diagnosis system based on a hybrid model of CNN and LSTM ordinal regression proposed in this invention. Detailed Implementation

[0065] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0066] refer to Figure 1-6 A millimeter-wave radar heart failure diagnostic system based on a hybrid model of CNN and LSTM ordinal regression includes:

[0067] The radar signal processing module acquires radar signals, performs mixing, sampling and mapping, and constructs a slow-time-range complex matrix.

[0068] The vital signs signal extraction module is used to select the slow-time complex sequence of the target corresponding range cell from the slow-time-range complex matrix, perform arctangent deharmonic phase unwinding processing, and generate vital signs time series signals.

[0069] The heartbeat waveform separation module is used to perform heart rate calculation on the vital sign time sequence signal. It extracts the high-frequency components of the heartbeat based on bandpass filtering, generates the heart rate value using the envelope autocorrelation method, and performs passband filtering and multi-band bandpass filtering based on the heart rate value to extract the heartbeat vibration waveform.

[0070] The deep feature learning module receives heartbeat vibration waveforms, extracts short-term local features through a convolutional neural network structure, extracts long-term dependencies through a long short-term memory network structure, and concatenates them into the output of the deep learning network to form a continuous score prediction.

[0071] The ordinal regression diagnostic module receives continuous score predictions, divides the output space of continuous feature representations through learnable thresholds, and generates systolic function grading results and diastolic function grading results.

[0072] In this embodiment, the modules are interconnected using the following method:

[0073] S1. Acquire radar signals, perform mixing, sampling and mapping, and construct a slow time-range complex matrix;

[0074] S2. Select the slow-time complex sequence of the target corresponding distance cell from the slow-time-range complex matrix, perform arctangent demodulation and phase dewinding processing to form a vital signs time series signal;

[0075] S3. Perform heart rate calculation on the vital signs time series signal, extract the high-frequency components of the heartbeat based on bandpass filtering, and use the envelope autocorrelation method to search for the peak position to generate the heart rate value;

[0076] S4. Perform passband filtering based on the heart rate value to filter out the respiratory fundamental frequency energy in the vital signs time sequence signal and obtain the preliminary filtered output;

[0077] S5. Based on the heart rate value, construct several frequency band filtering intervals, and sequentially perform multi-band bandpass filtering on the preliminary filtering output to extract the heartbeat vibration waveform;

[0078] S6. Convolutional neural network structure and long short-term memory network structure are constructed based on heartbeat vibration waveform input. The sequence extracts short-term local features through the convolutional neural network structure and extracts long-term dependencies through the long short-term memory network structure. They are then concatenated into the output of the deep learning network to form continuous score prediction.

[0079] S7. Input the continuous score prediction into the ordinal regression output structure, divide the output space of continuous feature representation according to the medical rating and through the learnable threshold, and generate the systolic function classification result and diastolic function classification result as the heart failure diagnosis output.

[0080] In this embodiment, S1 specifically includes:

[0081] S11. Transmit radar signals to the target area and receive echo signals. Perform frequency mixing processing on the echo signals to generate intermediate frequency signals. In this embodiment, FMCW radar is used.

[0082] S12. Perform sampling processing on the intermediate frequency signal and perform distance dimension transformation on the sampled data frame by frame. The sampled data of each frame forms several distance units in the distance dimension. In this embodiment, the sampling method for the intermediate frequency signal is ADC sampling; the distance dimension transformation is distance dimension Fourier transform.

[0083] S13. Arrange the range transformation results of several consecutively acquired frames in chronological order. The first dimension of the matrix consists of range units, and the second dimension consists of consecutively acquired frames, thus constructing a slow-time-range complex matrix. The target delay corresponds one-to-one with the range unit, with a total of M range units. By continuously acquiring N frames, the slow-time-range complex matrix S(M×N) can be constructed. Figure 2 Here is an example of an echo containing a human target: the target is located in the 3rd distance cell, the horizontal axis represents the distance cell, and the vertical axis represents the slow time in seconds.

[0084] In this embodiment, S2 specifically includes:

[0085] Visibility signals are extracted from the slow-time-distance matrix S; the slow-time dimension complex data of the distance cell where the target is located is selected, and the phase is extracted by arctangent demodulation; the phase is unwrapped to eliminate periodic jumps, resulting in a continuous phase sequence, which is the target's vital signs signal; such as Figure 3 As shown, the vital signs signals after untangling include both respiratory and heartbeat signals caused by chest cavity movement.

[0086] In this embodiment, the heart rate calculation process in S3 specifically includes:

[0087] S31. Construct a bandpass filter by inputting the vital signs time series signal into the bandpass filter to obtain the first filtered sequence. In this embodiment, the bandpass filter is an IIR bandpass filter with a passband of 3-30Hz.

[0088] S32. Perform envelope extraction processing on the first filtered sequence and perform autocorrelation calculation to generate an autocorrelation sequence; in this embodiment, the envelope extraction processing adopts Hilbert transform, and the absolute value is taken after the transform; the autocorrelation calculation is the autocorrelation function calculation commonly used in signal processing;

[0089] S33. Search for the target peak in the autocorrelation sequence to obtain the position and amplitude of the target peak. Calculate the candidate heart rate value based on the position of the target peak. The heart rate is the ratio of 60 times the slow time signal frame rate to the peak position value.

[0090] S34. Set the heart rate zone and peak amplitude threshold. When the candidate heart rate value is within the heart rate zone and the amplitude of the target peak is greater than the peak amplitude threshold, the candidate heart rate value is determined as the heart rate value and output to S4. When the candidate heart rate value is not within the heart rate zone or the amplitude of the target peak is not greater than the peak amplitude threshold, return to S1 to re-execute the radar signal acquisition process. In this embodiment, the heart rate zone is set to 48 to 160, and the threshold is set to 0.6.

[0091] In this embodiment, the passband filtering process in S4 specifically includes:

[0092] S41. Read the vital signs timing signal and heart rate value, organize the vital signs timing signal into a single-column sequence according to the sampling order, and input the heart rate value into the processing flow for passband parameter setting.

[0093] S42. Determine the start and end frequencies of the heart rate-related frequency band based on the heart rate value, and use the start and end frequencies as passband frequency band configuration parameters to generate a set of frequency parameters for passband filter construction.

[0094] S43. Establish a passband filter based on the set of frequency parameters, configure the coefficients, order and calculation structure of the passband filter, input the vital signs time sequence signal into the passband filter point by point and perform traversal operations according to the calculation order of the filter to form the first filtering sequence.

[0095] S44. The first filtered sequence is rearranged into a continuous sequence according to the sampling time order, which serves as the initial filtered output. In this embodiment, since the respiratory fundamental frequency is usually located in the frequency range [0.1Hz, 0.6Hz] and the heart rate frequency is usually located in [0.8Hz, 30Hz], a wide passband IIR filter is designed with a passband of [0.75Hz, 30Hz] to filter out the main respiratory fundamental frequency interference in the vital signs signal.

[0096] In this embodiment, S5 specifically includes:

[0097] S51. Read the preliminary filter output and heart rate value, use the heart rate value as a frequency construction parameter to input into the multi-band configuration process, and calculate the heart rate base frequency according to the heart rate value; the heart rate base frequency is heart rate value / 60;

[0098] S52. Based on the heartbeat fundamental frequency, generate several frequency band intervals. Use integer multiples of the heartbeat fundamental frequency as the center frequency of each frequency band interval, constructing a frequency band set consisting of several frequency bands. Then, map the frequency band set to the sampling parameters of the initial filtered output. The center frequency is an integer multiple of the heartbeat fundamental frequency, fluctuating by 0.1. If the heartbeat fundamental frequency is f... heart_base Then the frequency band range is:

[0099] [f heart_base -0.1, f heartbase +0.1],[2*f heart_base -0.1, 2*f heart_base +0.1],…,[n*f heart_base -0.1, n*f heart_base +0.1];

[0100] Where n is the harmonic order, which is 30 here, indicating that heartbeat signals within the 30th order are filtered out; the final heartbeat waveform after removing respiratory interference is obtained, as shown below. Figure 4 As shown;

[0101] S53. Establish several bandpass filters in sequence according to the frequency band set. Configure the filter coefficients, set the order and initialize the operation structure for each bandpass filter. Input the preliminary filter output to each bandpass filter and calculate the output point by point according to the operation order of the bandpass filters to generate the filter sequence corresponding to each frequency band.

[0102] S54. Traverse the filter sequences of each frequency band in frequency band order, align all filter sequences by time index and perform point-by-point superposition processing to form a heartbeat vibration waveform.

[0103] In this embodiment, S6 specifically includes:

[0104] S61. Read the heartbeat vibration waveform, divide the heartbeat vibration waveform into sequence segments according to a preset length, perform amplitude normalization processing on the sequence segments, and copy the processed sequence segments into two copies, and form a dataset with the corresponding ejection fraction and diastolic function; the diastolic function index corresponds to the ratio of the early diastolic blood flow velocity of the mitral valve to the early diastolic motion velocity of the tissue Doppler valve annulus;

[0105] S62. Input the two datasets into two sets of deep learning network structures with the same architecture. Both sets of deep learning network structures are composed of convolutional neural network structure, long short-term memory network structure and fully connected structure in the same order, and complete the structure initialization according to the same convolution configuration, pooling configuration, number of long short-term memory network layers and fully connected layer settings.

[0106] S63. In the two sets of deep learning network structures, the input sequences of each are sequentially input into the convolutional neural network structure to perform layer-by-layer convolution and pooling operations to generate local feature sequences. The local feature sequences are then input into the long short-term memory network structure to perform time-series calculations to form the feature vectors corresponding to the network.

[0107] S64. In the two sets of deep learning network structures, the feature vectors are input into their respective fully connected structures to perform matrix calculations, generating continuous score predictions for the corresponding ejection fraction and continuous score predictions for the corresponding diastolic function index, respectively.

[0108] S65. The two sets of continuous fractional predictions are used as the output of ejection function prediction and diastolic function prediction, respectively.

[0109] In this embodiment, since the two deep learning networks have the same structure, the blood-ejection network will be used as an example for description:

[0110] The dataset consists of heart rate waveform data and the corresponding LVEF and E / e'; LVEF is the ejection fraction, and E / e' is the diastolic function.

[0111] Constructing the input layer: The input data is copied into two copies and fed into two sets of networks with the same architecture. The ejection network is used to predict ejection function, and the diastolic network is used to predict diastolic function.

[0112] The CNN structure is as follows: Convolutional Layer 1, Pooling Layer 1, Convolutional Layer 2, Pooling Layer 2, Convolutional Layer 3, Pooling Layer 3; the number of convolutional kernels is set to 64, 128, and 256 respectively, the kernel size is set to 3, the stride is set to 1, and the activation function is set to ReLU; all pooling layers are set to max pooling, with a pooling window length of 2 and a stride of 2.

[0113] Construct an LSTM structure: first layer of neurons, second layer of neurons, first fully connected layer, second fully connected layer, SoftMax layer; set the number of neurons in the first and second layers to 128 and 64 respectively, and the number of neurons in the first and second fully connected layers to 128 and 64 respectively.

[0114] The constructed CNN and LSTM are concatenated into a deep learning network, which outputs continuous score predictions.

[0115] In this embodiment, S7 specifically includes:

[0116] S71. Read the continuous fraction prediction of ejection function and the continuous fraction prediction of diastolic function, and determine the number of ordinal regression thresholds according to the number of levels in the medical rating;

[0117] S72. Configure the corresponding number of learnable thresholds for the ordinal regression output structure in the order of the levels in the medical rating, and initialize the learnable thresholds;

[0118] S73. Input the continuous fraction prediction of ejection function into the ordinal regression output structure, perform subtraction operations on the continuous fraction prediction and each learnable threshold in sequence, calculate the cumulative probability for the result of each subtraction operation and generate a probability sequence, and perform step judgments on the probability sequence in a preset order to form the ordinal prediction output of ejection function;

[0119] S74. Input the continuous fraction prediction of diastolic function into the ordinal regression output structure, perform subtraction operations on the continuous fraction prediction and each learnable threshold in sequence, calculate the cumulative probability for the result of each subtraction operation and generate a probability sequence, and perform step judgments on the probability sequence in a preset order to form the ordinal prediction output of diastolic function;

[0120] S75. Use the ordinal prediction output of ejection function and the ordinal prediction output of diastolic function as the heart failure diagnosis result.

[0121] In this embodiment, according to the medical rating, the values of LVEF and E / e' are respectively divided into three intervals, corresponding to three functional categories of cardiac systolic and diastolic functions. This model correspondingly divides the systolic function and diastolic function into three categories. Systolic function: reduced ejection fraction type, preserved ejection fraction type, LVEF≥50%; mildly reduced ejection fraction type, LVEF is 41%-49%; LVEF: ≤40%. Respectively corresponding to ejection grades I, II, and III; Diastolic function: normal, E / e'<8; gray zone, E / e' is 8–13; dysfunction, E / e'≥13. Respectively corresponding to diastolic grades I, II, and III; Since both the systolic and diastolic functions are output as three levels, the ordinal regression structures of the two are also designed to be the same. The following design takes the systolic function ordinal regression model as an example: Suppose the output has 3 ordered categories, and 2 learnable thresholds are introduced: b1 < b2, dividing the continuous prediction space into three ordered intervals, and each interval corresponds to a category; then the cumulative probability of each category is:

[0122] P(y≤k∣x)=σ(b k -f(x));

[0123] Where σ is the sigmoid function and k is the corresponding category. Here, the difference between the model prediction value and the threshold is converted into a cumulative probability through the sigmoid function. Therefore, the probability of each category is: P(y=k|x)=P(y≤k)−P(y≤k−1), and in particular, P(y≤0)=0, P(y≤3)=1.

[0124] In this embodiment, the deep learning network update process specifically includes:

[0125] Random sampling algorithm is applied to the training set to generate disordered training data, and the network parameters are initialized according to the learning rate and Adam optimizer. The cross-entropy loss function is used as the loss calculation method; the learning rate is set to 0.001.

[0126] The training and validation sets are input into the model in batches for forward computation, loss calculation, and parameter updates. When the change in the loss value in consecutive iterations is lower than a preset threshold, the convergence condition is met, and training stops. In this embodiment, the maximum number of training iterations is set to 200, the loss monitoring window length is set to 10, and the maximum and minimum values ​​of the validation set loss in the most recent 10 iterations are counted. The loss change threshold is set to 0.0001. When the difference between the maximum and minimum values ​​is less than the loss change threshold, convergence is determined. In the most recent 10 iterations, the validation set loss has not been significantly less than the current minimum validation set loss, meaning there is no new minimum validation set loss update within this window. If any of the above convergence conditions are met before the end of 200 training iterations, training is immediately terminated, and the model parameters corresponding to the current iteration are retained as the training result. If the convergence condition is not met after reaching the maximum number of training iterations, the model parameters at the end of 200 iterations are used as the training result.

[0127] Example 1:

[0128] To verify the feasibility of this invention in practice, it was applied to a millimeter-wave radar measurement environment for monitoring changes in heart failure risk in a resting state at home. In this environment, the subject is in a quiet sitting or lying position, without wearing any equipment. The millimeter-wave radar is fixed at a position approximately one meter away from the chest to continuously collect micro-vibration signals from the chest cavity. Traditional contact-based heart failure detection relies on specific conditions and is not suitable for long-term continuous monitoring. This invention aims to solve this problem, enabling stable, continuous, and relatively accurate assessment of systolic and diastolic heart failure functions without contact.

[0129] In this scenario, data acquired by millimeter-wave radar is input into the signal processing link described in this invention. The radar signal processing module generates a slow-time-range complex matrix, which is then processed by the vital sign signal extraction module to form a continuous vital sign time-series signal. The heartbeat waveform separation module progressively filters out respiratory components and low-order interference, ultimately obtaining a relatively complete heartbeat vibration waveform. The deep feature learning module copies the input signal and feeds it into two sets of network structures for predicting ejection and diastole, respectively, to extract subtle vibrational differences between the systolic and diastolic cycles. Based on the continuous score prediction results, the ordinal regression diagnostic module outputs the systolic and diastolic function levels through learnable thresholds.

[0130] To verify the practical performance of this method, millimeter-wave radar data were collected from individuals of different age groups and body types, and ejection fraction and diastolic function indices measured using traditional clinical methods were obtained simultaneously. All radar signals were processed by the model to obtain heart failure function prediction results, which were then compared with reference results for consistency.

[0131] In practical applications, the monitored subject only needs to maintain natural breathing and does not need to perform any active actions. The millimeter-wave radar continuously collects micro-vibration signals from the chest wall, and the system automatically completes signal extraction, heartbeat separation, deep feature learning, and functional level output. This process is completely contactless, non-contact, and allows for continuous monitoring, making the monitored subject unaffected by environmental changes or behavioral limitations.

[0132] In data validation, a selection of typical data from a large sample was used to present the results. For ejection function level prediction, the predicted level obtained by this invention achieved a high matching rate with the reference level, and the mean square error between the continuous score prediction and the actual ejection score was within a reasonable range. For diastolic function level prediction, this invention also achieved good performance, with the predicted trend of the diastolic function ratio maintaining a stable relationship with the actual trend. Specific data are shown in the table below, including multiple indicators such as continuous score prediction, level determination, and matching accuracy with the reference level.

[0133] A comprehensive analysis of this data reveals that the present invention exhibits strong sensitivity to changes in heart failure function, particularly demonstrating stable feature extraction capabilities in distinguishing between low and medium-level heart failure. The ordinal rank and continuous score predictions output by the present invention maintain high consistency across samples of different body types and age groups, effectively reflecting the changing trends of the reference indicators. Overall, the present invention not only demonstrates excellent data accuracy but also features non-contact measurement, long-term continuous monitoring, and enhanced comfort, making it widely applicable in home and medical settings requiring long-term heart failure risk monitoring.

[0134] The structured data table of this embodiment is given below to compare the consistency between the output results of this invention and the reference results.

[0135] Table 1: Performance Evaluation Table for Millimeter-Wave Radar in Heart Failure Diagnosis

[0136] S01 62 60.8 I I 7.8 7.9 I I S02 55 53.4 I I 7.3 7.6 I I S03 48 51.2 II I 12.4 11.8 II II S04 42 44.1 II II 12.2 11.9 II II S05 48 46.7 II II 12.5 12.8 II II S06 32 30.2 III III 18.7 19.4 III III S07 29 26.1 III III 20.1 20.6 III III S08 52 50.6 I I 7.3 7.1 I I S09 44 47.8 II II 11.8 12.0 II II S10 45 43.9 II II 12.1 12.9 II II

[0137] According to Table 1, in predicting systolic function levels, the predicted ejection grade shows high consistency with the actual grade, especially in the boundary samples of levels I–II. The continuous score prediction of this invention accurately reflects the changing trend, ensuring the ordinal grade determination remains correct. In the diastolic function section, the continuous score trend output by this invention is consistent with the change in the actual E / e' ratio, correctly classifying diastolic grades in all samples, indicating that the deep feature learning structure of this invention can effectively capture diastolic vibration characteristics. Overall, the table data demonstrates the feasibility and stability of this invention for grading and assessing heart failure function under non-contact conditions.

[0138] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A millimeter-wave radar heart failure diagnostic system based on a hybrid model of CNN and LSTM ordinal regression, characterized in that, include: The radar signal processing module acquires radar signals, performs mixing, sampling and mapping, and constructs a slow-time-range complex matrix. The vital signs signal extraction module is used to select the slow-time complex sequence of the target corresponding range cell from the slow-time-range complex matrix, perform arctangent deharmonic phase unwinding processing, and generate vital signs time series signals. The heartbeat waveform separation module is used to perform heart rate calculation on the vital sign time sequence signal. It extracts the high-frequency components of the heartbeat based on bandpass filtering, generates the heart rate value using the envelope autocorrelation method, and performs passband filtering and multi-band bandpass filtering based on the heart rate value to extract the heartbeat vibration waveform. The deep feature learning module receives heartbeat vibration waveforms, extracts short-term local features through a convolutional neural network structure, extracts long-term dependencies through a long short-term memory network structure, and concatenates them into the output of the deep learning network to form a continuous score prediction. The ordinal regression diagnostic module receives continuous score predictions, divides the output space of continuous feature representations through a learnable threshold, and generates systolic function grading results and diastolic function grading results. The output space of the learnable threshold-divided continuous feature representation specifically includes: Read the continuous score predictions of ejection function and diastolic function, and determine the number of ordinal regression thresholds based on the number of medical rating levels; Configure a corresponding number of learnable thresholds for the ordinal regression output structure according to the medical rating order, and initialize the learnable thresholds; The continuous score prediction of the ejection function is input into the ordinal regression output structure. The difference operation is performed on the continuous score prediction and each learnable threshold in turn. The cumulative probability is calculated on the result of each difference operation and a probability sequence is generated. The probability sequence is then subjected to step judgment in a preset order to form the ordinal prediction output of the ejection function. The continuous score prediction of diastolic function is input into the ordinal regression output structure. The difference operation is performed on the continuous score prediction and each learnable threshold in turn. The cumulative probability is calculated on the result of each difference operation and a probability sequence is generated. The probability sequence is then subjected to step judgment in a preset order to form the ordinal prediction output of diastolic function. The ordinal prediction outputs of ejection function and diastolic function are used as diagnostic results for heart failure.

2. The millimeter-wave radar heart failure diagnostic system based on a hybrid model of CNN and LSTM ordinal regression according to claim 1, characterized in that, The modules are connected in the following way: S1. Acquire radar signals, perform mixing, sampling and mapping, and construct a slow time-range complex matrix; S2. Select the slow-time complex sequence of the target corresponding distance cell from the slow-time-range complex matrix, perform arctangent demodulation and phase dewinding processing to form a vital signs time series signal; S3. Perform heart rate calculation on the vital signs time series signal, extract the high-frequency components of the heartbeat based on bandpass filtering, and use the envelope autocorrelation method to search for the peak position to generate the heart rate value; S4. Perform passband filtering based on the heart rate value to filter out the respiratory fundamental frequency energy in the vital signs time sequence signal and obtain the preliminary filtered output. S5. Based on the heart rate value, construct several frequency band filtering intervals, and sequentially perform multi-band bandpass filtering on the preliminary filtering output to extract the heartbeat vibration waveform; S6. Convolutional neural network structure and long short-term memory network structure are constructed based on heartbeat vibration waveform input. The sequence extracts short-term local features through the convolutional neural network structure and extracts long-term dependencies through the long short-term memory network structure. They are then concatenated into the output of the deep learning network to form continuous score prediction. S7. Input the continuous score prediction into the ordinal regression output structure, divide the output space of continuous feature representation according to the medical rating and through the learnable threshold, and generate the systolic function classification result and diastolic function classification result as the heart failure diagnosis output.

3. The millimeter-wave radar heart failure diagnostic system based on a hybrid model of CNN and LSTM ordinal regression according to claim 2, characterized in that, S1 specifically includes: S11. Transmit radar signals to the target area and receive echo signals, perform frequency mixing on the echo signals, and generate intermediate frequency signals. S12. Perform sampling processing on the intermediate frequency signal and perform distance dimension transformation on the sampled data frame by frame. The sampled data of each frame forms several distance units in the distance dimension. S13. Arrange the distance dimension transformation results of several consecutively acquired frames in chronological order. The first dimension of the matrix consists of distance units, and the second dimension consists of consecutively acquired frames, thus constructing a slow-time-distance complex matrix.

4. The millimeter-wave radar heart failure diagnostic system based on a hybrid model of CNN and LSTM ordinal regression according to claim 3, characterized in that, The heart rate calculation process in S3 specifically includes: S31. Construct a bandpass filter by inputting the vital signs time series signal into the bandpass filter to obtain the first filtered sequence; S32. Perform envelope extraction processing on the first filtered sequence and calculate autocorrelation to generate an autocorrelation sequence; S33. Search for the target peak in the autocorrelation sequence, obtain the position and amplitude of the target peak, and calculate the candidate heart rate value based on the position of the target peak. S34. Set the heart rate zone and peak amplitude threshold. When the candidate heart rate value is within the heart rate zone and the amplitude of the target peak is greater than the peak amplitude threshold, the candidate heart rate value is determined as the heart rate value and output to S4. When the candidate heart rate value is not within the heart rate zone or the amplitude of the target peak is not greater than the peak amplitude threshold, return to S1 to re-execute the radar signal acquisition process.

5. The millimeter-wave radar heart failure diagnostic system based on a hybrid model of CNN and LSTM ordinal regression according to claim 4, characterized in that, The passband filtering process in S4 specifically includes: S41. Read the vital signs timing signal and heart rate value, organize the vital signs timing signal into a single-column sequence according to the sampling order, and input the heart rate value into the processing flow for passband parameter setting. S42. Determine the start and end frequencies of the heart rate-related frequency band based on the heart rate value, and use the start and end frequencies as passband frequency band configuration parameters to generate a set of frequency parameters for passband filter construction. S43. Establish a passband filter based on the set of frequency parameters, configure the coefficients, order and calculation structure of the passband filter, input the vital signs time sequence signal into the passband filter point by point and perform traversal operations according to the calculation order of the filter to form the first filtering sequence. S44. The first filtered sequence is rearranged into a continuous sequence according to the sampling time order, which is used as the initial filtered output.

6. The millimeter-wave radar heart failure diagnostic system based on a hybrid model of CNN and LSTM ordinal regression according to claim 5, characterized in that, S5 specifically includes: S51. Read the preliminary filter output and heart rate value, use the heart rate value as a frequency construction parameter to input into the multi-band configuration process, and calculate the heart rate base frequency according to the heart rate value. S52. Generate several frequency band intervals based on the heartbeat fundamental frequency, take the integer multiple of the heartbeat fundamental frequency as the center frequency of each frequency band interval, construct a frequency band set consisting of several frequency bands, and match the frequency band set with the sampling parameters of the preliminary filter output; S53. Establish several bandpass filters in sequence according to the frequency band set. Configure the filter coefficients, set the order and initialize the operation structure for each bandpass filter. Input the preliminary filter output to each bandpass filter and calculate the output point by point according to the operation order of the bandpass filters to generate the filter sequence corresponding to each frequency band. S54. Traverse the filter sequences of each frequency band in frequency band order, align all filter sequences by time index and perform point-by-point superposition processing to form a heartbeat vibration waveform.

7. The millimeter-wave radar heart failure diagnostic system based on a hybrid model of CNN and LSTM ordinal regression according to claim 6, characterized in that, S6 specifically includes: S61. Read the heartbeat vibration waveform, divide the heartbeat vibration waveform into sequence segments according to a preset length, perform amplitude normalization processing on the sequence segments, and copy the processed sequence segments into two copies, and form a dataset with the corresponding ejection fraction and diastolic function; the diastolic function index corresponds to the ratio of the early diastolic blood flow velocity of the mitral valve to the early diastolic motion velocity of the tissue Doppler valve annulus; S62. Input the two datasets into two sets of deep learning network structures with the same architecture. Both sets of deep learning network structures are composed of convolutional neural network structure, long short-term memory network structure and fully connected structure in the same order, and complete the structure initialization according to the same convolution configuration, pooling configuration, number of long short-term memory network layers and fully connected layer settings. S63. In the two sets of deep learning network structures, the input sequences of each are sequentially input into the convolutional neural network structure to perform layer-by-layer convolution and pooling operations to generate local feature sequences. The local feature sequences are then input into the long short-term memory network structure to perform time-series calculations to form the feature vectors corresponding to the network. S64. In the two sets of deep learning network structures, the feature vectors are input into their respective fully connected structures to perform matrix calculations, generating continuous score predictions for the corresponding ejection fraction and continuous score predictions for the corresponding diastolic function index, respectively. S65. The two sets of continuous fractional predictions are used as the output of ejection function prediction and diastolic function prediction, respectively.

8. The millimeter-wave radar heart failure diagnostic system based on a hybrid model of CNN and LSTM ordinal regression according to claim 7, characterized in that, The training process for the two sets of deep learning network structures in S6 specifically includes: Random sampling algorithm is applied to the training set to form disordered training data, and the network parameters are initialized according to the learning rate and Adam optimizer. The cross-entropy loss function is used as the loss calculation method. The training and validation sets are input into the model in batches to perform forward computation, loss calculation and parameter update. When the change in the loss value in consecutive iterations is lower than the preset threshold, the convergence condition is met and training is stopped.

Images