Cpr quality monitoring system based on double signal timing fusion and light-weight cnn

Abstract

This application discloses a cardiopulmonary resuscitation (CPR) quality monitoring system based on dual-signal temporal fusion and a lightweight CNN, relating to the field of signal processing technology. The system includes: a signal acquisition module for acquiring triaxial acceleration signals and transthoracic impedance signals; a preprocessing module for processing the triaxial acceleration signals and transthoracic impedance signals; a data repair module for determining interference intervals and repairing non-compression segments based on the temporal periodicity of the Z-axis acceleration sequence and the dominant frequency of the differential impedance sequence; a period segmentation and dynamic drift elimination module for periodically segmenting the repaired acceleration sequence and estimating and subtracting the drift DC component; an inference module for inputting the repaired data into a lightweight multi-branch convolutional neural network and simultaneously outputting compression depth, compression frequency, and the probability of sufficiency of chest recoil; and a feedback output module for smoothing the output results and providing feedback. This application improves the accuracy of CPR quality monitoring.

Description

Technical Field

[0001] This invention relates to the field of signal processing technology, specifically to a cardiopulmonary resuscitation quality monitoring system based on dual-signal temporal fusion and lightweight CNN. Background Technology

[0002] Cardiac arrest is a life-threatening emergency, and the quality of cardiopulmonary resuscitation (CPR) largely determines patient survival rates. International guidelines stipulate that high-quality chest compressions in adults should meet three criteria: a depth of 50-60 mm, a rate of 100-120 compressions per minute, and adequate chest recoil. However, rescuers are prone to fatigue, leading to improper techniques. Real-time monitoring and feedback of compression quality are essential, especially with automated external defibrillators (AEDs), a crucial piece of emergency equipment. Integrating CPR monitoring functionality can significantly improve resuscitation outcomes.

[0003] Current monitoring technologies primarily rely on accelerometers, but the double integration of acceleration signals leads to cumulative drift, causing depth measurement errors to increase over time. To suppress drift, traditional methods depend on external signal calibration, segmented integration reset, or pressure detection triggering integration reset upon complete chest recoil. However, these strategies have certain problems; for example, reset depends on the recoil state, and insufficient recoil can lead to an incorrect integration starting point, thus amplifying the error. While transthoracic impedance signals can indirectly reflect chest deformation, their amplitude does not have a stable linear relationship with compression depth, resulting in large errors when directly used for depth conversion. Furthermore, they are susceptible to influences from patient body shape, electrode placement, and motion interference, leading to insufficient robustness. Summary of the Invention

[0004] In order to solve the above-mentioned technical problems, this application proposes the following technical solution: This application provides a cardiopulmonary resuscitation quality monitoring system based on dual-signal temporal fusion and lightweight CNN, including: The signal acquisition module is used to synchronously acquire the triaxial acceleration timing signal and the transthoracic impedance timing signal at a preset sampling frequency, and to set a timing window of fixed duration for data buffering to obtain time-aligned triaxial acceleration signal and transthoracic impedance signal. The preprocessing module is used to preprocess the triaxial acceleration signal and the transthoracic impedance signal respectively to obtain the Z-axis acceleration sequence and the differential impedance sequence; The data repair module is used to determine the motion interference interval and repair the non-pressing segment within the interference interval based on the time-domain periodicity of the Z-axis acceleration sequence and the frequency dominance of the differential impedance sequence. The period segmentation and dynamic drift elimination module is used to adaptively segment the repaired Z-axis acceleration sequence. At the end of each pressing cycle, the integral drift DC component is estimated and the DC component is subtracted from the current cycle velocity sequence and displacement sequence to obtain the drift-corrected velocity sequence and displacement sequence. The inference module is used to input the repaired Z-axis acceleration sequence, the repaired differential impedance sequence, and the drift-corrected velocity and displacement sequences into a pre-built lightweight multi-branch convolutional neural network, and simultaneously output the compression depth, compression frequency, and the probability of sufficiency of chest recoil. The feedback output module is used to smooth the compression depth, compression frequency and the probability of sufficient chest recoil, and then provide real-time feedback on the compression quality through display and voice.

[0005] In one possible implementation, the triaxial acceleration signal and the transthoracic impedance signal are preprocessed to obtain a Z-axis acceleration sequence and a differential impedance sequence, including: After removing the gravity DC component of the triaxial acceleration signal, the Z-axis acceleration component perpendicular to the pressing direction is extracted. The extracted Z-axis acceleration components are bandpass filtered to obtain the Z-axis acceleration sequence after noise reduction and baseline correction. The transthoracic impedance signal is subjected to first-order differential and neighborhood smoothing filtering to obtain a differential impedance sequence that highlights the thoracic deformation characteristics. The formula for calculating the first-order difference is as follows: in, The result is the result of the difference calculation. This is the current transthoracic impedance value. The transthoracic impedance value at the previous moment. The sampling interval is denoted as .

[0006] In one possible implementation, based on the time-domain periodicity of the Z-axis acceleration sequence and the dominant frequency of the differential impedance sequence, the motion interference interval is determined, and data repair is performed on the non-pressing segments within the interference interval, including: Peak detection is performed on the Z-axis acceleration sequence to obtain the duration of multiple consecutive pressing cycles. The coefficient of variation is calculated based on the duration of the pressing cycle, and the coefficient of variation is used as a time-domain periodicity index. Perform a Fast Fourier Transform on the differential impedance sequence and extract the frequency corresponding to the point of maximum amplitude as the dominant frequency in the frequency domain. When the coefficient of variation is greater than the first preset value and the main frequency is less than the second preset value or greater than the third preset value, the current window is determined to be a motion interference zone. Adaptive notch filtering or linear interpolation methods are used to repair the data in the non-pressed segments within the motion interference range.

[0007] In one possible implementation, the step of using adaptive notch filtering or linear interpolation to repair the data of the non-pressing segment within the motion interference interval includes: determining the non-pressing segment within the motion interference interval; The Z-axis acceleration signal in the non-pressing segment is subjected to spectrum analysis to extract the interference main frequency, and the interference main frequency is set as the notch frequency of the adaptive notch filter. The signal is notched filtered using the adaptive notch filter. If the root mean square error of the non-pressed segment signal after notch filtering is greater than the preset tolerance threshold, then notch filtering is abandoned and linear interpolation is used instead. Take the corresponding position signal values ​​of one complete pressing cycle before and after the non-pressing segment, perform linear interpolation according to the time ratio, and reconstruct the non-pressing segment data; The reconstructed non-pressing segment data is spliced ​​with the undisturbed pressing segment data to obtain the repaired Z-axis acceleration sequence.

[0008] In one possible implementation, the corrected Z-axis acceleration sequence is adaptively periodically segmented. At the end of each pressing cycle, the integral drift DC component is estimated, and this DC component is subtracted from the current cycle's velocity and displacement sequence to obtain the drift-corrected velocity and displacement sequences, including: The dynamic threshold peak detection method is used to adaptively segment the Z-axis acceleration sequence after repair. The start and end positions of each pressing cycle are determined by identifying the dynamic threshold of the acceleration peak. At the end of each pressing cycle, a short-time window is selected, and the steady-state conditions of the short-time window are judged. When the short-time window meets the steady-state conditions, the integral drift DC component is estimated, and the calculation formulas are as follows: in, This is an estimate of the integral drift DC component of the velocity sequence. The number of sampling points within the short time window. This is the speed value. This is the estimate of the integral drift DC component of the displacement sequence. This is the displacement value; The DC component is subtracted from the current period velocity sequence and displacement sequence to obtain the drift-corrected velocity sequence and displacement sequence, calculated using the following formulas: in, , The velocity and displacement sequences after drift correction are respectively in The value at time, For the uncorrected integral velocity sequence, the first point, For the uncorrected integral displacement sequence, the first point; When the short-time window does not meet the steady-state condition, the drift rate is deduced from the difference between the displacement sequence or velocity sequence between adjacent cycles by using the integral drift accumulation law of multiple consecutive pressing cycles, and the drift-corrected velocity sequence and displacement sequence are obtained based on the drift rate.

[0009] In one possible implementation, when the short-time window does not meet the steady-state condition, the drift rate is deduced from the difference between the displacement or velocity sequences between adjacent cycles by utilizing the integral drift accumulation law of multiple consecutive pressing cycles, and the drift-corrected velocity and displacement sequences are obtained based on the drift rate, including: If the range of the displacement sequence within the short-time window of the current pressing cycle is greater than the preset threshold, it is determined that the steady-state condition is not met. Obtain the displacement or velocity sequence of multiple consecutive pressing cycles preceding the current pressing cycle; At the end of each cycle, a window of fixed length is selected, and the average displacement or average velocity within the window is calculated as the residual drift for the current cycle. Assuming that the drift changes linearly over multiple consecutive periods, the least squares method is used to fit the drift rate. After multiplying the drift rate by the time offset, subtract it from the velocity sequence and displacement sequence of the current period to obtain the drift-corrected velocity sequence and displacement sequence.

[0010] In one possible implementation, the lightweight multi-branch convolutional neural network includes an input layer connected to the inputs of an acceleration branch and a transthoracic impedance branch. The acceleration branch and the transthoracic impedance branch have a symmetrical structure, each including multiple one-dimensional convolutional layers, batch normalization layers, ReLU activation function layers, and global average pooling layers. The outputs of the acceleration branch and the transthoracic impedance branch are connected to the inputs of a feature fusion layer. The outputs of the feature fusion layer are connected to the inputs of a depth regression head, a frequency regression head, and a bounce binary classification head. The outputs of the depth regression head, the frequency regression head, and the bounce binary classification head are connected to the output layer. The lightweight multi-branch convolutional neural network uses the following loss function during training: in, This represents the total training loss function value for a lightweight multi-branch convolutional neural network. The mean squared error loss for the depth prediction task is used to measure the depth prediction value. Compared with the true value The differences between them The mean squared error loss for the frequency prediction task is used to measure the frequency prediction value. Compared with the true value The differences between them The binary cross-entropy loss is used for the chest recoil adequacy classification task to measure the predicted probability of recoil adequacy. With real labels The differences between them; , , These are the loss weight coefficients for depth regression, frequency regression, and bounce classification tasks, respectively.

[0011] In one possible implementation, the repaired Z-axis acceleration sequence, the repaired differential impedance sequence, and the drift-corrected velocity and displacement sequences are input into a pre-constructed lightweight multi-branch convolutional neural network, which simultaneously outputs the compression depth, compression frequency, and the probability of sufficiency of chest recoil, including: The repaired Z-axis acceleration sequence, the drift-corrected velocity sequence, and the displacement sequence are input into the acceleration branch of the lightweight multi-branch convolutional neural network, and the repaired differential impedance sequence is input into the transthoracic impedance branch of the lightweight multi-branch convolutional neural network. Each branch is processed sequentially through multiple layers of one-dimensional convolution, batch normalization layers, ReLU activation function layers, and global average pooling layers, outputting fixed-length acceleration branch feature vectors and impedance branch feature vectors respectively. The acceleration branch feature vector and the impedance branch feature vector are concatenated along the channel dimension to obtain the fused feature vector; The fused feature vectors are input into the depth regression head, frequency regression head, and rebound binary classification head of a lightweight multi-branch convolutional neural network, respectively, and the predicted values ​​of compression depth, compression frequency, and sufficiency of chest recoil are output simultaneously.

[0012] In one possible implementation, after smoothing the compression depth, compression frequency, and probability of thorax recoil adequacy, real-time feedback on compression quality is provided via display and voice, including: The compression depth, compression frequency, and thoracic recoil adequacy probability output by the lightweight multi-branch convolutional neural network are smoothed using an exponential moving average. The calculation formula is as follows: in, The result is smoothed and used for feedback. For smoothing coefficients, This represents the raw predicted values ​​directly output by a lightweight multi-branch convolutional neural network at the current time t, i.e., the unsmoothed compression depth, compression frequency, or probability of sufficiency of chest recoil. express Feedback results after real-time smoothing; The pressure quality is determined based on the smoothed result, and real-time feedback on the pressure quality is provided through display and voice.

[0013] In one possible implementation, the step of determining the pressing quality based on the smoothed result includes: Obtain the smoothed compression depth, compression frequency, and probability of sufficiency of chest recoil; The smoothed compression depth, compression frequency, and probability of sufficient chest recoil are compared with their respective preset threshold ranges and preset probability thresholds. If the smoothed press depth is less than the lower limit threshold, the press depth is determined to be too shallow; if the smoothed press depth is greater than the upper limit threshold, the press depth is determined to be too deep. If the smoothed pressing frequency is less than the lower frequency threshold, the pressing frequency is determined to be too slow; if the smoothed pressing frequency is greater than the upper frequency threshold, the pressing frequency is determined to be too fast. If the probability of sufficient rebound of the smoothed thoracic cavity is less than the preset sufficient rebound threshold, the thoracic cavity rebound state is determined to be insufficient rebound; if the probability of sufficient rebound of the smoothed thoracic cavity is greater than or equal to the preset sufficient rebound threshold, the thoracic cavity rebound state is determined to be sufficient rebound.

[0014] Compared with the prior art, the beneficial effects of this application are as follows: This application reconstructs the task division architecture of triaxial acceleration and transthoracic impedance signals. Triaxial acceleration independently calculates compression depth and frequency, while transthoracic impedance only assists in judging chest rebound, avoiding depth conversion errors caused by individual differences in impedance amplitude and significantly improving monitoring accuracy.

[0015] This application establishes a dynamic drift elimination mechanism within the cycle, reducing the reliance on accurate detection of chest recoil state and fundamentally solving the problem of cumulative drift due to acceleration quadratic integration. Depth measurement error does not increase over time, and integral correction and recoil discrimination are completely decoupled, eliminating conflicts between traditional reset and judgment logics. Furthermore, it designs a lightweight, multi-branch, one-dimensional convolutional neural network without manual features, resulting in a small parameter count that is easy to deploy on low-power embedded platforms. Simultaneously, it automatically learns spatiotemporal correlation features from the original time-series signal, exhibiting strong generalization ability. It employs dual-dimensional joint interference discrimination and selective data repair to effectively prevent interference from transport bumps and electrode slippage, adapting to scenarios such as on-site emergency care and pre-hospital transport. The entire process requires no external calibration, no operational condition segmentation, and no signal source switching, significantly reducing operational complexity. Real-time feedback via display and voice modes significantly improves the ease and standardization of operation at emergency sites.

[0016] This application has the advantages of good drift suppression, lightweight model, strong anti-interference, convenient deployment and user-friendly human-computer interaction, and is suitable for real-time monitoring of cardiopulmonary resuscitation quality in scenarios such as cardiac arrest emergency rescue and pre-hospital transport. Attached Figure Description

[0017] Figure 1 A schematic diagram of a cardiopulmonary resuscitation quality monitoring system based on dual-signal temporal fusion and lightweight CNN provided in this application embodiment; Figure 2 A block diagram of a lightweight multi-branch convolutional neural network structure provided in an embodiment of this application. Detailed Implementation

[0018] The present solution will now be described in conjunction with the accompanying drawings and specific embodiments.

[0019] Figure 1 A schematic diagram of a cardiopulmonary resuscitation quality monitoring system based on dual-signal temporal fusion and lightweight CNN provided in this application embodiment is shown below. Figure 1 This embodiment of a cardiopulmonary resuscitation quality monitoring system based on dual-signal temporal fusion and lightweight CNN includes: The system comprises a signal acquisition module, a preprocessing module, a data repair module, a period segmentation and dynamic drift elimination module, an inference module, and a feedback output module. The signal acquisition module synchronously acquires triaxial acceleration timing signals and transthoracic impedance timing signals at a preset sampling frequency, and sets a fixed-length timing window for data buffering to obtain time-aligned triaxial acceleration and transthoracic impedance signals.

[0020] In this embodiment, a sampling frequency of 500Hz is used. The triaxial acceleration timing signal and the transthoracic impedance timing signal are synchronously acquired through a triaxial accelerometer fixed directly above the pressure point and a transthoracic impedance electrode attached to a standard position. A 4-second timing window is set for data buffering. The two signals are synchronized from the same source to ensure strict timing alignment.

[0021] The preprocessing module is used to preprocess the triaxial acceleration signal and the transthoracic impedance signal respectively to obtain the Z-axis acceleration sequence and the differential impedance sequence.

[0022] In this embodiment, after removing the gravity DC component of the triaxial acceleration signal, the Z-axis acceleration component perpendicular to the pressing direction is extracted. The extracted Z-axis acceleration component is then bandpass filtered from 0.5Hz to 5Hz to obtain a noise-reduced and baseline-corrected Z-axis acceleration sequence. First-order difference and neighborhood smoothing filtering are applied to the transthoracic impedance signal to obtain a differential impedance sequence that highlights the thoracic deformation characteristics. The formula for calculating the first-order difference is: in, The result is the result of the difference calculation. This is the current transthoracic impedance value. The transthoracic impedance value at the previous moment. The sampling interval corresponds to a sampling frequency of 500Hz. It takes 2ms.

[0023] The data repair module is used to determine the motion interference range based on the time-domain periodicity of the Z-axis acceleration sequence and the frequency dominance of the differential impedance sequence, and to repair the non-pressing segments within the interference range. In this embodiment, the periodicity of the Z-axis acceleration sequence is detected to determine whether the period is stable. A Fast Fourier Transform is performed on the differential impedance sequence to extract the dominant frequency and determine if it falls within the effective pressing frequency band. When both indicators are abnormal simultaneously, it is determined to be motion interference, and adaptive notch filtering is applied only to the non-pressing intervals to repair the interference, retaining the effective pressing data. When calculating the periodicity, peak detection is performed on the Z-axis acceleration sequence to obtain the duration of three consecutive pressing cycles. The coefficient of variation is calculated based on the duration of the pressing cycles and used as the time-domain periodicity index.

[0024] A Fast Fourier Transform (FFT) is performed on the differential impedance sequence using a Hamming window. The frequency corresponding to the point of maximum amplitude is extracted as the dominant frequency in the frequency domain. When the coefficient of variation is greater than 0.2 and the dominant frequency is less than 1.5Hz or greater than 2.2Hz, the current window is determined to be a motion interference interval. Adaptive notch filtering or linear interpolation methods are used to repair the data of the non-pressing segments within the motion interference interval. First, the non-pressing segments within the motion interference interval are identified. The Z-axis acceleration signal within the non-pressing segments is subjected to spectral analysis to extract the interference dominant frequency, which is then set as the notch frequency of the adaptive notch filter. The signal is notched using the adaptive notch filter. If the root mean square error of the non-pressing segment signal after notch filtering is greater than a preset tolerance threshold, notch filtering is abandoned, and linear interpolation is used instead. The corresponding position signal values ​​of one complete pressing cycle before and after the non-pressing segment are taken and linearly interpolated according to the time ratio to reconstruct the non-pressing segment data. The reconstructed non-pressing segment data is then concatenated with the undisturbed pressing segment data to obtain the repaired Z-axis acceleration sequence.

[0025] The period segmentation and dynamic drift elimination module is used to adaptively segment the repaired Z-axis acceleration sequence. At the end of each pressing cycle, the integral drift DC component is estimated and subtracted from the current cycle velocity sequence and displacement sequence to obtain the drift-corrected velocity sequence and displacement sequence.

[0026] In this embodiment, a dynamic threshold peak detection method is used to adaptively segment the repaired Z-axis acceleration sequence. First, the local mean and standard deviation of the repaired Z-axis acceleration sequence are calculated. Based on this, an adaptive amplitude threshold and a time interval threshold are set. The positive acceleration peak (the start point of the press) and the negative acceleration peak (the end point of the rebound) of each pressing cycle are identified. The interval between two adjacent start points constitutes a complete pressing cycle. Simultaneously, the time interval threshold is dynamically updated based on the duration of 3 to 5 consecutive cycles detected in real time to track changes in the pressing frequency. When an abnormal discontinuity is detected, linear interpolation correction is performed using historical cycle durations to maintain the continuity of the segmentation.

[0027] A short window is selected at the end of each compression cycle, with a preferred window length of 50–150 ms. This is primarily based on the following: the compression rate range for high-quality CPR is 100–120 compressions / min, corresponding to a cycle length of 0.5–0.6 s; and the steady-state window after complete chest recoil typically occupies 15%–25% of the cycle length, resulting in a calculated steady-state window length of approximately 75–150 ms. Furthermore, considering the accelerometer noise level and a 500 Hz sampling rate, to ensure statistical stability of drift estimation, the number of sample points within the window should not be less than 25 (corresponding to 50 ms). Therefore, the short window is calculated based on the compression rate, and a range of 50–150 ms is determined to be effective and robust. In practical applications, the optimal value can be selected within this range through offline experiments or adaptive search.

[0028] The mean values ​​of velocity and displacement are used as estimates of the DC component of the drift, and the calculation formulas are as follows: in, This is an estimate of the integral drift DC component of the velocity sequence. The number of sampling points within the short time window. This is the speed value. This is the estimate of the integral drift DC component of the displacement sequence. This represents the displacement value.

[0029] Subtracting the DC component from the current periodic velocity and displacement sequences yields the drift-corrected velocity and displacement sequences, calculated using the following formulas: in, , The velocity and displacement sequences after drift correction are respectively in The value at time, For the uncorrected integral velocity sequence, the first point, For the uncorrected integral displacement sequence, the first point.

[0030] The pressing depth is taken as the maximum relative displacement within the cycle, which does not depend on the absolute zero point and ensures long-term stability.

[0031] When the short-time window of the above method does not meet the steady-state condition, the drift generated by the second integral of the acceleration signal usually exhibits a slow change that accumulates over time. The drift rate can be deduced by analyzing the difference between displacement or velocity sequences between adjacent periods, without relying on whether an ideal steady-state window appears within any period.

[0032] Specifically, using the displacement sequence of the three consecutive pressing cycles preceding the current pressing cycle, the average displacement within a 50ms window is taken at the end of each cycle as the residual displacement of that cycle. Assuming that the drift changes approximately linearly within the three cycles, the drift rate is fitted using the least squares method. The drift rate is then multiplied by time and subtracted from the velocity and displacement sequence of the current cycle to complete the drift correction.

[0033] The inference module is used to input the repaired Z-axis acceleration sequence, the repaired differential impedance sequence, and the drift-corrected velocity and displacement sequences into a pre-built lightweight multi-branch convolutional neural network, and simultaneously output the compression depth, compression frequency, and the probability of sufficiency of chest recoil.

[0034] In this embodiment, the dual signals within the 4-second window are each truncated into vectors of length 2000, with zero padding for any insufficient portion, forming dual-channel samples. No manual periodic features are extracted, and the period segmentation is used only for alignment and drift correction.

[0035] Figure 2 This paper illustrates the specific structure of a lightweight multi-branch convolutional neural network in an embodiment of the present invention. The network employs a two-branch symmetric architecture specifically designed to process dual-channel input samples constructed from Z-axis acceleration sequences and differential impedance sequences.

[0036] In this embodiment, the lightweight multi-branch convolutional neural network includes: an input layer, which is connected to the input ends of the acceleration branch and the transthoracic impedance branch, respectively. The acceleration branch and the transthoracic impedance branch adopt a symmetrical structure, each including multiple one-dimensional convolutional layers, batch normalization layers, ReLU activation function layers, and global average pooling layers. The output ends of the acceleration branch and the transthoracic impedance branch are connected to the input ends of the feature fusion layer. The output ends of the feature fusion layer are connected to the input ends of the depth regression head, the frequency regression head, and the bounce binary classification head, respectively. The output ends of the depth regression head, the frequency regression head, and the bounce binary classification head are connected to the output layer.

[0037] Input layer: Receives dual-channel samples with dimensions [4, 2000], where 4 corresponds to the Z-axis acceleration sequence, differential impedance sequence, and drift-corrected velocity and displacement sequences, and 2000 corresponds to the number of timing points within a 4-second window at a sampling rate of 500Hz.

[0038] Acceleration branch and transthoracic impedance branch: These two branches have a completely symmetrical structure and independently extract the characteristics of their respective input signals. Each branch contains, in sequence: Three-layer one-dimensional convolution: Used to automatically learn local spatiotemporal correlation features from raw time-series signals, eliminating the need for manually designed periodic features. The number of channels and kernel size of each convolutional layer can be configured according to the computing power of the actual deployment platform. For example, each layer can be set with 16, 32, or 64 convolutional kernels, all with a kernel size of 3. Batch normalization layer: Accelerates network convergence and improves model generalization ability. ReLU activation function layer: Introduces nonlinearity and enhances feature representation ability. Global average pooling layer: Compresses the feature maps extracted from each branch into fixed-length feature vectors, significantly reducing the number of network parameters and effectively preventing overfitting. Feature fusion layer: Concatenates the feature vectors output from the acceleration branch and the impedance branch to form a joint feature representation that integrates the temporal correlation information of the two signals.

[0039] Multi-task output head: Based on the fused features, three parallel output sub-networks are connected to achieve multi-task learning. The depth regression head consists of one or more fully connected layers, outputting continuous compression depth predictions. The frequency regression head has a similar structure to the depth regression head, outputting continuous compression frequency predictions. The rebound binary classification head consists of a fully connected layer followed by a sigmoid activation function, outputting the probability of sufficient chest rebound.

[0040] Network training uses a weighted loss function: in, This represents the total training loss function value for a lightweight multi-branch convolutional neural network. The mean squared error loss for the depth prediction task is used to measure the depth prediction value. Compared with the true value The differences between them The mean squared error loss for the frequency prediction task is used to measure the frequency prediction value. Compared with the true value The differences between them The binary cross-entropy loss is used for the chest recoil adequacy classification task to measure the predicted probability of recoil adequacy. With real labels The differences between them; , , These are the loss weight coefficients for depth regression, frequency regression, and bounce classification tasks, respectively.

[0041] The entire network structure is lightweight with a small number of parameters, making it suitable for deployment on low-power embedded platforms such as automated external defibrillators (AEDs), enabling real-time inference.

[0042] The feedback output module is used to smooth the compression depth, compression frequency and the probability of sufficient chest recoil, and then provide real-time feedback on the compression quality through display and voice.

[0043] In this embodiment, an exponential moving average is used to smooth the compression depth, compression frequency, and chest recoil adequacy probability output by the lightweight multi-branch convolutional neural network. The calculation formula is as follows: in, The result, after being smoothed and processed by an exponential moving average, is the final press quality index value output to the display and voice modules. The smoothing coefficient typically ranges from [0,1] and is used to control the weight distribution between the current original output value and the historical smoothed value. This represents the raw prediction value directly output by the lightweight multi-branch convolutional neural network at the current time t, i.e., the unsmoothed compression depth, compression frequency, or probability of sufficiency of chest recoil. express The feedback result after time-smoothing is used as the historical smoothed value in the recursive calculation of the current time.

[0044] The pressure quality is judged based on the smoothed result, and real-time feedback on the pressure quality is provided through display and voice. The voice content includes issues such as improper pressure, insufficient depth, excessive depth, too slow frequency, too fast frequency, insufficient rebound, and interference, providing real-time operation guidance in all scenarios.

[0045] In this embodiment, when determining the quality of chest compressions, firstly, the smoothed compression depth, compression frequency, and probability of sufficient chest recoil are obtained. These values ​​are then compared with their respective preset threshold ranges and preset probability thresholds. If the smoothed compression depth is less than the lower depth threshold, the compression is determined to be too shallow; if it is greater than the upper depth threshold, the compression is determined to be too deep; if the smoothed compression frequency is less than the lower frequency threshold, the compression is determined to be too slow; if it is greater than the upper frequency threshold, the compression is determined to be too fast; if the smoothed probability of sufficient chest recoil is less than the preset sufficient recoil threshold, the chest recoil is determined to be insufficient; if the smoothed probability of sufficient chest recoil is greater than or equal to the preset sufficient recoil threshold, the chest recoil is determined to be sufficient. Based on these three determinations, a comprehensive chest compression quality feedback command is generated and provided in real-time through display and voice feedback.

[0046] In this embodiment, the triaxial acceleration signal is used independently for the calculation of compression depth and frequency. The transthoracic impedance signal is not involved in the depth and frequency calculation, but is only used in conjunction with the acceleration signal to determine the chest wall rebound. Dynamic drift elimination reduces the dependence on accurate detection of the chest wall rebound state, and period segmentation is only used for signal alignment and drift estimation.

[0047] The monitoring system provided in this application embodiment can be implemented using a combination of hardware and software. As an example, the monitoring system provided in this application embodiment can be a processor in the form of a hardware decoding processor. For example, the processor in the form of a hardware decoding processor can be one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0048] In this embodiment, "multiple" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, the simultaneous existence of A and B, or the existence of B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0049] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0050] The above description is merely a specific embodiment of this application. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application.

Claims

1. A cardiopulmonary resuscitation quality monitoring system based on dual-signal temporal fusion and lightweight CNN, characterized in that, include: The signal acquisition module is used to synchronously acquire the triaxial acceleration timing signal and the transthoracic impedance timing signal at a preset sampling frequency, and to set a timing window of fixed duration for data buffering to obtain time-aligned triaxial acceleration signal and transthoracic impedance signal. The preprocessing module is used to preprocess the triaxial acceleration signal and the transthoracic impedance signal respectively to obtain the Z-axis acceleration sequence and the differential impedance sequence; The data repair module is used to determine the motion interference interval and repair the non-pressing segment within the interference interval based on the time-domain periodicity of the Z-axis acceleration sequence and the frequency dominance of the differential impedance sequence. The period segmentation and dynamic drift elimination module is used to adaptively segment the repaired Z-axis acceleration sequence. At the end of each pressing cycle, the integral drift DC component is estimated and the DC component is subtracted from the current cycle velocity sequence and displacement sequence to obtain the drift-corrected velocity sequence and displacement sequence. The inference module is used to input the repaired Z-axis acceleration sequence, the repaired differential impedance sequence, and the drift-corrected velocity and displacement sequences into a pre-built lightweight multi-branch convolutional neural network, and simultaneously output the compression depth, compression frequency, and the probability of sufficiency of chest recoil. The feedback output module is used to smooth the compression depth, compression frequency and the probability of sufficient chest recoil, and then provide real-time feedback on the compression quality through display and voice.

2. The cardiopulmonary resuscitation quality monitoring system based on dual-signal temporal fusion and lightweight CNN according to claim 1, characterized in that, The triaxial acceleration signal and the transthoracic impedance signal are preprocessed respectively to obtain the Z-axis acceleration sequence and the differential impedance sequence, including: After removing the gravity DC component of the triaxial acceleration signal, the Z-axis acceleration component perpendicular to the pressing direction is extracted. The extracted Z-axis acceleration components are bandpass filtered to obtain the Z-axis acceleration sequence after noise reduction and baseline correction. The transthoracic impedance signal is subjected to first-order differential and neighborhood smoothing filtering to obtain a differential impedance sequence that highlights the thoracic deformation characteristics. The formula for calculating the first-order difference is as follows: in, The result is the result of the difference calculation. This is the current transthoracic impedance value. The transthoracic impedance value at the previous moment. The sampling interval is denoted as .

3. The cardiopulmonary resuscitation quality monitoring system based on dual-signal temporal fusion and lightweight CNN according to claim 1, characterized in that, Based on the time-domain periodicity regularity of the Z-axis acceleration sequence and the dominant frequency of the differential impedance sequence, the motion interference interval is determined, and data repair is performed on the non-pressing segments within the interference interval, including: Peak detection is performed on the Z-axis acceleration sequence to obtain the duration of multiple consecutive pressing cycles. The coefficient of variation is calculated based on the duration of the pressing cycle, and the coefficient of variation is used as a time-domain periodicity index. Perform a Fast Fourier Transform on the differential impedance sequence and extract the frequency corresponding to the point of maximum amplitude as the dominant frequency in the frequency domain. When the coefficient of variation is greater than the first preset value and the main frequency is less than the second preset value or greater than the third preset value, the current window is determined to be a motion interference zone. Adaptive notch filtering or linear interpolation methods are used to repair the data in the non-pressed segments within the motion interference range.

4. The cardiopulmonary resuscitation quality monitoring system based on dual-signal temporal fusion and lightweight CNN according to claim 3, characterized in that, The method of using adaptive notch filtering or linear interpolation to repair data in the non-pressing segment within the motion interference range includes: The non-pressing segment within the motion interference zone is determined; The Z-axis acceleration signal in the non-pressing segment is subjected to spectrum analysis to extract the interference main frequency, and the interference main frequency is set as the notch frequency of the adaptive notch filter. The signal is notched filtered using the adaptive notch filter. If the root mean square error of the non-pressed segment signal after notch filtering is greater than the preset tolerance threshold, then notch filtering is abandoned and linear interpolation is used instead. Take the corresponding position signal values ​​of one complete pressing cycle before and after the non-pressing segment, perform linear interpolation according to the time ratio, and reconstruct the non-pressing segment data; The reconstructed non-pressing segment data is spliced ​​with the undisturbed pressing segment data to obtain the repaired Z-axis acceleration sequence.

5. The cardiopulmonary resuscitation quality monitoring system based on dual-signal temporal fusion and lightweight CNN according to claim 1, characterized in that, Adaptive period segmentation is performed on the repaired Z-axis acceleration sequence. At the end of each pressing cycle, the integral drift DC component is estimated and subtracted from the current cycle's velocity and displacement sequence to obtain the drift-corrected velocity and displacement sequences, including: The dynamic threshold peak detection method is used to adaptively segment the Z-axis acceleration sequence after repair. The start and end positions of each pressing cycle are determined by identifying the dynamic threshold of the acceleration peak. At the end of each pressing cycle, a short-time window is selected, and the steady-state conditions of the short-time window are judged. When the short-time window meets the steady-state conditions, the integral drift DC component is estimated, and the calculation formulas are as follows: in, This is an estimate of the integral drift DC component of the velocity sequence. The number of sampling points within the short time window. This is the speed value. This is the estimate of the integral drift DC component of the displacement sequence. This is the displacement value; The DC component is subtracted from the current period velocity sequence and displacement sequence to obtain the drift-corrected velocity sequence and displacement sequence, calculated using the following formulas: in, , The velocity and displacement sequences after drift correction are respectively in The value at time, For the uncorrected integral velocity sequence, the first point, For the uncorrected integral displacement sequence, the first point; When the short-time window does not meet the steady-state condition, the drift rate is deduced from the difference between the displacement sequence or velocity sequence between adjacent cycles by using the integral drift accumulation law of multiple consecutive pressing cycles, and the drift-corrected velocity sequence and displacement sequence are obtained based on the drift rate.

6. The cardiopulmonary resuscitation quality monitoring system based on dual-signal temporal fusion and lightweight CNN according to claim 5, characterized in that, When the short-time window does not meet the steady-state condition, the drift rate is deduced from the difference between the displacement or velocity sequences between adjacent cycles by utilizing the integral drift accumulation law of multiple consecutive pressing cycles. Based on the drift rate, the drift-corrected velocity and displacement sequences are obtained, including: If the range of the displacement sequence within the short-time window of the current pressing cycle is greater than the preset threshold, it is determined that the steady-state condition is not met. Obtain the displacement or velocity sequence of multiple consecutive pressing cycles preceding the current pressing cycle; At the end of each cycle, a window of fixed length is selected, and the average displacement or average velocity within the window is calculated as the residual drift for the current cycle. Assuming that the drift changes linearly over multiple consecutive periods, the least squares method is used to fit the drift rate. After multiplying the drift rate by the time offset, subtract it from the velocity sequence and displacement sequence of the current period to obtain the drift-corrected velocity sequence and displacement sequence.

7. The cardiopulmonary resuscitation quality monitoring system based on dual-signal temporal fusion and lightweight CNN according to claim 1, characterized in that, The lightweight multi-branch convolutional neural network includes an input layer, which is connected to the inputs of an acceleration branch and a transthoracic impedance branch. The acceleration branch and the transthoracic impedance branch adopt a symmetrical structure, each including multiple one-dimensional convolutional layers, batch normalization layers, ReLU activation function layers, and global average pooling layers. The outputs of the acceleration branch and the transthoracic impedance branch are connected to the inputs of a feature fusion layer. The outputs of the feature fusion layer are connected to the inputs of a depth regression head, a frequency regression head, and a bounce binary classification head, respectively. The outputs of the depth regression head, the frequency regression head, and the bounce binary classification head are connected to the output layer. The lightweight multi-branch convolutional neural network uses the following loss function during training: in, This represents the total training loss function value for a lightweight multi-branch convolutional neural network. The mean squared error loss for the depth prediction task is used to measure the depth prediction value. Compared with the true value The differences between them The mean squared error loss for the frequency prediction task is used to measure the frequency prediction value. Compared with the true value The differences between them The binary cross-entropy loss is used for the chest recoil adequacy classification task to measure the predicted probability of recoil adequacy. With real labels The differences between them; , , These are the loss weight coefficients for depth regression, frequency regression, and bounce classification tasks, respectively.

8. The cardiopulmonary resuscitation quality monitoring system based on dual-signal temporal fusion and lightweight CNN according to claim 1, characterized in that, The repaired Z-axis acceleration sequence, the repaired differential impedance sequence, and the drift-corrected velocity and displacement sequences are input into a pre-constructed lightweight multi-branch convolutional neural network, which simultaneously outputs compression depth, compression frequency, and the probability of sufficiency of chest recoil, including: The repaired Z-axis acceleration sequence, the drift-corrected velocity sequence, and the displacement sequence are input into the acceleration branch of the lightweight multi-branch convolutional neural network, and the repaired differential impedance sequence is input into the transthoracic impedance branch of the lightweight multi-branch convolutional neural network. Each branch is processed sequentially through multiple layers of one-dimensional convolution, batch normalization layers, ReLU activation function layers, and global average pooling layers, outputting fixed-length acceleration branch feature vectors and impedance branch feature vectors respectively. The acceleration branch feature vector and the impedance branch feature vector are concatenated along the channel dimension to obtain the fused feature vector; The fused feature vectors are input into the depth regression head, frequency regression head, and rebound binary classification head of a lightweight multi-branch convolutional neural network, respectively, and the predicted values ​​of compression depth, compression frequency, and sufficiency of chest recoil are output simultaneously.

9. The cardiopulmonary resuscitation quality monitoring system based on dual-signal temporal fusion and lightweight CNN according to claim 1, characterized in that, After smoothing the compression depth, compression frequency, and probability of sufficiency of chest recoil, real-time feedback on compression quality is provided through display and voice, including: The compression depth, compression frequency, and thoracic recoil adequacy probability output by the lightweight multi-branch convolutional neural network are smoothed using an exponential moving average. The calculation formula is as follows: in, The result is smoothed and used for feedback. For smoothing coefficients, This represents the raw predicted values ​​directly output by a lightweight multi-branch convolutional neural network at the current time t, i.e., the unsmoothed compression depth, compression frequency, or probability of sufficiency of chest recoil. express Feedback results after real-time smoothing; The pressing quality is judged based on the smoothed result, and real-time feedback on the pressing quality is provided through display and voice.

10. The cardiopulmonary resuscitation quality monitoring system based on dual-signal temporal fusion and lightweight CNN according to claim 9, characterized in that, The step of determining the pressing quality based on the smoothed result includes: Obtain the smoothed compression depth, compression frequency, and probability of sufficiency of chest recoil; The smoothed compression depth, compression frequency, and probability of sufficient chest recoil are compared with their respective preset threshold ranges and preset probability thresholds. If the smoothed press depth is less than the lower limit threshold, the press depth is determined to be too shallow; if the smoothed press depth is greater than the upper limit threshold, the press depth is determined to be too deep. If the smoothed pressing frequency is less than the lower frequency threshold, the pressing frequency is determined to be too slow; if the smoothed pressing frequency is greater than the upper frequency threshold, the pressing frequency is determined to be too fast. If the probability of sufficient rebound of the smoothed thoracic cavity is less than the preset sufficient rebound threshold, the thoracic cavity rebound state is determined to be insufficient rebound; if the probability of sufficient rebound of the smoothed thoracic cavity is greater than or equal to the preset sufficient rebound threshold, the thoracic cavity rebound state is determined to be sufficient rebound.

Images