Waveform subtype-based impedance cardiography feature point detection and quality control method

Abstract

The application provides an impedance cardiogram feature point detection and quality control method based on waveform subtypes, and the dZ / dt signal, the base impedance Z0 signal and the electrocardiogram ECG signal of an impedance cardiogram ICG are synchronously collected; the dZ / dt signal is pretreated, a heart beat segment is intercepted, a quality feature vector is extracted for each heart beat segment, and a quality score is calculated; the comparative learning loss is used for training, and the homologous enhancement view is narrowed; the streaming prototype clustering is performed; the point-level uncertainty of each feature point is calculated, if the point-level uncertainty exceeds a set threshold, then the constrained relocation is performed; the low-confidence counter is established, and the low-confidence closed-loop control is realized based on the quality score and the point-level uncertainty. The application can trigger the closed-loop control such as a prompt or calibration in a continuous low-confidence state, and improves the robustness and availability of the ICG feature point detection and the blood dynamics parameter estimation in a real scene.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of biomedical technology, and particularly relates to a waveform subtype-based impedance cardiography feature point detection and quality control method. BACKGROUND

[0002] Impedance Cardiography (ICG) is a technology for non-invasive and continuous hemodynamic monitoring by measuring the thoracic impedance change with the cardiac cycle and analyzing the derivative signal dZ / dt. In applications such as bedside monitoring, wearable continuous monitoring and remote health management, ICG signals and electrocardiogram (ECG) signals are usually collected synchronously, the R wave of ECG is used to locate the cardiac event and analyze the ICG beat segment, and then parameters such as pre-ejection period (PEP), left ventricular ejection time (LVET) and stroke volume (SV) are calculated. The accuracy of ICG parameter calculation depends largely on the stable detection and time positioning of feature points such as B point, C point, X point and O point, so the robustness and quality control ability of feature point detection are the key to the engineering landing of ICG.

[0003] In the prior art, ICG feature point detection methods mostly use fixed threshold rules, fixed template matching or single supervised models for positioning. Such methods usually assume that ICG waveforms have relatively stable morphological structures, but in actual acquisition conditions, ICG waveform morphology will change significantly with individual differences, body position changes, respiratory status, electrode contact changes and motion artifacts, etc., showing changes in the main rising slope, peak-valley number and relative timing, and even the weakening or absence of feature points, which leads to false detection, missed detection or point drift of algorithms based on single morphological assumptions, further causing the cumulative estimation deviation of hemodynamic parameters such as PEP, LVET and SV, and affecting the usability and reliability of continuous monitoring.

[0004] Furthermore, common issues in real-world signals, such as baseline drift, power frequency interference, superposition of low-frequency respiratory components, and transient motion artifacts, can significantly reduce the signal-to-noise ratio (SNR) or morphological distortion in some cardiac segments. Existing solutions often only use simple SNR thresholds or empirical rules to eliminate segments, lacking a quantifiable scoring mechanism for cardiac segment quality and a point-level evaluation and output control strategy for the reliability of feature point detection results. Without reliable quality control, the system may continuously output incorrect feature points and parameter trends, making it difficult to trigger timely error correction actions such as electrode checks, re-acquisition, or calibration. Further, existing solutions have attempted to introduce clustering or classification to handle waveform differences, but common distance-based clustering methods have high computational costs and are sensitive to noise, making them unsuitable for wearable and edge real-time deployment requirements. Purely supervised classification or regression models rely on labeled data and are sensitive to distribution drift; when waveforms drift over time or new waveform subtypes emerge, the uncertainty of the model output increases, and an interpretable error correction path is still lacking. Furthermore, existing technologies rarely integrate the processes of "waveform subtype identification, feature point detection, point-level uncertainty assessment, constraint relocation, and continuous low-confidence closed-loop triggering" into a unified chain, resulting in insufficient stability and security of the system in complex scenarios.

[0005] Therefore, there is an urgent need for a technical solution that can achieve adaptive feature point detection under multi-subtype ICG waveform conditions, quantitatively evaluate the quality and point-level reliability of cardiac segments, relocate under uncertain conditions through physiological timing constraints, and trigger closed-loop control such as prompts or calibration under continuous low reliability conditions, so as to improve the robustness and usability of ICG feature point detection and hemodynamic parameter estimation in real-world scenarios. Summary of the Invention

[0006] This invention proposes a waveform subtype-based impedance cardiography feature point detection and quality control method, which can achieve adaptive feature point detection under multi-subtype ICG waveform conditions, and quantitatively evaluate the quality of cardiac segments and point-level reliability. When uncertain, it can perform relocation through physiological timing constraints, and trigger closed-loop control such as prompts or calibration in continuous low reliability states, thereby improving the robustness and usability of ICG feature point detection and hemodynamic parameter estimation in real-world scenarios.

[0007] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0008] A method for impedance echocardiography feature point detection and quality control based on waveform subtypes includes:

[0009] S1. Simultaneously acquire the dZ / dt signal, the base impedance Z0 signal, and the ECG signal from impedance cardiograph (ICG);

[0010] S2. Preprocess the dZ / dt signal. The preprocessing steps are as follows: baseline correction, notch filtering, bandpass filtering, and amplitude normalization.

[0011] S3. Detect the R-wave position of the ECG signal, extract the heartbeat segment of the preprocessed ICG based on the R-wave position, and resample each heartbeat segment and align it with the R-wave. The purpose of aligning with the R-wave is to keep all heartbeat segments consistent in terms of time dimension and feature position.

[0012] S4. Extract the quality feature vector for each heartbeat segment, and calculate the quality score based on the quality feature vector according to the quality scoring formula.

[0013] S5. Construct two augmented views for heartbeat segments that pass the quality score. After encoding and normalization by the encoder, obtain the embedding vector h. Use contrastive learning loss for training to bring the augmented views of the same source closer together and separate the heartbeats of different sources.

[0014] S6. Perform streaming prototype clustering on the embedded vector h using cosine distance. If the cosine distance exceeds the threshold for creating a new subtype, create a new ICG waveform subtype and initialize the prototype. Otherwise, it is assigned to the nearest ICG waveform subtype and the prototype is updated.

[0015] S7. Based on the ICG waveform subtype, call the detection network to identify feature points, and calculate the point-level uncertainty of each feature point through multiple inactivation inferences;

[0016] S8. If the point-level uncertainty of a feature point exceeds the set threshold, then within the candidate interval of that feature point, constrained relocation is performed by combining the first and second derivatives of the heartbeat segment with the time prior penalty of the relative R wave.

[0017] S9. Establish a low-confidence counter, and statistically analyze low-confidence beat segments based on quality scores and point-level uncertainty. If the limit is exceeded, trigger an electrode contact check and output a low-confidence flag. If a set number of qualified beat segments appear consecutively, output the time position and confidence flag of each feature point.

[0018] Furthermore, in step S1, the dZ / dt signal, Z0 signal, and ECG signal are sampled synchronously using the same clock source; the ECG signal provides a time reference for the localization of cardiac events, and the Z0 signal is used for scale correction in subsequent calculations of hemodynamic parameters.

[0019] Furthermore, in step S2, the discrete sequence of the dZ / dt signal is processed by sliding median filtering to obtain the baseline and then the baseline is corrected. The corrected sequence is then filtered by a second-order IIR notch filter, the notched sequence is then filtered by a bandpass filter, and the amplitude of the bandpass filtered sequence is then normalized.

[0020] Furthermore, in step S3, a time window is extracted based on the R-wave position to obtain a cardiac segment. The time window covers the electromechanical coupling preparation segment before the R-wave and the ejection-related segment after the R-wave, so that feature points including points B, C, X, and O all fall within the cardiac segment.

[0021] Furthermore, the quality feature vector in step S4 includes: peak-to-peak amplitude of the cardiac segment, number of effective peak-to-valley pairs within the cardiac segment, slope of the main rising edge, spectral entropy, energy or energy percentage within the target physiological frequency band, and time interval deviation between adjacent cardiac segments.

[0022] Furthermore, the enhancement methods in step S5 for constructing the enhanced view include amplitude scaling, temporal micro-warping, additive Gaussian noise, and random segment occlusion.

[0023] Furthermore, the detection network mentioned in step S7 is a subtype expert feature point detection network corresponding to the ICG waveform subtype. Different subtype expert feature point detection networks learn different feature point localization strategies for the morphological differences of the corresponding ICG waveform subtype.

[0024] Furthermore, the constraint relocation described in step S8 is achieved by minimizing the objective function containing the first derivative, second derivative, and time prior penalty term of the relative R wave within the candidate interval of the feature point.

[0025] Furthermore, in step S9, a heartbeat segment with a qualified quality score and a point-level uncertainty greater than a set threshold at any feature point is a low-confidence heartbeat segment; a heartbeat segment with a qualified quality score and a point-level uncertainty at no feature point greater than a set threshold is a qualified heartbeat segment.

[0026] In another aspect, this invention proposes an impedance cardiography feature point detection and quality control system based on waveform subtypes, comprising:

[0027] Synchronous acquisition module: Synchronously acquires the dZ / dt signal, base impedance Z0 signal, and ECG signal from impedance cardiography (ICG);

[0028] Preprocessing module: preprocesses the dZ / dt signal. The preprocessing process is as follows: baseline correction, notch filtering, bandpass filtering, and amplitude normalization.

[0029] R-wave detection and heartbeat segmentation module: detects the R-wave position of the ECG signal, extracts heartbeat segments from the preprocessed ICG based on the R-wave position, and resamples each heartbeat segment and aligns it with the R-wave to ensure that all heartbeat segments are consistent in terms of time dimension and feature position;

[0030] Quality scoring module: Extracts quality feature vectors for each heartbeat segment, and calculates the quality score based on the quality feature vectors according to the quality scoring formula;

[0031] Contrastive learning encoder module: Construct two augmented views for heartbeat segments with qualified quality scores, encode and normalize them to obtain the embedding vector h, and use contrastive learning loss for training to bring the augmented views of the same source closer together and separate the heartbeats of different sources.

[0032] Prototype Clustering / Subtype Management Module: Performs streaming prototype clustering on the embedded vector h using cosine distance. If the cosine distance exceeds the threshold for creating a new subtype, a new ICG waveform subtype is created and the prototype is initialized. Otherwise, it is assigned to the nearest ICG waveform subtype and the prototype is updated.

[0033] Uncertainty calculation module: Based on the ICG waveform subtype, the detection network is called to identify feature points, and the point-level uncertainty of each feature point is calculated through multiple inactivation inferences;

[0034] Constraint relocation module: If the point-level uncertainty of a feature point exceeds a set threshold, then within the candidate interval of that feature point, constraint relocation is performed by combining the first and second derivatives of the cardiac segment with the time prior penalty of the relative R wave.

[0035] Low-confidence closed-loop control module: Establish a low-confidence counter, statistically analyze low-confidence heartbeat segments based on quality score and point-level uncertainty, and trigger electrode contact check if the limit is exceeded; if a set number of qualified heartbeat segments appear consecutively, output the time position and confidence mark of each feature point.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] 1. This invention extracts quality feature vectors from cardiac beat segments and calculates quality scores. When the quality score is below a threshold, the segment is directly discarded, preventing low-quality cardiac beats from entering the subsequent detection chain. This reduces false detections, missed detections, and position drift caused by noise artifacts from the source, thereby improving the stability of continuous monitoring output.

[0038] 2. This invention constructs an enhanced view and uses contrastive learning to train the encoder to obtain robust embedding representations of amplitude changes, temporal micro-distortions, and local occlusion. It also combines streaming prototype clustering to achieve online creation and updating of waveform subtypes, which can adapt to multi-subtype waveforms and long-term drift scenarios, avoiding mismatch between fixed templates or offline models.

[0039] 3. After determining the subtype, this invention calls the corresponding subtype expert feature point detection network to output each feature point B, C, X, and O. The point-level uncertainty is calculated through deactivation and repeated inference to achieve a quantitative assessment of the reliability of the point, thereby providing a clear basis for error correction and output control in abnormal scenarios.

[0040] 4. When the point-level uncertainty exceeds a set threshold, this invention performs constraint relocation by minimizing the objective function containing derivative terms and time prior penalty terms within the candidate interval. It also combines continuous low-confidence counts to trigger electrode checks and output low-confidence markers, forming a closed-loop control. This significantly improves the physiological rationality of feature point results and the reliability and engineering usability of hemodynamic parameter (PEP, LVET, SV) estimation. Attached Figure Description

[0041] Figure 1 This is a flowchart of the method in Embodiment 1 of the present invention;

[0042] Figure 2 This is a schematic diagram of ICG signal heartbeat segment segmentation and feature point definition in Embodiment 1 of the present invention;

[0043] Figure 3 This is a schematic diagram of ICG signal contrast learning embedding training and streaming prototype clustering in Embodiment 1 of the present invention;

[0044] Figure 4 This is a schematic diagram of the ICG signal point-level uncertainty calculation and constraint relocation mechanism in Embodiment 1 of the present invention;

[0045] Figure 5 This is a schematic diagram of ICG signal closed-loop quality control and hemodynamic parameter output in Embodiment 1 of the present invention;

[0046] Figure 6 This is a system structure block diagram of Embodiment 2 of the present invention;

[0047] Figure 7 This is a comparison diagram of feature point annotations in Embodiment 3 of the present invention;

[0048] (a) is a comparison of the annotations at point B, (b) is a comparison of the annotations at point C, and (c) is a comparison of the annotations at point X.

[0049] Figure 8 This is a comparison diagram of the ICG signals before and after filtering in Embodiment 3 of the present invention;

[0050] (a) is the ICG signal diagram before filtering, and (b) is the ICG signal diagram after filtering.

[0051] Figure 9 This is a comparison chart of the median of the LVET algorithm in Embodiment 3 of the present invention. Detailed Implementation

[0052] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0053] To make the purpose and features of this invention patent more apparent and understandable, the invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0054] Example 1:

[0055] This embodiment provides a detailed description and technical implementation of the impedance cardiograph feature point detection and quality control method based on waveform subtype proposed in this invention.

[0056] like Figure 1 As shown, the method includes the following steps:

[0057] I. Synchronous signal acquisition:

[0058] The dZ / dt signal, base impedance Z0 signal, and ECG signal of impedance cardiography (ICG) are acquired simultaneously. The ECG signal is used to provide a time reference for the localization of cardiac events, and the Z0 signal is used for scale correction in subsequent hemodynamic parameter calculations. The dZ / dt signal, Z0 signal, and ECG signal are sampled synchronously using the same clock source to reduce cross-channel delay errors.

[0059] II. Signal Preprocessing:

[0060] The dZ / dt signal is preprocessed, including obtaining the baseline by using a sliding median filter with a length of 0.8 seconds. And perform baseline correction ,right Obtained through a second-order IIR notch filter with a center frequency of 50Hz. ,right Obtained through a 0.7Hz to 20Hz bandpass filter And perform amplitude normalization. ;

[0061] in, The original dZ / dt signal discrete sequence, The baseline for the moving median filter estimation. This is the baseline-corrected sequence. The sequence after notch filtering, This is the sequence after bandpass filtering. This is the sequence after amplitude normalization, where n is the index of the sampling point.

[0062] The sliding median filter described in this step is used to suppress the influence of non-Gaussian spike artifacts and slow drift on the baseline. The notch filter is used to suppress power frequency interference. The bandpass filter is used to retain the main frequency bands related to heartbeat and suppress low-frequency and high-frequency respiratory noise. Amplitude normalization reduces the bias caused by amplitude scale changes in different individuals and different electrode contact states on subsequent embedding characterization and clustering.

[0063] III. R-wave detection and heartbeat segmentation:

[0064] In this embodiment, the Pan-Tompkins algorithm is used to detect the R-wave position of the ECG signal. , with each Cut off based on the benchmark The time window yields heartbeat segments, and each heartbeat segment is resampled into 256 points and aligned with the R wave.

[0065] The time window covers the electromechanical coupling preparation segment before the R wave and the ejection-related segment after the R wave, ensuring that feature points including points B, C, X, and O all fall within the cardiac segment; resampling is used to maintain the consistency of network input dimensions under different heart rate conditions, thereby facilitating comparative learning and subtype expert network inference.

[0066] In this step, the resampling is exemplarily set to 256 points, but it is not limited to 256 points. The specific setting of sampling points is related to the data processing method of resampling. For example, directly extracting the first 100 points and the last 156 points centered on the R-wave position results in a resampling of 256 points. However, if a 1-second derivative segment of 200 ms before and 800 ms after the R-wave position is extracted, and then subjected to notch filtering, bandpass filtering, smoothing, high-pass drift removal, normalization, and quality screening, the effective segment is trimmed to 80% of its original length, ultimately resulting in 160 points, then the resampling is set to 160 points. Different data processing methods result in different numbers of resampling points. As long as all heartbeat segments remain consistent in the time dimension and feature position, and the network input dimension is consistent under different heart rate conditions, the number of resampling points can be set in any way.

[0067] like Figure 2 The diagram shows the segmentation of cardiac beats in an ICG signal and the definition of feature points. The upper part of the diagram shows the ECG signal with the R wave position marked. Below is an impedance echocardiogram dZ / dt heartbeat segment aligned with the R wave; truncated based on each R wave. A time window is used to obtain a cardiac segment, which is then resampled to 256 points to obtain a uniform length input. Further, impedance cardiography feature points are defined within the cardiac segment, including points B, C, X, and O, corresponding to the initial feature, main peak feature, ejection-related inflection point / trough feature, and early diastolic recovery feature of the ejection-related waveform, respectively. Their time positions are denoted as follows: The schematic diagram can also provide the candidate search intervals for each feature point. ,in Set to 0.05s–0.25s after the R wave. Set the time to 0.02s to 0.20s after point B. Set the time to 0.05s to 0.35s after point C. The time interval is set to 0.05s to 0.45s after point X, thus providing a physiological temporal constraint basis for subsequent feature point detection and constraint relocation.

[0068] IV. Quality Score:

[0069] Extract quality feature vectors for each heartbeat segment .

[0070] in, It represents the peak-to-peak amplitude of a cardiac segment, used to characterize the strength of the pulse wave amplitude; This is the difference between the maximum and minimum values ​​of the cardiac segment.

[0071] It indicates the number of effective peak-valley pairs or the completeness of peak-valley pairing within a segment, used to characterize whether the waveform periodic structure is complete; This represents the total number of local maxima and local minima.

[0072] Indicates the slope of the main rising edge. , The first derivative of a heartbeat segment; Used to characterize the steepness of the rising limb of a waveform.

[0073] Spectral entropy is used to characterize the complexity of frequency domain distribution and the degree of noise pollution. , where m is the frequency discrete index. This represents the normalized power spectral density at the m-th frequency point.

[0074] It represents the energy or energy percentage within the target physiological frequency band, used to characterize whether the effective pulse components are concentrated in the target frequency band; ;in This is the frequency domain representation of a heartbeat segment.

[0075] It indicates the deviation between adjacent heartbeat intervals or the deviation of the current segment time interval, and is used to characterize rhythm stability. ,in R-wave time position, for The time position when the maximum value is reached.

[0076] Then based on the quality feature vector According to the quality scoring formula Calculate the quality score; where , is the Sigmoid activation function; is the weight vector, representing the weights of each feature in the quality feature vector; T denotes transpose; This is a quality bias term.

[0077] parameter By minimizing cross-entropy loss Learning is gained, among which The training loss is for quality scoring; For the first Quality score of each cardiac segment; For the first Quality labels for individual heartbeat segments; For sample index; , These are the upper threshold for spectral entropy and the lower threshold for in-band energy, respectively. Quality label. Generated according to the following rules: when the heartbeat fragment satisfies , , , Time to take Otherwise take ,and , .

[0078] when If the heartbeat segment fails to meet the quality score, the segment is discarded and no feature points are output.

[0079] The quality feature vector characterizes the fragment quality from three aspects: temporal morphology, extreme value structure, and frequency domain energy distribution. This enables the screening process to constrain baseline drift, power frequency residue, and motion artifacts. By removing low-quality fragments before embedding learning and subtype modeling, the cascading impact of junk fragments on prototype creation and feature point output can be reduced, thereby improving the overall stability of the entire process.

[0080] V. Comparative Learning Encoding:

[0081] when The quality score of the cardiac segment is qualified, and two enhanced views are constructed for the cardiac segment. , The enhancement methods include amplitude scaling, temporal micro-warping, additive Gaussian noise, and random segment occlusion. The enhanced view is input into a self-supervised contrastive learning encoder to obtain a 128-dimensional embedding, which is then L2 normalized. The normalization formula is as follows: , to obtain the embedding vector Therefore, based on the two enhanced views We can obtain the two corresponding embedding vectors. The encoder is trained using a contrastive learning loss.

[0082] The contrastive learning loss is implemented using the following formula:

[0083] ;

[0084] in This represents the comparison loss function value. express For feature similarity calculation, this embodiment uses cosine similarity; This represents a temperature coefficient used to adjust the smoothness of the similarity distribution; for example, The numerator represents the positive sample pairs. The similarity, the denominator term represents the sample Overall similarity with other samples within the batch. By minimizing this loss, different enhanced views of the same heartbeat segment can be made closer in feature space, while remaining distinguishable from the feature representations of other heartbeat segments.

[0085] In this step, amplitude scaling is used to simulate amplitude changes caused by differences in electrode contact and individual impedance; temporal micro-warping is used to simulate small heart rate changes and local morphological stretching; additive Gaussian noise is used to simulate measurement noise; and random segment occlusion is used to improve the model's robustness to local missing data and short-term artifacts. Through contrastive learning, the embedding of the same heartbeat under reasonable perturbation is kept consistent, so that the embedding space can better reflect the differences in the real waveform morphology, providing a stable and separable representation basis for online prototype clustering.

[0086] Figure 3 This diagram illustrates the ICG signal contrastive learning embedding training and streaming prototype clustering of the present invention. The left side of the diagram shows the contrastive learning embedding training process: constructing two enhanced views for the same heartbeat segment that has passed quality screening. and The enhancement methods include amplitude scaling, temporal micro-warping, additive Gaussian noise, and random segment occlusion; the enhanced views are then input into the self-supervised encoder Enc to obtain embedding vectors. , Furthermore, the embedding vector is L2 normalized to make the embedding robust to amplitude changes and local missing values. The encoder is trained by contrastive learning loss function so that the embeddings of the same heartbeat segment under different enhancements maintain high similarity while maintaining separability from the embeddings of other heartbeat segments, thereby obtaining a stable representation for waveform subtype recognition.

[0087] VI. Prototype Clustering / Subtype Management:

[0088] For embedding vectors Perform streaming prototype clustering and use cosine distance Distribute;

[0089] when Create a new waveform subtype and initialize a new prototype. ;

[0090] when The prototype is assigned the subtype with the smallest distance and updated according to the prototype formula. Update the prototype;

[0091] in It is the first Each subtype prototype, Is it a new prototype? 0.35 is the subtype label index, 0.35 is the threshold for creating a new subtype, and 0.90 / 0.10 are the prototype update coefficients.

[0092] In this step, after accumulating 200 satisfactory results... After a valid cardiac segment, a subtype merging is triggered, and merging is only performed on two subtypes with a sample size of not less than 30; when two subtype prototypes exist... Satisfying cosine similarity And satisfy the average embedding distance difference of the two subtypes' most recent 200 effective fragments At that time, the two subtypes were merged into the same subtype and then... Update the merged prototype, in which The number of valid fragments has been allocated to the two subtypes. and The average embedding distance from the most recent 200 valid fragments to their respective prototypes. For the prototypes of the two subtypes to be merged, This is the merged prototype; the merging strategy is used to suppress prototype fragmentation and improve the long-term stability of the subtype model.

[0093] This step enables the system to automatically absorb unknown new subtypes, and at the same time, through exponential sliding updates, the prototype can smoothly change with long-term drift, thereby avoiding prototype mutations caused by a single anomaly. This step changes the definition of waveform subtypes from a fixed template to an updatable prototype embedded in space, improving adaptability to different populations, different states, and long-term continuous monitoring.

[0094] Figure 3 The diagram on the right illustrates the streaming prototype clustering process: the embedding vector of the current heartbeat segment. With existing prototype collection Calculate cosine distance and according to Subtype assignment or new subtype creation is performed based on a preset creation threshold, where It is the first Each subtype prototype, Is it a new prototype? This is the subtype label index, 0.35 is the new subtype creation threshold, and 0.90 / 0.10 are the prototype update coefficients. When... When the value is greater than the threshold, a new waveform subtype is created and used. Initialize new prototype ,when When less than or equal to the threshold, The prototype is assigned to the subtype corresponding to the smallest distance, and the corresponding prototype is updated according to the online update formula. Through the above combination mechanism of "contrastive learning embedding - prototype creation threshold - online update", the adaptive establishment and continuous tracking of ICG waveform subtypes are realized, providing a basis for subsequent subtype-based invocation of expert networks for feature point detection.

[0095] VII. Uncertainty Calculation:

[0096] For a given waveform subtype, the corresponding subtype expert feature point detection network is invoked to output the time positions of points B, C, X, and O. Expert networks for different subtypes learn different feature point localization strategies based on their morphological differences, thereby reducing the mismatch risk of a single detection network under morphological conditions such as weak B points, bimodal C points, or unclear X points, and improving the accuracy of feature point detection.

[0097] The subtype expert feature point detection network is a feature point localization model for one-dimensional physiological time-series signals. Different ICG waveform subtypes correspond to separate subtype expert feature point detection networks with independent parameters. Each subtype expert feature point detection network uses the same basic network structure, but completes parameter learning through its own subtype training samples to adapt to the differences in waveform morphology, peak-valley distribution, and relative temporal sequence of points B, C, X, and O under different subtypes. For example, the subtype expert feature point detection network adopts a structure combining a one-dimensional convolutional neural network and a recurrent temporal modeling network. Through convolution operations, it extracts local temporal features related to the main rising edge, local peaks and valleys, inflection point changes, and the diastolic rise segment layer by layer. After convolutional feature extraction, a bidirectional gated recurrent unit (Bi-GRU) network is used to set the temporal feature modeling layer to utilize the sequential dependencies between different feature points. Finally, a regression method is used to directly output the temporal positions of the four feature points. Each subtype expert feature point detection network is trained using supervised learning, and the training samples consist of ICG heartbeat segments that have undergone quality screening, subtype classification, and feature point location annotation.

[0098] The inference process was repeated 30 times with the detection network deactivated to obtain the time output set for each feature point. ,in , indicating the feature point category; Indicates the first Secondary inactivation reasoning; Indicates the first The time of the next output point.

[0099] Then calculate the point-level uncertainty using the point-level uncertainty formula. ;

[0100] ;

[0101] in to The temporal output for the same feature point under 30 inactivation inferences. It is a function of standard deviation. This is a truncation function; by normalizing the standard deviation by 20ms and truncating it to the 0-1 interval, it facilitates unified threshold triggering constraint relocation and closed-loop quality control.

[0102] In this step, repeated inference can reflect the fluctuation of the model output under the randomness of parameter deactivation. The greater the output fluctuation, the more uncertain the model is about the point. This invention uses this uncertainty to trigger subsequent constraint relocation and closed-loop control, thereby realizing a closed-loop chain of "detection-measurement-error correction-control".

[0103] Figure 4 This is a schematic diagram of the ICG signal point-level uncertainty assessment and constraint relocation mechanism of the present invention. In the point-level uncertainty assessment process on the left side of the diagram, for the same heartbeat segment with an assigned waveform subtype, the corresponding subtype expert feature point detection network is invoked, and the inference is repeated a preset number of times under inactivation conditions to obtain the feature points. Time output set Then, the discreteness of the output set is calculated and the point-level uncertainty is obtained. ,in The larger the value, the greater the output fluctuation at that point and the lower the reliability.

[0104] 8. Constraint Relocation:

[0105] When the point-level uncertainty of any feature point satisfies When this occurs, constraint relocation is performed on the feature point, the constraint relocation being performed within the candidate interval. Minimize the objective function T accomplish;

[0106] The objective function is: ;

[0107] The candidate interval for point B is set to 0.05s–0.25s after the R wave, the candidate interval for point C is set to 0.02s–0.20s after point B, the candidate interval for point X is set to 0.05s–0.35s after point C, and the candidate interval for point O is set to 0.05s–0.45s after point X, and the parameters are... , t Candidate time positions, and These are the first and second derivatives of the cardiac segment, respectively; This is the time prior penalty term relative to the R-wave.

[0108] The time prior penalty Calculated using a piecewise function: Feature points to be relocated Preset center With scale ,in , , , ;when Time to take ,when Time to take This results in a secondary penalty being imposed on candidate points that exceed the reasonable physiological time delay range.

[0109] The first and second derivatives of the cardiac segment can be calculated as follows: for the cardiac segment First, a smoothed sequence is obtained by using a 5-point moving average. ,in, This indicates the first segment of the resampled heartbeat fragment. Each sampling time, , This represents the sampling period after resampling, i.e., the time interval between two adjacent sampling moments. Central difference is used to calculate the internal sampling points. and For boundary points and The derivative is calculated using both forward and backward differencing, and then expressed in the objective function as... and correspond and .

[0110] In this step, the candidate interval is used to solidify physiological time-series constraints and prevent points from crossing unreasonable regions; in the objective function... Used for locations with significant changes in preferred morphology. Used to suppress local extremum misselection caused by spike noise. This is used to limit relocation to a region with a reasonable time delay relative to the R-wave, thereby providing interpretable and more physiologically consistent location results even in high uncertainty scenarios.

[0111] Figure 4 During the constraint relocation process shown on the right side of the diagram, when any feature point satisfies... When the threshold is exceeded, the system selects the candidate region for the corresponding feature point. The internal search for the objective function is performed by combining the first derivative term, the second derivative term, and the time prior penalty term relative to the R-wave. Achieve interpretable point relocation and obtain the relocated feature point temporal position. It outputs the data, thereby suppressing point drift and improving the physiological rationality and robustness of feature point localization in uncertain scenarios.

[0112] IX. Low-reliability closed-loop control:

[0113] Establish a low-reliability counter In the heartbeat segment And there exists any feature point At that time, the heartbeat fragment was marked as low confidence, and then... , accumulated to The system triggers an electrode contact check prompt and marks subsequent quantitative heartbeat segments as low confidence; for example, in this embodiment, the quantitative value is 10.

[0114] The condition is met when 10 consecutive quantitative heartbeat segments are present. And all feature points of the cardiac segment satisfy At that time, the low-reliability counter will be... Zero out; output the time positions and reliable markers of points B, C, X, and O after constraint relocation and closed-loop quality control.

[0115] The continuous low confidence triggering rule described in this step distinguishes between short-term uncertainty and continuous degradation, avoiding frequent prompts caused by occasional noise; when the continuous low confidence reaches the threshold, it triggers an electrode check prompt and marks subsequent outputs as low confidence, thereby reducing the impact of error trends on upper-level decisions at the system level and improving the security and availability of continuous monitoring.

[0116] Figure 5 This is a schematic diagram of the ICG signal closed-loop quality control and hemodynamic parameter output of the present invention. The left side of the diagram shows the low-confidence closed-loop control logic: the system establishes a low-confidence counter. When a heartbeat segment meets the quality scoring threshold but any feature point's consistency / reliability index is lower than the preset requirement (corresponding point-level uncertainty exceeds the threshold), the heartbeat segment is marked as low reliability and set to... Accumulation; when When the preset trigger threshold is reached, an output electrode contact check prompt is issued, and several subsequent heartbeat segments are marked as low confidence; when several consecutive heartbeat segments meet the quality score and consistency requirements, the counter is activated. By resetting the system to zero, a closed-loop identification, alerting, and output degradation control for continuously deteriorating conditions can be achieved.

[0117] Based on the above steps, this embodiment can also output hemodynamic parameters, as explained below:

[0118] Based on the hemodynamic parameters output from the feature points after quality control, the parameter calculation satisfies: pre-contraction ejection phase Left ventricular ejection time ,in For R-wave time position, and The time positions of points B and X are determined after constraint repositioning and closed-loop quality control; and the validity of the parameter output is judged: when or The corresponding heartbeat segment and Marked as low confidence; the trend value is obtained by applying median filtering with a window length of 10 heartbeats to the continuous monitoring output. The trend value and the confidence marker for each bet are output together.

[0119] Individualization coefficients are determined through a single reference calibration. Then calculate stroke volume SV The reference calibration continuously collects 60 data points that meet the following conditions while the user is at rest: And the uncertainty of all feature points satisfies The effective heartbeat segments are calculated, and each effective heartbeat is analyzed. and And denoted by their medians as and ,in ; obtain reference stroke volume at calibration time And calculate:

[0120] ;

[0121] in Take the median of the base impedance corresponding to the 60 valid heartbeats. ,and Based on user height pass Calculation; then, for each condition met during continuous monitoring Heartbeat calculate SV The trend is not updated when the heartbeat has a low-confidence marker. SV When the heartbeat does not have a low-confidence marker, SV The trend is obtained by median filtering with an input window length of 10 heartbeats. and output , And trusted tags.

[0122] Figure 5 The right side shows the stroke output and trend update strategy for hemodynamic parameters: the system calculates the pre-systolic ejection phase based on the time location of characteristic points after quality control. With left ventricular ejection time The validity of each beat parameter was determined; the trend value was obtained by median filtering with a fixed window length on the continuous monitoring results. and The system outputs the trend value along with the bet confidence marker. Furthermore, the system obtains individualized coefficients after completing one reference calibration. Then, according to the formula for calculating stroke volume, the following is obtained: It outputs the trend update using the same method as described above. Among them, parameters with low confidence levels can be labeled but not used for trend updates to reduce the contamination of trend curves by abnormal segments, thereby improving the stability and engineering usability of continuous monitoring parameter output.

[0123] This embodiment organically combines cardiac segment quality scoring screening, contrastive learning embedding representation, online creation and updating of waveform subtypes through streaming prototype clustering, subtype expert feature point detection, point-level uncertainty quantification for inactivated repeated inference, objective function constraint relocation based on candidate intervals and time prior penalty terms, and closed-loop hints and output degradation marking triggered by continuous low-confidence counts. This significantly reduces false positives, false negatives, and point drift of feature points under multi-subtype ICG waveforms and noise artifacts, achieving quantifiable and reliable output and interpretable error correction of feature point results, and further improving… PEP , LVET and the calibration coefficient Corrected SV Stability, cross-individual consistency and engineering usability of hemodynamic parameters in continuous monitoring.

[0124] Example 2:

[0125] This embodiment proposes an impedance echocardiography feature point detection and quality control system based on waveform subtypes, such as... Figure 6 As shown, it includes:

[0126] Synchronous acquisition module: Synchronously acquires the dZ / dt signal, base impedance Z0 signal, and ECG signal from impedance cardiography (ICG);

[0127] Preprocessing module: preprocesses the dZ / dt signal. The preprocessing process is as follows: baseline correction, notch filtering, bandpass filtering, and amplitude normalization.

[0128] R-wave detection and heartbeat segmentation module: detects the R-wave position of the ECG signal, extracts heartbeat segments from the preprocessed ICG based on the R-wave position, and resamples each heartbeat segment and aligns it with the R-wave.

[0129] Quality scoring module: Extracts quality feature vectors for each heartbeat segment, and calculates the quality score based on the quality feature vectors according to the quality scoring formula;

[0130] Contrastive learning encoder module: Construct two augmented views for heartbeat segments with qualified quality scores, encode and normalize them to obtain the embedding vector h, and use contrastive learning loss for training to bring the augmented views of the same source closer together and separate the heartbeats of different sources.

[0131] Prototype Clustering / Subtype Management Module: Performs streaming prototype clustering on the embedded vector h using cosine distance. If the cosine distance exceeds the threshold for creating a new subtype, a new ICG waveform subtype is created and the prototype is initialized. Otherwise, it is assigned to the nearest ICG waveform subtype and the prototype is updated.

[0132] Uncertainty calculation module: Based on the ICG waveform subtype, the detection network is called to identify feature points, and the point-level uncertainty of each feature point is calculated through multiple inactivation inferences;

[0133] Constraint relocation module: If the point-level uncertainty of a feature point exceeds a set threshold, then within the candidate interval of that feature point, constraint relocation is performed by combining the first and second derivatives of the cardiac segment with the time prior penalty of the relative R wave.

[0134] Low-confidence closed-loop control module: Establish a low-confidence counter, statistically analyze low-confidence heartbeat segments based on quality score and point-level uncertainty, and trigger electrode contact check if the limit is exceeded; if a set number of qualified heartbeat segments appear consecutively, output the time position and confidence mark of each feature point.

[0135] In the synchronous acquisition module, the dZ / dt signal, Z0 signal and ECG signal are sampled synchronously using the same clock source; the ECG signal provides a time reference for the localization of cardiac events, and the Z0 signal is used for scale correction in subsequent calculations of hemodynamic parameters.

[0136] In the preprocessing module, the discrete sequence of the dZ / dt signal is processed by sliding median filtering to obtain the baseline and then the baseline is corrected. The corrected sequence is then filtered by a second-order IIR notch filter. The notched sequence is then filtered by a bandpass filter. Finally, the amplitude of the bandpass filtered sequence is normalized.

[0137] In the R-wave detection and heartbeat segmentation module, a heartbeat segment is obtained by truncating a time window based on the R-wave position. The time window covers the electromechanical coupling preparation segment before the R-wave and the ejection-related segment after the R-wave, so that feature points including points B, C, X, and O all fall within the heartbeat segment.

[0138] In the quality scoring module, the quality feature vector includes: peak-to-peak amplitude of the cardiac segment, number of effective peak-to-valley pairs within the cardiac segment, slope of the main rising edge, spectral entropy, energy or energy percentage within the target physiological frequency band, and time interval deviation between adjacent cardiac segments.

[0139] In the contrastive learning encoder module, the enhancement methods in the enhanced view construction include amplitude scaling, temporal micro-warping, additive Gaussian noise, and random segment occlusion.

[0140] The detection network in the uncertainty calculation module is a subtype expert feature point detection network corresponding to the ICG waveform subtype. Different subtype expert feature point detection networks learn different feature point localization strategies for the morphological differences of the corresponding ICG waveform subtype.

[0141] In the constrained relocation module, the constrained relocation is achieved by minimizing the first derivative, second derivative, and time prior penalty term of the relative R wave within the candidate interval of the feature point using an objective function.

[0142] In a low-confidence closed-loop control module, a heartbeat segment with a qualified quality score and a point-level uncertainty greater than a set threshold at any feature point is considered a low-confidence heartbeat segment; a heartbeat segment with a qualified quality score and a point-level uncertainty at no feature point greater than a set threshold is considered a qualified heartbeat segment.

[0143] The impedance cardiography feature point detection and quality control system based on waveform subtype proposed in this embodiment can achieve the impedance cardiography feature point detection and quality control method based on waveform subtype described in Embodiment 1, and has the same technical effect.

[0144] Example 3:

[0145] This embodiment is a verification example of the method described in this invention.

[0146] To verify the feasibility and effectiveness of the impedance cardiograph feature point detection and quality control method based on waveform subtype described in this invention, a representative verification was performed using publicly available ICG data.

[0147] First, the class separability of heartbeat segments from the publicly available Heartcycle Database (ICG derivatives) was verified. When the heartbeat segments were divided into 2, 3, 4, and 5 classes, the corresponding subtype recognition accuracies reached 91.62%, 88.48%, 87.43%, and 83.77%, respectively. In the baseline state of the ReBeatICG Database, when the heartbeat segments were divided into 2, 3, 4, and 5 classes, the corresponding subtype recognition accuracies reached 94.72%, 94.38%, 93.23%, and 93.75%, respectively. These results indicate that there are distinguishable waveform subtype structures among different ICG heartbeat segments. Using a temporal network to identify waveform subtypes is feasible and provides a basis for the present invention to call an expert feature point detection network based on subtype.

[0148] To verify the effectiveness of the feature point detection chain, a comparative experiment on the publicly available ReBeatICGDatabase dataset with expert-annotated points was conducted. Since the annotations for points B, C, and X in the publicly available validation data are highly complete, points B, C, and X are used as representative feature points in the following explanation. Point O can be located and validated using the same method. In the experiment, the expert annotation results were used as the ground truth, and the algorithm's recognition result falling within ±40ms of the expert-annotated point was considered a correct recognition. Figure 7 As shown in Figures (a), (b), and (c), "+" indicates the algorithm annotation of this invention, and circles indicate expert annotations. The annotation results show that, under baseline conditions, the recognition accuracy of points B, C, and X reaches 97.58%, 99.11%, and 92.12%, respectively; under cognitive load conditions, the recognition accuracy of points B, C, and X reaches 95.08%, 99.07%, and 90.29%, respectively. Among them, the recognition accuracy of point C reaches about 99% in both conditions, and point X maintains a recognition accuracy of over 90% even under conditions of significant noise interference and morphological changes, indicating that the feature point detection and trusted control link described in this invention has high robustness in complex scenarios.

[0149] Regarding signal quality improvement and handling of complex interference conditions, supplementary verification was performed on ICG signals from publicly available datasets under noisy scenarios. The ICG signals before and after filtering using the method of this invention are shown below. Figure 8 Figures (a) and (b) show the results. The results indicate that after processing 48 ICG signals, the average signal-to-noise ratio (SNR) increased by 5.82 dB, the mean square error of the denoised signal decreased to 0.011, and the normalized correlation coefficient increased to 0.978. Simultaneously, the average energy of the high-frequency band decreased by more than 40%, indicating that the processing significantly suppressed high-frequency artifacts and noise interference while preserving the main morphological characteristics of the heartbeat. When the signal-to-noise ratio was 10 dB, the recognition accuracies for points B, C, and X were 95.62%, 98.23%, and 90.12%, respectively; when the signal-to-noise ratio was 5 dB, the corresponding accuracies were 94.27%, 97.54%, and 87.53%, respectively; and when motion artifacts were superimposed, the corresponding accuracies were 95.12%, 98.52%, and 89.87%, respectively. These results demonstrate that the method described in this invention can still maintain high stability in feature point recognition under low SNR and motion artifact scenarios, providing support for low-confidence judgment and output control under continuous monitoring conditions.

[0150] Then, the results of feature point identification and parameter evaluation were compared and analyzed. The results show that the hemodynamic parameters automatically identified by the algorithm of this invention have a high degree of consistency with the expert annotation results. Taking the median of left ventricular ejection time (LVET) as an example, ... Figure 9As shown, the median of the algorithm of this invention for LVET is 0.247, the median of expert annotation is 0.246, and the difference is only 0.001.

[0151] Other parameters, such as the maximum rate of change of impedance (dz / dt)max, had an algorithmic median of 2.081 and an expert median of 2.079, with a difference of 0.002. The algorithmic median for the combined LVET and (dz / dt)max index was 0.358, and the expert median was 0.354, with a difference of 0.004. Under cognitive load, the differences between LVET, (dz / dt)max, and their combined indexes and expert annotations were 0.012, 0.04, and 0.02, respectively. This demonstrates that the feature point results obtained based on the method of this invention can be used for subsequent hemodynamic parameter calculations and can maintain good parameter output stability and consistency.

[0152] In summary, this invention reduces the probability of low-quality segments entering subsequent detection processes by screening cardiac segment quality scores; it achieves online establishment and updating of ICG waveform subtypes through contrastive learning embedding representations and streaming prototype clustering; it quantifies point-level uncertainty by calling expert feature point detection networks according to subtypes and combining this with deactivated repeated inference; and it reduces the risks of false detections, missed detections, and point drift by performing constrained relocation and introducing continuous low-confidence closed-loop control in high-uncertainty scenarios, thereby improving the stability and engineering usability of hemodynamic parameter outputs. The above implementation details and verification results further demonstrate the feasibility and effectiveness of the technical solution of this invention.

[0153] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for impedance cardiography feature point detection and quality control based on waveform subtype, characterized in that, include: S1. Simultaneously acquire the dZ / dt signal, the base impedance Z0 signal, and the ECG signal from impedance cardiograph (ICG); S2. Preprocess the dZ / dt signal. The preprocessing steps are as follows: baseline correction, notch filtering, bandpass filtering, and amplitude normalization. S3. Detect the R-wave position of the ECG signal, extract the heartbeat segment of the preprocessed ICG based on the R-wave position, and resample each heartbeat segment and align it with the R-wave to ensure that all heartbeat segments are consistent in time dimension and feature position. S4. Extract the quality feature vector for each heartbeat segment, and calculate the quality score based on the quality feature vector according to the quality scoring formula. S5. Construct two augmented views for heartbeat segments that pass the quality score. After encoding and normalization by the encoder, obtain the embedding vector h. Use contrastive learning loss for training to bring the augmented views of the same source closer together and separate the heartbeats of different sources. S6. Perform streaming prototype clustering on the embedded vector h using cosine distance. If the cosine distance exceeds the threshold for creating a new subtype, create a new ICG waveform subtype and initialize the prototype. Otherwise, it is assigned to the nearest ICG waveform subtype and the prototype is updated. S7. Based on the ICG waveform subtype, call the detection network to identify feature points, and calculate the point-level uncertainty of each feature point through multiple inactivation inferences; S8. If the point-level uncertainty of a feature point exceeds the set threshold, then within the candidate interval of that feature point, constrained relocation is performed by combining the first and second derivatives of the heartbeat segment with the time prior penalty of the relative R wave. S9. Establish a low confidence counter, statistically analyze low confidence beat segments based on quality score and point-level uncertainty, and trigger electrode contact check and output low confidence flag if the limit is exceeded. If a set number of qualified heartbeat segments appear consecutively, the time position and credibility marker of each feature point will be output.

2. The method for impedance cardiograph feature point detection and quality control based on waveform subtype according to claim 1, characterized in that, In step S1, the dZ / dt signal, Z0 signal and ECG signal are sampled synchronously using the same clock source; the ECG signal provides a time reference for the localization of cardiac events, and the Z0 signal is used for scale correction in subsequent calculations of hemodynamic parameters.

3. The method for impedance cardiograph feature point detection and quality control based on waveform subtype according to claim 1, characterized in that, In step S2, the discrete sequence of the dZ / dt signal is processed by sliding median filtering to obtain the baseline and then the baseline is corrected. The corrected sequence is then filtered by a second-order IIR notch filter. The notched sequence is then filtered by a bandpass filter. Finally, the amplitude of the bandpass filtered sequence is normalized.

4. The method for impedance cardiograph feature point detection and quality control based on waveform subtype according to claim 1, characterized in that, In step S3, a time window is extracted based on the R-wave position to obtain a cardiac segment. The time window covers the electromechanical coupling preparation segment before the R-wave and the ejection-related segment after the R-wave, so that feature points including points B, C, X, and O all fall within the cardiac segment.

5. The method for impedance cardiograph feature point detection and quality control based on waveform subtype according to claim 1, characterized in that, The quality feature vector in step S4 includes: peak-to-peak amplitude of the cardiac segment, number of effective peak-to-valley pairs within the cardiac segment, slope of the main rising edge, spectral entropy, energy or energy percentage within the target physiological frequency band, and time interval deviation between adjacent cardiac segments.

6. The method for impedance cardiograph feature point detection and quality control based on waveform subtype according to claim 1, characterized in that, In step S5, the enhancement methods in the enhanced view construction include amplitude scaling, temporal micro-warping, additive Gaussian noise, and random segment occlusion.

7. The method for impedance cardiograph feature point detection and quality control based on waveform subtype according to claim 1, characterized in that, The detection network mentioned in step S7 is a subtype expert feature point detection network corresponding to the ICG waveform subtype. Different subtype expert feature point detection networks learn different feature point localization strategies for the morphological differences of the corresponding ICG waveform subtype.

8. The method for impedance cardiograph feature point detection and quality control based on waveform subtype according to claim 1, characterized in that, The constraint relocation described in step S8 is achieved by minimizing the objective function containing the first derivative, second derivative, and time prior penalty term of the relative R wave within the candidate interval of the feature point.

9. The method for impedance cardiograph feature point detection and quality control based on waveform subtype according to claim 1, characterized in that, In step S9, a heartbeat segment that has passed the quality score and whose point-level uncertainty at any feature point is greater than the set threshold is a low-confidence heartbeat segment. A heartbeat segment that has passed the quality score and whose point-level uncertainty of all feature points is not greater than the set threshold is a qualified heartbeat segment.

10. A quality control system for impedance cardiograph feature point detection based on waveform subtype, characterized in that, include: Synchronous acquisition module: Synchronously acquires the dZ / dt signal, base impedance Z0 signal, and ECG signal from impedance cardiography (ICG); Preprocessing module: preprocesses the dZ / dt signal. The preprocessing process is as follows: baseline correction, notch filtering, bandpass filtering, and amplitude normalization. R-wave detection and heartbeat segmentation module: detects the R-wave position of the ECG signal, extracts heartbeat segments from the preprocessed ICG based on the R-wave position, and resamples each heartbeat segment and aligns it with the R-wave to ensure that all heartbeat segments are consistent in terms of time dimension and feature position; Quality scoring module: Extracts quality feature vectors for each heartbeat segment, and calculates the quality score based on the quality feature vectors according to the quality scoring formula; Contrastive learning encoder module: Construct two augmented views for heartbeat segments with qualified quality scores, encode and normalize them to obtain embedding vector h, and use contrastive learning loss for training to bring the augmented views of the same source closer together and separate the heartbeats of different sources. Prototype Clustering / Subtype Management Module: Performs streaming prototype clustering on the embedding vector h using cosine distance. If the cosine distance exceeds the threshold for creating a new subtype, a new ICG waveform subtype is created and the prototype is initialized. Otherwise, it is assigned to the nearest ICG waveform subtype and the prototype is updated. Uncertainty calculation module: Based on the ICG waveform subtype, the detection network is called to identify feature points, and the point-level uncertainty of each feature point is calculated through multiple inactivation inferences; Constraint relocation module: If the point-level uncertainty of a feature point exceeds a set threshold, then within the candidate interval of that feature point, constraint relocation is performed by combining the first and second derivatives of the cardiac segment with the time prior penalty of the relative R wave. Low-confidence closed-loop control module: Establish a low-confidence counter, statistically analyze low-confidence beat segments based on quality score and point-level uncertainty, and trigger electrode contact check if the limit is exceeded; if a set number of qualified beat segments appear consecutively, output the time position and confidence mark of each feature point.

Images