An organ biological age evaluation device based on electrocardiogram signals

Abstract

The application discloses an organ biological age evaluation device based on an electrocardiogram signal, and the device comprises an electrocardiogram signal preprocessing module, an electrocardiogram feature extraction module, an organ state representation generation module, an organ biological age conversion module and a result output module. The electrocardiogram signal preprocessing module is used for carrying out lead standardization, filtering, resampling, fixed-length clipping or padding and quality control on original electrocardiogram signals; the electrocardiogram feature extraction module and the organ state representation generation module are used for extracting electrocardiogram representations and outputting age representation parameters related to multiple organs; the organ biological age conversion module is used for mapping the age representation parameters into biological age values, age difference values and age classification results of corresponding organs; and the result output module is used for generating an evaluation report. The device can realize biological age evaluation of organs such as a heart, a liver and kidneys, and is suitable for health management and physiological state evaluation.

Description

Technical Field

[0001] This invention relates to the fields of medical artificial intelligence and biomedical signal processing technology, and in particular to an organ biological age assessment device based on electrocardiogram signals. Background Technology

[0002] Electrocardiography (ECG), as a non-invasive, low-cost, easy-to-operate, and easily repeatable physiological signal, has wide application value in clinical examinations, primary healthcare, and wearable monitoring scenarios. Compared with imaging examinations and some biochemical tests, ECG acquisition is convenient, highly standardized, and has good universality and scalability, thus becoming one of the important data sources for mining information about human physiological states.

[0003] Traditional electrocardiogram (ECG) analysis primarily relies on doctors' manual interpretation of waveforms or processing based on manually designed algorithms. While these methods are effective for routine rhythm recognition and identification of typical abnormal patterns, their ability to represent complex, subtle, and long-term evolving electrophysiological changes remains limited, and they struggle to meet the demands of large-scale automated analysis and continuous evaluation. With the development of deep learning, researchers have begun to directly utilize raw ECG waveforms to construct data-driven models, thereby improving the automation level and feature representation capabilities of ECG signal analysis.

[0004] Existing research indicates that models such as convolutional neural networks, recurrent neural networks, and Transformers have been widely applied in the field of electrocardiogram analysis. Hannun et al. proposed a deep neural network-based arrhythmia detection model, which enables automatic identification of multiple rhythm types in dynamic electrocardiograms (Hannun AY, Rajpurkar P, Haghpanahi M, et al. Cardiologist-level arrhythmia detection and classification in ambulatoryelectrocardiograms using a deep neural network. Nat Med. 2019;25(1):65-69.). El-Ghaish et al. further introduced a bidirectional Transformer structure on the basis of multi-scale convolutional feature extraction to enhance the ability to model temporally dependent information (M. Hammad, AM Iliyasu, A. Subasi, ESL Ho and A. AA El-Latif, "A Multitier Deep Learning Model for ArrhythmiaDetection," in IEEE Transactions on Instrumentation and Measurement, vol. 70,pp. 1-9, 2021, Art no. 2502809.). The above research demonstrates that deep learning models can extract relatively rich morphological and temporal representations from electrocardiograms (ECGs), providing a technical foundation for further mining the high-level physiological information contained in ECGs.

[0005] Beyond routine classification tasks, recent research has revealed that electrocardiograms (ECGs) not only contain rhythm and conduction information but also implicitly reveal age-related physiological characteristics. Lima et al. trained a deep neural network based on large-scale raw ECG data to estimate the age of subjects, proposing the concept of ECG-age and pointing out that the difference between ECG-predicted age and actual age can reflect an individual's physiological aging characteristics (Lima, EM, Ribeiro, AH, Paixão, GMM). et al. Deep neural network-estimated electrocardiographic age as a mortalitypredictor. Nat Commun12, 5117 (2021). https: / / doi.org / 10.1038 / s41467-021-25351-7; Chang et al. further proposed a deep learning-based heart age estimation method, showing that age information inferred from electrocardiograms can serve as a useful representation for assessing cardiac physiological state (Chang CH, Lin CS, Luo YS, LeeY-T and Lin C (2022) Electrocardiogram-Based Heart Age Estimation by a DeepLearning Model Provides More Information on the Incidence of CardiovascularDisorders. Front. Cardiovasc. Med. 9:754909. doi: 10.3389 / fcvm.2022.754909). Related reviews also point out that artificial intelligence can extract biological age-related information from electrocardiograms, and the deviation between AI-ECG age and actual age can serve as a substitute indicator of biological aging.

[0006] Furthermore, some multi-task learning studies have shown that a shared backbone network combined with a multi-output head structure can learn multiple relevant representations based on a single ECG input, improving the model's ability to characterize complex physiological features. The multi-task deep learning model proposed by Lin et al. utilizes a resting 12-lead ECG to simultaneously model multiple cardiovascular-related endpoints, demonstrating the existence of multi-dimensional representational information in the ECG that can be extracted by a unified backbone network and decomposed by a multi-task output layer. This type of structure provides a valuable network foundation for the technical route of "shared feature extraction—multi-organ state representation output—organ age conversion" in this invention.

[0007] However, most existing AI research based on electrocardiograms (ECGs) focuses primarily on disease classification, anomaly identification, or specific event prediction, with less emphasis on the systematic utilization of ECG signals from the perspectives of organ state representation and biological age assessment. In particular, current technologies remain relatively inadequate in the area of ​​using single or short-duration ECG signals as input to output unified age-related representation parameters for multiple organs, and further converting these parameters into organ biological age assessment values. How to establish a device that extracts multi-organ state representation parameters from ECG signals and converts them into organ biological age assessment values ​​while retaining the advantages of convenient ECG acquisition remains a technical problem to be solved in this field. Summary of the Invention

[0008] The purpose of this invention is to address the shortcomings of existing technologies by providing an organ biological age assessment device based on electrocardiogram (ECG) signals. The device uses ECG signals as the main input and takes the risk of chronic diseases in multiple different organ systems as the prediction target, thereby expanding from "only looking at the heart" to "reflecting the overall aging state of organs". Based on this, the device calculates the health risk scores of each major organ in the human body.

[0009] The objective of this invention is achieved through the following technical solution: an organ biological age assessment device based on electrocardiogram signals, comprising: The ECG signal preprocessing module is used to perform lead standardization, filtering and noise reduction, resampling, fixed-length truncation or padding, lead-by-lead standardization, and outlier detection on the acquired ECG signals to generate standardized ECG data for subsequent analysis. The electrocardiogram feature extraction module is used to extract feature information characterizing the characteristics of cardiac electrical activity from the preprocessed electrocardiogram data. The feature information includes heart rate variability features, PQRST band morphological features, and electrocardiogram amplitude features. An organ state characterization generation module is used to output characterization parameters related to different organ states based on the feature information and using a machine learning model. The characterization parameters include one or more of organ aging-related characterization values, organ age index, or organ age deviation values. The organ biological age conversion module is used to convert the characterization results of different organs into organ biological age assessment values ​​based on the organ state characterization parameters and in combination with a preset age scale model, reference population distribution parameters or calibration parameters. The results output module is used to output the organ biological age assessment value, organ age deviation value, organ age level or organ status assessment report for each organ.

[0010] Furthermore, the electrocardiogram signal preprocessing module is used to convert the original electrocardiogram waveform record into a standardized array format for model training or inference. The electrocardiogram signal preprocessing module includes: The record reading unit is used to locate the base name path of the corresponding WFDB record in the dataset directory based on the patient ID and ECG record ID of the sample to be processed, and read the WFDB record to obtain the original waveform matrix sig, the original sampling rate fs, and the lead name list names; The lead mapping unit is used to map the original waveform matrix to a preset 12-lead set according to the lead name. The 12-lead set includes at least I, II, III, aVR, aVL, aVF, V1, V2, V3, V4, V5 and V6. Missing leads are filled in with a zero-value sequence to obtain a standardized 12-lead waveform matrix x. A bandpass filter unit is used to perform bandpass filtering on the standardized 12-lead waveform matrix. The bandpass frequency band is set to 0.5Hz to 45Hz, the filter order is set to 4, and the forward-backward zero-phase filtering method is used to filter the time axis to suppress low-frequency baseline drift and high-frequency interference. The notch filter unit is used to perform notch filtering on the waveform matrix after bandpass filtering to suppress power frequency noise. When the notch frequency is higher than the Nyquist frequency threshold, the corresponding notch processing is not performed. The resampling unit is used to resample the waveform after notch filtering from the original sampling rate fs to the target sampling rate TARGET_FS; The length unification unit is used to unify the resampled waveform to a fixed number of sampling points TARGET_FS×DURATION_SEC corresponding to a fixed time length DURATION_SEC. When the resampled length is greater than the target length, it is truncated at the center of the waveform. When the resampled length is less than the target length, it is symmetrically zero-padded at both ends of the waveform to obtain a waveform matrix with a fixed shape. The standardization unit is used to perform lead-by-lead standardization on a fixed-length waveform matrix. It calculates the mean and standard deviation for each lead, converts the waveform of that lead into a zero-mean, unit-variance form, introduces a stable term into the standard deviation, and truncates the amplitude of the standardized waveform. The quality control unit is used to detect the validity of the standardized waveform matrix. When there are abnormal values, it uses a preset replacement rule to replace the abnormal values ​​with zero values, and sets the corresponding leads to a zero sequence when the standard deviation of the leads is lower than a preset threshold. The data output unit is used to save the fixed-specification waveform matrix that has passed quality control as an array file in floating-point format, and record metadata such as sample identifier, path, sampling rate, number of leads and number of sampling points for subsequent training, verification or inference.

[0011] Furthermore, the organ state representation generation module takes the preprocessed 12-lead ECG segment as input, extracts representation vectors through a shared ECG feature backbone network, and outputs organ state representation parameters corresponding to different organs through multiple organ representation output heads. The organ state representation parameters include one or more of the following: organ aging-related representation values, organ age index, or organ age deviation values. Furthermore, the organ state characterization generation module includes: The input building unit is used to represent the preprocessed ECG signal as a three-dimensional tensor X∈R^(B×12×T), where B is the batch size, 12 is the number of leads, and T is the number of fixed sampling points; The multi-scale convolutional feature extraction unit is used to perform multiple sets of one-dimensional convolution operations on the input tensor X in parallel to obtain morphological features at different time scales, obtain multiple intermediate features, and obtain fused features X_ms through scale dimension fusion. The temporal downsampling unit is used to sequentially perform batch normalization, nonlinear activation, max pooling and random deactivation on the fused feature X_ms to improve training stability and achieve temporal downsampling to obtain feature X_0; The convolutional transformation unit is used to transform the feature X_0 sequentially through at least two convolutional blocks. Each convolutional block contains one-dimensional convolution, batch normalization, non-linear activation, and max pooling operations to progressively improve the expressive power of the channel dimension and compress the length of the time dimension to obtain the feature X_c. The channel recalibration unit is used to input the feature X_c into the channel recalibration module. The channel recalibration module is composed of several residual blocks with SE structure connected in series, which is used to adaptively weight the importance of different channels to obtain the enhanced feature X_crm. The bidirectional temporal coding unit is used to transform the enhanced feature X_crm into a sequence form and input it into the Transformer encoder to learn long-range dependencies. This includes performing forward encoding on the temporal sequence to obtain Z_fwd, performing reverse encoding on the sequence after time dimension inversion to obtain Z_bwd, and fusing the forward encoding result and the reverse encoding result to obtain the sequence representation Z. The feature aggregation unit is used to perform adaptive average pooling on the sequence representation Z, and aggregate it in the time dimension to obtain a fixed-dimensional ECG feature vector f. The multi-organ representation output unit is used to construct multiple organ representation output heads based on the fixed-dimensional electrocardiogram feature vector f, and output the corresponding organ state representation parameters for different organs. Each organ representation output head maps the feature vector of the shared backbone output to the representation output value corresponding to the organ through a linear mapping layer, and concatenates the representation output values ​​corresponding to multiple organs in the task dimension to obtain the model output.

[0012] Furthermore, the organ biological age conversion module includes: (1) Single organ characterization quantification unit, used to generate single organ quantification results for the characterization output value corresponding to each organ, wherein the single organ quantification results include one or more of the following: organ aging-related characterization value, organ age index, organ age deviation value, age percentile or age score. (2) Scale mapping unit, used to determine at least one mapping threshold, quantile threshold or scale conversion parameter based on the distribution of the corresponding organ representation output value in the training sample, calibration sample or reference population, and to map the organ representation output value of the target object to the age percentile, age score, organ age index or organ age deviation value of the corresponding organ. (3) Age conversion unit, used to convert the characterization results of each organ into organ biological age assessment values ​​based on the single organ quantification results and in combination with the preset age scale model, reference population distribution parameters or calibration parameters. (4) Fusion assessment unit, used to fuse organ biological age assessment values, organ age deviation values ​​or organ age indices corresponding to multiple organs to generate overall organ status assessment results; (5) Report generation unit, used to generate and output organ status assessment reports based on the organ biological age assessment value, organ age deviation value, organ age index and overall organ status assessment results of each organ.

[0013] To achieve the above objectives, the present invention also provides an electronic device, including a memory and a processor, wherein the memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to realize the functions of each module in the above-mentioned organ biological age assessment device based on electrocardiogram signals.

[0014] To achieve the above objectives, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the functions of each module in the above-described organ biological age assessment device based on electrocardiogram signals.

[0015] The technical solution of this invention can be summarized as follows: 1. A device for preprocessing and extracting features of electrocardiogram (ECG) signals is proposed. The ECG signal data is filtered, denoised, resampled, length unified, standardized and quality controlled. Feature information characterizing the ECG activity is extracted from the preprocessed ECG signal. 2. An organ state representation generation device with electrocardiogram signal features as the main input was constructed, and the organ aging-related representation values, organ age index or organ age deviation value corresponding to multiple organs were output through machine learning model. 3. An organ biological age conversion device is proposed. Through characterization quantization, scale mapping and age conversion modules, the organ state characterization parameters output by the model are converted into organ biological age assessment values ​​corresponding to each organ, and further organ state assessment reports are generated.

[0016] The beneficial effects of this invention are as follows: This invention proposes an organ biological age assessment device based on electrocardiogram (ECG) signals. By preprocessing ECG signals, extracting features, generating organ state representations, and converting organ biological ages, it can extract representation information related to different organ states from conventional ECG data and output organ biological age assessment values, organ age deviation values, organ age indices, or organ state assessment results for each organ.

[0017] Compared with existing technologies, this invention makes full use of the convenient, non-invasive, low-cost and easy-to-standardize characteristics of electrocardiogram (ECG) signal acquisition, and transforms the temporal, morphological and statistical features contained in the ECG into characterization parameters that can reflect the differences in organ states. This enables the expansion from single ECG signals to multi-organ state characterization analysis, and improves the utilization efficiency of ECG signals in the quantitative assessment of physiological states.

[0018] Furthermore, this invention employs a network structure that combines multi-scale convolutional feature extraction, channel recalibration, and bidirectional temporal coding, which can more fully extract local morphological features and long-range dependency information from electrocardiogram signals, thereby improving the stability and accuracy of organ state characterization parameter generation. Through age scale models, reference population distribution parameters, or calibration parameters, the model output is further converted into organ biological age assessment values ​​that are easy to understand and apply, improving the interpretability and readability of the assessment results.

[0019] This invention has a high degree of automation, real-time performance, and scalability, and can be applied to scenarios such as health status assessment, long-term physiological state change analysis, physical examination auxiliary analysis, and individualized physiological age assessment. It provides a new technical means for conducting multi-organ state characterization analysis and organ biological age assessment based on electrocardiogram signals. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of an organ biological age assessment device based on electrocardiogram (ECG) signals according to the present invention. The diagram shows that the device includes an ECG input section, a tabular data input section, a feature encoding section, a feature fusion section, and a multi-organ representation output section. The ECG input section is used to input 12-lead ECG data, the tabular data input section is used to input auxiliary tabular data, the feature encoding section encodes the ECG data and tabular data respectively to generate corresponding tokens, the feature fusion section is used to concatenate tokens from different sources and perform cross-attention fusion with task query vectors, and the multi-organ representation output section is used to output organ state representation parameters corresponding to multiple organs and generate organ biological age assessment results based on the organ state representation parameters. Figure 2 This is a schematic diagram of an electronic device as illustrated in an exemplary embodiment of the present invention. Detailed Implementation

[0021] The core technology of this invention is to jointly encode and fuse the electrocardiogram (ECG) signals and optional auxiliary tabular data of the target object to construct a multi-organ state representation generation device, and to perform age scale conversion based on the organ state representation parameters output by the device, thereby realizing the biological age assessment of multiple organs. The device includes an ECG signal preprocessing module, a tabular data preprocessing module, an ECG feature extraction module, a multimodal fusion module, a multi-organ representation output module, an organ biological age conversion module, and a result output module.

[0022] (1) Electrocardiogram signal preprocessing module The ECG signal preprocessing module is used to uniformly convert the original ECG waveform records into standardized input data that can be directly called upon for model training or inference. Specifically, the original ECG waveforms are read from WFDB format files and mapped to a standard 12-lead structure, which includes at least leads I, II, III, aVR, aVL, aVF, and V1 to V6; when individual leads are missing, zero-value sequences are used to fill in the missing leads to maintain the consistency of the input structure.

[0023] Furthermore, a standardized denoising process is performed on the signals of each lead. Preferably, a fourth-order Butterworth bandpass filter is first used to process the signal, with the bandpass frequency set to 0.5Hz to 45Hz to reduce baseline drift and high-frequency noise; then a 50Hz or 60Hz notch filter is used to suppress power frequency interference. The filtered signal is then uniformized to 250Hz through anti-aliasing resampling, and a fixed-length processing is performed on each record to uniformize it to a duration of 10 seconds, thereby obtaining an input tensor of size 12×2500.

[0024] To reduce dimensional differences between samples and improve computational stability, Z-score standardization was performed on each lead. After standardization, the waveform amplitude was truncated to within ±5 standard deviations. For waveforms containing NaN or Inf values, zero values ​​were used for replacement, and lead-by-lead standardization was re-performed. If outliers still existed after repair, or if the waveform still had abnormal amplitude values ​​after preprocessing, the sample was deemed invalid and discarded. Quality-controlled ECG data were saved in array format for subsequent model calls.

[0025] Specifically, the electrocardiogram signal preprocessing module includes the following units.

[0026] (1.1) A record reading unit, used to locate the base name path of the corresponding WFDB record in the dataset directory based on the subject_id (patient ID) and ecg_id or study_id (ECG record ID) of the sample to be processed, and read the WFDB record to obtain the original waveform matrix sig, the original sampling rate fs, and the lead name list namess. For example, when batch processing ECG data in a physical examination center, the 12-lead ECG record of the same subject can be located according to the subject_id in the physical examination system and the study_id generated by the ECG device.

[0027] (1.2) Lead mapping unit, used to map the multi-lead waveforms obtained by the recording and reading unit to a preset 12-lead set according to the lead name. The 12-lead set includes at least I, II, III, aVR, aVL, aVF, V1, V2, V3, V4, V5, and V6. Missing leads in the original record are filled in using a zero-value sequence to obtain a standardized 12-lead waveform matrix x. This process differs from simply inputting the model according to the original lead order in the file, and can avoid model misreading caused by differences in lead naming or lead arrangement between different devices.

[0028] (1.3) Bandpass filtering unit, used to perform bandpass filtering on the standardized 12-lead waveform matrix. The bandpass frequency band is set to 0.5Hz to 45Hz, the filter order is set to 4, and a forward-backward zero-phase filtering method is used to filter the time axis to suppress low-frequency baseline drift and high-frequency electromyography or electromagnetic interference. This zero-phase filtering method can reduce noise while reducing the phase shift of the P wave, QRS complex, and T wave positions.

[0029] (1.4) Notch filter unit, used to perform notch filtering on the waveform matrix after bandpass filtering to suppress power frequency noise; when the notch frequency is higher than the Nyquist frequency threshold, the corresponding notch processing is not performed. For example, in a 50Hz mains power environment, a 50Hz notch is preferentially set, and in a 60Hz mains power environment, a 60Hz notch is preferentially set; if the original sampling rate is too low and the notch frequency is unavailable, this step is skipped to avoid introducing unstable filtering results.

[0030] (1.5) A resampling unit is used to resample the notch-filtered waveform from the original sampling rate fs to the target sampling rate TARGET_FS. In one embodiment, TARGET_FS is set to 250Hz, and a multiphase resampling method is used to achieve resampling. The upsampling and downsampling ratio is approximately determined based on the rational number of TARGET_FS / fs, thereby achieving anti-aliasing filtering during the resampling process. This process enables ECG recordings from 250Hz, 500Hz, or 1000Hz devices to be converted to a uniform specification.

[0031] (1.6) A length unification unit is used to unify the resampled waveform to a fixed number of sampling points L = TARGET_FS × DURATION_SEC corresponding to a fixed time length DURATION_SEC. In one embodiment, DURATION_SEC is set to 10 seconds and L is 2500. When the resampled length is greater than the target length, it is truncated at the center of the waveform; when the resampled length is less than the target length, symmetrical zero padding is performed at both ends of the waveform to obtain a waveform matrix x with a fixed shape. Centering can preserve relatively stable ECG segments in the middle of the recording, and symmetrical padding can avoid overall offset of the effective waveform.

[0032] (1.7) The standardization unit is used to perform lead-by-lead standardization on a fixed-length waveform matrix; the mean and standard deviation are calculated for each lead, and the waveform of that lead is converted into a zero-mean, unit-variance form; to avoid division by zero, a stabilizing term epsilon is introduced into the standard deviation. Furthermore, the amplitude of the standardized waveform is truncated, for example, the amplitude is limited to the range of -5 to 5, in order to suppress the influence of abnormal spikes on model training or inference.

[0033] (1.8) The quality control unit is used to detect the validity of the standardized waveform matrix. When there are abnormal values ​​such as NaN or Inf, the abnormal values ​​are replaced with zero values ​​using a preset replacement rule, and normalization is optionally performed again for each lead. If the standard deviation of a certain lead is too small, the lead is set to a zero sequence. After the repair is completed, the abnormal values ​​are detected again. If there are still abnormal values, the sample is determined to be invalid and skipped. Furthermore, when the absolute amplitude of the waveform exceeds the preset abnormal threshold, it is determined to be an abnormal sample and skipped. For example, in the scenario of wearable device acquisition, if a certain chest lead presents an approximately constant sequence due to electrode detachment, the lead can be set to zero, and other leads can still participate in model inference.

[0034] (1.9) Data output unit, used to save the fixed-specification waveform matrix that has passed quality control as an array file in floating-point format, and record metadata such as sample identifier, path, sampling rate, number of leads and number of sampling points in the manifest file for subsequent training, validation or inference calls. For example, the manifest file may record subject_id, ecg_id, array file path, TARGET_FS, number of leads 12 and number of sampling points L so that the model training program can read it in batches.

[0035] After the above processing, the ECG input of a single target object can be represented as a 12×L matrix; the batch input can be represented as a three-dimensional tensor X_ecg, where X_ecg belongs to R^(B×12×L).

[0036] (2) Tabular data preprocessing module The tabular data preprocessing module is used to organize and standardize the auxiliary structured variables of the target object. These auxiliary structured variables may include one or more of demographic variables, basic vital signs, laboratory test indicators, or other structured physiological parameters. Standardization is performed on continuous variables, and missing or outlier values ​​are transformed or corrected according to preset rules to ensure the consistency and compatibility of subsequent model inputs. The preprocessed tabular variables are then associated with the corresponding electrocardiogram data according to the target object identifier, forming a unified analysis sample.

[0037] For example, in a health checkup scenario, auxiliary structured variables may include gender, actual age, height, weight, blood pressure, blood glucose, blood lipids, creatinine, liver function indicators, or inflammatory markers. For missing laboratory test indicators, the median of the training set, specific missing markers, or a missing mask that the model can recognize can be used for processing. The technical advantage of this optional module is that when structured data is available in the target application scenario, the model can simultaneously utilize dynamic physiological information from the electrocardiogram and static physiological information from the tabular data; when tabular data is unavailable, the device can still complete the evaluation based solely on the electrocardiogram signal.

[0038] (3) Electrocardiogram feature extraction module The electrocardiogram (ECG) feature extraction module is used to extract feature information characterizing cardiac electrical activity from the preprocessed 12-lead ECG signal. The feature information includes one or more of the following: heart rate variability features, PQRST band morphological features, ECG amplitude features, and high-dimensional ECG representations generated by a neural network. In one embodiment, the ECG feature extraction module employs a pre-trained ECG feature backbone network, which includes multi-scale convolutional layers, residual blocks with compression excitation structures, and a bidirectional Transformer encoder to simultaneously extract local waveform morphological features and long-range temporal dependencies. The improvement of this structure compared to a single convolutional network is that multi-scale convolution can simultaneously sense short-term QRS morphology and longer ST-T changes; the compression excitation structure can adaptively adjust the importance of features from different channels; and the bidirectional Transformer encoder can capture long-range relationships across the cardiac cycle.

[0039] In one specific embodiment, the electrocardiogram feature extraction and multi-organ state characterization generation process includes the following units.

[0040] (3.1) Input construction unit, used to represent the preprocessed electrocardiogram signal as a three-dimensional tensor X_ecg, X_ecg belongs to R^(B×12×L), where B is the batch size, 12 is the number of leads, and L is the number of fixed sampling points.

[0041] (3.2) Multi-scale convolution feature extraction unit, used to perform multiple sets of one-dimensional convolution operations on the input tensor X_ecg in parallel to obtain morphological features at different time scales, obtain multiple intermediate features, and obtain fused features X_ms through scale fusion. For example, different convolution kernel sizes can be set to focus on the rapid changes of QRS wave groups, the intermediate scale changes of P waves and T waves, and the trend changes of ST segments respectively.

[0042] (3.3) Temporal downsampling unit, used to perform batch normalization, nonlinear activation, max pooling and random deactivation on the fused feature X_ms in sequence to improve training stability and realize temporal downsampling to obtain feature X_0.

[0043] (3.4) Convolutional transformation unit, used to transform the feature X_0 sequentially through at least two convolutional blocks. Each convolutional block contains one-dimensional convolution, batch normalization, nonlinear activation and max pooling operations to progressively improve the expressive power of the channel dimension and further compress the length of the time dimension to obtain the feature X_c.

[0044] (3.5) Channel recalibration unit, used to input feature X_c into channel recalibration module. Channel recalibration module is composed of several residual blocks with SE structure connected in series, used to adaptively weight the importance of different channels, thereby obtaining enhanced feature X_crm. This step can enhance key waveform channels related to organ state and reduce the interference of noise channels or weakly correlated channels on the output.

[0045] (3.6) Bidirectional temporal coding unit, used to transform the enhanced feature X_crm into a sequence form and input it into the Transformer encoder to learn long-range dependencies, including: performing Transformer encoding on the temporal sequence to obtain the forward encoding result Z_fwd; performing Transformer encoding on the sequence after time dimension inversion to obtain the reverse encoding result Z_bwd; and fusing the forward encoding result and the reverse encoding result to obtain the final sequence representation Z_seq.

[0046] (3.7) Feature aggregation unit: Performs adaptive average pooling on the sequence representation Z_seq, and aggregates it in the time dimension to obtain a fixed-dimensional ECG feature vector f_ecg. To avoid symbol confusion, Z_seq represents the ECG sequence representation in this paper, and N represents its sequence length, that is, Z_seq belongs to R^(B×N×D).

[0047] (3.8) Multi-organ representation output unit, used to construct multiple organ representation output heads based on the fixed-dimensional ECG feature vector f_ecg, and output corresponding organ state representation parameters for different organs respectively. Each organ representation output head maps the feature vector of the shared backbone output to the representation output value corresponding to the organ through a linear mapping layer, and concatenates the representation output values ​​corresponding to multiple organs in the task dimension to obtain the model output.

[0048] For each ECG input sample, the ECG representation sequence output by the backbone network is denoted as: in, Indicates the length of the feature sequence. This indicates the embedding dimension. The... This serves as the ECG input representation for subsequent multimodal fusion modules.

[0049] (4) Multimodal fusion module The multimodal fusion module is used to fuse electrocardiogram representations and tabular variable representations. Specifically, the preprocessed tabular variables are input into a nonlinear embedding layer, which preferably includes LayerNorm, linear transformation, and the GELU activation function, thereby obtaining the tabular representation: in, This indicates the number of tokens after the table variable is embedded. This indicates the embedding dimension consistent with the electrocardiogram representation.

[0050] Furthermore, the electrocardiogram representation and the tabular representation are concatenated along the sequence dimension to form a unified multimodal context sequence: in, This represents a unified multimodal context sequence.

[0051] To achieve targeted aggregation of information from different organs, pre-set Learnable task query vectors: in, This represents the number of organ state representation tasks to be output. The task query vector and the multimodal context sequence interact through a multi-head cross-attention module to obtain an organ-specific fusion representation: in, This represents the fusion representation of multiple organs. This structure allows the query vector for each organ to selectively aggregate information related to that organ from both electrocardiogram and tabular representations.

[0052] For example, in a kidney condition assessment task, the corresponding task query vector can focus more on contextual information related to blood pressure, creatinine, heart rate variability, and specific electrocardiogram morphologies; in a lung or metabolic condition assessment task, the corresponding task query vector can focus more on information related to respiratory heart rate variability, right ventricular load-related electrocardiogram performance, body mass index, or blood glucose and lipid indicators.

[0053] (5) Multi-organ characterization output module The multi-organ state characterization generation module is used to generate organ state characterization parameters related to different organ states based on electrocardiogram feature information and, in conjunction with auxiliary structured data, using a machine learning model. These organ state characterization parameters include one or more of the following: organ aging-related characterization values, organ age index, predicted organ age deviation value, predicted age percentile, or predicted health risk score.

[0054] In one specific embodiment, the multi-organ state characterization generation module generates a characterization based on the fused characterization. Generate organ state representation parameters for multiple organs. For the ... Fusion characterization of each organ The corresponding representation output value of the organ is obtained by mapping through independent linear output heads: in, and They represent the first Weight parameters and bias parameters of the output head of each organ. Indicates the first The organ state representation parameters corresponding to each organ. The outputs of multiple organs can be concatenated as follows: Where S represents the set of organ state characterization parameters corresponding to K organs. The organ state characterization parameters... It is not used to output disease diagnosis conclusions, but to characterize the age-related state differences, aging-related characterization values, organ age index, or organ age deviation trend of the corresponding organ.

[0055] During the model training phase, the risk of chronic diseases, organ dysfunction, age-related phenotypes, or calibrated organ age labels from multiple organ systems can be used as evaluation targets. This allows the model to learn shared and organ-specific representations related to the states of multiple organs from electrocardiogram (ECG) signals. For example, output heads can be constructed for the heart, kidneys, liver, lungs, cerebrovascular system, or metabolic system, and supervised training can be performed using the corresponding organ system's chronic disease risk labels or physiological function indicators. This training method expands the invention from traditional ECG analysis that "only looks at the heart" to a multi-organ assessment device that can reflect the overall aging state of organs.

[0056] (6) Organ Biology Age Conversion Module The organ biological age conversion module is used to convert the organ state characterization parameters generated by the multi-organ characterization output module into organ biological age assessment results. Specifically, the organ biological age conversion module includes the following units.

[0057] (6.1) Single organ characterization quantification unit, used to generate single organ quantification results for the organ state characterization parameter s_k corresponding to each organ, wherein the single organ quantification results include one or more of the following: organ aging related characterization value, organ age index, organ age deviation value prediction, age percentile or age score.

[0058] (6.2) Scale mapping unit, used to determine at least one mapping threshold, quantile threshold or scale conversion parameter based on the distribution of corresponding organ state characterization parameters in training samples, calibration samples or reference population, and to map the organ state characterization parameter s_k of the target object to the age percentile, age score, organ age index or organ age deviation value of the corresponding organ.

[0059] (6.3) Age Conversion Unit: This unit converts the characterization results of each organ into an organ biological age assessment value based on the single-organ quantification results and in conjunction with a preset age scale model, reference population distribution parameters, or calibration parameters. Specifically, it establishes an age scale function for each organ based on preset training samples, calibration samples, or a reference population. And the organ status characterization parameters are mapped to organ biological age assessment values: in, Indicates the first The organ biological age assessment value corresponding to each organ. It can be constructed using linear mapping, nonlinear regression mapping, quantile mapping, lookup table mapping, or a combination thereof.

[0060] Furthermore, if the actual age of the target is denoted as... Then the first one can be calculated. Organ age deviation value for each organ: in, This is used to characterize the degree of deviation of the organ's biological age from its actual age. Preferably, it can also be based on the distribution of a reference population. or Further conversion to age percentiles, age scores, or age groups enhances the interpretability of the results.

[0061] (6.4) Fusion assessment unit, used to fuse organ biological age assessment values, organ age deviation values, or organ age indices corresponding to multiple organs to generate an overall organ status assessment result R: in, Indicates the first The fusion weights corresponding to each organ can be determined by using preset weights, equal weights, or based on organ importance.

[0062] (6.5) Report generation unit, used to generate and output organ status assessment reports based on organ biological age assessment values, organ age deviation values, organ age index, health risk scores and overall organ status assessment results for each organ.

[0063] For example, in a personal health management scenario, if the actual age of a target individual... Given an age of 50, the model obtains the organ biological age assessment value corresponding to its k-th organ. If the age is 58, then the age deviation value for this organ is... For children aged 8 years, the report may mark the organ as above the age-matched reference level and provide an age-appropriate grading or risk assessment based on the reference population quantile. This result is used for health status assessment and follow-up management and does not directly replace clinical diagnosis.

[0064] (7) Result Output Module The result output module is used to output the organ biological age assessment value, organ age deviation value, organ age level, health risk score, or organ status assessment report for each organ. In one embodiment, the result output module outputs at least the following: target object identification information, assessment time, and organ biological age assessment value for each organ. Organ age deviation value Organ age index, age percentile, age class, health risk score, and overall organ status assessment results (R).

[0065] For example, in a hospital physical examination report generation scenario, the results output module can list the heart, kidneys, liver, lungs, cerebrovascular system, or metabolic system as separate report items, and display the organ biological age assessment value, deviation from the actual age, percentile of the same age group, and rank description for each organ. In a remote health management scenario, the results output module can also output only structured JSON results for subsequent use by the health management platform.

[0066] like Figure 1 As shown, after the 12-lead electrocardiogram (ECG) data is input into the ECG encoding module, an ECG feature sequence is generated. After the auxiliary table data is input into the table data encoding module, a table feature sequence is generated. The two are concatenated to form a unified multimodal context sequence. Multiple task query vectors After inputting into the fusion module, cross-attention operations are performed with the context sequence to obtain fusion representations corresponding to multiple organs. Subsequently, the multi-organ characterization output module outputs organ state characterization parameters based on the fused characterization of each organ. It is then mapped to organ biological age assessment values ​​for multiple organs through the organ biological age conversion module. and organ age deviation value Finally, the results output module generates an organ status assessment report.

[0067] Corresponding to the aforementioned embodiments of the organ biological age assessment device based on electrocardiogram signals, this application also provides an electronic device, including: one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors perform the functions of each module in the organ biological age assessment device based on electrocardiogram signals as described above. Figure 2 The diagram shown is a hardware structure diagram of any data processing-capable device, including the organ biological age assessment device based on electrocardiogram signals provided in this application embodiment. Except for... Figure 2 In addition to the processor, memory, DMA controller, disk, and non-volatile memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0068] Corresponding to the aforementioned embodiments of the organ biological age assessment device based on electrocardiogram signals, this embodiment of the invention also provides a computer-readable storage medium storing a program thereon. When the program is executed by a processor, it implements the functions of each module in the organ biological age assessment device based on electrocardiogram signals in the above embodiments.

[0069] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0070] 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 scope of protection of the present invention.

[0071] The above embodiments are only used to illustrate the design concept and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, all equivalent changes or modifications made based on the principles and design ideas disclosed in the present invention are within the protection scope of the present invention.

Claims

1. A device for assessing organ biological age based on electrocardiogram signals, characterized in that, include: The ECG signal preprocessing module is used to perform lead standardization, filtering and noise reduction, resampling, fixed-length truncation or padding, lead-by-lead standardization, and outlier detection on the acquired ECG signals to generate standardized ECG data for subsequent analysis. The electrocardiogram feature extraction module is used to extract feature information characterizing the characteristics of cardiac electrical activity from the preprocessed electrocardiogram data. The feature information includes heart rate variability features, PQRST band morphological features, and electrocardiogram amplitude features. An organ state characterization generation module is used to output characterization parameters related to different organ states based on the feature information and using a machine learning model. The characterization parameters include one or more of organ aging-related characterization values, organ age index, or organ age deviation values. The organ biological age conversion module is used to convert the characterization results of different organs into organ biological age assessment values ​​based on the organ state characterization parameters and in combination with a preset age scale model, reference population distribution parameters or calibration parameters. The results output module is used to output the organ biological age assessment value, organ age deviation value, organ age level or organ status assessment report for each organ.

2. The organ biological age assessment device based on electrocardiogram signals according to claim 1, characterized in that, The electrocardiogram (ECG) signal preprocessing module is used to convert the original ECG waveform records into a standardized array format for model training or inference. The ECG signal preprocessing module includes: The record reading unit is used to locate the base name path of the corresponding WFDB record in the dataset directory based on the patient ID and ECG record ID of the sample to be processed, and read the WFDB record to obtain the original waveform matrix sig, the original sampling rate fs, and the lead name list names; The lead mapping unit is used to map the original waveform matrix to a preset 12-lead set according to the lead name. The 12-lead set includes at least I, II, III, aVR, aVL, aVF, V1, V2, V3, V4, V5 and V6. Missing leads are filled in with a zero-value sequence to obtain a standardized 12-lead waveform matrix x. A bandpass filter unit is used to perform bandpass filtering on the standardized 12-lead waveform matrix. The bandpass frequency band is set to 0.5Hz to 45Hz, the filter order is set to 4, and the forward-backward zero-phase filtering method is used to filter the time axis to suppress low-frequency baseline drift and high-frequency interference. The notch filter unit is used to perform notch filtering on the waveform matrix after bandpass filtering to suppress power frequency noise. When the notch frequency is higher than the Nyquist frequency threshold, the corresponding notch processing is not performed. The resampling unit is used to resample the waveform after notch filtering from the original sampling rate fs to the target sampling rate TARGET_FS; The length unification unit is used to unify the resampled waveform to a fixed number of sampling points TARGET_FS×DURATION_SEC corresponding to a fixed time length DURATION_SEC. When the resampled length is greater than the target length, it is truncated at the center of the waveform. When the resampled length is less than the target length, it is symmetrically zero-padded at both ends of the waveform to obtain a waveform matrix with a fixed shape. The standardization unit is used to perform lead-by-lead standardization on a fixed-length waveform matrix. It calculates the mean and standard deviation for each lead, converts the waveform of that lead into a zero-mean, unit-variance form, introduces a stable term into the standard deviation, and truncates the amplitude of the standardized waveform. The quality control unit is used to detect the validity of the standardized waveform matrix. When there are abnormal values, it uses a preset replacement rule to replace the abnormal values ​​with zero values, and sets the corresponding leads to a zero sequence when the standard deviation of the leads is lower than a preset threshold. The data output unit is used to save the fixed-specification waveform matrix that has passed quality control as an array file in floating-point format, and record metadata such as sample identifier, path, sampling rate, number of leads and number of sampling points for subsequent training, verification or inference.

3. The organ biological age assessment device based on electrocardiogram signals according to claim 2, characterized in that, The organ state representation generation module takes the preprocessed 12-lead ECG segment as input, extracts the representation vector through the shared ECG feature backbone network, and outputs the organ state representation parameters corresponding to different organs through multiple organ representation output heads. The organ state representation parameters include one or more of the following: organ aging-related representation values, organ age index, or organ age deviation value.

4. The organ biological age assessment device based on electrocardiogram signals according to claim 3, characterized in that, The organ state characterization generation module includes: The input building unit is used to represent the preprocessed ECG signal as a three-dimensional tensor X∈R^(B×12×T), where B is the batch size, 12 is the number of leads, and T is the number of fixed sampling points; The multi-scale convolutional feature extraction unit is used to perform multiple sets of one-dimensional convolution operations on the input tensor X in parallel to obtain morphological features at different time scales, obtain multiple intermediate features, and obtain fused features X_ms through scale dimension fusion. The temporal downsampling unit is used to sequentially perform batch normalization, nonlinear activation, max pooling and random deactivation on the fused feature X_ms to improve training stability and achieve temporal downsampling to obtain feature X_0; The convolutional transformation unit is used to transform the feature X_0 sequentially through at least two convolutional blocks. Each convolutional block contains one-dimensional convolution, batch normalization, non-linear activation, and max pooling operations to progressively improve the expressive power of the channel dimension and compress the length of the time dimension to obtain the feature X_c. The channel recalibration unit is used to input the feature X_c into the channel recalibration module. The channel recalibration module is composed of several residual blocks with SE structure connected in series, which is used to adaptively weight the importance of different channels to obtain the enhanced feature X_crm. The bidirectional temporal coding unit is used to transform the enhanced feature X_crm into a sequence form and input it into the Transformer encoder to learn long-range dependencies. This includes performing forward encoding on the temporal sequence to obtain Z_fwd, performing reverse encoding on the sequence after time dimension inversion to obtain Z_bwd, and fusing the forward encoding result and the reverse encoding result to obtain the sequence representation Z. The feature aggregation unit is used to perform adaptive average pooling on the sequence representation Z, and aggregate it in the time dimension to obtain a fixed-dimensional ECG feature vector f. The multi-organ representation output unit is used to construct multiple organ representation output heads based on the fixed-dimensional electrocardiogram feature vector f, and output the corresponding organ state representation parameters for different organs. Each organ representation output head maps the feature vector of the shared backbone output to the representation output value corresponding to the organ through a linear mapping layer, and concatenates the representation output values ​​corresponding to multiple organs in the task dimension to obtain the model output.

5. The organ biological age assessment device based on electrocardiogram signals according to claim 1, characterized in that, The organ biological age conversion module includes: (1) Single organ characterization quantification unit, used to generate single organ quantification results for the characterization output value corresponding to each organ, wherein the single organ quantification results include one or more of the following: organ aging-related characterization value, organ age index, organ age deviation value, age percentile or age score. (2) Scale mapping unit, used to determine at least one mapping threshold, quantile threshold or scale conversion parameter based on the distribution of the corresponding organ representation output value in the training sample, calibration sample or reference population, and to map the organ representation output value of the target object to the age percentile, age score, organ age index or organ age deviation value of the corresponding organ. (3) Age conversion unit, used to convert the characterization results of each organ into organ biological age assessment values ​​based on the single organ quantification results and in combination with the preset age scale model, reference population distribution parameters or calibration parameters. (4) Fusion assessment unit, used to fuse organ biological age assessment values, organ age deviation values ​​or organ age indices corresponding to multiple organs to generate overall organ status assessment results; (5) Report generation unit, used to generate and output organ status assessment reports based on the organ biological age assessment value, organ age deviation value, organ age index and overall organ status assessment results of each organ.

6. An electronic device comprising a memory and a processor, characterized in that, The memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the functions of each module in the organ biological age assessment device based on electrocardiogram signals according to any one of claims 1-5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the functions of each module in the organ biological age assessment device based on electrocardiogram signals as described in any one of claims 1-5.

Images