A system for assessing a state of reduced cardiopulmonary function
By collecting multi-dimensional physiological indicators and monitoring with wearable devices, combined with unsupervised clustering analysis and predictive models, a cardiopulmonary function decline index was constructed. This solved the problems of single input and static modeling in existing central lung function assessment systems, enabling early identification and personalized intervention, and improving the accuracy and effectiveness of cardiopulmonary function management.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XIANGYA HOSPITAL CENT SOUTH UNIV
- Filing Date
- 2026-04-23
- Publication Date
- 2026-05-29
Smart Images
Figure CN122117415A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of health management technology, and in particular to a system for assessing the decline in cardiopulmonary function. Background Technology
[0002] With the accelerating aging of the population and changes in lifestyle, declining cardiopulmonary function has become a significant hidden danger affecting the health of the population. Decreased cardiopulmonary function is typically characterized by insidious onset and slow progression; therefore, effective quantitative analysis and status prediction are of great importance for health management.
[0003] Existing data processing systems for cardiopulmonary function status mainly suffer from the following technical problems: First, the limited dimension of input features leads to insufficient representational capabilities in the constructed data models. Existing data processing systems often rely on single functional indicators (such as ejection fraction or vital capacity) as model inputs, lacking comprehensive feature extraction from multi-dimensional physiological indicators such as cardiac and pulmonary function. This single-dimensional feature input approach results in data models that cannot comprehensively represent the coordinated changes and overall state of the cardiopulmonary system, have limited ability to capture early and subtle functional changes, and suffer from insufficient accuracy and comprehensiveness in model output.
[0004] Second, there is a lack of a modeling mechanism for the gradual evolution of dynamic time-series data. Current data collection methods primarily rely on hospitals or specialized testing institutions, resulting in long sampling cycles, high costs, and discrete, static snapshots that cannot form continuous time-series data suitable for dynamic modeling. This makes it difficult for existing technologies to construct time-series prediction models capable of capturing subtle trends and gradual evolution of individual cardiopulmonary function, thus hindering dynamic quantitative analysis of functional state changes.
[0005] Third, existing data processing architectures lack the capability to generate personalized outputs based on risk stratification. Current data processing systems primarily focus on recording and managing diagnosed patients. For healthy or sub-healthy individuals, there is a lack of a mechanism to adaptively generate personalized output plans based on their dynamically changing risk indices. This makes it difficult for existing data processing systems to provide individuals with scientific quantitative guidance and to achieve dynamic data output based on changes in status. Summary of the Invention
[0006] The purpose of this invention is to provide a cardiopulmonary function decline assessment system, which, through the collection and comprehensive analysis of multi-dimensional physiological indicators, combined with real-time monitoring by wearable devices and personalized adaptive intervention, achieves early and accurate identification and dynamic health management of cardiopulmonary function decline, thereby solving at least one of the aforementioned problems in the prior art.
[0007] This invention provides a system for assessing cardiopulmonary function decline, the system specifically comprising: The data acquisition module is used to collect multi-dimensional physiological indicator data of individuals and preprocess the multi-dimensional physiological indicator data; The index construction module is used to classify the cardiopulmonary function decline status of individuals based on preprocessed multi-dimensional physiological index data and to construct a cardiopulmonary function decline index using an unsupervised clustering analysis model. The model building module is used to build a cardiopulmonary function decline index prediction model based on multi-source clinical data. It collects individual physiological data in real time or periodically through wearable devices and inputs it into the cardiopulmonary function decline index prediction model, and outputs the individual's current cardiopulmonary function decline index. The plan generation module is used to generate personalized health intervention plans based on the output cardiopulmonary function decline index and a pre-set tiered intervention plan sample library through a large model, and adaptively optimize the health intervention plan as the index changes.
[0008] Compared with the prior art, the present invention has at least one of the following technical effects: 1. This invention achieves early and accurate identification and dynamic health management of cardiopulmonary function decline by collecting and comprehensively analyzing multi-dimensional physiological indicators, combined with real-time monitoring by wearable devices and personalized adaptive intervention.
[0009] 2. This invention employs a multi-dimensional joint analysis of cardiac function, pulmonary function, and metabolic indicators, overcoming the limitations of previous single-dimensional assessments. Based on multi-dimensional physiological indicators, an unsupervised clustering analysis model is used to construct a cardiopulmonary function decline index, which is divided into four levels: Level 1 indicates good cardiopulmonary function, Level 2 indicates mild cardiopulmonary function decline, Level 3 indicates moderate cardiopulmonary function decline, and Level 4 indicates severe cardiopulmonary function decline. The index construction breaks through the traditional binary division of "healthy / ill" and provides a crucial assessment tool for risk stratification of cardiopulmonary function decline, helping to more comprehensively and accurately identify the early risk of cardiopulmonary function decline.
[0010] 3. This invention establishes a predictive model for cardiopulmonary function decline index based on multi-source clinical data. By wearing mobile monitoring devices such as smart bracelets, smartwatches, portable pulmonary function instruments, and portable electrocardiographs, human physiological data is continuously and regularly collected and input into the predictive model. The model outputs the individual's current cardiopulmonary function decline index in real time or periodically, which helps to achieve dynamic home monitoring of early decline risk.
[0011] 4. This invention adaptively optimizes the intervention plan based on the dynamic changes of the cardiopulmonary function decline index. The plan includes drug types, dosage and method of administration, health behavior guidance, and follow-up management recommendations. By continuously assessing the individual's cardiopulmonary function status, it improves the identification and intervention of early cardiopulmonary function decline, thereby effectively delaying disease progression and reducing the risk of cardiopulmonary function-related diseases.
[0012] 5. Based on unsupervised clustering analysis of core feature parameters, this invention combines clinical event rate grading to construct a scientific cardiopulmonary function decline index system, thereby improving the accuracy of early condition identification.
[0013] 6. This invention optimizes the number of cluster classifications through comprehensive evaluation of multi-dimensional stability and clinical discriminability, ensuring the reliability and clinical applicability of the grading of cardiopulmonary function decline status.
[0014] 7. This invention constructs a prediction model based on time-series Transformer, and uses physiological data from multiple time points to achieve dynamic prediction of the cardiopulmonary function decline index, which can capture the gradual evolution pattern.
[0015] 8. This invention uses a multi-layer Transformer encoder and a global feature extraction structure to deeply mine long-term dependencies in time-series physiological data, thereby improving the feature expression capability of the prediction model.
[0016] 9. This invention employs a multi-head self-attention mechanism and a feedforward neural network module to effectively capture the interaction features between physiological indicators at different time points, thereby enhancing the comprehensiveness of temporal feature extraction.
[0017] 10. This invention achieves dynamic adjustment and rapid convergence of model parameters through joint training of the cross-entropy loss function and the Adam optimizer, thereby improving the accuracy of the prediction of the cardiopulmonary function decline index.
[0018] 11. This invention combines individual static attributes with dynamic monitoring data, and generates personalized intervention plans through large-scale model semantic reasoning, thereby realizing adaptive health management based on risk changes.
[0019] 12. This invention constructs a sample library of graded intervention programs covering the entire index range by mining historical medical data and labeling at the index level, providing data support for personalized health guidance. Attached Figure Description
[0020] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0021] Figure 1 This is a schematic diagram of the structure of a cardiopulmonary function decline assessment system provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the algorithm framework of a cardiopulmonary function decline assessment system provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a cardiopulmonary function decline index prediction model provided in an embodiment of the present invention. Detailed Implementation
[0022] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.
[0023] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.
[0024] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.
[0025] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."
[0026] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.
[0027] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.
[0028] Reference Figure 1 , Figure 1 A schematic diagram of the structure of a cardiopulmonary function decline assessment system according to an embodiment of the present invention is shown below, in detail: The data acquisition module is used to collect multi-dimensional physiological indicator data of individuals and preprocess the multi-dimensional physiological indicator data; The index construction module is used to classify the cardiopulmonary function decline status of individuals based on preprocessed multi-dimensional physiological index data and to construct a cardiopulmonary function decline index using an unsupervised clustering analysis model. The model building module is used to build a cardiopulmonary function decline index prediction model based on multi-source clinical data. It collects individual physiological data in real time or periodically through wearable devices and inputs it into the cardiopulmonary function decline index prediction model, and outputs the individual's current cardiopulmonary function decline index. The plan generation module is used to generate personalized health intervention plans based on the output cardiopulmonary function decline index and a pre-set tiered intervention plan sample library through a large model, and adaptively optimize the health intervention plan as the index changes.
[0029] In this embodiment, the data acquisition module is responsible for collecting and preprocessing multi-dimensional physiological indicator data of an individual. In practice, data is collected using various professional medical testing equipment and wearable devices. For professional medical testing equipment, an electrocardiograph (ECG) is used to collect data related to cardiac electrical activity, such as heart rate and rhythm; an echocardiogram is used to obtain structural and functional indicators of the heart, such as ejection fraction and ventricular wall thickness; and a pulmonary function analyzer is used to measure lung function indicators such as vital capacity and forced expiratory volume. For wearable devices, smart bracelets or smartwatches can collect data such as an individual's heart rate, steps taken, and sleep quality in real time. The collected multi-dimensional physiological indicator data may contain noise, missing values, etc., therefore preprocessing is necessary. For noisy data, a smoothing filter method is used, such as a moving average filter, which smooths the data by averaging the data within a certain window, reducing the interference of random noise. For missing values, different processing methods are adopted according to the characteristics of the data. If the data has temporal continuity, linear interpolation can be used to estimate missing values based on data from adjacent time points. If the data is highly correlated, missing values can be filled using data from other relevant indicators through regression analysis and other methods. The preprocessed data is more accurate and complete, providing a reliable foundation for subsequent modules.
[0030] The index construction module, based on preprocessed multidimensional physiological indicator data, employs an unsupervised clustering analysis model to classify the cardiopulmonary function decline status of individuals and construct a cardiopulmonary function decline index. First, the preprocessed multidimensional physiological indicator data is standardized to ensure that different indicators have the same dimensions and scale, avoiding the impact of different indicator dimensions on the clustering results. For example, for heart rate and vital capacity data, whose numerical ranges and units vary significantly, standardization converts them into data with a mean of 0 and a standard deviation of 1. Then, a suitable unsupervised clustering algorithm, such as K-means clustering, is selected. The number of clusters K is determined; the value of K can be determined based on clinical experience and data analysis. For example, classifying cardiopulmonary function decline status into four levels—healthy, mild decline, moderate decline, and severe decline—results in K=4. The standardized multidimensional physiological indicator data is then input into the K-means clustering algorithm, which classifies individuals into different categories based on the similarity between data points. After clustering, each category is assigned a specific score; for example, a score of 0 for the healthy category, 1 for the mildly impaired category, 2 for the moderately impaired category, and 3 for the severely impaired category. Based on the individual's category, the corresponding score is used as the initial value for that individual's cardiopulmonary function decline index. Furthermore, to more accurately reflect an individual's cardiopulmonary function status, the weights of each indicator during the clustering process can be considered, and the initial values can be weighted and adjusted to obtain the individual's accurate cardiopulmonary function decline index.
[0031] The model building module constructs a cardiopulmonary function decline index (CPR) prediction model based on multi-source clinical data. Individual physiological data is collected in real-time or periodically via wearable devices and input into the model, outputting the individual's current CPR index. Multi-source clinical data includes a large amount of cardiopulmonary function-related data from diagnosed patients, such as multi-dimensional physiological indicators and corresponding CPR indices for patients of different ages, genders, and lifestyles. This multi-source clinical data is organized and analyzed to extract feature variables related to the CPR index. For example, analysis reveals a correlation between age, exercise frequency, and smoking status and the CPR index; these factors are then used as feature variables. Next, a suitable machine learning algorithm, such as the support vector machine (SVM) algorithm, is selected to build the prediction model. The multi-source clinical data is divided into training and testing sets. The training set is used to train the model, adjusting its parameters to accurately fit the training data. The testing set is used to evaluate the model's performance, calculating metrics such as prediction accuracy and recall. Based on the evaluation results, the model is optimized and adjusted until satisfactory performance is achieved. In practical applications, wearable devices collect an individual's physiological data in real time or periodically, such as real-time heart rate and steps, and periodically (e.g., daily or weekly) more comprehensive physiological indicators. After organizing and preprocessing this collected data according to the model's required format, it is input into a pre-constructed cardiopulmonary function decline index prediction model. The model calculates and predicts based on the input data, outputting the individual's current cardiopulmonary function decline index.
[0032] The intervention module generates personalized health intervention plans based on the output cardiopulmonary function decline index and a pre-set tiered intervention plan sample library. The plan is then adaptively optimized as the index changes. The pre-set tiered intervention plan sample library contains health intervention measures for different cardiopulmonary function decline index levels. For example, for an individual with a cardiopulmonary function decline index of 0 (healthy), the intervention plan might include maintaining existing healthy lifestyle habits such as a balanced diet and moderate exercise; for an individual with a cardiopulmonary function decline index of 1 (mild decline), the intervention plan might add targeted exercise suggestions, such as 3-5 aerobic exercise sessions per week for about 30 minutes each time; for an individual with a cardiopulmonary function decline index of 2 (moderate decline), the intervention plan might include dietary adjustments in addition to exercise suggestions, such as reducing the intake of high-fat and high-sugar foods and increasing the intake of fruits and vegetables; for an individual with a cardiopulmonary function decline index of 3 (severe decline), the intervention plan might recommend seeking medical attention promptly for further examination and treatment. After outputting an individual's cardiorespiratory function decline index, the corresponding initial intervention plan is retrieved from the tiered intervention plan sample library. Then, a large-scale model is used to personalize and optimize the initial intervention plan. The large-scale model can consider more individual information, such as age, gender, lifestyle habits, and medical history, adjusting and refining the initial intervention plan based on this information. For example, for an older individual with mild cardiorespiratory function decline who has a smoking habit, the large-scale model might add smoking cessation suggestions to the initial intervention plan and adjust the exercise intensity and frequency to better suit the individual's actual situation. As the individual's cardiorespiratory function decline index changes, the plan generation module monitors the index changes in real time. When the index changes, the corresponding plan is retrieved again from the tiered intervention plan sample library based on the new index, and the large-scale model is used to adaptively optimize the health intervention plan by combining the individual's latest information, ensuring that the individual always receives a health intervention plan that is most suitable for their current condition.
[0033] As a brief description of the aforementioned cardiopulmonary function decline assessment system, refer to Figure 2 The algorithm framework diagram shown is as follows: First, the intelligent monitoring device is responsible for collecting multi-dimensional physiological data of individuals, specifically including heart rate, blood oxygen saturation, EF (ejection fraction), MMEF (maximum mid-expiratory flow rate), and FEV1 / FVC (forced expiratory volume in one second to forced vital capacity). These data are an important foundation for assessing cardiopulmonary function.
[0034] In the stage of the cardiopulmonary function decline index prediction model, after acquiring physiological data, core features are selected from these data to construct time-series data. These core features include EF, FEV1 / FVC, and MMEF. Constructing time-series data helps to observe the changing trends of these indicators over time. An unsupervised clustering analysis model is used to classify the individual's cardiopulmonary function decline status, and a cardiopulmonary function decline index is constructed and output based on the classification results. This index can intuitively reflect the degree of cardiopulmonary function decline in an individual.
[0035] In the modeling phase of intervention programs for cardiopulmonary function decline, the corresponding intervention program is searched in a pre-defined tiered intervention program sample library based on the cardiopulmonary function decline index output by the prediction model. The large-scale model API is then invoked for inference, leveraging the powerful reasoning capabilities of the large-scale model and combining individual gender, age, heart rate, blood oxygen saturation, and pre-defined prompts to infer from the sample library. This step aims to generate more accurate and personalized intervention programs. After the large-scale model's inference, health recommendations matching the individual's current cardiopulmonary function status are output. These recommendations cover multiple aspects such as diet, exercise, and lifestyle, aiming to help individuals improve their cardiopulmonary function.
[0036] This completes a full identification and intervention process.
[0037] In some embodiments, the index construction module is specifically used for: Based on the preprocessed multidimensional physiological index data, ejection fraction (cardiac function index), forced expiratory volume in one second (fEV1) to forced vital capacity (fEV1) and maximum mid-expiratory flow rate (MPV) were selected as core feature parameters to characterize the state of cardiopulmonary function decline. The selected core feature parameters are input into the K-means clustering model, and clustering iteration is performed with a preset number of candidate categories to generate clustering results under different category divisions; Based on the stability and clinical interpretability of the clustering results, the optimal number of categories is selected from the candidate categories, and individuals are divided into the corresponding number of cardiopulmonary function decline status categories. The event rate of patients with cardiopulmonary function decline in each category was statistically analyzed. Each decline category was assigned a value in order of increasing event rate, forming a cardiopulmonary function decline index grading system from low to high, representing good cardiopulmonary function to severe decline.
[0038] Furthermore, based on the stability and clinical interpretability of the clustering results, the optimal number of categories is selected from the candidate categories to classify individuals into the corresponding number of cardiopulmonary function impairment categories, specifically including: Based on the continuous characteristics of the progressive decline in cardiopulmonary function over time and the clinical grading method, multiple candidate clustering categories are set. Cluster analysis is performed for each number of candidate categories to evaluate the stability of the clustering results under different numbers of candidate categories and obtain cluster stability evaluation results. The stability evaluation includes determining whether similar samples are continuously classified into the same category and the changes in cluster centers under different numbers of categories. Based on the clustering results, the differences in core feature parameters and clinical interpretability among different candidate categories with different numbers of categories are evaluated to obtain clinical discrimination assessment results. Based on a comprehensive comparison of the cluster stability assessment results and the clinical discrimination assessment results, the number of target categories that meet the preset requirements for both cluster stability and clinical discrimination is selected as the optimal number of categories. Based on the optimal number of classifications, determine the category of cardiopulmonary function decline corresponding to each individual.
[0039] In this embodiment, based on commonly used clinical evaluation indicators and relevant literature research results, cardiac function indicators ejection fraction (EF), pulmonary function indicators forced expiratory volume in one second to forced vital capacity ratio (FEV1 / FVC), and maximum mid-expiratory flow (MMEF) are significantly correlated with the occurrence and development of cardiopulmonary function decline. Therefore, the three functional indicators EF, FEV1 / FVC, and MMEF were selected as core feature parameters of the sample for cluster analysis of cardiopulmonary function decline status.
[0040] The K-means clustering model was used, taking into account the continuous decline of cardiopulmonary function over time and commonly used clinical functional grading methods. Cardiopulmonary function decline states were categorized into 3 to 6 candidate categories (i.e., k=3, k=4, k=5, or k=6) for cluster analysis. The clustering process was repeated for different candidate k values to evaluate the stability and discriminative power of the clustering results. Stability included whether similar samples were consistently grouped into the same category and the changes in cluster centers under different k values. Experimental results showed that when k=3, different cardiopulmonary function decline states were easily merged, resulting in insufficient category discrimination; when k>4, the category division was too granular, and cluster stability decreased significantly; while when k=4, the clustering results showed good stability and interpretability, with clear differences in cardiopulmonary function decline between categories.
[0041] Therefore, the samples were ultimately divided into four categories, corresponding to four states of cardiopulmonary function decline. Based on the aforementioned collected sample data, the number of samples with cardiopulmonary function-related diseases in each category was further counted according to individual disease information, and the event rate for each category was calculated, i.e.: Similarly, we can conclude that The events were sorted from lowest to highest frequency and assigned cardiopulmonary function decline indices from 1 to 4. Level 1 indicates good cardiopulmonary function, Level 2 indicates mild decline, Level 3 indicates moderate decline, and Level 4 indicates severe decline. This 4-level classification aligns with clinical understanding of cardiopulmonary function, balancing the precision of risk stratification with model stability, and is beneficial for subsequent prediction of cardiopulmonary function decline and the generation of intervention plans.
[0042] In some embodiments, the model building module is specifically used for: Collect multidimensional physiological index data of individuals at multiple time points in the order of examination time, construct time series samples with fixed time steps, and assign the supervised learning label of the corresponding time step to the cardiopulmonary function decline index corresponding to the end of the time window. Divide the data into training set and test set according to a preset ratio to form a sample dataset. Construct a predictive model for cardiopulmonary function decline index with time-series Transformer as the main structure; The sample dataset is input into the cardiopulmonary function decline index prediction model. The cross-entropy loss function is used for multi-class training. The model parameters are iteratively updated through the optimizer until the model converges, so that the classification probability distribution predicted by the model gradually approaches the true label distribution.
[0043] In this embodiment, multidimensional physiological index data (including EF, FEV1 / FVC, and MMEF) for each individual are collected in chronological order of examination time. Time-series samples are constructed at fixed time steps, such as continuous data from the most recent T weeks. Each sample is then represented as: in, This represents a time series sample, which is a collection of multidimensional physiological index data of multiple individuals collected at fixed time steps, used for subsequent analysis and model building. This represents the data collected over T consecutive time steps.
[0044] Based on the above clustering analysis results, the supervised learning label (i.e., the cardiopulmonary function decline index) at the corresponding time step is assigned the value at the end of the time window (i.e., The cardiopulmonary function decline index corresponding to the data was used to construct a sample dataset. The processed sample dataset was then divided into a training set and a test set in a 7:3 ratio.
[0045] Cardiopulmonary function decline is characterized by long-term, gradual evolution, and physiological indicators at a single time point are insufficient to accurately reflect the true risk status. The Transformer structure can capture temporal variations, therefore, a prediction model with a temporal Transformer as its primary structure was chosen.
[0046] A fully connected layer is used to linearly map multidimensional physiological indicators for each time period, converting them into feature embedding vectors of a uniform dimension. Simultaneously, temporal position encoding is introduced to represent the sequential relationship between different time periods. The feature embeddings and temporal position encodings are then fused to form the model input sequence. This input sequence is fed into a TransformerEncoder for temporal feature modeling. The encoder includes a multi-head self-attention module for modeling dependencies between different time points and a feedforward neural network module for nonlinear feature transformation. Through multi-layered encoder structures, features of cardiopulmonary function changing over time are extracted.
[0047] Based on this, a global average pooling structure is used to aggregate the temporal features output by the Encoder. The resulting feature vector is then input into a fully connected classification layer, which outputs the classification probability distribution corresponding to the cardiopulmonary function decline index. The one with the highest probability is then used as the final predicted cardiopulmonary function decline index.
[0048] Multi-class classification training is performed using the cross-entropy loss function, and the Adam optimizer is used to update the model parameters. The cross-entropy loss function is defined as follows: in, This represents the true label of a sample in the i-th class, when the sample belongs to the i-th class. ,otherwise ; Let represent the probability value that the model predicts for a sample to belong to the i-th class, and satisfy . By minimizing the aforementioned cross-entropy loss function, the model's predicted classification probability distribution gradually approximates the true label distribution, thereby achieving effective classification prediction of the cardiopulmonary function decline index.
[0049] After training, the model's structural parameters and weight information are stored for later retrieval and updates. The trained model is then deployed to the application environment, enabling it to perform inference and prediction on newly collected data, thereby supporting the practical application of the cardiopulmonary function decline index.
[0050] Individual physiological data collected in real time or periodically by intelligent monitoring devices (e.g., weekly collection) is input into a trained model, which outputs a cardiopulmonary function decline index for the corresponding time period. Based on the prediction results, dynamic updates and continuous home monitoring of an individual's cardiopulmonary function status are achieved, thereby identifying potential risks of cardiopulmonary function decline in advance.
[0051] Furthermore, the cardiopulmonary function decline index prediction model includes a fully connected layer, a multi-layer Transformer encoder, a global average pooling layer, and a fully connected classification layer; The fully connected layer is used to linearly transform the multidimensional physiological index data contained in each time step of the input sequence into a feature embedding vector with a unified dimension, and then fuses the feature embedding vector and the time position encoding element by element to form a model input sequence carrying temporal position information. The multi-layer Transformer encoder is used to extract deep temporal features from the model input sequence by sequentially stacking multiple Transformer encoders, thereby capturing the gradual evolution of cardiopulmonary function over time. The global average pooling layer is used to generate a global feature vector by averaging the temporal feature vectors of all time steps output by the last Transformer encoder layer over the time dimension. The fully connected classification layer is used to calculate the probability distribution of each category corresponding to the current cardiopulmonary function decline index of an individual based on the global feature vector through linear transformation and Softmax activation function, and the index level corresponding to the maximum probability is used as the final prediction result.
[0052] Furthermore, each Transformer encoder includes a multi-head self-attention module and a feedforward neural network module; The multi-head self-attention module is used to calculate the attention weights between time steps within the model input sequence through the multi-head self-attention mechanism, capture the long-term dependencies and interaction features between physiological indicators at different time points, and obtain the time-series features after weighted aggregation. The feedforward neural network module is used to perform deep transformation and mapping on the weighted aggregated temporal features through multi-layer nonlinear activation functions, and output a temporal feature vector.
[0053] In this embodiment, reference is made to Figure 3 First, the cardiopulmonary function decline index prediction model receives input sequence data, which contains multidimensional physiological indicators of an individual at multiple time steps, specifically including EF, FEV1 / FVC, and MMEF. These physiological indicators reflect the state of cardiopulmonary function at different time points.
[0054] The fully connected layer performs a linear transformation on the multidimensional physiological indicator data contained in each time step of the input sequence. Its function is to map physiological indicator data of different dimensions into feature embedding vectors with a unified dimension. Simultaneously, the feature embedding vectors and temporal position encodings are fused element-wise. The temporal position encoding adds positional information to the data at each time step, forming a model input sequence carrying temporal positional information so that the model can recognize the temporal order of the data.
[0055] A multi-layer Transformer encoder consists of multiple Transformer encoders stacked sequentially. Each Transformer encoder contains a multi-head self-attention module and a feedforward neural network module.
[0056] The multi-head self-attention module calculates the attention weights between time steps within the model's input sequence using a multi-head self-attention mechanism. This mechanism can capture the long-term dependencies and interaction features between physiological indicators at different time points. For example, a change in a physiological indicator at an early time point may affect indicators at later time points. The multi-head self-attention module can identify and quantify this relationship, obtaining the weighted aggregated temporal features.
[0057] The feedforward neural network module utilizes multi-layer nonlinear activation functions to perform deep transformation and mapping on the weighted aggregated temporal features. This process can further extract complex features from the data and output a more representative temporal feature vector. Through sequential processing by multiple Transformer encoders, the model performs layer-by-layer abstraction of deep temporal features from the input sequence, thereby capturing the gradual evolution of cardiopulmonary function over time.
[0058] like Figure 3 As shown, the multi-head attention layer is the core component of the multi-head self-attention module. When the input sequence enters the multi-head attention layer, this layer focuses on the input data from multiple different representation subspaces. It maps the input sequence data to multiple different heads, each head independently calculating attention weights. Each head calculates the attention weights between different time steps within the input sequence based on its own parameter settings, thereby capturing the correlation between physiological indicators at different time points. In this way, the multi-head attention layer can mine long-term dependencies and interaction features in the data from multiple perspectives, obtaining weighted aggregated temporal features from multiple different perspectives. After the multi-head attention layer completes its calculations, the output results enter a normalization layer. The main function of this normalization layer is to standardize the feature data output by the multi-head attention layer. It adjusts the distribution of the data to a relatively stable range, for example, making the mean of the data 0 and the variance 1. The advantage of doing so is that it can accelerate the model training process and improve the model's stability and convergence. At the same time, normalization can also reduce the impact of the difference in the units of different features on the model, enabling the model to learn the feature patterns in the data more effectively. After processing by the normalization layer, the temporal features output by the multi-head self-attention module are more standardized and stable, providing a good foundation for subsequent module processing.
[0059] like Figure 3As shown, the pre-feedback layer is a crucial component of the feedforward neural network module. It receives temporal feature data processed by the multi-head self-attention module. The pre-feedback layer performs deep transformations on the input data through a series of linear transformations and non-linear activation functions. For example, it might use activation functions such as ReLU (Rectified Linear Unit) to perform non-linear mapping on the input data, thereby increasing the model's expressive power. In this process, the pre-feedback layer can further extract complex features from the data and uncover hidden patterns and regularities. Through multiple layers of pre-feedback operations, useful information in the data is continuously refined and strengthened, preparing for the generation of more representative temporal feature vectors.
[0060] Similar to multi-head self-attention modules, the feedforward neural network module also inputs the results into a normalization layer after completing the operations of the feedback layer. This normalization layer standardizes and adjusts the feature vector output by the feedback layer, ensuring that the data distribution meets the training requirements of the model. This step further stabilizes the model's training process and prevents problems such as vanishing or exploding gradients during training. After processing by the normalization layer, the temporal feature vector output by the feedforward neural network module is more reliable and stable, and can be better used for subsequent processing by global average pooling layers and fully connected classification layers, ultimately achieving accurate prediction of the cardiopulmonary function decline index.
[0061] The global average pooling layer receives the temporal feature vectors from all time steps output by the last Transformer encoder layer. It integrates the feature information from multiple time steps into a single global feature vector by averaging over the time dimension. This step effectively reduces data dimensionality while preserving key features, generating a global feature vector that represents the entire input sequence.
[0062] The fully connected classification layer calculates the cardiopulmonary function index (CPR) based on the global feature vector output from the global average pooling layer using a linear transformation and a softmax activation function. The linear transformation maps the global feature vector to different class spaces, while the softmax activation function converts these mappings into probability distributions, outputting the probability of each class corresponding to the individual's current CPR level, such as P(Level 1), P(Level 2), P(Level 3), and P(Level 4). Finally, the level corresponding to the highest probability is used as the final predicted CPR level. For example, if P(Level 3) has the highest probability, the predicted CPR level for the individual is 3.
[0063] Furthermore, the step of inputting the sample dataset into the cardiopulmonary function decline index prediction model, using the cross-entropy loss function for multi-class training, and iteratively updating the model parameters through an optimizer until the model converges specifically includes: A cross-entropy loss function is constructed to measure the difference between the model's predicted probability distribution and the true label distribution. The cross-entropy loss function is defined as the weighted negative sum of the logarithm of the predicted probability of each category output by the cardiopulmonary function decline index prediction model and the true label. The true label is used to represent the true cardiopulmonary function decline index category to which the sample belongs in one-hot encoding form. The sample dataset is input into the cardiopulmonary function decline index prediction model. After forward propagation calculation in each network layer, the predicted probability distribution of the sample belonging to various cardiopulmonary function decline indices is output. The predicted probability distribution is then substituted into the cross-entropy loss function to calculate the loss value of the current batch. Based on the calculated loss value, the gradient information of each network parameter of the cardiopulmonary function decline index prediction model with respect to the loss value is calculated layer by layer through the backpropagation algorithm, and the error signal is propagated from the output layer to the input layer. The preset Adam optimizer is invoked, and the model network parameters are iteratively updated based on gradient information, momentum estimation, and adaptive learning rate adjustment strategies to gradually reduce the loss function value.
[0064] In this embodiment, the core function of the cross-entropy loss function is to measure the difference between the model's predicted probability distribution and the true label distribution. Specifically, the cross-entropy loss function is defined as the weighted negative sum of the logarithmic values of the predicted probabilities for each category output by the cardiopulmonary function decline index (CPF) prediction model and the true label. Here, the true label uses a one-hot encoding form to represent the true CPF category to which the sample belongs. For example, if the CPF is divided into four levels, the one-hot encoding form of the true label for a sample belonging to level 3 might be [0, 0, 1, 0]. This cross-entropy loss function, defined in this way, can accurately quantify the degree of deviation between the model's prediction results and the actual situation, providing a clear basis for subsequent model parameter adjustments.
[0065] The prepared sample dataset is input into the cardiopulmonary function decline index prediction model. The sample dataset consists of multidimensional physiological indicator data collected from individuals at multiple time points according to the examination time sequence. Time-series samples are constructed using fixed time steps and divided into training and test sets according to a preset ratio. After the data enters the model, it undergoes forward propagation calculations through each network layer. During this process, the model performs a series of transformations and processing on the input sample data, ultimately outputting the predicted probability distribution of the sample belonging to various cardiopulmonary function decline indices. For example, the model might output that the probability of a sample belonging to level 1, 2, 3, and 4 cardiopulmonary function decline indices is 0.1, 0.05, 0.8, and 0.05, respectively. This predicted probability distribution is then substituted into the previously constructed cross-entropy loss function to calculate the loss value for the current batch. This loss value reflects the magnitude of the model's prediction error for the sample data under the current parameter settings.
[0066] Based on the calculated loss value, the backpropagation algorithm is used to calculate the gradient information of each network parameter of the cardiopulmonary function decline index prediction model with respect to the loss value layer by layer. The principle of the backpropagation algorithm is to propagate the error signal from the output layer to the input layer. In this way, the influence of each network parameter on the loss value can be accurately calculated, that is, the gradient information. This gradient information is crucial for subsequent adjustment of model parameters to reduce the loss value.
[0067] The pre-defined Adam optimizer is invoked, and based on the calculated gradient information, combined with momentum estimation and adaptive learning rate adjustment strategies, the model network parameters are iteratively updated. The Adam optimizer is a commonly used optimization algorithm that adaptively adjusts the learning rate of each parameter based on gradient information, while using momentum estimation to accelerate the convergence process. In each iteration, the optimizer fine-tunes the model's network parameters based on gradient information, gradually reducing the loss function value. As the number of iterations increases, the model's predictions become increasingly closer to reality. When the loss function value no longer decreases significantly or reaches the pre-defined convergence condition, the model has reached convergence, at which point it can accurately predict the individual's cardiopulmonary function decline index.
[0068] This embodiment can effectively train the cardiopulmonary function decline index prediction model, enabling it to have good predictive performance and providing reliable technical support for the early identification of cardiopulmonary function decline.
[0069] In some embodiments, the scheme generation module is specifically used for: Based on the cardiopulmonary function decline index, the corresponding level of the intervention program sample library is matched and called from the pre-set graded intervention program sample library; Collect static attribute data and dynamic monitoring data of individuals, and integrate the static attribute data, dynamic monitoring data and preset prompt words to construct large model input information. The static attribute data includes the individual's gender and age, and the dynamic monitoring data includes heart rate and blood oxygen saturation indicators collected in real time or periodically through smart wearable devices. Based on the input information of the large model, the system performs retrieval, matching, and semantic reasoning in the corresponding level of the solution sample library to generate a personalized health intervention plan that matches the individual's current cardiopulmonary function status and individual characteristics.
[0070] In this embodiment, after obtaining the prediction result of the individual cardiopulmonary function decline index, the corresponding graded intervention program sample library is called according to the cardiopulmonary function decline index, and individualized health intervention suggestions are generated by calling large model interfaces (such as deepseek, chatgpt, etc.). The specific implementation method is as follows.
[0071] The model takes an individual’s gender, age, and physiological data (including heart rate and blood oxygen saturation) collected in real time or periodically by intelligent monitoring devices, along with preset prompts (“Provide health advice based on physiological indicators”) as input. The large model searches and infers based on the sample library of solutions corresponding to the predicted decline index, and outputs health advice (including drug type, dosage and method of administration, health behavior guidance and follow-up management suggestions) that match the individual’s current cardiopulmonary function status.
[0072] As the individual's cardiopulmonary function decline index changes dynamically, the large model adaptively matches and optimizes the intervention plan, thereby providing individuals with personalized and phased prevention and intervention guidance, effectively reducing the risk of cardiopulmonary function decline.
[0073] Furthermore, the steps for constructing the tiered intervention program sample library include: Collect historical medical data and extract historical health advice and its corresponding individual information; Based on the cardiopulmonary function decline index of individuals at corresponding time points in historical medical data, each historical health recommendation and its corresponding individual information are labeled with an index level to determine the cardiopulmonary function decline index level to which each historical health recommendation belongs. Based on the labeled cardiopulmonary function decline index level, all historical health recommendations belonging to the same index level and their corresponding individual information are collected and integrated to form an intervention program sample library corresponding to the cardiopulmonary function decline index of each index level.
[0074] In this embodiment, the collected medical data is used to construct a tiered intervention protocol sample library. The data includes an individual's gender, age, heart rate, blood oxygen saturation, and corresponding health recommendations. These health recommendations include medication type, dosage and method of administration, health behavior guidance, and follow-up management recommendations. The samples are grouped according to the cardiopulmonary function decline index, with samples corresponding to the same decline index grouped into the same protocol sample library. This forms a tiered intervention protocol sample library with a one-to-one correspondence to the cardiopulmonary function decline index, resulting in a total of four protocol sample libraries.
[0075] Specifically, historical medical data includes hospital patient medical records and records from health management institutions. After collecting a sufficient amount of historical medical data, it is meticulously sorted and extracted. The focus is on extracting historical health advice and corresponding individual information. Historical health advice covers various suggestions given by doctors regarding an individual's cardiopulmonary function, such as dietary adjustments, exercise plans, and guidance on sleep schedules. Individual information includes a person's basic information, medical history, physiological indicators, and other aspects. Through this step, the raw data materials for constructing the sample database are obtained.
[0076] After extracting historical health recommendations and their corresponding individual information, each historical health recommendation and its corresponding individual information are labeled with an index level based on the individual's cardiopulmonary function decline index at the corresponding time point in the historical medical data. Specifically, it is necessary to clarify the cardiopulmonary function decline index of each individual when receiving the health recommendation, and determine the cardiopulmonary function decline index level to which each historical health recommendation belongs based on a pre-defined cardiopulmonary function decline index grading standard. For example, if the cardiopulmonary function decline index is divided into 1-4 levels, when an individual's cardiopulmonary function decline index at a certain time point is assessed as level 2, then the corresponding historical health recommendation is labeled as level 2. This step provides a clear classification basis for subsequent sample collection.
[0077] Based on the labeled cardiopulmonary function decline index levels, samples were collected and integrated. All historical health recommendations belonging to the same index level and their corresponding individual information were centrally organized. For example, all historical health recommendations and their individual information labeled as Level 1 cardiopulmonary function decline were collected together; similarly, a similar process was performed for Levels 2, 3, 4, and so on. In this way, an intervention protocol sample library corresponding to each cardiopulmonary function decline index level was formed. These sample libraries contain rich historical health recommendations and individual information, providing ample reference samples for subsequently generating personalized health intervention plans based on real-time cardiopulmonary function decline indices.
[0078] This embodiment can systematically construct a sample library of graded intervention programs, providing strong data support for personalized intervention for cardiopulmonary function decline, and helping to improve the pertinence and effectiveness of health interventions.
[0079] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application.
[0080] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.
[0081] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.
Claims
1. A system for assessing cardiopulmonary function decline, characterized in that, The system specifically includes: The data acquisition module is used to collect multi-dimensional physiological indicator data of individuals and preprocess the multi-dimensional physiological indicator data; The index construction module is used to classify the cardiopulmonary function decline status of individuals based on preprocessed multi-dimensional physiological index data and to construct a cardiopulmonary function decline index using an unsupervised clustering analysis model. The model building module is used to build a cardiopulmonary function decline index prediction model based on multi-source clinical data. It collects individual physiological data in real time or periodically through wearable devices and inputs it into the cardiopulmonary function decline index prediction model, and outputs the individual's current cardiopulmonary function decline index. The plan generation module is used to generate personalized health intervention plans based on the output cardiopulmonary function decline index and a pre-set tiered intervention plan sample library through a large model, and adaptively optimize the health intervention plan as the index changes.
2. The system according to claim 1, characterized in that, The index construction module is specifically used for: Based on the preprocessed multidimensional physiological index data, ejection fraction (cardiac function index), forced expiratory volume in one second (fEV1) to forced vital capacity (fEV1) and maximum mid-expiratory flow rate (MPV) were selected as core feature parameters to characterize the state of cardiopulmonary function decline. The selected core feature parameters are input into the K-means clustering model, and clustering iteration is performed with a preset number of candidate categories to generate clustering results under different category divisions; Based on the stability and clinical interpretability of the clustering results, the optimal number of categories is selected from the candidate categories, and individuals are divided into the corresponding number of cardiopulmonary function decline status categories. The event rate of patients with cardiopulmonary function decline in each category was statistically analyzed. Each decline category was assigned a value in order of increasing event rate, forming a cardiopulmonary function decline index grading system from low to high, representing good cardiopulmonary function to severe decline.
3. The system according to claim 2, characterized in that, Based on the stability and clinical interpretability of the clustering results, the optimal number of categories is selected from the candidate categories, and individuals are divided into the corresponding number of cardiopulmonary function impairment categories, specifically including: Based on the continuous characteristics of the progressive decline in cardiopulmonary function over time and the clinical grading method, multiple candidate clustering categories are set. Cluster analysis is performed for each number of candidate categories to evaluate the stability of the clustering results under different numbers of candidate categories and obtain cluster stability evaluation results. The stability evaluation includes determining whether similar samples are continuously classified into the same category and the changes in cluster centers under different numbers of categories. Based on the clustering results, the differences in core feature parameters and clinical interpretability among different candidate categories with different numbers of categories are evaluated to obtain clinical discrimination assessment results. Based on a comprehensive comparison of the cluster stability assessment results and the clinical discrimination assessment results, the number of target categories that meet the preset requirements for both cluster stability and clinical discrimination is selected as the optimal number of categories. Based on the optimal number of classifications, determine the category of cardiopulmonary function decline corresponding to each individual.
4. The system according to claim 1, characterized in that, The model building module is specifically used for: Collect multidimensional physiological index data of individuals at multiple time points in the order of examination time, construct time series samples with fixed time steps, and assign the supervised learning label of the corresponding time step to the cardiopulmonary function decline index corresponding to the end of the time window. Divide the data into training set and test set according to a preset ratio to form a sample dataset. Construct a predictive model for cardiopulmonary function decline index with time-series Transformer as the main structure; The sample dataset is input into the cardiopulmonary function decline index prediction model. The cross-entropy loss function is used for multi-class training. The model parameters are iteratively updated through the optimizer until the model converges, so that the classification probability distribution predicted by the model gradually approaches the true label distribution.
5. The system according to claim 4, characterized in that, The cardiopulmonary function decline index prediction model includes a fully connected layer, a multi-layer Transformer encoder, a global average pooling layer, and a fully connected classification layer. The fully connected layer is used to linearly transform the multidimensional physiological index data contained in each time step of the input sequence into a feature embedding vector with a unified dimension, and then fuses the feature embedding vector and the time position encoding element by element to form a model input sequence carrying temporal position information. The multi-layer Transformer encoder is used to extract deep temporal features from the model input sequence by sequentially stacking multiple Transformer encoders, thereby capturing the gradual evolution of cardiopulmonary function over time. The global average pooling layer is used to generate a global feature vector by averaging the temporal feature vectors of all time steps output by the last Transformer encoder layer over the time dimension. The fully connected classification layer is used to calculate the probability distribution of each category corresponding to the current cardiopulmonary function decline index of an individual based on the global feature vector through linear transformation and Softmax activation function, and the index level corresponding to the maximum probability is used as the final prediction result.
6. The system according to claim 5, characterized in that, Each Transformer encoder includes a multi-head self-attention module and a feedforward neural network module; The multi-head self-attention module is used to calculate the attention weights between time steps within the model input sequence through the multi-head self-attention mechanism, capture the long-term dependencies and interaction features between physiological indicators at different time points, and obtain the time-series features after weighted aggregation. The feedforward neural network module is used to perform deep transformation and mapping on the weighted aggregated temporal features through multi-layer nonlinear activation functions, and output a temporal feature vector.
7. The system according to claim 4, characterized in that, The process of inputting the sample dataset into the cardiopulmonary function decline index prediction model, using the cross-entropy loss function for multi-class training, and iteratively updating the model parameters through an optimizer until the model converges specifically includes: A cross-entropy loss function is constructed to measure the difference between the model's predicted probability distribution and the true label distribution. The cross-entropy loss function is defined as the weighted negative sum of the logarithm of the predicted probability of each category output by the cardiopulmonary function decline index prediction model and the true label. The true label is used to represent the true cardiopulmonary function decline index category to which the sample belongs in one-hot encoding form. The sample dataset is input into the cardiopulmonary function decline index prediction model. After forward propagation calculation in each network layer, the predicted probability distribution of the sample belonging to various cardiopulmonary function decline indices is output. The predicted probability distribution is then substituted into the cross-entropy loss function to calculate the loss value of the current batch. Based on the calculated loss value, the gradient information of each network parameter of the cardiopulmonary function decline index prediction model with respect to the loss value is calculated layer by layer through the backpropagation algorithm, and the error signal is propagated from the output layer to the input layer. The preset Adam optimizer is invoked, and the model network parameters are iteratively updated based on gradient information, momentum estimation, and adaptive learning rate adjustment strategies to gradually reduce the loss function value.
8. The system according to claim 1, characterized in that, The scheme generation module is specifically used for: Based on the cardiopulmonary function decline index, the corresponding level of the intervention program sample library is matched and called from the pre-set graded intervention program sample library; Collect static attribute data and dynamic monitoring data of individuals, and integrate the static attribute data, dynamic monitoring data and preset prompt words to construct large model input information. The static attribute data includes the individual's gender and age, and the dynamic monitoring data includes heart rate and blood oxygen saturation indicators collected in real time or periodically through smart wearable devices. Based on the input information of the large model, the system performs retrieval, matching, and semantic reasoning in the corresponding level of the solution sample library to generate a personalized health intervention plan that matches the individual's current cardiopulmonary function status and individual characteristics.
9. The system according to claim 8, characterized in that, The steps for constructing the sample library of the tiered intervention program include: Collect historical medical data and extract historical health advice and its corresponding individual information; Based on the cardiopulmonary function decline index of individuals at corresponding time points in historical medical data, each historical health recommendation and its corresponding individual information are labeled with an index level to determine the cardiopulmonary function decline index level to which each historical health recommendation belongs. Based on the labeled cardiopulmonary function decline index level, all historical health recommendations belonging to the same index level and their corresponding individual information are collected and integrated to form an intervention program sample library corresponding to the cardiopulmonary function decline index of each index level.