Human system emergent property measurement method integrating ensemble learning and scoring card

Abstract

The application discloses a kind of human system overall emergent property measurement method of fusion integrated learning and score card, specifically include: (1) index system construction based on analytic hierarchy process, (2) abnormal feature extraction of time series signal, (3) physical examination dataset feature dimension expansion, (4) human system overall emergent property measurement method based on "integrated learning-score card": for the index data obtained after the above processing, integrated learning model is constructed to process ten subsystems respectively, to obtain the health status of subsystem and the contribution weight to overall health.For the multi-source heterogeneous data such as physical examination static index and time series physiological signal generated by wearable device, a health evaluation scheme combining static and dynamic indicators is proposed;Under the realistic condition that the physical examination data is incomplete, the rationality of feature generation is ensured, and the prediction performance of the model and the interpretability of the score result are also considered, which is convenient for actual health management and risk stratification application.

Description

Technical Field

[0001] This invention relates to the field of medical health monitoring and assessment technology, specifically to the interdisciplinary technologies of medical big data analysis, wearable device time-series physiological signal processing, and artificial intelligence health risk assessment, and in particular to a method for measuring the overall emergent nature of the human system based on ensemble learning and scorecard models for multi-source heterogeneous health data. Background Technology

[0002] With the continuous advancement of modern medicine and technology, health monitoring and assessment methods are gradually evolving from single-indicator approaches to multi-dimensional and comprehensive methods. Traditional health assessment methods typically rely on single or a small number of physiological indicators, such as blood pressure, blood sugar, and cholesterol. While these methods have a certain degree of accuracy in diagnosing specific conditions, they often fail to comprehensively and accurately reflect an individual's overall health status due to a lack of holistic understanding of human health. These methods neglect the complex systemic nature of health and fail to fully capture its emergent characteristics—that is, health status depends not only on the independent performance of individual indicators but also on the interaction and coordination between various physiological systems. Therefore, traditional health assessment methods struggle to comprehensively and accurately reflect an individual's overall health status.

[0003] To address the above issues, this research adopts a holistic emergent perspective, utilizing hospital big data resources and combining them with machine learning techniques to establish a comprehensive health measurement model. This model treats human health as a dynamic and complex system, where the interactions and global characteristics among various indicators have a crucial impact on health status. The construction of this holistic emergent model not only helps to grasp health status holistically but also reveals potential health trends or risks from multiple individual health indicators, providing more accurate and comprehensive health assessments for individuals and medical institutions.

[0004] Machine learning technology has significant advantages in processing complex and massive amounts of data, efficiently extracting potential patterns and correlations, and avoiding the biases and subjective prejudices that may arise from manual analysis. Meanwhile, a health scoring system based on big data can monitor health status in real time and provide personalized health risk assessments, promptly identifying potential health problems. This study uses machine learning to deeply mine and analyze large-scale physical examination data accumulated in hospitals and time-series physiological index data collected by wearable sensors, identifying key indicators highly correlated with overall health, constructing an emergent health model, and designing a health measurement and scoring system based on this model to achieve a comprehensive assessment of human health status. Summary of the Invention

[0005] Purpose of the Invention: This invention addresses the need for holistic assessment of multi-source heterogeneous health data by proposing an emergent measurement method for the human body system based on "ensemble learning-scorecard". Specifically, it first constructs a three-layer indicator system (target layer / criteria layer / sub-criteria layer) driven by the analytic hierarchy process (AHP). Based on static physical examination indicators, it introduces time-series physiological signals generated by wearable devices. Through patched self-attention and sequence autoencoding reconstruction errors, it extracts dynamic features such as abnormal frequency, duration, and amplitude of motion and rhythm signals, which are then fused with the static indicators at the sub-criteria layer. For the static physical examination indicator dataset, it further estimates the health probability of each physiological system using an ensemble learning model with decision trees as weak learners. Finally, the scorecard model maps the logarithmic odds to a unified health score, achieving a comprehensive, interpretable, quantifiable, and stratified measurement of individual health status.

[0006] Ensemble learning-scorecard-based methods for measuring the overall emergent nature of the human system include:

[0007] (1) Construction of an indicator system based on the analytic hierarchy process (AHP): Based on the AHP, overall health is divided into a target layer, a criterion layer, and a sub-criterion layer. The criterion layer is organized according to ten physiological systems, and the sub-criterion layer includes both static physical examination indicators and dynamic features such as abnormal frequency, duration, and amplitude extracted from time-series signals. Among them, static indicators are used to characterize an individual's basic health status, while time-series indicators are used to reflect changes in health status over a long period of time. Static indicators provide an individual's basic health structure, while time-series indicators reflect changes in health status over a long period of time. The combination of the two forms a comprehensive multi-level health assessment system.

[0008] (2) Extraction of abnormal features of time series signals: Based on the inherent features of motion signals and rhythmic physiological signals, the patching self-attention method is used to process motion signals and the sequence autoencoder reconstruction error method is used to process rhythmic physiological signals to identify abnormal behavior. Features such as abnormal frequency, duration, amplitude and stability are extracted to construct an individual-oriented abnormal behavior feature vector as a supplementary input to physical examination data.

[0009] (3) Extension of the feature dimensions of the physical examination dataset: Since a complete physical examination dataset may be lacking, it is necessary to extend the feature dimensions of the dataset. Different methods are used to generate data for numerical indicators and binary indicators respectively. By setting reasonable generation rules, it is ensured that the generation process conforms to statistical laws and health judgment standards.

[0010] (4) Based on the indicator data obtained after the above processing, an ensemble learning model is constructed to process the ten subsystems respectively, and the health status of the subsystems and their contribution weights to the overall health are obtained. The health status of the ten subsystems is fused according to the contribution weights and input into the scoring card model to obtain the overall health status score.

[0011] Furthermore, the construction of the indicator system based on the analytic hierarchy process in step (1) specifically includes:

[0012] (11) Based on the Analytic Hierarchy Process (AHP), the overall health assessment of the human body is divided into three levels:

[0013] 1) Target layer: Health status score of the human body.

[0014] 2) Criterion layer: Key factors affecting human health.

[0015] 3) Sub-criteria layer: specific health indicators.

[0016] (12) The key factors within the quasi-test layer are systematically classified according to the eight major physiological systems to determine the human health status included in the physical examination data, and the auditory system and visual system are added in addition, for a total of ten categories.

[0017] (13) The sub-criteria layer categorizes common indicators in the physical examination report according to their respective physiological systems, and introduces time-series health data collected by wearable devices and physiological sensors to dynamically supplement the health assessment of each system. Specifically, the sub-criteria layer adds corresponding feature vector dimensions to describe abnormal behaviors exhibited by individuals during continuous monitoring. These time-series indicators are transformed into structured parameters, such as abnormal frequency, duration, intensity, and stability, through the anomaly detection and feature extraction module, and then input into the hierarchical analysis framework along with static physical examination indicators.

[0018] The indicator system based on the analytic hierarchy process (AHP) can hierarchically model the human health status from the target layer, criterion layer, and sub-criterion layer, compared with existing technologies that only rely on a single physical examination indicator or simple weighting. It can also be systematically divided according to the ten major physiological systems, making the health assessment structure clearer, more comprehensive, and the results more interpretable.

[0019] Furthermore, the extraction of abnormal features from the time-series signal in step (2) specifically includes:

[0020] (21) Resample, denoise and standardize each channel, and slice to form fixed-length samples;

[0021] (22) Detecting abnormal behavior of motion signals. Motion signals usually manifest as rapid dynamic changes and sudden events, such as violent fluctuations in acceleration and attitude angle signals. By inputting motion signals into the model and learning the features of each window, significant time-varying changes or irregular patterns can be identified. When the feature value of a certain window exceeds the preset abnormal threshold, the window can be determined as an abnormal behavior window;

[0022] (23) Detect abnormal behavior of physiological signals. Physiological signals are usually characterized by periodicity and low-frequency fluctuations, such as regular fluctuations in heartbeat signals and slight fluctuations in blood oxygen signals. Input physiological signals into the model to obtain reconstructed signals and calculate the reconstruction error. The reconstruction error is defined as the error value between the original signal and the reconstructed signal. If the mean square error within the window is greater than the preset abnormal threshold, it is determined to be an abnormal behavior window.

[0023] (24) Extract features such as frequency, duration, amplitude and stability of abnormal behavior windows;

[0024] (25) Align the aggregated dynamic features to the most recent physical examination record according to the physical examination time, and output the individual-oriented dynamic health feature vector as a supplementary input to the static physical examination data for the sub-criteria layer.

[0025] Compared to existing technologies that rely solely on static physical examination data for evaluation, the extraction of abnormal features from time-series signals can integrate continuous monitoring information acquired by wearable devices and physiological sensors to extract dynamic features such as frequency, duration, amplitude, and stability of abnormalities, thereby improving the ability to identify short-term fluctuations, early abnormalities, and trend risks.

[0026] Furthermore, the expansion of the feature dimensions of the physical examination dataset in step (3) specifically includes:

[0027] (31) Methods for expanding numerical indicators: Numerical indicators are quantifiable indicators whose values ​​change continuously within a certain range, such as blood pressure, blood sugar, and heart rate. Assume that the distribution of indicators in healthy samples conforms to a normal distribution with a mean within the reference range, while the distribution of indicators in unhealthy samples conforms to a normal distribution with a mean deviating from the reference range. Expanding the original dataset by selecting healthy or unhealthy indicators based on the system health labels can fill in the missing feature dimensions in the original dataset.

[0028] (32) Extension method of binary indicators: Binary indicators are non-quantifiable discrete indicators with fixed values ​​of 0 or 1, such as whether one smokes or has a history of illness. Assume that the distribution of indicators of healthy and unhealthy samples follows a Bernoulli distribution. Select appropriate parameters to generate indicators based on the system health label status of the original dataset, and expand the missing feature dimensions in the original dataset.

[0029] The feature dimension expansion of the physical examination dataset, compared with the existing technology that has insufficient feature dimensions and many missing features, which leads to the limited expressive ability of the evaluation model, can supplement missing features by using different expansion methods for numerical and binary indicators, improve data integrity and sample usability, and enhance the model training effect and the stability of health assessment.

[0030] Furthermore, the method for measuring the overall emergent nature of the human system based on "ensemble learning-scoring card" in step (4) specifically includes:

[0031] (41) Use an ensemble learning model to calculate the health probability and contribution weight of the ten subsystems. The contribution weight measures the impact of the health status of the subsystem on the overall health status.

[0032] (42) Input the estimation results of the ensemble learning model into the scorecard model to obtain the overall health score.

[0033] The human system-wide emergent measurement method based on "ensemble learning-scoring card" can simultaneously provide the health probability, contribution weight and overall health score of each subsystem, compared with existing technologies that directly output a single classification result or a single probability value. This not only improves the overall assessment accuracy, but also enhances the quantitative ability and interpretability of the assessment results, making it easier for individual health management and risk warning.

[0034] Beneficial effects: Compared with the prior art, the significant advantages of this invention are:

[0035] 1. Under the framework of the Analytic Hierarchy Process (AHP), this invention integrates the static physical examination indicators of ten physiological systems with the time-series physiological signals collected by wearable devices into the same indicator system. The static indicators reflect the basic health level, while the dynamic indicators reflect the fluctuations and abnormalities over a period of time, thus upgrading the overall health assessment from "looking at a few indicators" to a comprehensive evaluation "by system and by level".

[0036] 2. This invention designs two anomaly identification methods for motion signals and rhythmic physiological signals, namely patched self-attention detection and sequence autoencoder reconstruction error. These methods transform the original waveforms, which were originally difficult to use directly, into features such as anomaly frequency, duration, amplitude, and stability that can be directly incorporated into the model. This effectively improves the utilization rate of time-series data and provides quantifiable dynamic supplementary information for physical examination data.

[0037] 3. This invention expands the feature dimensions of numerical and binary / non-numerical physical examination indicators by combining statistical distribution assumptions, and uses an ensemble learning model with decision tree as a weak learner and a scorecard model in series to output system-level health probability and overall health score. Under the realistic condition of incomplete physical examination data, it ensures the rationality of feature generation, while taking into account the predictive performance of the model and the interpretability of the scoring results, which is convenient for practical health management and risk stratification applications. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the human system overall emergent measurement method based on ensemble learning-scoring card according to the present invention. Detailed Implementation

[0039] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0040] Example: This example provides a method for measuring the overall emergent nature of the human system by integrating ensemble learning and scoring cards. Utilizing...

[0041] (1) Construction of an indicator system based on the analytic hierarchy process:

[0042] Based on the analytic hierarchy process (AHP), this study divides the overall health assessment of the human body into three levels:

[0043] Target layer: Health status score of the human body.

[0044] Criterion level: Key factors affecting human health.

[0045] Sub-criteria layer: specific health indicators.

[0046] The key factors within the quasi-test layer are systematically categorized according to the eight major physiological systems, which are used to classify the human health status included in the physical examination data. Additionally, the auditory and visual systems are added, resulting in a total of ten categories.

[0047] For the digestive system health assessment, eight indicators were selected: past medical history, internal medicine, surgery, eight items of tumor screening, gastric disease screening, 25 items of biochemical examination, blood routine test, and color Doppler ultrasound results of liver, gallbladder, pancreas and spleen.

[0048] For the assessment of musculoskeletal health, five indicators were selected: medical history, internal medicine, surgery, body mass index, and the feature vector of the time-series data covered.

[0049] For the assessment of circulatory system health, six indicators were selected: past medical history, internal medicine, surgery, 25 items of biochemical examination, blood routine test, and the feature vector of the time series data covered.

[0050] For respiratory health assessment, five indicators were selected: past medical history, internal medicine, chest CT scan, eight tumor-related items in physical examination, and the feature vector of the time-series data covered.

[0051] For the assessment of neurological health, three indicators were selected: past medical history, internal medicine, and eight items related to tumors detected during physical examination.

[0052] For the assessment of reproductive system health, five indicators were selected: past medical history, female reproductive organs, TCT cytology test, eight tumor markers in physical examination, and vaginal secretion test.

[0053] For the assessment of urinary system health, six indicators were selected: past medical history, internal medicine, color Doppler ultrasound of both kidneys, complete blood count, urinalysis, and 25 biochemical tests.

[0054] For the health assessment of the auditory system, one indicator was selected: hearing test.

[0055] For the assessment of visual system health, one indicator was selected: visual acuity test.

[0056] (2) Extraction of abnormal features of time-series signals: Due to the different dynamic characteristics of the two types of signals: motion signals change rapidly and have obvious abrupt changes, which are suitable for capturing short-term pattern deviations using Transformer; physiological signals change slowly and have strong rhythm, which are more suitable for modeling normal waveforms with autoencoders and identifying anomalies using reconstruction errors.

[0057] For motion signals, a Patch-Transformer-based timing detection method is used, employing accelerometers and angular velocity meters for fall detection. The signal consists of triaxial acceleration and triaxial angular velocity.

[0058] (1)

[0059] Where X is the acquired six-bit signal. These represent the accelerations along the x-axis, y-axis, and z-axis, respectively. These are the angular velocities in the x-axis, y-axis, and z-axis directions, respectively.

[0060] To ensure consistent sample duration, all samples are uniformly framed into a 12-second window. Long-sequence activities are segmented using a sliding window, while fall samples are downsampled to maintain the same length.

[0061] At the input end, the triaxial acceleration and triaxial angular velocity signals are divided into three-dimensional data blocks through a sliding window. , This represents the i-th sliding window block; C represents the number of input channels; n 2 This represents the length of a single flattened block. Local motion features are extracted using two CNN layers, flattened, and then concatenated to form a sequence feature map. F ND This represents a sequence feature map consisting of all blocks, where D represents the feature dimension of a single block, and N represents the total number of blocks.

[0062] To model the correlation between different convolutional channels, a multi-head self-attention mechanism is introduced. Let the input feature matrix be X, then the query matrix Q, key matrix K, and value matrix V are obtained from the input features through linear mapping, i.e.

[0063] (2)

[0064] in , , These are the learnable projection matrices used to generate the query, key, and value, respectively. The attention output is represented as...

[0065] (3)

[0066] in Given the dimension of the key vector, the encoding layer employs an L-layer stacked structure and introduces residual connections and layer normalization to avoid gradient degradation issues in deep training. The final output undergoes global average pooling and is then processed by a linear layer for fall and non-fall classification. (4)

[0067] Where GAP is a global average pooling layer, Linear is a fully connected layer, and X is a... L This represents the final high-level feature obtained after stacking the preceding L layers of encoders. This represents the sample-level global semantic features output by the global average pooling layer, where P is the final output judgment result. After determining that abnormal behavior has occurred within a certain window, the number of abnormal windows per unit time is counted, and the energy peak within the abnormal window describes the characteristics of the abnormal behavior.

[0068] For physiological signals, an anomaly detection method based on reconstruction error is adopted. The model selected is a Long Short-Term Memory Autoencoder (LSTM-AE). The model reconstructs normal physiological rhythms and uses the reconstruction error to determine anomalies. The reconstruction error is defined as:

[0069] (5)

[0070] in The original signal, For the reconstructed signal output by the model, when When the threshold is exceeded, it is determined that there is an abnormal physiological event in that time period, and the maximum and average reconstruction errors of the abnormal windows, as well as the number and proportion of abnormal windows, are calculated.

[0071] Finally, the detected abnormal window information is aggregated and standardized over time to form a dynamic health feature vector for each individual.

[0072] Compared with existing technologies that only use static physical examination data for evaluation, the above-mentioned method for extracting abnormal features of time-series signals can integrate continuous monitoring information obtained from wearable devices and physiological sensors to extract dynamic features such as abnormal frequency, duration, amplitude and stability, thereby improving the ability to identify short-term fluctuations, early abnormalities and trend risks.

[0073] (3) Extension of feature dimensions of physical examination dataset: The extension of feature dimensions of dataset in this embodiment mainly includes two aspects: generation of numerical indicators and generation of non-numerical indicators. Numerical indicators are generated based on normal distribution, and the specific rules are as follows.

[0074] Healthy sample generation:

[0075] For each metric Given its reference normal range [lower, upper], the formula for generating healthy samples is:

[0076] (6)

[0077] in, This represents a normal distribution with mean μ and variance σ.

[0078] Unhealthy sample generation: Unhealthy samples are generated based on the mean offset of healthy samples. If a sample is marked as unhealthy, a low or high value is selected, with each sample having a 50% probability of being selected.

[0079] Low-value samples follow the following distribution:

[0080] (7)

[0081] High-value samples follow the following distribution:

[0082] (8)

[0083] Non-numerical indicators are generated based on the Bernoulli distribution, with the following specific rules.

[0084]

[0085] in X represents the value of the i-th non-numerical physical examination indicator in the final generated sample. i ~Bernoulli(0.1) is a random variable that follows a Bernoulli distribution.

[0086] Compared to existing techniques that suffer from insufficient or missing feature dimensions, which limit the expressive power of evaluation models, the aforementioned method for expanding the feature dimensions of physical examination datasets can supplement missing features by using different expansion methods for numerical and binary indicators, thereby improving data integrity and sample availability, enhancing model training effectiveness, and improving the stability of health assessment.

[0087] (4) Emergent measurement method of human system based on “ensemble learning-scorecard”: The weak learner of the ensemble learning model in this embodiment is a decision tree, which can capture the nonlinear relationship and interaction effect between indicators.

[0088] After labeling the detection index data of the sub-criteria layer with health status and performing necessary preprocessing, the data is input into the learning module. The goal is to minimize the error between the output result and the label result.

[0089] (10)

[0090] (11)

[0091] Where L represents the total loss function of the entire model, used to measure the sum of the model's fitting error and complexity penalty on the training set; θ represents all the parameters to be learned by the model. i This represents the true label of the i-th sample. This represents the model's predicted output for the i-th sample. Let represent the loss function of the model for the i-th sample. It is the sum of the loss functions for n samples, used to measure the difference between the model and the true value. k It is the k-th decision tree. The regularization term Ω(f) is applied to all K trees. k The summation is performed, where K represents the number of decision trees, used to control model complexity. In equation (11), T represents the number of leaf nodes in the k-th tree; γ represents the penalty coefficient for the number of leaf nodes, used to suppress excessive tree structure complexity; w j Let λ represent the weight of the j-th leaf node; λ represents the L2 regularization coefficient of the leaf node weight, where L2 is the squared L2 norm of the parameter vector. This represents the sum of squared weights of all leaf nodes in the tree. This represents the L2 norm regularization penalty term for leaf weights, used to limit excessively large leaf weights.

[0092] For the t-th iteration, a new weak learner f is added based on the weak learners from the previous t-1 iterations. t To optimize the current model, let the predicted value of the model for sample i after the (t-1)th iteration be... Then, at the prediction point, the first and second derivatives of the loss function with respect to the predicted values ​​are defined as follows: (12) in, Indicates the true label, Let represent the model's prediction value for the i-th sample after the (t-1)-th iteration. and Let be the first-order gradient and the second-order gradient of sample i in the t-th iteration, respectively, used to characterize the local change trend and curvature information of the current sample with respect to the objective function. Based on the gradient statistics obtained by equation (13), the t-th weak learner learns by minimizing the following equation. (13)

[0093] in, For candidate weak learners f on sample x i The output value, This is a first-order term, reflecting the linear constraint of the sample gradient on the current weak learner output. This is a second-order term used to describe the local curvature properties of the objective function. This is the regularization term for the weak learner f, used to control model complexity and suppress overfitting.

[0094] When determining the tree structure of the weak learner in round t, a greedy strategy is used to traverse the candidate split points and calculate the split gain: (14)

[0095] Where D is the set of samples contained in the current node to be split. and These are the sample sets assigned to the left and right child nodes of the current node to be split after candidate partitioning; and Let represent the first and second gradients corresponding to sample i in the t-th iteration; γ represents the penalty coefficient for the number of leaf nodes, and λ represents the L2 regularization coefficient for the weights of the leaf nodes. By calculating the gain corresponding to each candidate split, the splitting method with the largest gain is selected to generate the left and right child nodes; when the gain is less than a preset threshold or the maximum depth is reached, the splitting stops. After the new tree is built, it is added to the model as the weak learner built in this iteration. Finally, the output of the entire model is the weighted sum of all weak learners.

[0096] Define benchmark score Calculated with double odds points (PDO) The following equation holds true: (15)

[0097] Parameter A controls the baseline for scoring, and parameter B controls the range of scoring. As the baseline score, Let be the initial logarithmic odds, representing the baseline unhealthy proportion of the dataset. PDO refers to the double odds point, indicating the fractional change required to double the probability. Solving this equation yields the values ​​of the two parameters:

[0098] (16)

[0099] Since the overall health status needs to be evaluated by combining the health status of each subsystem, let p be the health probability output by the i-th subsystem. i Then odds are determined by the following formula:

[0100] (17)

[0101] Where w i Let represent the proportion of the i-th subsystem's impact on overall health, and odds represent the probability of being unhealthy, which is the output of the ensemble learning model.

[0102] The human health score output by the rating card model can be calculated using the following formula:

[0103] (18)

[0104] The human system overall emergent measurement method based on "ensemble learning-scorecard" adopts an overall modeling approach of "ensemble learning + scorecard". It can effectively characterize the nonlinear relationship and interaction effect between various detection indicators by using decision tree ensemble model, thereby improving the accuracy and robustness of the overall health status assessment of the human system. At the same time, it further maps the model output into an interpretable health score, realizing the quantitative expression and intuitive presentation of the assessment results. Thus, it takes into account the predictive performance, stability and interpretability, and has better practical application value.

[0105] The above description is merely a preferred embodiment of the present invention and should not be construed as limiting the scope of the invention. Therefore, any equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A method for measuring the overall emergent nature of the human system by integrating ensemble learning and rating cards, characterized in that, The method includes: (1) Construct an indicator system based on the analytic hierarchy process (AHP). According to the AHP, overall health is divided into a target layer, a criterion layer and a sub-criterion layer. The criterion layer is organized according to ten physiological systems. The sub-criterion layer includes both static physical examination indicators and dynamic features of abnormal frequency, duration and amplitude extracted from time-series signals. Static indicators are used to characterize an individual's basic health status, while time-series indicators are used to reflect changes in health status over a long period of time. The combination of the two forms a comprehensive multi-level health assessment system. (2) Extract abnormal features of time-series signals. Based on the inherent features of motion signals and rhythmic physiological signals, the patching self-attention method is used to process motion signals and the sequence autoencoder reconstruction error method is used to process rhythmic physiological signals to identify abnormal behavior. Abnormal frequency, duration, amplitude and stability features are extracted and an individual-oriented abnormal behavior feature vector is constructed as a supplementary input to the physical examination data. (3) Expand the feature dimensions of the physical examination dataset. Use different methods to generate data for numerical and non-numerical indicators respectively. By setting reasonable generation rules, ensure that the data conforms to statistical laws and health judgment standards during the generation process. (4) For the processed index data, an integrated learning model is constructed to process the ten subsystems respectively, and the health status of the subsystems and their contribution weights to the overall health are obtained. The health status of the ten subsystems is fused according to the contribution weights and input into the scoring card model to obtain the overall health status score.

2. The method for measuring the overall emergent nature of the human system by integrating ensemble learning and scoring cards as described in claim 1, characterized in that: Step (2) specifically includes (21) Resample, denoise and standardize each channel, and slice to form fixed-length samples; (22) Detect abnormal behavior of motion signals. Motion signals are characterized by rapid dynamic changes and sudden events, including violent fluctuations in acceleration and attitude angle signals. By inputting motion signals into the model and learning the features of each window, significant time-period changes or irregular patterns can be identified. When the feature value of a certain window exceeds the preset abnormal threshold, the window can be determined to be an abnormal behavior window. (23) Detect abnormal behavior of physiological signals. The physiological signals are characterized by periodicity and low-frequency fluctuations, including regular fluctuations of heartbeat signals and slight fluctuations of blood oxygen signals. Input physiological signals into the model to obtain reconstructed signals and calculate the reconstruction error. The reconstruction error is defined as the error value between the original signal and the reconstructed signal. If the mean square error within the window is greater than the preset abnormal threshold, it is determined to be an abnormal behavior window. (24) Extract the abnormal frequency, duration, amplitude and stability features from the abnormal behavior window; (25) Align the aggregated dynamic features to the most recent physical examination record according to the physical examination time, and output the individual-oriented dynamic health feature vector as a supplementary input to the static physical examination data for the sub-criteria layer.

3. The method for measuring the overall emergent nature of the human system by integrating ensemble learning and scoring cards as described in claim 2, characterized in that: Step (3) specifically includes: (31) Methods for expanding numerical indicators: Numerical indicators are quantified indicators whose values ​​change continuously within a certain range, including blood pressure, blood sugar, and heart rate. Assume that the distribution of indicators in healthy samples conforms to a normal distribution with the mean within the reference range, and the distribution of indicators in unhealthy samples conforms to a normal distribution with the mean deviating from the reference range. Based on the system health labels of the original dataset, methods for expanding healthy or unhealthy indicators are selected to expand the missing feature dimensions in the original dataset. (32) Methods for extending non-numerical indicators: Non-numerical indicators are indicators that cannot be quantified and take discrete values. Assume that the distribution of indicators of healthy and unhealthy samples conforms to the Bernoulli distribution. Select appropriate parameters to generate indicators based on the system health label status of the original dataset, and expand the missing feature dimensions in the original dataset.

4. The method for measuring the overall emergent nature of the human system by integrating ensemble learning and scoring cards as described in claim 3, characterized in that: Step (4) specifically includes: (41) The health probability and contribution weight of the ten subsystems are calculated using an ensemble learning model. The contribution weight measures the impact of the health status of the system on the overall health status. Then, the health probabilities of each subsystem are fused according to their weights and input into the scoring card model to obtain the final overall health score.

5. The method for measuring the overall emergent nature of the human system by integrating ensemble learning and scoring cards as described in claim 1, characterized in that: In step (2) of extracting abnormal features from time-series signals, the two types of signals have different dynamic characteristics: motion signals change rapidly and have obvious abrupt changes, making them suitable for capturing short-term pattern deviations using Transformer; physiological signals change smoothly and have strong rhythmicity, making them suitable for modeling normal waveforms using autoencoders and identifying anomalies using reconstruction errors. For motion signals, a Patch-Transformer-based timing detection method is used, employing accelerometers and angular velocity meters for fall detection. The signal consists of triaxial acceleration and triaxial angular velocity. (1) Where X is the acquired six-bit signal. These represent the accelerations along the x-axis, y-axis, and z-axis, respectively. These represent the angular velocities along the x-axis, y-axis, and z-axis, respectively. To ensure consistent sample duration, all samples are framed into a 12-second window. Long activity sequences are segmented using a sliding window, while fall samples are downsampled to maintain the same length. At the input end, the triaxial acceleration and triaxial angular velocity signals are divided into three-dimensional data blocks through a sliding window. , This represents the i-th sliding window block; C represents the number of input channels; n 2 The length of a single block after flattening is represented by the sequence feature map, which is obtained by extracting local motion features through two layers of CNN, flattening, and then concatenating them. F ND This represents the sequence feature map consisting of all blocks, where D represents the feature dimension of a single block, and N represents the total number of blocks. To model the correlation between different convolutional channels, a multi-head self-attention mechanism is introduced. Let the input feature matrix be X, then the query matrix Q, key matrix K, and value matrix V are obtained from the input features through linear mapping, i.e. （2）， in, , , These are the learnable projection matrices used to generate the query, key, and value, respectively. The attention output is represented as... （3） in Given the dimension of the key vector, the encoding layer employs an L-layer stacked structure and introduces residual connections and layer normalization to avoid gradient degradation issues in deep training. The final output undergoes global average pooling and is then processed by a linear layer for fall and non-fall classification. （4） Where GAP is a global average pooling layer, Linear is a fully connected layer, and X is a... L This represents the final high-level feature obtained after stacking the preceding L layers of encoders. This represents the sample-level global semantic features output by the global average pooling layer, where P is the final output judgment result. After determining that abnormal behavior has occurred within a certain window, the number of abnormal windows is counted per unit time. The energy peak within the abnormal window describes the characteristics of the abnormal behavior. For physiological signals, an anomaly detection method based on reconstruction error is adopted. The model selected is a Long Short-Term Memory Autoencoder (LSTM-AE). The model reconstructs normal physiological rhythms and uses the reconstruction error to determine anomalies. The reconstruction error is defined as: （5） in The original signal, For the reconstructed signal output by the model, when When the threshold is exceeded, it is determined that there is an abnormal physiological event in that time period, and the maximum and average reconstruction errors of the abnormal windows, as well as the number and proportion of abnormal windows, are calculated. Finally, the detected abnormal window information is aggregated and standardized over time to form a dynamic health feature vector for each individual.

6. The method for measuring the overall emergent nature of the human system by integrating ensemble learning and scoring cards as described in claim 1, characterized in that: In step (3), the feature dimension expansion of the physical examination dataset includes two aspects: the generation of numerical indicators and the generation of non-numerical indicators. The numerical indicators are generated based on the normal distribution, and the specific rules are as follows. Healthy sample generation: For each metric Given its reference normal range [lower, upper], the formula for generating healthy samples is: (6) in, This represents a normal distribution with mean μ and variance σ. Unhealthy sample generation: Unhealthy samples are generated based on the mean offset of healthy samples. If a sample is marked as unhealthy, a low or high value is selected, with each sample having a 50% probability of being selected. Low-value samples follow the following distribution: (7) High-value samples follow the following distribution: (8) Non-numerical indicators are generated based on the Bernoulli distribution, with the following specific rules. in X represents the value of the i-th non-numerical physical examination indicator in the final generated sample. i ~Bernoulli(0.1) is a random variable that follows a Bernoulli distribution.

7. The method for measuring the overall emergent nature of the human system by integrating ensemble learning and scoring cards as described in claim 1, characterized in that: (4) The weak learner in the ensemble learning model is a decision tree, which captures the nonlinear relationships and interaction effects between indicators. After labeling the detection index data of the sub-criteria layer with health status and performing necessary preprocessing, the data is input into the learning module. The goal is to minimize the error between the output result and the label result. Where L represents the total loss function of the entire model, used to measure the sum of the model's fitting error and complexity penalty on the training set; θ represents all the parameters to be learned in the model, y i This represents the true label of the i-th sample. This represents the model's predicted output for the i-th sample. Let represent the loss function of the model for the i-th sample. It is the sum of the loss functions of n samples, used to measure the difference between the model and the true value, f k It is the k-th decision tree. The regularization term Ω(f) is applied to all K trees. k ) sum, where K represents the number of decision trees, used to control model complexity, T in equation (11) represents the number of leaf nodes of the k-th tree; γ represents the penalty coefficient for the number of leaf nodes, used to suppress excessive tree structure complexity; w j This shows the weight of the j-th leaf node; λ represents the L2 regularization coefficient of the leaf node weights, where L2 is the squared L2 norm of the parameter vector; This represents the sum of squared weights of all leaf nodes in the tree. This represents the L2 norm regularization penalty term for the leaf weights, used to limit the leaf weights from becoming too large. For the t-th iteration, a new weak learner f is added based on the weak learners from the previous t-1 iterations. t To optimize the current model, let the predicted value of the model for sample i after the (t-1)th iteration be... Then, at the prediction point, the first and second derivatives of the loss function with respect to the predicted values ​​are defined as follows: ， in, Indicates the true label, Let represent the model's prediction value for the i-th sample after the (t-1)-th iteration. and Let be the first-order gradient and the second-order gradient of sample i in the t-th iteration, respectively, used to characterize the local change trend and curvature information of the current sample with respect to the objective function. Based on the gradient statistics obtained by equation (13), the t-th weak learner learns by minimizing the following equation: ， in, For candidate weak learners f on sample x i The output value, This is a first-order term, reflecting the linear constraint of the sample gradient on the current weak learner output. This is a second-order term used to describe the local curvature properties of the objective function. This is the regularization term for the weak learner f, used to control model complexity and suppress overfitting. When determining the tree structure of the weak learner in round t, a greedy strategy is used to traverse the candidate split points and calculate the split gain: Where D is the set of samples contained in the current node to be split. and These are the sample sets assigned to the left and right child nodes of the current node to be split after candidate partitioning; and Let be the first-order gradient and the second-order gradient of sample i in the t-th iteration, respectively; γ represents the penalty coefficient for the number of leaf nodes, and λ represents the L2 regularization coefficient for the weights of the leaf nodes. By calculating the gain corresponding to each candidate split, the splitting method with the largest gain is selected to generate the left and right child nodes. When the gain is less than a preset threshold or the maximum depth is reached, the splitting stops. After the new tree is built, it is added to the model as the weak learner built in this iteration. Finally, the output of the entire model is the weighted sum of all weak learners. Define benchmark score Calculated with double odds points (PDO) The following equation holds true: Where parameter A is the baseline point for controlling the score, and B is the score range. As the baseline score, Let PDO be the initial logarithmic odds, representing the baseline unhealthy proportion of the dataset, and PDO be the double odds point, which represents the fractional change required to double the probability. Solving this equation yields the values ​​of the two parameters: Since the overall health status needs to be evaluated by combining the health status of each subsystem, consider the health probability output by the i-th subsystem as follows: Then odds are determined by the following formula: ， Where w i Let represent the proportion of the i-th subsystem's impact on overall health, and let odds represent the probability of being unhealthy, which is the output of the ensemble learning model. The human health score output by the rating card model can be calculated using the following formula: 。 8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the human system overall emergent measurement method that integrates ensemble learning and scoring cards as described in any one of claims 1 to 7.

9. A computer-readable storage medium storing computer instructions thereon, characterized in that, When executed by the processor, the computer instructions implement the human system overall emergent measurement method that integrates fusion learning and scoring cards as described in any one of claims 1-7.

Images