Physical examination data analysis method and system based on artificial intelligence

Abstract

The invention discloses a health examination data analysis method and system based on artificial intelligence, and the method comprises the steps: receiving multi-modal health examination original data of a user, and generating a multi-modal feature matrix fused with time-space correlation; inputting the multi-modal feature matrix into a lightweight dual-channel network to obtain a decoupled dual-modal feature group; constructing a hidden Markov chain based on the bimodal feature group, and generating a health state transition trajectory diagram with a timestamp; inputting the health state transition trajectory diagram into a discriminator of the generative adversarial network, and outputting a confidence score and a pathology trigger threshold of a high-risk node; and activating a rule engine according to a pathological trigger threshold, and generating a hierarchical health intervention instruction set in combination with individual living habit data. By using the embodiment of the invention, the physical examination data can be efficiently analyzed, and the accuracy of disease early warning and health intervention is optimized.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and in particular to an artificial intelligence-based health examination data analysis method and system. Background Art

[0002] With the multimodality of health examination data, traditional analysis methods struggle to effectively integrate structured reports with unstructured imaging data, resulting in problems such as a lack of spatiotemporal correlation and delayed detection of dynamic anomalies. Existing technologies typically employ single-modality analysis or simple feature splicing, resulting in inaccurate predictions of organ function degradation pathways and a lack of adversarial verification mechanisms against real clinical cases. Furthermore, health intervention programs are often triggered based on static thresholds, failing to consider the dynamic relationship between individual lifestyle habits and pathological development, resulting in insufficient timeliness and targeted interventions. Summary of the Invention

[0003] The purpose of the present invention is to provide a health examination data analysis method and system based on artificial intelligence to address the deficiencies in the existing technology, to efficiently analyze health examination data, and to optimize the accuracy of early disease warning and health intervention.

[0004] One embodiment of the present application provides a method for analyzing health examination data based on artificial intelligence, the method comprising: Receive the user's multimodal health checkup raw data, perform spatiotemporal alignment processing on the structured report data and unstructured medical imaging data according to the preset cross-modal alignment rules, and generate a multimodal feature matrix that integrates spatiotemporal correlations; Inputting the multimodal feature matrix into a lightweight dual-channel network, synchronously decoupling the static physiological baseline feature vector and the dynamic abnormal fluctuation feature vector, and obtaining a decoupled dual-modal feature group; A hidden Markov chain is constructed based on the bimodal feature set, and the organ function degradation path is calculated through the state transition probability to generate a health state transition trajectory diagram with a time stamp; Input the health state transition trajectory graph into the discriminator of the generative adversarial network, perform adversarial matching with the real clinical case database, and output the confidence score and pathology trigger threshold of the high-risk node; The rule engine is activated according to the pathological trigger threshold, and a hierarchical health intervention instruction set is generated in combination with individual lifestyle data.

[0005] Optionally, the receiving of the user's multimodal health checkup raw data, performing spatiotemporal alignment processing on the structured report data and the unstructured medical imaging data according to a preset cross-modal alignment rule, and generating a multimodal feature matrix integrating spatiotemporal correlations, includes: According to the timestamp of the DICOM header file of the medical image and the acquisition time of the physical examination report, the time axis is aligned through the dynamic time warping algorithm to generate a time synchronization tag set; Based on the time-synchronized label set, the organ anatomical landmarks in the image sequence are spatially mapped with the physiological indicators in the report, and a multi-resolution B-spline deformation field is used to generate an anatomical-physiological correlation matrix; According to the anatomical-physiological correlation matrix, Lagrangian interpolation constraints are imposed on the structured report data, and texture spectrum features of the image are extracted to output a spatiotemporal fusion feature matrix with unified dimension.

[0006] Optionally, the multimodal feature matrix is ​​input into a lightweight dual-channel network, and a static physiological baseline feature vector and a dynamic abnormal fluctuation feature vector are synchronously decoupled to obtain a decoupled dual-modal feature group, including: According to the entropy distribution of the spatiotemporal fusion feature matrix, the initial convolution kernel weights are assigned to the static channel and the dynamic channel respectively through the back propagation greedy algorithm; In the static channel, a sliding window mean filter is used to generate a baseline feature vector based on the normal fluctuation range of physiological indicators, and a steady-state feature tensor is output. The Mahalanobis distance between the current eigenvalue and the steady-state eigentensor is calculated in the dynamic channel, the excessive fluctuation pattern is captured through the gated recurrent unit, and the abnormal fluctuation eigenvector is output; Apply Gram-Schmidt orthogonalization to the steady-state eigentensor and the abnormal fluctuation eigenvector to eliminate the characteristic coupling interference and generate the initial decoupled characteristic group; The knowledge distillation technology is used to compress the dual-channel network to 1 / 8 the number of parameters, retaining the preset percentage of feature decoupling accuracy, and finally outputting the decoupled bimodal feature group.

[0007] Optionally, the step of constructing a hidden Markov chain based on the bimodal feature group, calculating the organ function degradation path through state transition probability, and generating a health state transition trajectory diagram with a timestamp includes: The bimodal feature group is divided into multiple health status dimensions according to organ system, and a discrete latent state set is generated through K-means clustering; Based on the historical clinical follow-up data set, Bayesian inference is used to calculate the transition probability between each latent state and construct the organ state transition matrix; performing constrained Markov chain Monte Carlo sampling according to the organ state transition matrix to generate multiple possible state transition paths; Eliminate the paths in the state transition path that have not experienced abnormal states for a preset number of consecutive times, and retain the path subset containing risk transfer nodes; Map the path subset to the physical examination timeline, predict the time point of each state transition through exponential smoothing, and output the health state transition trajectory diagram with timestamps.

[0008] Optionally, the health state transition trajectory graph is input into the discriminator of the generative adversarial network, and adversarially matched with a real clinical case database to output the confidence score and pathology trigger threshold of the high-risk node, including: The diagnostic records in the real clinical case database are converted into case feature tensors through ICD-11 coding; Cut the health state transition trajectory graph by quarterly time window to obtain a trajectory slice sequence; The trajectory slices are input into the GAN discriminator, and the cosine similarity is matched with the case feature tensor to output the authenticity score of each slice; When the authenticity score is lower than the preset score threshold, it is marked as a high-risk node, and the confidence level is calculated based on the proportion of similar nodes in the real clinical case database; According to the pathological progression rate distribution of high-risk nodes in the real clinical case database, the pathological trigger threshold curve is fitted.

[0009] Optionally, activating a rule engine according to the pathological trigger threshold and generating a hierarchical health intervention instruction set in combination with individual lifestyle data may include: Correlate and analyze pathological trigger thresholds with user lifestyle data, and identify key risk factors through a decision tree. Retrieve intervention templates matching risk factors from medical guideline-related databases to generate initial draft instructions; Convert the instruction strength level in the initial instruction draft according to the confidence level, where high confidence corresponds to mandatory instruction strength and medium confidence corresponds to recommended instruction strength; Instructions are sorted in reverse order of risk node timestamps to generate a hierarchical health intervention instruction set with execution time windows.

[0010] Another embodiment of the present application provides a health examination data analysis system based on artificial intelligence, the system comprising: The receiving module is used to receive the user's multimodal health examination raw data, perform spatiotemporal alignment processing on the structured report data and the unstructured medical imaging data according to the preset cross-modal alignment rules, and generate a multimodal feature matrix that integrates spatiotemporal correlation; a decoupling module, configured to input the multimodal feature matrix into a lightweight dual-channel network, synchronously decouple a static physiological baseline feature vector and a dynamic abnormal fluctuation feature vector, and obtain a decoupled dual-modal feature group; A construction module is used to construct a hidden Markov chain based on the bimodal feature group, calculate the organ function degradation path through state transition probability, and generate a health state transition trajectory diagram with a time stamp; An adversarial module is used to input the health state transition trajectory graph into the discriminator of the generative adversarial network, perform adversarial matching with the real clinical case database, and output the confidence score and pathology trigger threshold of the high-risk node; A generation module is used to activate the rule engine according to the pathological trigger threshold and generate a layered health intervention instruction set in combination with individual lifestyle data.

[0011] Yet another embodiment of the present application provides a storage medium, wherein the storage medium stores a computer program, wherein the computer program is configured to execute any of the above methods when run.

[0012] Yet another embodiment of the present application provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute any of the above methods.

[0013] Compared with the existing technology, the present invention provides an artificial intelligence-based health examination data analysis method, which receives the user's multimodal health examination raw data and generates a multimodal feature matrix that integrates spatiotemporal correlation; inputs the multimodal feature matrix into a lightweight dual-channel network to obtain a decoupled bimodal feature group; constructs a hidden Markov chain based on the bimodal feature group to generate a health state transition trajectory graph with a timestamp; inputs the health state transition trajectory graph into the discriminator of the generative adversarial network to output the confidence score and pathological trigger threshold of the high-risk node; activates the rule engine according to the pathological trigger threshold, and generates a layered health intervention instruction set in combination with individual lifestyle data, thereby enabling efficient analysis of health examination data and optimizing the accuracy of early disease warning and health intervention. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 A hardware structure block diagram of a computer terminal for an artificial intelligence-based health examination data analysis method provided in an embodiment of the present invention; Figure 2 A flowchart of a health examination data analysis method based on artificial intelligence provided by an embodiment of the present invention; Figure 3 A schematic structural diagram of an artificial intelligence-based health examination data analysis system provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0015] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and are not to be construed as limiting the present invention.

[0016] The embodiment of the present invention first provides a health examination data analysis method based on artificial intelligence, which can be applied to electronic devices such as computer terminals, specifically ordinary computers.

[0017] The following describes it in detail by taking running on a computer terminal as an example. Figure 1 The hardware structure block diagram of a computer terminal for a health examination data analysis method based on artificial intelligence provided by an embodiment of the present invention. Figure 1 As shown, the computer device includes a processor, a memory, and a network interface connected via a system bus, wherein the memory may include a non-volatile storage medium and an internal memory.

[0018] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions that, when executed, enable the processor to perform any one of the artificial intelligence-based health examination data analysis methods.

[0019] The processor is used to provide computing and control capabilities and support the operation of the entire computer equipment.

[0020] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium. When the computer program is executed by the processor, the processor can execute any health examination data analysis method based on artificial intelligence.

[0021] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art will understand that Figure 1 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0022] It should be understood that the processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0023] See also Figure 2 , an embodiment of the present invention provides a health examination data analysis method based on artificial intelligence, which may include the following steps: S201, receiving the user's multimodal health checkup raw data, performing spatiotemporal alignment processing on the structured report data and the unstructured medical imaging data according to a preset cross-modal alignment rule, and generating a multimodal feature matrix integrating spatiotemporal correlation; Specifically, the time axis can be aligned using a dynamic time warping algorithm based on the timestamp of the DICOM header file of the medical image and the acquisition time of the physical examination report to generate a time synchronization tag set; Accurate extraction and verification of multi-source timestamps‌ The system first parses the DICOM header file of the medical image and extracts the key time parameters: Sequence acquisition time: parameter SeriesTime (hours, minutes, and seconds of sequence acquisition, such as 14:35:22.123); Image generation time: parameter AcquisitionTime (single image generation timestamp, millisecond accuracy); The parameter ReportCollectionTime (report collection time, e.g., 2023-05-21 14:40) is also extracted from the structured medical report. To address time format differences, a standardized conversion is performed: DICOM time is converted to the parameter UnixTimestamp (the number of seconds since January 1, 1970). Verification steps include detecting time reversal errors (e.g., triggering an alarm if the image time is > time_gap_max = 5 minutes later than the report time) and compensating for device clock drift (dynamically corrected using the parameter ClockDriftRate = 0.5 seconds / day).

[0024] Medical Adaptation of Dynamic Time Warping Algorithms Use the DTW algorithm to align the time axis: Constraint settings: Set the maximum time stretch ratio to MaxWarpRatio = 1.8 (allowing the image timeline to stretch 1.8 times); set the path slope constraint parameter to SlopeConstraint = 0.5-2.0 (avoiding excessive time distortion); Cost function design: Time difference cost: Cost_time = |T_img - T_report|; Event correlation cost: Assign a weight of parameter EventWeight = 3.0 to the ECG gated image.

[0025] Execution process: Construct the image time series {t_img1, t_img2,...} and the report time points {t_rep1, t_rep2,...}; calculate the cumulative cost matrix (size MatrixSize = M × N, where M / N is the sequence length); backtrack the optimal regularized path to generate the parameter SyncLabelPair (synchronous label pair, such as (CT_slice_23, BloodReport_202305211440)).

[0026] ‌Standardized output of time-synchronized tag sets‌ Generate a set of three types of labels, as shown in Table 1: Table 1 Tag Type Data content example Usage Image frame labels‌ CT_Series2_Frame105@14:35:22 Locating a specific image frame Report Entry Tags‌ Blood_GLU@14:40 Correlate blood glucose test results Synchronization relationship tags SyncLevel=B (A: precise synchronization, B: tolerance synchronization) Mark alignment quality Synchronization quality grading standards: Grade A: time difference ≤ 1 minute, number of spatially associated organs ≥ 3; Grade B: time difference ≤ 3 minutes, number of spatially associated organs ≥ 1; Grade C: synchronization failure requires manual verification.

[0027] Based on the time-synchronized label set, the organ anatomical landmarks in the image sequence are spatially mapped with the physiological indicators in the report, and a multi-resolution B-spline deformation field is used to generate an anatomical-physiological correlation matrix; Intelligent recognition and registration of anatomical landmarks Based on the image frames with time-synchronized label positioning, perform organ segmentation and landmark extraction: Deep Learning Segmentation: Use parameter OrganSegModel=U-Net3+ (improved 3D U-Net network) to segment organ regions.

[0028] Key point positioning: Liver: portal vein bifurcation point (parameter LandmarkID=L1); Heart: left ventricular apex point (parameter LandmarkID=H3); Kidney: renal hilum center point (parameter LandmarkID=K2); The positioning accuracy reaches the parameter LocError≤1.2mm (millimeter), and the corresponding physiological indicator values ​​are extracted from the physical examination report (such as liver-related parameter ALT=38U / L, heart-related parameter BNP=125pg / mL).

[0029] ‌Spatial transformation of multi-resolution B-spline deformation fields‌ Establish a spatial mapping model between anatomical points and physiological indicators: Control grid construction: Coarse resolution layer: grid size parameter GridSize_Coarse = 20mm (20 mm cube); fine resolution layer: grid size parameter GridSize_Fine = 5mm (5 mm cube).

[0030] Deformation field generation: Calculate the displacement vector of the landmark point: DisplacementVector=(Δx,Δy,Δz) (three-dimensional offset); solve the global deformation at the coarse grid layer and fit the parameters AffineTransform (affine transformation matrix); apply local corrections at the fine grid layer, using parameter B-SplineOrder=3 (cubic B-spline basis function).

[0031] Physiological indicator mapping: Physiological values ​​are assigned to corresponding spatial coordinates through the deformation field; for example, the liver region coordinate (x=102, y=87, z=45) is mapped to the value ALT=38U / L.

[0032] ‌Generation Rules for Anatomical-Physiological Correlation Matrix‌ The output dimension is a three-dimensional matrix with parameter MatrixDim=256×256×128 (length×width×depth): Matrix cell values: Organ internal unit: stores the corresponding physiological index value (such as Matrix

[120]

[80]

[40] =HDL-C 1.8mmol / L); organ boundary unit: stores the parameter BoundaryType=1 (1: soft tissue boundary, 2: bone boundary); empty area unit: fills the parameter NullValue=-1000.

[0033] Spatial resolution: Conventional CT data: parameter VoxelSize = 0.8 mm³ (0.8 mm cubic voxel); High-resolution MRI: parameter VoxelSize = 0.5 mm³; The matrix comes with metadata: organ volume (OrganVolume), physiological value maximum (MaxValue), and other statistics.

[0034] According to the anatomical-physiological correlation matrix, Lagrangian interpolation constraints are imposed on the structured report data, and texture spectrum features of the image are extracted to output a spatiotemporal fusion feature matrix with unified dimension.

[0035] ‌Missing Data Handling for Structured Reports‌ For missing items in the physical examination report (such as some hospitals did not test the parameter Lp(a)), perform interpolation compensation: Interpolation constraints: Spatial constraints: weighting of existing indicators in the same organ region (weight w_dist = 1 / d², d is the spatial distance); Temporal constraints: linear extrapolation of data from adjacent time points.

[0036] Lagrange interpolation implementation: Set the interpolation node: select the known value point in the space sphere with radius parameter Radius=15mm; Construct interpolation polynomial: order parameter PolyOrder=2 (second-order polynomial); Calculate missing values: For example, based on liver area parameters AST=45U / L, GGT=32U / L, interpolate parameter ALT≈38U / L.

[0037] ‌Texture Spectrum Feature Extraction for Medical Images‌ Four types of texture features were extracted from the original image: gray-level co-occurrence matrix, wavelet packet energy, fractal dimension, and local binary pattern. The extraction window size parameter ROI_Size = 32×32 pixels was used, and the inter-layer fusion parameter MaxPooling_3D (three-dimensional maximum pooling) was used.

[0038] Generation of spatiotemporal fusion feature matrix The final output parameter FeatureMatrixSize=512×512×8 (length×width×feature layer) fusion matrix: Bottom layer (Layer 1-3): stores anatomical-physiological correlation matrix data (resolution downsampled to 1.6 mm); Middle layer (Layer 4-6): stores texture spectrum features (wavelet feature layer Layer 4, fractal feature layer Layer 5); ‌Top layer (Layer7-8): stores spatiotemporal metadata; Layer 7: timestamp (conversion parameter TimeEncode = UnixTimestamp / 1e6); Layer 8: spatial coding (parameter SpaceCode = organ ID × 1000 + region number).

[0039] The matrix comes with normalization rules: physiological values ​​are normalized to the parameter NormRange=[0,1], and image texture values ​​are normalized to the parameter Z-Score (standard score).

[0040] The method first integrates physical examination data from different testing devices, including numerical test reports and imaging images. Through timeline calibration algorithms and spatial registration technology, a precise correspondence is established between structured data such as blood indicators and anatomical structures in CT / MRI images, ensuring that physiological indicators and organ morphological changes collected at the same time can be correlated and analyzed. This breaks the data silos in traditional physical examinations and constructs a unified feature expression system in time and space. This fusion process can capture the potential correlation between abnormal biochemical indicators and changes in organ structure, providing high-quality input for subsequent in-depth analysis and avoiding misjudgments caused by data fragmentation.

[0041] S202, inputting the multimodal feature matrix into a lightweight dual-channel network, synchronously decoupling a static physiological baseline feature vector and a dynamic abnormal fluctuation feature vector, and obtaining a decoupled dual-modal feature group; Specifically, the initial convolution kernel weights can be assigned to the static channel and the dynamic channel respectively through the back propagation greedy algorithm according to the entropy distribution of the spatiotemporal fusion feature matrix; The system first calculates the entropy distribution of the spatiotemporal fusion feature matrix. Entropy (Entropy_Value) is used to quantify the degree of information chaos within the feature matrix. For example, after slicing the 128-dimensional feature vector according to the time window, the Shannon Entropy (Shannon_Entropy) of each slice is calculated. If a slice contains a sudden change in image texture (e.g., a sudden increase in the liver CT value from 50 HU to 90 HU) combined with abnormal blood sugar fluctuations (from 5.6 mmol / L to 11.2 mmol / L), its entropy value will be significantly higher than that of a slice with a stable physiological state. The entropy distribution results are stored as a histogram, with the horizontal axis divided into 10 entropy value intervals (e.g., 0-0.5, 0.5-1.0, etc.) and the vertical axis showing the proportion of slices in each interval.

[0042] Based on the entropy distribution, the convolution kernel weights are assigned using the Backpropagation Greedy Algorithm. The algorithm operates in two stages: Backpropagation: Build the initial architecture of a lightweight two-channel network, using a 3×3 convolution kernel for the static channel (Static_Channel) and a 5×5 convolution kernel for the dynamic channel (Dynamic_Channel). High-entropy slices (entropy > 1.2) are used as training samples, and feature importance scores are calculated backwards. For example, the temporal fluctuation feature importance score in the dynamic channel is 0.92, higher than the anatomical structure feature score of 0.75 in the static channel.

[0043] Greedy allocation phase: Initial weights are assigned in descending order of feature importance. Dynamic channels are prioritized for large convolution kernel weights (the weight of a 5×5 kernel is initialized to 0.35) to capture long-term anomalies. Static channels are assigned small convolution kernels (the weight of a 3×3 kernel is initialized to 0.28) to focus on local steady-state features. After allocation, weight normalization is performed to ensure that the sum of all channel weights is 1.0.

[0044] The final output is a dual-channel structure with initial weights. For example, in a patient's liver data, the convolution kernel weights of the static channel are concentrated in the uniform liver parenchyma (peak weight 0.41), while the dynamic channel weights are biased toward the area with varying portal vein blood flow velocity (peak weight 0.39). This step lays the foundation for the weight configuration for subsequent feature decoupling.

[0045] In the static channel, a sliding window mean filter is used to generate a baseline feature vector based on the normal fluctuation range of physiological indicators, and a steady-state feature tensor is output. The core of static channel processing is to extract steady-state features. The system uses sliding window mean filtering to smooth the input features: Dynamic Window Size Adjustment: Set the window length based on the organ's metabolic cycle. For example, liver metabolism indicators (such as ALT and AST) use a 30-day window (Window_Size=30days), while cardiovascular indicators (such as blood pressure) use a 7-day window. The default window sliding step (Sliding_Step) is 1 day to ensure continuous coverage.

[0046] Filtering: Take the arithmetic mean of all eigenvalues ​​within a window. For example, suppose a window contains 30 blood glucose measurements (in mmol / L): [5.1, 5.3, 5.0, ..., 6.2]. After filtering, the output is the window mean of 5.4 mmol / L. If there are missing values ​​within the window, interpolation between the previous and next windows is used to fill in the missing values.

[0047] When generating a baseline feature vector, it is necessary to consider the normal fluctuation range of physiological indicators: Individualized thresholds are loaded from a medical knowledge base. For example, the normal range of serum creatinine for a 60-year-old male user is 0.7-1.3 mg / dL. The system converts this range into boundary constraints for the feature vector.

[0048] The filtered results were corrected for out-of-bounds conditions. If the mean value in a window exceeded the normal range (e.g., a creatinine value of 1.5 mg / dL), the value was clipped to the upper limit of 1.3 mg / dL to ensure that the baseline reflected the steady-state level.

[0049] The final output Steady_State_Feature_Tensor is a three-dimensional structure: Dimension 1: time axis (aggregated by monthly granularity); Dimension 2: organ system (such as hepatobiliary system, cardiovascular system); Dimension 3: feature type (such as image texture mean, biochemical index baseline).

[0050] For example, the tensor element [2023-06, Liver, CT_Homogeneity] = 0.89, indicating that the liver CT homogeneity was at a stable high value in that month.

[0051] The Mahalanobis distance between the current eigenvalue and the steady-state eigentensor is calculated in the dynamic channel, the excessive fluctuation pattern is captured through the gated recurrent unit, and the abnormal fluctuation eigenvector is output; The dynamic channel specializes in anomaly detection, and the key technology is the Mahalanobis distance calculation: Extract reference distribution parameters from the steady-state feature tensor, including the mean vector and covariance matrix. For example, the mean resting heart rate of the cardiovascular system is 72 beats per minute (bpm), and the covariance of systolic blood pressure is 2.5.

[0052] Calculate the distance from the current eigenvalue to the reference distribution: Assume the current heart rate is 89 bpm (exceeding the steady-state mean of 72 bpm). The standard deviation of heart rate based on historical data is 8.2; Mahalanobis distance = |89-72| / 8.2 = 2.07 (dimensionless); Set the distance threshold: a Mahalanobis distance > 2.0 is considered a significant anomaly and triggers subsequent processing.

[0053] The Gated Recurrent Unit (GRU) is responsible for capturing the fluctuation pattern: Input layer: Receives time series segments with Mahalanobis distances exceeding the limit. For example, the Mahalanobis distance sequence of blood glucose for 5 consecutive days is: [1.8, 2.1, 3.0, 2.7, 2.3].

[0054] Memory Gate Mechanism: Identifies persistent anomalies. When the distance between three consecutive time points is greater than 2.0, the GRU's Update Gate weight is increased to 0.9 to strengthen memory of the current pattern.

[0055] Reset Gate Application: Filters out transient interference. If the anomaly only lasts for a single moment (such as a single heart rate spike), reset the gate to zero.

[0056] The output abnormal fluctuation feature vector contains three types of key information: Fluctuation range: such as the maximum deviation of blood glucose value +4.2 mmol / L; Fluctuation duration: if the duration of continuous abnormality = 120 hours; Organ correlation: For example, the liver and kidney abnormality synchronization index = 0.78; The vector dimension is fixed at 64 dimensions, covering all monitoring indicators.

[0057] Apply Gram-Schmidt orthogonalization to the steady-state eigentensor and the abnormal fluctuation eigenvector to eliminate the characteristic coupling interference and generate the initial decoupled characteristic group; To eliminate the redundant coupling between the dual-channel features, Gram-Schmidt orthogonalization is used: ‌Construct feature space basis vectors‌: Expand the steady-state feature tensor into a base vector group Base_Vectors_Static, for example, vector V1 = [liver CT value, blood glucose baseline]; The abnormal fluctuation feature vector group is Base_Vectors_Dynamic, for example, vector W1=[blood sugar fluctuation amplitude, heart rate variability].

[0058] ‌Perform an orthographic projection‌: Calculate the projection component (Projection_Component) of the dynamic vector on the static vector, for example, the projection value of W1 on V1 Proj=dot(W1,V1) / |V1|^2; Subtract this component from the original dynamic vector: W1_ortho = W1 - Proj × V1.

[0059] Normalization: Scale the orthogonalized vector to unit length (Unit_Length) to ensure that all features are in the same dimension.

[0060] This process addresses two types of coupled interference: Spatial coupling: For example, spurious correlations between liver image texture and blood glucose fluctuations (correlation coefficient decreased from 0.68 to 0.05); Temporal Coupling: Eliminates the effects of baseline drift on anomaly detection (e.g., false positives caused by morning blood pressure rhythm).

[0061] The generated Initial Decoupled Feature Set contains two independent matrices: Static matrix: 32-dimensional steady-state characteristics, each dimension represents the long-term stable state of the organ; Dynamic matrix: 32-dimensional abnormal features, each dimension describes short-term pathological fluctuations; The orthogonality between the two is verified by cosine similarity (Cosine_Similarity), with a target value of <0.1 (close to a vertical relationship).

[0062] The knowledge distillation technology is used to compress the dual-channel network to 1 / 8 the number of parameters, retaining the preset percentage of feature decoupling accuracy, and finally outputting the decoupled bimodal feature group.

[0063] Knowledge distillation technology enables lightweight models: ‌Build a teacher-student model‌: Teacher model (Teacher_Model): original two-channel network (8M parameters); Student model (Student_Model): streamlined architecture (target parameter size 1M).

[0064] Distillation process: The teacher model outputs "soft labels" (Soft_Labels): the class probability distribution of static features (e.g., probability of normal liver = 0.92, probability of fibrosis = 0.07); the student model learns to fit the soft labels rather than the original data hard labels (Hard_Labels); the temperature parameter (Temperature_Parameter) is set to T = 5 to increase the difference in probabilities between different classes.

[0065] Precision preservation mechanism during compression: Accuracy monitoring: After each round of compression, the feature decoupling accuracy is tested and the orthogonality loss value is calculated; Dynamic adjustment: If the accuracy drops below a preset threshold (Preset_Percentage=5%), the compression ratio will be regressed (e.g., from 1 / 8 to 1 / 6). Fine-tuning strategy: Targeted enhancement of training on abnormal sample sets (such as cancer patient data); The resulting bimodal feature set meets the following requirements: 1 / 8 the number of parameters (e.g., from 8 million to 1 million); feature decoupling accuracy ≥ 95% (e.g., correlation between static and dynamic features ≤ 0.05); and real-time processing speed increased by 400% (single analysis time reduced from 500ms to 125ms). This feature set serves as input to a hidden Markov chain to model organ degeneration pathways.

[0066] A unique neural network architecture is designed, with one channel extracting stable features reflecting an individual's long-term health baseline, while the other channel focuses on capturing short-term abnormal fluctuations. Orthogonalization constraints ensure the independence of these two feature types, and model compression techniques are employed to maintain computational efficiency. This achieves a "steady-state-transient" separation of health status, enabling both overall health assessment and identification of sudden abnormalities. This decoupling approach significantly increases sensitivity to early pathology while avoiding misclassification of normal physiological fluctuations as abnormalities.

[0067] S203, constructing a hidden Markov chain based on the bimodal feature group, calculating the organ function degradation path through state transition probability, and generating a health state transition trajectory diagram with a time stamp; Specifically, the bimodal feature group can be divided into multiple health status dimensions according to the organ system, and the discrete latent state set can be generated through K-means clustering; First, the input bimodal feature group is divided into organ systems. The system presets an organ classification mapping table (Organ_Class_Mapping), which divides the 128-dimensional feature vector into 8 organ subsystems according to anatomical location, including: Respiratory System (Respiratory_System), which includes 15 features: measured vital capacity, blood oxygen saturation, chest CT ground-glass opacity ratio, FEV1 / FVC ratio (rate per second), bronchial dilation index, etc.; example features include the average density value of the right upper lobe CT (range: -850 to -700 HU) and the fluctuation range of carbon dioxide partial pressure (PaCO2); Digestive System, which includes 14 features: pepsinogen I / II ratio, fecal occult blood test value, intestinal flora diversity index, abdominal ultrasound echo intensity distribution, etc.; example features include gastric antral mucosal fold thickness (ultrasound measurement) and pancreatic duct diameter change rate; Endocrine System, which includes 12 features: fasting blood glucose fluctuation coefficient, glycated hemoglobin HbA1c, thyroid nodule TI-RADS grade, cortisol circadian rhythm deviation, etc.; example features include pancreatic beta cell function index (HOMA-β) and TSH receptor antibody concentration; Nervous_System, which includes 10 features: EEG alpha wave power ratio, cognitive function score decline rate, MRI white matter high signal volume, peripheral nerve conduction velocity, etc.; example features include annual hippocampal volume atrophy rate (measured by MRI sequence), tremor frequency power spectrum peak, etc. ‌Musculoskeletal_System‌, which includes 9 features: bone density T-score, joint space width, dynamic changes in creatine kinase, body fat distribution asymmetry index, etc.; example features include average bone density of lumbar vertebrae L1-L4 (DXA scan) and quadriceps cross-sectional area (MRI measurement).

[0068] Each subsystem is individually subjected to the K-means clustering algorithm: Initialize cluster centers: Use the Max-Min Distance Method. For example, in the hepatobiliary system, select the three sample points with the farthest distance in the feature space as the initial centers: Center point 1 represents healthy status (e.g., ALT < 40 U / L and CT homogeneity > 0.85); Center point 2 represents mild abnormality (ALT 40-80 U / L); The center point 3 represents severe abnormality (ALT>80 U / L and presence of fatty liver markings).

[0069] ‌Iterative Optimization Process‌: Calculate the Euclidean distance (i.e., the straight-line distance in multidimensional space) from all sample points to each center. Assign samples to the cluster with the nearest center; Recalculate the sample mean within the cluster as the new center; Stop when the center point moves less than the threshold Converge_Threshold=0.01.

[0070] Finally, a discrete hidden state set is generated, and each subsystem outputs 3-5 state codes: Cardiovascular system: S1_healthy state, S2_arrhythmia risk state, S3_coronary ischemia state; hepatobiliary system: L1_normal metabolic state, L2_hepatocellular inflammation state, L3_fibrosis tendency state; the total number of states is fixed at 32, covering all organ subsystems.

[0071] Based on the historical clinical follow-up data set, Bayesian inference is used to calculate the transition probability between each latent state and construct the organ state transition matrix; The historical clinical follow-up dataset contains five-year physical examination series for 100,000 patients. The system processes the data in the following steps: Data alignment: Map each physical examination result to a set of hidden states. For example, the sequence of liver and gallbladder states for a patient's three physical examinations is [L1, L1, L2]. State transition frequency statistics: Calculate the number of L1→L1 transitions (e.g., 12,540 times); calculate the number of L1→L2 transitions (e.g., 893 times); set the statistical period to quarterly (90 days as the transition time window).

[0072] Bayesian inference is used to calculate transition probabilities: Prior probability setting: Use the basic transfer rate in medical guidelines. For example, the recommended value from L1 to L2 is 0.5% / year; Likelihood function construction: Based on the actual transfer frequency, define the existing binomial distribution likelihood function; Posterior probability calculation: Update the existing conjugate prior distribution and output the posterior transition probability value (e.g. 0.72%). The constructed organ state transition matrix is ​​a 32×32 square matrix: the row index (i) represents the starting state; the column index (j) represents the target state; the matrix element M ij Represents the probability of transitioning from state i to j; for example, in the matrix M L1,L2 =0.0072, indicating that the probability of transitioning from a healthy liver and gallbladder state to an inflammatory state within a quarter is 0.72%.

[0073] performing constrained Markov chain Monte Carlo sampling according to the organ state transition matrix to generate multiple possible state transition paths; Markov chain Monte Carlo sampling is used to simulate future state changes: Initialization path: Start with the current physical examination status. For example, the patient's current cardiovascular status is S1 (healthy state) and the liver and gallbladder status is L2 (inflammation state). State transition rule: Read the possible transition targets and corresponding probabilities of the current state from the transition matrix; use the roulette wheel selection method to randomly determine the next state.

[0074] Key constraints ensure medical rationality: One-way degradation constraint: prohibiting a direct jump from an abnormal state back to a healthy state (e.g., not allowing L2→L1); Organ coordination constraint: When the cardiovascular system enters S3 (ischemic state), the renal state must not be better than K2 (mild renal function decline); ‌Time continuity constraint‌: The interval between adjacent state transitions must be ≥ 30 days.

[0075] The sampling process is executed 1000 times to generate a state transition path example: Pathway 23: [(Month1: Cardiovascular S1, Hepatobiliary L2), (Month4: Cardiovascular S1, Hepatobiliary L3), (Month7: Cardiovascular S2, Hepatobiliary L3)]; Pathway 47: [(Month1: Cardiovascular S1, Hepatobiliary L2), (Month5: Cardiovascular S2, Hepatobiliary L2), (Month8: Cardiovascular S2, Hepatobiliary L3)]; ... (1000 paths in total).

[0076] Each path contains 3-5 state nodes, covering the next 12-month forecast period.

[0077] Eliminate the paths in the state transition path that have not experienced abnormal states for a preset number of consecutive times, and retain the path subset containing risk transfer nodes; The default number is defined as three consecutive physical examinations (9 months) without any abnormal conditions. System execution path filtering: Abnormal status definition: Cardiovascular system: S2 (arrhythmia) and S3 (ischemia); Hepatobiliary system: L2 (inflammation) and L3 (fibrosis); other systems are similar (a total of 18 abnormal status codes are defined).

[0078] Path scanning algorithm: Check each status node in the path one by one; mark the location where the abnormal status first appears (for example, path 23 appears in liver and gallbladder L3 in Month 4); Continuity judgment: If three consecutive nodes from the starting point of a path are in a healthy state (such as path 89: [S1 / L1, S1 / L1, S1 / L1]), it is classified as a low-risk path.

[0079] Path rejection rule: Automatically discard all low-risk paths (about 35% of the total sample number); only retain paths containing at least one abnormal node (for example, path 23 is retained because it contains L3).

[0080] The output path subset must meet the following requirements: include risk transfer nodes, such as the state transition from S1 to S2; label the risk type, such as progression of hepatobiliary fibrosis (L2→L3); and record the duration of abnormalities, such as continuous abnormal blood sugar for ≥6 months.

[0081] Map the path subset to the physical examination timeline, predict the time point of each state transition through exponential smoothing, and output the health state transition trajectory diagram with timestamps.

[0082] Operations mapped to the physical examination timeline: The last actual physical examination date is time zero (T0) Align the simulation nodes in the path by the sampling time offset: the path node "Month4" corresponds to T0+120 days; the path node "Month7" corresponds to T0+210 days.

[0083] Exponential smoothing forecasting optimizes time point accuracy: Weighting mechanism: Recent paths are given higher weights. For example, paths generated in the first 30 days have a weight of 0.6; paths generated between 31 and 60 days have a weight of 0.3; and paths generated between 61 and 90 days have a weight of 0.1.

[0084] Time point calculation: Assume that there are 200 pathways predicting the transition from cardiovascular state S1 to S2: 50 pathways occur at T0 + 90 ± 5 days; 100 pathways occur at T0 + 120 ± 8 days; and 50 pathways occur at T0 + 150 ± 6 days. The weighted average gives the most likely transition time: T0 + (90 × 0.6 + 120 × 0.3 + 150 × 0.1) = T0 + 108 days.

[0085] Finally, the health state transfer trajectory diagram is generated: The horizontal axis shows precise timestamps (e.g., 2024-03-15, 2024-06-22, etc.); the vertical axis shows organ subsystem status codes; node styles include: green circles (healthy) (e.g., S1 / L1); yellow triangles (mildly abnormal) (e.g., S2 / L2); and red diamonds (severely abnormal) (e.g., S3 / L3). Connecting arrows indicate the direction of state progression (e.g., L2 → L3). This chart visually illustrates the time series of degeneration risk for each organ system over the next 12 months, for example, by noting "2024-08-17 Hepatobiliary system may progress to L3 (confidence probability 72%)."

[0086] The health status of each organ system is modeled as a Markov state node, and a state transition model is trained using clinical big data. Probabilistic calculations are used to deduce the most likely path of health deterioration, and key turning points are marked. Visualizations of evolving health risk trends help doctors predict critical stages of disease progression. This time-series modeling can identify progressive lesions that are difficult to detect with conventional examinations, providing a time window for early intervention.

[0087] S204: Input the health state transition trajectory graph into the discriminator of the generative adversarial network, perform adversarial matching with the real clinical case database, and output the confidence score and pathology trigger threshold of the high-risk node; Specifically, the diagnostic records in the real clinical case database can be converted into case feature tensors through ICD-11 coding; The real clinical case database contains 500,000 anonymized patient data, each record contains diagnosis text, laboratory indicators, imaging reports, etc. The conversion process is divided into three stages: ICD-11 Code Mapping: Use natural language processing tools (such as clinical term parsers) to extract diagnostic keywords. For example, "acute anterior myocardial infarction" is mapped to the ICD-11 code "BA41.0" (a specific code under the myocardial infarction category). Multiple coding associations are performed for comorbidities. For example, when a diabetic patient has concurrent kidney disease, both "5A10" (diabetes code) and "GB60" (chronic kidney disease code) are marked at the same time.

[0088] ‌Eigenvectorization‌: Each ICD-11 code is converted into a 128-dimensional vector (Vector_Dimension_128, which uses 128 numbers to represent the disease characteristics). The clinical correlation between codes is reflected by the vector angle: Similar diseases (e.g., BA41.0 and BA41.1 inferior wall myocardial infarction) have a vector angle of <15 degrees; The angle between the vectors of unrelated diseases (e.g., BA41.0 and 2F30 lung cancer) is >75 degrees; Laboratory indicators (such as troponin I values) are normalized to the interval [0,1] and then appended to the vector.

[0089] Construct a three-dimensional tensor (Tensor_3D): Dimension 1: patient ID (anonymized number); Dimension 2: timeline (sorted by admission date); Dimension 3: feature layer (including ICD vector + indicator value, a total of 256 dimensions).

[0090] Finally, the case feature tensor (Case_Feature_Tensor) is generated with a size of [500,000 patients × 10 time points × 256 feature dimensions], occupying approximately 3.2TB (Terabyte) of storage space.

[0091] Cut the health state transition trajectory graph by quarterly time window to obtain a trajectory slice sequence; The health state transition trajectory diagram is a time-stamped organ state evolution diagram. The cutting operation is as follows: Time window division rules: The quarterly time window is fixed at 90 days. The sliding step is set to 30 days to ensure that adjacent slices have a 60-day overlap. The starting point of the cut is the current date, and the end point is the last timestamp of the forecast period.

[0092] Slice generation logic: A single slice covers a time range, for example, from 2024-01-01 to 2024-03-31. All state nodes and transition arrows within this period are extracted. If there is no state change within a slice (e.g., the cardiovascular S1 state is maintained throughout), it is marked as a steady-state slice.

[0093] Sequence assembly requirements: Each slice is converted into a 512×512 pixel (pixel, the smallest unit of the image) raster image; the image channel (Image_Channel) contains: red channel: abnormal status marker (such as liver and gallbladder L3 status displayed as dark red); green channel: healthy status marker (such as kidney K1 status displayed as bright green); blue channel: metastasis path arrow.

[0094] The final output trajectory slice sequence, for example, patient P123 generates [slice Q1-2024, slice Q2-2024...] a total of 8 images.

[0095] The trajectory slices are input into the GAN discriminator, and the cosine similarity is matched with the case feature tensor to output the authenticity score of each slice; The structure of the Generative Adversarial Network Discriminator is as follows: Dual Input Processing: Path A: Receives trajectory slice images and extracts spatial features through the convolutional layer (Convolutional_Layer, image feature extraction layer); Path B: Receives case feature tensors and extracts statistical features through the fully connected layer (Fully_Connected_Layer, multidimensional data processing layer).

[0096] ‌Cosine similarity matching‌: Map image features (1024-dimensional vectors) and case features (256-dimensional) to the same dimension; calculate the cosine of the angle between the vectors: the value range is [-1, 1], where 1 indicates complete similarity.

[0097] Example matching process: Input: A slice shows "2024Q2 Cardiovascular S2→S3 metastasis"; matching case features: ICD code BA41 (myocardial infarction) + troponin > 0.5 ng / mL; calculated cosine similarity = 0.87.

[0098] ‌Authenticity score conversion‌: The similarity value is converted to a score using the Sigmoid function.

[0099] Scoring threshold setting: 0.85: strong authenticity (scoring level A); 0.65-0.85: moderate authenticity (scoring level B); <0.65: questionable authenticity (scoring level C).

[0100] Output format: slice ID + score value (e.g. Slice_Q2_2024: 0.87 → score B).

[0101] When the authenticity score is lower than the preset score threshold, it is marked as a high-risk node, and the confidence level is calculated based on the proportion of similar nodes in the real clinical case database; The default rating threshold is fixed at 0.65: High-risk node marking: Scan all slice authenticity scores and filter slices with a score of 0.65 or less; locate specific abnormal conditions within the slice (e.g., "2024Q3 Hepatobiliary L3 status persisting for 90 days").

[0102] Generate high-risk node data structure: { "time_window": "2024-Q3", "organ_system": "Hepatobiliary system", "abnormal_state": "L3 fibrosis tendency state", "slice_score": 0.58 }.

[0103] Confidence calculation process: Search for nodes of the same class (Same_Class_Node) in the case database. The conditions include: same organ system (such as hepatobiliary system); same abnormal state (such as L3 state); same duration (such as 90±15 days); and count the total number of nodes of this class in the case database (such as 12,580 cases).

[0104] Calculate the proportion of high-risk nodes: Confidence = (Number of similar nodes / Total number of nodes in the case database) × 100%; Example: 12,580 / 3,200,000 = 0.393% → Confidence 39.3%.

[0105] Confidence level classification: high confidence: ≥60% (corresponding to high medical risk); medium confidence: 30%-60% (medium risk); low confidence: <30% (rare risk).

[0106] According to the pathological progression rate distribution of high-risk nodes in the real clinical case database, the pathological trigger threshold curve is fitted.

[0107] Pathological progression rate is defined as the rate of organ status deterioration: Case database data analysis: Extract subsequent evolution records of similar high-risk nodes (e.g., patients with hepatobiliary L3 status); Calculate the time to progression to the terminal state (such as cirrhosis ICD code DB90): Case A: 180 days from L3 to cirrhosis (just an assumption); Case B: 730 days from L3 to cirrhosis (just an assumption); generate a time distribution histogram with the horizontal axis as days and the vertical axis as the number of cases.

[0108] Curve fitting method: Use kernel density estimation; set the bandwidth parameter to 30 days (to control the smoothness of the curve); and fit the output probability density curve.

[0109] Trigger threshold delineation: The 25th percentile of the cumulative distribution of the density curve is taken as the warning threshold; example: if the progression time of 25% of patients is ≤200 days, then the threshold T_warning is set to 200 days; the 75th percentile is taken as the intervention threshold; example: if the progression time of 75% of patients is ≤500 days, then T_intervention is set to 500 days.

[0110] The final pathological trigger threshold curve contains two boundary lines: the orange dotted line: T_warning (recommended intervention line); the red solid line: T_intervention (mandatory intervention line); the horizontal axis of the curve is the duration of the current risk state, and the vertical axis is the cumulative probability of deterioration.

[0111] By intelligently comparing predicted health trajectories with the clinical course of confirmed patients, an adversarial training mechanism is used to identify pathologically significant feature nodes and quantify their risk level. This significantly improves the accuracy of risk warnings and prevents over-treatment. Confidence scores help doctors identify high-risk individuals who truly require intervention, while pathology trigger thresholds provide a scientific basis for developing personalized monitoring plans.

[0112] S205 , activating a rule engine according to the pathological triggering threshold, and generating a hierarchical health intervention instruction set in combination with individual lifestyle data.

[0113] Specifically, the pathology trigger threshold can be correlated with the user's lifestyle data and analyzed, and the key risk factors can be located through the decision tree; The pathological trigger threshold contains two key boundary lines: the warning threshold (Warning_Threshold, orange dotted line): reaching this line requires the initiation of recommended-level intervention; the intervention threshold (Intervention_Threshold, red solid line): reaching this line requires the initiation of mandatory-level intervention.

[0114] Association analysis is performed in three steps: Data fusion and matching: User lifestyle data comes from wearable devices and questionnaires, including 32 indicators such as sleep duration, average daily steps, and smoking index.

[0115] Align the time scale of the threshold curve with the lifestyle timeline. For example, when the user's hepatobiliary system risk is at the warning threshold (such as March 2024), extract the lifestyle data for that month; focus on the temporal relationship between alcohol intake and the breakthrough of the liver risk threshold.

[0116] Decision tree node construction: The root node (Root_Node) is set to the target pathology threshold (such as cardiovascular warning threshold breakthrough).

[0117] The branching rules are based on clinical research evidence: first-level branch: sleep quality score < 60 points (low-quality sleep) and lasting for more than 30 days; second-level branch: sodium intake > 5 g / day, and diastolic blood pressure fluctuation > 15 mmHg.

[0118] Leaf nodes (Leaf_Node) identify key factors: Example output: Key factors = frequent late nights (falling asleep after 1 a.m. > 4 times a week) + high-salt diet (urine sodium test > 200 mmol / L).

[0119] ‌Cause significance verification‌: The contribution of the inducement is calculated using a permutation test: the user's living habit data is randomly disrupted 1,000 times; if the correlation strength between the inducement and the threshold breakthrough in the original data exceeds 95% of the random samples, it is judged to be significant.

[0120] The final output is a list of key risk factors with confidence marks, for example: 1. Factor: Weekday coffee intake > 400 mg (confidence mark P < 0.05); 2. Factor: Sedentary time > 10 hours / day (confidence mark P < 0.01).

[0121] Retrieve intervention templates matching risk factors from medical guideline-related databases to generate initial draft instructions; The retrieval and generation process is as follows: Multi-level search strategy: First-level search: Use key factors as keywords, such as entering "high-sodium diet cardiovascular intervention"; Second-level screening: Filter by user demographic attributes (age, gender), for example: 50-year-old male users automatically exclude "gestational hypertension intervention template"; Third-level matching: Associate target organ systems (for example, search results must contain the "hepatobiliary system" tag).

[0122] Template feature extraction: Extracts structured fields from the guideline: intervention type (diet / exercise / drug); frequency of implementation (e.g., aerobic exercise 3 times a week); quantitative target (e.g., daily sodium intake ≤ 1500 mg).

[0123] Example template snippet: { Source: WHO Guidelines for Primary Prevention of Cardiovascular Disease 2023, "Measure": "DASH Dietary Pattern", "Parameters": Daily vegetables ≥ 500g, saturated fat accounts for <10% of total calories }.

[0124] Draft is dynamically generated: Inject user data into template parameters: basal metabolic rate to calculate daily calorie needs; current sodium intake as a baseline reference. Generate personalized initial instruction draft, for example: Draft 1: Adopt the DASH diet and reduce current sodium intake from 3200 mg / day to 1500 mg / day (based on user baseline data); Draft 2: Perform 40 minutes of moderate-intensity exercise three times a week (reference user physical fitness assessment score of 65 points).

[0125] Conflict detection: When multiple template requirements conflict (such as an exercise plan with knee joint load restrictions), orthopedic expert rule verification is automatically triggered.

[0126] Convert the instruction strength level in the initial instruction draft according to the confidence level, where high confidence corresponds to mandatory instruction strength and medium confidence corresponds to recommended instruction strength; Confidence grading: High confidence: ≥60% possibility of pathological progression (red warning); Medium confidence: 30%-60% possibility (yellow warning); Low confidence: <30% possibility (blue observation).

[0127] The intensity conversion rules are as follows: ‌Command Strength Definition‌: Mandatory level: Medical instructions that must be executed. Failure to comply will trigger an alarm. Characteristic words: "Stop immediately", "Strictly prohibited", "Must be completed within 72 hours"; Recommendation level: Non-mandatory recommendations recommended for execution. Characteristic words: "Recommended adjustment", "Recommended attempt", "Can be achieved in stages".

[0128] ‌Intensity Mapping Logic‌: High confidence risk + core trigger → mandatory level instruction. For example: User's cardiovascular risk confidence level is 72% + Smoking Index > 400 → the instruction is upgraded to "Immediately activate smoking cessation clinic service (mandatory level)"; Medium confidence risk + minor contributing factors → recommendation-level instructions. For example: Metabolic risk confidence level 45% + BMI=28 → retain "Recommendation to lose 1 kg per month (recommendation level)".

[0129] Dynamic Intensity Calibration: Continuously monitor user compliance: If recommended-level instructions are not executed for two consecutive weeks and the risk threshold continues to approach, the policy will be automatically upgraded to the mandatory level. If the mandatory-level instruction execution rate is >90% and the risk subsides, the policy will be downgraded to the recommended level.

[0130] Strength indicator visualization: Mandatory-level instructions are displayed as a red shield icon on mobile devices; recommended-level instructions are displayed as a green compass icon.

[0131] Instructions are sorted in reverse order of risk node timestamps to generate a hierarchical health intervention instruction set with execution time windows.

[0132] The risk node timestamp marks the occurrence time of the high-risk state in the health trajectory diagram. Example timestamp sequence: 2024-03-15 (hepatobiliary risk), 2024-02-20 (cardiovascular risk).

[0133] Hierarchical instruction set generation process: ‌Reverse sort logic‌: Prioritize the most recent risks: Sort by timestamp from most recent to most recent. This sorting is based on the medical principle of "urgency first," meaning that newly emerging risks require a more rapid response. Example sorting results: 1. 2024-03-15 Hepatobiliary System L3 Risk (Priority 1); 2. 2024-02-20 Cardiovascular System S2 Risk (Priority 2).

[0134] Time window binding rules: Each instruction is bound to an execution time window: emergency intervention window: 0-7 days after the risk node (such as mandatory smoking cessation instructions); mid-term adjustment window: 8-30 days (such as dietary structure adjustment); long-term maintenance window: 31-90 days (such as developing exercise habits).

[0135] Dynamic window width adjustment: High-confidence risk: The window width is compressed by 50% (for example, the original 30-day window width is compressed to 15 days). Multi-system concurrent risk: The total window width is automatically extended (for example, 30 days for a single system → 45 days for dual systems).

[0136] ‌Layered Instruction Set Encapsulation‌: The final structured instruction set is generated in JSON-LD format (Lightweight Data Interchange Format): { "Command sequence": [ { "Risk type": "Hepatobiliary system L3 fibrosis tendency", "Execution Time Window": "2024-03-16 to 2024-04-15", "measure": [ { "Command content": "Complete abstinence (mandatory level)", "Supervision Mechanism": "Weekly alcohol metabolism testing" }, { "Instruction content": "Daily supplementation of silymarin 500mg (recommended level)", "Monthly liver function review" } ] }, { "Risk type": "Decreased elasticity of S2 arteries in the cardiovascular system", "Execution Time Window": "2024-02-21 to 2024-05-21", "measure": [...] } ] }.

[0137] Output delivery: Simultaneously pushed to the user's mobile app and doctor management platform.

[0138] Based on risk levels, the system automatically matches evidence-based medical guidelines and integrates patient data on exercise, diet, and other lifestyle habits to generate a step-by-step plan of recommendations, from emergency medical treatment to lifestyle adjustments. This creates a closed-loop management system from disease prediction to intervention execution. These tiered instructions ensure timely treatment for high-risk patients while also improving the health management of the general population through practical recommendations, forming a precise prevention system.

[0139] It can be seen that the user's multimodal health check-up original data is received to generate a multimodal feature matrix that integrates spatiotemporal correlation; the multimodal feature matrix is ​​input into a lightweight dual-channel network to obtain a decoupled bimodal feature group; a hidden Markov chain is constructed based on the bimodal feature group to generate a health state transition trajectory graph with a timestamp; the health state transition trajectory graph is input into the discriminator of the generative adversarial network to output the confidence score and pathological trigger threshold of the high-risk node; the rule engine is activated according to the pathological trigger threshold, and a layered health intervention instruction set is generated in combination with individual lifestyle data, so that health check-up data can be efficiently analyzed and the accuracy of early disease warning and health intervention can be optimized.

[0140] Another embodiment of the present invention provides a health examination data analysis system based on artificial intelligence, see Figure 3 , the system may include: The receiving module 301 is used to receive the user's multimodal health checkup raw data, perform spatiotemporal alignment processing on the structured report data and the unstructured medical image data according to the preset cross-modal alignment rules, and generate a multimodal feature matrix integrating spatiotemporal correlation; A decoupling module 302 is configured to input the multimodal feature matrix into a lightweight dual-channel network, synchronously decouple the static physiological baseline feature vector and the dynamic abnormal fluctuation feature vector, and obtain a decoupled dual-modal feature group; A construction module 303 is configured to construct a hidden Markov chain based on the bimodal feature group, calculate the organ function degradation path through the state transition probability, and generate a health state transition trajectory diagram with a time stamp; The adversarial module 304 is configured to input the health state transition trajectory graph into the discriminator of the generative adversarial network, perform adversarial matching with the real clinical case database, and output the confidence score and pathology trigger threshold of the high-risk node; The generation module 305 is used to activate the rule engine according to the pathological trigger threshold and generate a hierarchical health intervention instruction set in combination with individual lifestyle data.

[0141] An embodiment of the present invention further provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps of any one of the above method embodiments when running.

[0142] Specifically, in this embodiment, the above-mentioned storage medium may be configured to store a computer program for performing the following steps: S201, receiving the user's multimodal health checkup raw data, performing spatiotemporal alignment processing on the structured report data and the unstructured medical imaging data according to a preset cross-modal alignment rule, and generating a multimodal feature matrix integrating spatiotemporal correlation; S202, inputting the multimodal feature matrix into a lightweight dual-channel network, synchronously decoupling a static physiological baseline feature vector and a dynamic abnormal fluctuation feature vector, and obtaining a decoupled dual-modal feature group; S203, constructing a hidden Markov chain based on the bimodal feature group, calculating the organ function degradation path through state transition probability, and generating a health state transition trajectory diagram with a time stamp; S204: Input the health state transition trajectory graph into the discriminator of the generative adversarial network, perform adversarial matching with the real clinical case database, and output the confidence score and pathology trigger threshold of the high-risk node; S205 , activating a rule engine according to the pathological triggering threshold, and generating a hierarchical health intervention instruction set in combination with individual lifestyle data.

[0143] An embodiment of the present invention further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any one of the above method embodiments.

[0144] Specifically, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0145] Specifically, in this embodiment, the processor may be configured to execute the following steps through a computer program: S201, receiving the user's multimodal health checkup raw data, performing spatiotemporal alignment processing on the structured report data and the unstructured medical imaging data according to a preset cross-modal alignment rule, and generating a multimodal feature matrix integrating spatiotemporal correlation; S202, inputting the multimodal feature matrix into a lightweight dual-channel network, synchronously decoupling a static physiological baseline feature vector and a dynamic abnormal fluctuation feature vector, and obtaining a decoupled dual-modal feature group; S203, constructing a hidden Markov chain based on the bimodal feature group, calculating the organ function degradation path through state transition probability, and generating a health state transition trajectory diagram with a time stamp; S204: Input the health state transition trajectory graph into the discriminator of the generative adversarial network, perform adversarial matching with the real clinical case database, and output the confidence score and pathology trigger threshold of the high-risk node; S205 , activating a rule engine according to the pathological triggering threshold, and generating a hierarchical health intervention instruction set in combination with individual lifestyle data.

[0146] The above describes in detail the structure, features and effects of the present invention based on the embodiments shown in the drawings. The above is only a preferred embodiment of the present invention, but the scope of implementation of the present invention is not limited to what is shown in the drawings. Any changes made in accordance with the concept of the present invention, or modifications to equivalent embodiments with equivalent changes, which do not exceed the spirit covered by the description and drawings, should be within the scope of protection of the present invention.

Claims

1. A health examination data analysis method based on artificial intelligence, characterized in that: The method comprises: Receive the user's multimodal health checkup raw data, perform spatiotemporal alignment processing on the structured report data and unstructured medical imaging data according to the preset cross-modal alignment rules, and generate a multimodal feature matrix that integrates spatiotemporal correlations; Inputting the multimodal feature matrix into a lightweight dual-channel network, synchronously decoupling the static physiological baseline feature vector and the dynamic abnormal fluctuation feature vector, and obtaining a decoupled dual-modal feature group; A hidden Markov chain is constructed based on the bimodal feature set, and the organ function degradation path is calculated through the state transition probability to generate a health state transition trajectory diagram with a time stamp; Input the health state transition trajectory graph into the discriminator of the generative adversarial network, perform adversarial matching with the real clinical case database, and output the confidence score and pathology trigger threshold of the high-risk node; The rule engine is activated according to the pathological trigger threshold, and a hierarchical health intervention instruction set is generated in combination with individual lifestyle data.

2. The method according to claim 1, characterized in that The receiving of the user's multimodal health checkup raw data, performing spatiotemporal alignment processing on the structured report data and the unstructured medical imaging data according to a preset cross-modal alignment rule, and generating a multimodal feature matrix integrating spatiotemporal correlations, includes: According to the timestamp of the DICOM header file of the medical image and the acquisition time of the physical examination report, the time axis is aligned through the dynamic time warping algorithm to generate a time synchronization tag set; Based on the time-synchronized label set, the organ anatomical landmarks in the image sequence are spatially mapped with the physiological indicators in the report, and a multi-resolution B-spline deformation field is used to generate an anatomical-physiological correlation matrix; According to the anatomical-physiological correlation matrix, Lagrangian interpolation constraints are imposed on the structured report data, and texture spectrum features of the image are extracted to output a spatiotemporal fusion feature matrix with unified dimension.

3. The method according to claim 2, characterized in that The multimodal feature matrix is ​​input into a lightweight dual-channel network, and a static physiological baseline feature vector and a dynamic abnormal fluctuation feature vector are synchronously decoupled to obtain a decoupled dual-modal feature group, including: According to the entropy distribution of the spatiotemporal fusion feature matrix, the initial convolution kernel weights are assigned to the static channel and the dynamic channel respectively through the back propagation greedy algorithm; In the static channel, a sliding window mean filter is used to generate a baseline feature vector based on the normal fluctuation range of physiological indicators, and a steady-state feature tensor is output. The Mahalanobis distance between the current eigenvalue and the steady-state eigentensor is calculated in the dynamic channel, the excessive fluctuation pattern is captured through the gated recurrent unit, and the abnormal fluctuation eigenvector is output; Apply Gram-Schmidt orthogonalization to the steady-state eigentensor and the abnormal fluctuation eigenvector to eliminate the characteristic coupling interference and generate the initial decoupled characteristic group; The knowledge distillation technology is used to compress the dual-channel network to 1 / 8 the number of parameters, retaining the preset percentage of feature decoupling accuracy, and finally outputting the decoupled bimodal feature group.

4. The method according to claim 3, characterized in that The method comprises: constructing a hidden Markov chain based on the bimodal feature group, calculating the organ function degradation path through the state transition probability, and generating a health state transition trajectory diagram with a time stamp, including: The bimodal feature group is divided into multiple health status dimensions according to organ system, and a discrete latent state set is generated through K-means clustering; Based on the historical clinical follow-up data set, Bayesian inference is used to calculate the transition probability between each latent state and construct the organ state transition matrix; performing constrained Markov chain Monte Carlo sampling according to the organ state transition matrix to generate multiple possible state transition paths; Eliminate the paths in the state transition path that have not experienced abnormal states for a preset number of consecutive times, and retain the path subset containing risk transfer nodes; Map the path subset to the physical examination timeline, predict the time point of each state transition through exponential smoothing, and output the health state transition trajectory diagram with timestamps.

5. The method according to claim 4, characterized in that The health state transition trajectory graph is input into the discriminator of the generative adversarial network, and adversarially matched with the real clinical case database to output the confidence score and pathology trigger threshold of the high-risk node, including: The diagnostic records in the real clinical case database are converted into case feature tensors through ICD-11 coding; Cut the health state transition trajectory graph by quarterly time window to obtain a trajectory slice sequence; The trajectory slices are input into the GAN discriminator, and the cosine similarity is matched with the case feature tensor to output the authenticity score of each slice; When the authenticity score is lower than the preset score threshold, it is marked as a high-risk node, and the confidence level is calculated based on the proportion of similar nodes in the real clinical case database; According to the pathological progression rate distribution of high-risk nodes in the real clinical case database, the pathological trigger threshold curve is fitted.

6. The method according to claim 5, characterized in that The method of activating a rule engine according to the pathological trigger threshold and generating a hierarchical health intervention instruction set in combination with individual lifestyle data includes: Correlate and analyze pathological trigger thresholds with user lifestyle data, and identify key risk factors through a decision tree. Retrieve intervention templates matching risk factors from medical guideline-related databases to generate initial draft instructions; Convert the instruction strength level in the initial instruction draft according to the confidence level, where high confidence corresponds to mandatory instruction strength and medium confidence corresponds to recommended instruction strength; Instructions are sorted in reverse order of risk node timestamps to generate a hierarchical health intervention instruction set with execution time windows.

7. A health examination data analysis system based on artificial intelligence, characterized in that: The system comprises: The receiving module is used to receive the user's multimodal health examination raw data, perform spatiotemporal alignment processing on the structured report data and the unstructured medical imaging data according to the preset cross-modal alignment rules, and generate a multimodal feature matrix that integrates spatiotemporal correlation; a decoupling module, configured to input the multimodal feature matrix into a lightweight dual-channel network, synchronously decouple a static physiological baseline feature vector and a dynamic abnormal fluctuation feature vector, and obtain a decoupled dual-modal feature group; A construction module is used to construct a hidden Markov chain based on the bimodal feature group, calculate the organ function degradation path through state transition probability, and generate a health state transition trajectory diagram with a time stamp; An adversarial module is used to input the health state transition trajectory graph into the discriminator of the generative adversarial network, perform adversarial matching with the real clinical case database, and output the confidence score and pathology trigger threshold of the high-risk node; A generation module is used to activate the rule engine according to the pathological trigger threshold and generate a layered health intervention instruction set in combination with individual lifestyle data.

8. The system according to claim 7, characterized in that The receiving module is specifically configured to: According to the timestamp of the DICOM header file of the medical image and the acquisition time of the physical examination report, the time axis is aligned through the dynamic time warping algorithm to generate a time synchronization tag set; Based on the time-synchronized label set, the organ anatomical landmarks in the image sequence are spatially mapped with the physiological indicators in the report, and a multi-resolution B-spline deformation field is used to generate an anatomical-physiological correlation matrix; According to the anatomical-physiological correlation matrix, Lagrangian interpolation constraints are imposed on the structured report data, and texture spectrum features of the image are extracted to output a spatiotemporal fusion feature matrix with unified dimension.

9. A storage medium, characterized in that: The storage medium stores a computer program, wherein the computer program is configured to execute the method according to any one of claims 1 to 6 when executed.

10. An electronic device comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to run the computer program to perform the method according to any one of claims 1 to 6.

Images