Intelligent physical test health management method and system integrated with AI large model, and physical test instrument

Abstract

The invention relates to the technical field of health management, in particular to an intelligent physical measurement health management method and system integrated with an AI large model and a physical measurement instrument, and the method comprises the steps: collecting multi-dimensional physiological sign time sequence data and analyzing a biological rhythm, and generating an individualized physiological steady state baseline track; tissue microstructure texture features are extracted in combination with continuous dynamic body measurement image data, and an organ function reserve dynamic map is constructed. And fusing the data to establish a multi-level chronic disease risk gradient model, and outputting a multi-time-scale disease activity prediction curve. High-risk target organs are recognized on the basis of images, personalized intervention window periods are deduced in combination with prediction curves, and organ-divided and staged accurate intervention task sets are formed. By dynamically scheduling individualized health intervention resources and collecting feedback data, a physiological steady state baseline is reconstructed in real time based on curative effect attribution and threshold drift detection, a closed-loop dynamic health management strategy is formed, and accurate early warning and individualized intervention of chronic disease risks are achieved.

Description

Technical Field

[0001] This invention relates to the field of health management technology, and in particular to an intelligent physical fitness testing health management method, system, and physical fitness testing instrument that integrates a large AI model. Background Technology

[0002] With the continuous rise in the prevalence of chronic diseases, traditional health management models, relying on single-dimensional physiological indicators and lacking dynamic physiological rhythm analysis and cross-modal correlation analysis of organ function reserves, struggle to achieve precise stratification and individualized intervention for chronic disease risks. Current technologies primarily rely on static thresholds for physiological homeostasis assessment, failing to adequately consider individual differences in biorhythms and the temporal characteristics of homeostasis shifts; body imaging analysis is limited to qualitative observation of structures, lacking coupled modeling of tissue microstructure texture and functional status; risk prediction models fail to effectively integrate physiological baseline trajectories with dynamic organ function maps, resulting in a single timescale for disease activity prediction and coarse risk gradient classification. Furthermore, intervention strategies often employ a one-size-fits-all approach, failing to incorporate individual characteristics and dynamic extrapolation of disease progression to determine precise intervention windows, and the allocation of health intervention resources lacks a dynamic adaptation mechanism under spatiotemporal constraints, making it difficult to form a closed-loop management system based on intervention response feedback.

[0003] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0004] The main objective of this invention is to provide an intelligent physical fitness testing and health management method, system, and physical fitness testing instrument that integrates a large AI model. This aims to solve the technical problems of traditional health management models, such as reliance on single-dimensional physiological indicators, lack of dynamic physiological rhythm analysis and cross-modal correlation analysis of organ function, homeostasis assessment based on static thresholds, insufficient integration of risk prediction models, one-size-fits-all intervention strategies, and lack of dynamic adaptation and closed-loop management in health intervention resource scheduling, which makes it difficult to achieve precise risk stratification and individualized intervention for chronic diseases.

[0005] To achieve the above objectives, this invention provides an intelligent physical fitness assessment and health management method integrating a large AI model, the method comprising the following steps: Collect time-series data of individual multidimensional physiological characteristics; perform biorhythm analysis and steady-state shift modeling on the time-series data of individual multidimensional physiological characteristics to generate individualized physiological steady-state baseline trajectories; Acquire continuous dynamic body measurement image data of an individual; perform tissue microstructure texture characterization extraction and cross-modal functional coupling analysis based on the continuous dynamic body measurement image data of the individual to generate a dynamic atlas of organ function reserve; By integrating individualized physiological homeostasis baseline trajectory and dynamic atlas of organ function reserve, a multi-level chronic disease progression risk gradient model is constructed, and multi-timescale disease activity prediction curves are output. High-risk target organ regions are identified based on individual continuous dynamic body imaging; combined with multi-timescale disease activity prediction curves, personalized intervention window periods are extrapolated for each target organ region, generating a set of precise intervention tasks by organ and stage. Identify available individualized health intervention resources; based on the precise intervention task set, dynamically adapt and schedule the individualized health intervention resources under spatiotemporal constraints, and simultaneously collect intervention response feedback data; Based on intervention response feedback data, efficacy attribution analysis and threshold drift detection are performed to drive real-time reconstruction of the physiological homeostatic baseline trajectory, forming a closed-loop dynamic threshold health management strategy.

[0006] Optionally, the step of collecting individual multidimensional physiological sign time-series data; performing biorhythm analysis and homeostasis shift modeling on the individual multidimensional physiological sign time-series data to generate an individualized physiological homeostasis baseline trajectory includes: The individual multidimensional physiological signs time series data are collected by the collaborative collection of wearable sensor array and non-contact body measurement terminal. The individual multidimensional physiological signs time series data include at least heart rate variability, skin conductance response, respiratory tidal volume, microcirculatory blood oxygen saturation and sublingual microvascular flow velocity. The individual multidimensional physiological signs time series data are subjected to diurnal rhythm phase decoupling and autonomic nerve tension spectrum decomposition to generate an individualized biological rhythm parameter matrix; Based on the individualized biological rhythm parameter matrix, a physiological homeostatic dynamic equation is constructed, and the topology of the homeostatic attractor is solved. Long-term steady-state shift quantization is performed using steady-state attractor topology to generate a steady-state shift entropy value sequence. By integrating the steady-state offset entropy sequence with the individualized biological rhythm parameter matrix, an individualized physiological homeostasis baseline generator is trained. The individualized physiological homeostasis baseline generator outputs an individualized physiological homeostasis baseline trajectory, which includes a dynamic confidence interval and a pathological perturbation sensitive zone.

[0007] Optionally, the step of acquiring individual continuous dynamic body measurement image data; and performing tissue microstructure texture characterization extraction and cross-modal functional coupling analysis based on the individual continuous dynamic body measurement image data to generate a dynamic atlas of organ functional reserves includes: Continuous dynamic body measurement data of an individual were acquired simultaneously using a multispectral skin-mucosal imaging system and optical coherence tomography of the tongue surface microstructure. Multi-scale vascular network skeleton extraction was performed on the individual's continuous dynamic body imaging data to construct a microcirculation topology connection map; By integrating microcirculation topology maps with infrared thermo-metabolic images, tissue perfusion-metabolic coupling modeling is performed to generate dynamic atlases of organ-level functional reserves. Based on the dynamic map of organ-level functional reserves, the functional redundancy index and compensatory attenuation slope of each target organ are calculated. Cross-organ correlation analysis of functional redundancy index and compensatory attenuation slope was conducted to identify key nodes of functional compensation imbalance. Using the hub node as the anchor point, the data is back-mapped to the anatomical structure space to generate a dynamic atlas of organ function reserves, which includes a set of spatial positioning labels.

[0008] Optionally, the step of extracting the multi-scale vascular network skeleton from the individual's continuous dynamic body imaging data and constructing a microcirculation topology map includes: Perform multispectral channel light scattering consistency calibration on continuous dynamic body measurement image data of individuals; An adaptive local contrast enhancement algorithm is used to sharpen the texture of the calibrated image, generating an enhanced texture image. A multi-scale Gabor filter bank was constructed based on enhanced texture images to extract directional texture features of tissue microstructures. A lightweight U-Net segmentation model is trained using directional texture features to achieve pixel-level segmentation of microvascular networks; Graph theory optimization is performed on the segmentation results to remove pseudo-connected branches and complete the broken blood vessel segments, outputting a microcirculation topology connection graph.

[0009] Optionally, the method of fusing individualized physiological homeostasis baseline trajectories and dynamic organ function reserve maps to construct a multi-level chronic disease progression risk gradient model and outputting multi-timescale disease activity prediction curves includes: By mapping individualized physiological homeostasis baseline trajectories to the spatiotemporal coordinate system of organ functional reserve dynamic map, a cross-modal homeostasis-function joint representation space is constructed. Within the cross-modal steady-state-functional joint characterization space, a multidimensional risk potential field for the progression of chronic diseases is defined, which is composed of the inflammatory factor fluctuation entropy, the metabolite accumulation gradient, and the amplitude of neuroendocrine axis perturbation. Based on the multidimensional risk potential field, a Lagrange mechanical framework is used to simulate the disease progression path and generate a disease activity evolution manifold. Input a pre-trained multimodal large model, which integrates clinical guideline knowledge graphs, real-world cohort survival data, and individual gene polymorphism information to enhance causal inference of the disease activity evolution manifold; The output includes disease activity prediction curves covering multiple time granularities, with each curve accompanied by uncertainty quantification indicators and key inflection point warning indicators.

[0010] Optionally, the method involves identifying high-risk target organ regions based on individual continuous dynamic body imaging; combining multi-timescale disease activity prediction curves to perform personalized intervention window period simulations for each target organ region, generating a precise intervention task set by organ and stage, including: Based on the spatial localization tag set of the dynamic atlas of organ function reserve in continuous dynamic body measurement images, four types of high-risk target organ regions, namely liver, kidney, pancreas and retina, are identified. Pathological imaging feature transfer learning was performed on each target organ region to identify subclinical lesion markers of early fibrosis, microaneurysms, and β-cell apoptosis; Based on the detection intensity, spatial distribution density, and deviation from the physiological homeostatic baseline trajectory of subclinical lesion markers, the intervention urgency score of each target organ was calculated. Based on individual daily activity trajectories and environmental exposure data, a spatiotemporal accessibility map of interventionable resources is constructed, and feasible intervention paths that conform to physiological rhythm constraints are extracted. The intervention urgency score was Pareto optimally matched with feasible intervention pathways to deduce the optimal intervention initiation time window and minimum effective intervention intensity for each target organ. By integrating the target organ intervention time window, intensity threshold, and resource constraints, a set of precise intervention tasks with execution priority, dose gradient, and efficacy monitoring nodes is generated.

[0011] Optionally, the process involves identifying available individualized health intervention resources; dynamically adapting and scheduling these resources under spatiotemporal constraints based on the precise intervention task set; and simultaneously collecting intervention response feedback data, including: Identify available personalized health intervention resources, including nutrient sustained-release microcapsules, transcutaneous electrical stimulation parameter sets, personalized exercise prescription libraries, and combinations of gut microbiota modulators; A digital twin is established for the individualized health intervention resources, and its pharmacokinetic response is simulated under the individual's physiological homeostatic baseline trajectory to generate a resource efficacy-toxicity trade-off matrix. Based on the resource efficiency-toxicity trade-off matrix and the precise intervention task set, a multi-objective reinforcement learning algorithm is used for dynamic resource scheduling, outputting an intervention execution instruction sequence with spatiotemporal coordinates, and simultaneously collecting intervention response feedback data.

[0012] Optionally, the step of performing efficacy attribution analysis and threshold drift detection based on intervention response feedback data to drive real-time reconstruction of the physiological homeostasis baseline trajectory and form a closed-loop dynamic threshold health management strategy includes: Multi-source heterogeneous signal alignment was performed on the intervention response feedback data, including continuous blood glucose fluctuation spectrum, dynamic changes in the urine microalbumin / creatinine ratio, time-series concentrations of inflammatory factors, and subjective symptom diary text; We constructed an attribution graph neural network for intervention efficacy to analyze the contribution weights of each intervention to the functional reserve of different target organs. When the functional redundancy index of any target organ falls below a preset safety threshold and the duration exceeds a preset time, the threshold drift detection mechanism is triggered. Based on the threshold drift detection results, the individualized physiological homeostasis baseline generator is retrained, the physiological homeostasis baseline trajectory and its dynamic confidence interval are updated, and the closed-loop dynamic threshold health management strategy is iterated.

[0013] Furthermore, to achieve the above objectives, the present invention also provides an intelligent physical fitness assessment and health management system integrating a large AI model, the system comprising: The data modeling module is used to collect time-series data of individual multidimensional physiological characteristics; perform biorhythm analysis and steady-state shift modeling on the time-series data of individual multidimensional physiological characteristics, and generate individualized physiological steady-state baseline trajectory. The image analysis module is used to acquire continuous dynamic body measurement image data of an individual; based on the continuous dynamic body measurement image data of the individual, it performs tissue microstructure texture characterization extraction and cross-modal functional coupling analysis to generate a dynamic atlas of organ function reserves; The risk modeling module is used to integrate individualized physiological homeostasis baseline trajectories and dynamic maps of organ function reserves to construct a multi-level chronic disease progression risk gradient model and output disease activity prediction curves at multiple time scales. The intervention simulation module is used to identify high-risk target organ regions based on individual continuous dynamic body imaging; combined with multi-timescale disease activity prediction curves, it performs personalized intervention window simulations for each target organ region, generating a set of precise intervention tasks by organ and stage. The resource scheduling module is used to identify available individualized health intervention resources; based on the precise intervention task set, it performs dynamic adaptation scheduling of individualized health intervention resources under spatiotemporal constraints, and simultaneously collects intervention response feedback data; The closed-loop management module is used to perform efficacy attribution analysis and threshold drift detection based on intervention response feedback data, drive real-time reconstruction of the physiological homeostasis baseline trajectory, and form a closed-loop dynamic threshold health management strategy.

[0014] In addition, to achieve the above objectives, the present invention also provides a body composition analyzer, the body composition analyzer comprising: a memory, a processor, and an intelligent body composition analysis and health management program integrating an AI large model stored in the memory and executable on the processor, the intelligent body composition analysis and health management program integrating an AI large model configured to implement the steps of the intelligent body composition analysis and health management method integrating an AI large model as described above.

[0015] This invention provides an intelligent physical fitness assessment and health management method integrating a large AI model. The method collects and models multidimensional physiological sign time-series data to generate individualized physiological homeostasis baseline trajectories, overcoming the limitations of traditional static thresholds and achieving dynamic and accurate assessment of physiological states. Based on continuous dynamic physical fitness images, it performs tissue microstructure texture characterization and cross-modal coupling analysis to construct a dynamic atlas of organ function reserves, compensating for the shortcomings of traditional image analysis that only focuses on anatomical structures, and achieving in-depth analysis of structure-function correlations. Furthermore, it integrates the physiological homeostasis baseline and organ function atlas to construct a multi-level risk model, outputting multi-timescale prediction curves, improving the precision and timeliness of chronic disease risk stratification. By combining image recognition and risk prediction to extrapolate personalized intervention windows, a precise intervention task set by organ and stage is generated, solving the blindness of the traditional one-size-fits-all intervention model and achieving dynamic and precise matching of intervention timing and plan. Based on spatiotemporal constraints, health intervention resources are dynamically adapted and scheduled and feedback data is collected. Combined with efficacy attribution and threshold drift detection, a closed-loop management strategy is formed to optimize resource utilization efficiency. At the same time, by driving model iteration through real-time data, the dynamism of health management and the sustainability of intervention effects are improved. Finally, an intelligent health management system with a full process from data collection and risk modeling to precise intervention and closed-loop optimization is constructed, which significantly improves the effectiveness of early warning and individualized management of chronic diseases. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating an embodiment of the intelligent physical fitness assessment and health management method integrating a large AI model according to the present invention.

[0017] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] Reference Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the intelligent physical fitness testing and health management method integrating a large AI model according to the present invention.

[0020] In one embodiment, the intelligent physical fitness testing and health management method integrating a large AI model includes: Step S100: Collect time-series data of individual multidimensional physiological characteristics. Perform biorhythm analysis and steady-state shift modeling on the time-series data of individual multidimensional physiological characteristics to generate individualized physiological steady-state baseline trajectories.

[0021] The multidimensional physiological sign time-series data can be a high-dimensional sequence set containing continuous changes in multiple physiological parameters, which can be used to provide a dynamic observation basis for individual physiological states. Furthermore, the multidimensional physiological sign time-series data can include, but is not limited to, one or more of cardiovascular time-series signals, metabolic dynamic signals, and autonomic nervous activity signals. The individualized physiological homeostasis baseline trajectory can be a dynamic evolution path reflecting the deviation trend of physiological parameters over a long time scale, and can be used to replace fixed thresholds to achieve personalized assessment. It is understood that the individualized physiological homeostasis baseline trajectory can be constructed into a non-parametric probability distribution trajectory through periodic decomposition and trend separation. In this embodiment, the individualized physiological homeostasis baseline trajectory serves as the core input of the risk model, and together with the organ function reserve dynamic map, constitutes the two-dimensional representation basis for chronic disease progression. The biological rhythm analysis module can be a signal processing unit for separating periodic rhythm components, which can be implemented using Fourier transform, wavelet analysis, or periodic Gaussian process regression. The homeostasis deviation modeling module can be a modeling unit for extracting long-term trend deviations, which can be implemented using non-parametric regression or change point detection methods.

[0022] Step S200: Acquire individual continuous dynamic body measurement image data. Based on the individual continuous dynamic body measurement image data, perform tissue microstructure texture characterization extraction and cross-modal functional coupling analysis to generate a dynamic atlas of organ functional reserves.

[0023] Continuous dynamic body measurement image data can be a multimodal image sequence containing structural and functional dynamic information, which can be used to provide a dynamic visualization data source for organ microstructure and local functional status. Furthermore, continuous dynamic body measurement image data can include, but is not limited to, one or more of the following: tissue elasticity dynamic sequences, microcirculation thermal distribution sequences, and light scattering texture sequences. Tissue microstructure texture representation can be a quantitative feature vector describing the microscopic spatial distribution pattern within a tissue, which can be used to quantify tissue microstructure changes that cannot be identified by traditional imaging. It is understood that tissue microstructure texture representation can be achieved through multi-scale image analysis algorithms such as gray-level co-occurrence matrix, local binary mode, wavelet entropy, and structural tensor analysis. In this embodiment, tissue microstructure texture representation serves as the structural input for cross-modal functional coupling analysis, working in conjunction with multidimensional physiological sign time-series data. Organ function reserve dynamic atlas can be a multidimensional function space representing the evolution of organ functional response potential over time, which can be used to overcome the limitations of traditional imaging that only describes morphology. It is understood that organ function reserve dynamic atlas can be constructed by nonlinearly mapping tissue microstructure texture representation with physiological sign data to create a dynamic matrix. In this embodiment, the dynamic atlas of organ function reserves serves as one of the input modalities of the multi-level risk model, forming the basis for prediction together with the individualized physiological homeostasis baseline trajectory. The cross-modal functional coupling analysis module can be an analysis unit that establishes a nonlinear mapping relationship between image texture features and physiological signals, and can be implemented using deep canonical correlation analysis, multimodal contrastive learning, or cross-modal attention mechanisms.

[0024] Step S300: Integrate individualized physiological homeostasis baseline trajectory and organ function reserve dynamic map to construct a multi-level chronic disease progression risk gradient model and output disease activity prediction curves at multiple time scales.

[0025] The multi-level chronic disease progression risk gradient model can be a hierarchical prediction system, used to output risk evolution paths at different time granularities. Furthermore, the multi-level chronic disease progression risk gradient model can include, but is not limited to, one or more of short-term, medium-term, and long-term risk gradient models. The multi-timescale disease activity prediction curve can be a multi-resolution probability output sequence representing the change in the probability of disease progression over time, used to provide the evolutionary trend of disease activity under different prediction windows. It is understood that the multi-timescale disease activity prediction curve can output probability distribution curves at multiple time granularities through a hierarchical neural network structure. In this embodiment, the multi-timescale disease activity prediction curve serves as the direct input to the intervention window period extrapolation module, determining the triggering sequence of the intervention task.

[0026] Step S400: Identify high-risk target organ regions based on individual continuous dynamic body imaging. Combined with multi-timescale disease activity prediction curves, perform personalized intervention window simulations for each target organ region, generating a precise intervention task set by organ and stage.

[0027] High-risk target organ regions can be local tissue regions with significant structural abnormalities and functional reserve decline, which can be used to pinpoint the anatomical location of the intervention target. Further, high-risk target organ regions can include, but are not limited to, one or more of the following: liver fibrosis regions, coronary artery plaque regions, and insulin-resistant adipose tissue domains. Personalized intervention windows can be the optimal intervention initiation and duration interval for a specific target organ region, which can be used to avoid ineffective intervention or missing the intervention opportunity. Further, personalized intervention windows can include, but are not limited to, one or more of the following: early warning windows, functional compensation windows, and pre-failure windows. Organ-specific and stage-specific precise intervention task sets can be a combination of task sequences consisting of multiple organ-specific and stage-adaptive intervention actions, which can be used to achieve individualized intervention transformation. It is understood that organ-specific and stage-specific precise intervention task sets can generate task lists by matching a preset intervention strategy library. In this embodiment, organ-specific and stage-specific precise intervention task sets serve as the execution instruction source for the dynamic adaptation scheduling module, directly determining the type and priority of resource scheduling. The image recognition module can be a computer vision unit for locating structurally abnormal regions, which can be implemented using a semantic segmentation network or an anomaly detection autoencoder. The intervention window simulation module can be a decision-making unit for calculating the optimal timing for intervention initiation, and can be implemented using dynamic programming or reinforcement learning.

[0028] Step S500: Identify available individualized health intervention resources. Based on the precision intervention task set, dynamically adapt and schedule individualized health intervention resources under spatiotemporal constraints, and simultaneously collect intervention response feedback data.

[0029] Personalized health intervention resources can be a diverse set of services and equipment that can be scheduled to perform health intervention tasks, and can be used to provide entities or service carriers for performing intervention tasks. Furthermore, personalized health intervention resources can include, but are not limited to, one or more of the following: remote physiological monitoring equipment, AI-driven personalized nutrition plans, and community-level rehabilitation service nodes. The dynamic adaptation scheduling mechanism under spatiotemporal constraints can be a scheduling system that automatically matches the optimal execution plan, and can be used to ensure that intervention tasks are performed by appropriate resources at the right time and place. Furthermore, the dynamic adaptation scheduling mechanism under spatiotemporal constraints can adopt one or more of the following: geographical coverage scheduling strategy, response delay priority scheduling strategy, and cost-benefit balance scheduling strategy. Intervention response feedback data can be a set of subsequent observational data on individual physiological state and behavioral response after intervention execution, and can be used to provide empirical evidence for efficacy attribution and threshold drift detection. Furthermore, intervention response feedback data can include, but are not limited to, one or more of the following: physiological indicator response sequences, behavioral compliance records, and subjective experience scores. In this embodiment, personalized health intervention resources identify available instances by accessing a cloud-based intervention resource pool, and their spatiotemporal distribution affects the feasibility of task execution. The dynamic adaptation scheduling module can be an operational unit that performs resource allocation and task scheduling, and can be implemented using multi-objective optimization algorithms or constraint satisfaction solvers.

[0030] Step S600: Based on the intervention response feedback data, perform efficacy attribution analysis and threshold drift detection to drive real-time reconstruction of the physiological homeostatic baseline trajectory, forming a closed-loop dynamic threshold health management strategy.

[0031] The efficacy attribution analysis module can be an analytical unit that identifies significant changes in physiological state during intervention tasks, and can be used to distinguish between effective and ineffective interventions. Furthermore, the efficacy attribution analysis module can be implemented using Shapley value decomposition, causal inference models, or counterfactual reasoning methods. The threshold drift detection module can be an online change detection unit that detects systematic shifts in an individual's physiological baseline, and can be used to identify long-term evolutionary trends in an individual's physiological state. Furthermore, the threshold drift detection module can be implemented using online Bayesian updates, change point detection algorithms, or sliding window statistical control charts. The closed-loop dynamic threshold health management strategy can be an adaptive management framework that enables the health management model to have continuous evolution capabilities. Furthermore, the closed-loop dynamic threshold health management strategy can include, but is not limited to, one or more of the following: baseline trajectory update strategies, risk model retraining strategies, and intervention strategy iteration strategies. In this embodiment, the efficacy attribution analysis module and the threshold drift detection module collaboratively output a correction signal that drives the baseline trajectory update, and its output serves as the input to the real-time reconstruction module. The real-time reconstruction module can be an online learning unit that dynamically updates the physiological steady-state baseline trajectory, implemented using incremental neural network updates or online Gaussian process regression.

[0032] Taking the dynamic management of individuals with prediabetes as an example, the intelligent health management method integrating an AI big model in this embodiment can be as follows: A middle-aged individual continuously wears a multi-parameter wearable device to collect time-series data on their blood glucose, heart rate variability, and activity level; and undergoes abdominal infrared thermal imaging and ultrasound elastography once a week. The AI ​​big model identifies abnormalities in the nocturnal blood glucose fluctuation rhythm through the biorhythm analysis module, and detects a long-term downward trend in insulin sensitivity through the steady-state shift modeling module, generating an individualized physiological steady-state baseline trajectory. The cross-modal functional coupling analysis module correlates liver texture heterogeneity with fasting insulin levels to construct an insulin resistance tissue atlas. The multi-level risk model outputs hourly blood glucose fluctuation risk and weekly insulin sensitivity decline trend. The image recognition module locates high-risk areas for liver fibrosis. The intervention window period projection module determines that the individual will enter a compensatory window period within 3 days and generates a task set: increase morning aerobic exercise, adjust carbohydrate intake ratio, and activate daily reminders from an AI nutritionist. The dynamic adaptation and scheduling module matches the intelligent exercise guidance terminal and nutrition plan push service in the individual's community and executes the plan the following morning. Following the intervention, the system collected continuous 72-hour data on blood glucose fluctuations and sleep quality. The efficacy attribution module found that exercise intervention significantly contributed to blood glucose stability, and the threshold drift detection module identified a shift in the insulin resistance baseline. The real-time reconstruction module updated the physiological homeostasis baseline, the risk model was recalibrated, and a new morning exercise + low-carb breakfast combination strategy was added to the intervention strategy library, forming a closed loop.

[0033] In one embodiment, time-series data of individual multidimensional physiological characteristics are collected; biorhythm analysis and steady-state shift modeling are performed on the time-series data of individual multidimensional physiological characteristics to generate an individualized physiological steady-state baseline trajectory, including: The wearable sensor array and the non-contact body measurement terminal work together to collect time-series data of individual multidimensional physiological signs. The time-series data of individual multidimensional physiological signs include at least heart rate variability, skin conductance response, respiratory tidal volume, microcirculatory blood oxygen saturation and sublingual microvascular flow velocity. Wearable sensor arrays can be non-invasive systems used to simultaneously acquire multimodal physiological signals. It is understood that wearable sensor arrays, by integrating photoplethysmography (PPG) wave sensor arrays, skin conductance microelectrode arrays, etc., can achieve continuous monitoring of heart rate variability and skin conductance response. Furthermore, as a source of multidimensional physiological sign time-series data, wearable sensor arrays work in conjunction with non-contact body measurement terminals to fill the gaps in high-frequency signal acquisition.

[0034] The non-contact body measurement terminal can be a physiological data acquisition device discreetly embedded in smart clothing or an oral patch. In this embodiment, the non-contact body measurement terminal uses an invisible tidal volume sensing fabric and an oral micro-blood flow patch to achieve non-contact acquisition of tidal volume and sublingual microvascular flow velocity. Furthermore, the non-contact body measurement terminal obtains microcirculatory blood oxygen saturation through infrared microvascular imaging technology, forming data complementarity with the wearable sensor array.

[0035] Diurnal rhythm phase decoupling and autonomic nerve tension spectrum decomposition are performed on individual multidimensional physiological sign time series data to generate individualized biological rhythm parameter matrices.

[0036] The circadian rhythm phase decoupling can be a signal processing procedure that separates the rhythmic components of multidimensional physiological time-series data into phase angle and amplitude parameters. This process employs Fourier harmonic analysis combined with a nonlinear phase synchronization algorithm to extract the periodic components of heart rate variability and respiratory tidal volume. Furthermore, the circadian rhythm phase decoupling serves as a direct input to the personalized biorhythm parameter matrix, and its output parameters are used as initial conditions for constructing the physiological homeostatic dynamic equation. Autonomic tension spectrum decomposition can be a process of nonnegatively linearly decomposing the sympathetic and parasympathetic active components in multidimensional physiological signals. In this embodiment, this process assesses signal randomness based on Lempel-Ziv complexity and combines nonnegative matrix decomposition to perform low-rank decomposition of heart rate variability and skin conductance response. Furthermore, the results of autonomic tension spectrum decomposition and circadian rhythm phase decoupling together constitute the core components of the personalized biorhythm parameter matrix. The personalized biorhythm parameter matrix can be a high-dimensional parameter set integrating circadian rhythm phase parameters and autonomic tension components. Understandably, this matrix is ​​formed by time-aligning and stitching together the outputs of rhythm phase decoupling and tension spectrum decomposition. Furthermore, the individualized biological rhythm parameter matrix serves as the driving input for the steady-state attractor topology, and its changes directly affect the generation of the steady-state shift entropy value sequence.

[0037] Physiological homeostatic dynamic equations are constructed based on individualized biological rhythm parameter matrices, and the topology of the homeostatic attractor is solved.

[0038] The physiological homeostatic dynamics equation can be a set of nonlinear differential equations or a graph-structured model describing the long-term evolutionary behavior of an individual physiological system. Understandably, this equation simulates the chaotic behavior of the steady-state trajectory or the regulatory relationships between organ functional modules by using an individualized biorhythm parameter matrix as the driving variable. Furthermore, the physiological homeostatic dynamics equation receives the individualized biorhythm parameter matrix as input and outputs a steady-state attractor topology as a geometric representation of its long-term stable state. The steady-state attractor topology can be the geometric form in which the physiological homeostatic dynamics equation converges in the state space. Understandably, this structure is identified through numerical integration or graph embedding analysis, identifying its Lyapunov exponent, fractal dimension, and stable domain boundary. Furthermore, the steady-state attractor topology serves as the basis for calculating the steady-state offset entropy sequence, and changes in its topological parameters are quantified as offsets.

[0039] Long-term steady-state offset quantization is performed using the steady-state attractor topology to generate a sequence of steady-state offset entropy values.

[0040] The steady-state shift entropy sequence can be an entropy evolution sequence reflecting the degree of attractor structure shift over time. Understandably, this sequence is calculated using Kullback-Leibler divergence or topological entropy algorithms to determine the distributional difference between the attractor structure and the baseline. Furthermore, the steady-state shift entropy sequence is fused with the individualized circadian rhythm parameter matrix as one of the training objectives for an individualized physiological steady-state baseline generator.

[0041] By fusing steady-state offset entropy sequences with individualized biological rhythm parameter matrices, an individualized physiological homeostasis baseline generator is trained.

[0042] The personalized physiological homeostasis baseline generator can be a trajectory generation model based on a deep learning architecture. Understandably, this generator employs a Transformer or graph autoencoder architecture, taking a rhythm parameter matrix and entropy sequence as input to output a dynamic confidence interval and a pathological perturbation sensitive band. Furthermore, the personalized physiological homeostasis baseline generator receives a steady-state shift entropy sequence and a personalized biological rhythm parameter matrix as input, and outputs a personalized physiological homeostasis baseline trajectory.

[0043] The individualized physiological homeostasis baseline trajectory is output through the individualized physiological homeostasis baseline generator.

[0044] The dynamic confidence interval can be a probability region reflecting the boundary of normal physiological fluctuations. Understandably, this interval is calculated using Monte Carlo Dropout or Bayesian confidence propagation to determine the quantile boundaries of multiple trajectories. Furthermore, the dynamic confidence interval, together with the pathological perturbation sensitive zone, constitutes a component of the individualized physiological homeostatic baseline trajectory.

[0045] A pathological disturbance sensitive zone can serve as a warning time interval indicating impending decompensation of the physiological system. This zone is generated by determining whether the steady-state deviation entropy value sequence continuously exceeds a preset threshold and spans 30% of the diurnal cycle. Furthermore, the pathological disturbance sensitive zone and the dynamic confidence interval together constitute a two-dimensional criterion for intervention triggering. Identifying the pathological disturbance sensitive zone within the individualized physiological steady-state baseline trajectory can be achieved by using a change point detection algorithm to locate the starting point of a continuous increase in entropy value and extend the sensitive zone's range. For example, this operation can set the entropy threshold to the baseline mean + 2 standard deviations, marking the sensitive zone when it exceeds the threshold for 8 consecutive hours and spans 30% of the diurnal cycle. This operation further enhances the technical effectiveness of identifying early warning zones and triggering precise interventions.

[0046] Taking dynamic baseline modeling of individuals at early risk of metabolic syndrome as an example, the method in this embodiment can be as follows: Tidal volume and skin conductance are collected using smart clothing; microcirculatory blood oxygenation and blood flow velocity are collected using a sublingual microvascular imaging patch; and heart rate variability is collected simultaneously using a wearable wristband. The biorhythm analysis module constructs an individualized biorhythm parameter matrix through Fourier harmonic analysis and NMF decomposition, revealing a continuous decrease in the nocturnal parasympathetic recovery rate. The steady-state shift modeling module solves for the decreasing trend of attractor fractal dimension based on the Lorenz equation, generating a steady-state shift entropy value sequence. The baseline generation network fuses the rhythm parameters and entropy values, outputting a trajectory containing dynamic confidence intervals and sensitive zones. When the entropy value exceeds the standard for five consecutive days and spans 30% of the diurnal cycle, the system identifies the autonomic nervous system imbalance sensitive zone, marks the decrease in microcirculatory blood oxygenation in the liver area as high risk, and triggers AI nutritionist intervention.

[0047] In one embodiment, continuous dynamic body measurement image data of an individual is acquired; based on the continuous dynamic body measurement image data of the individual, tissue microstructure texture characterization extraction and cross-modal functional coupling analysis are performed to generate a dynamic atlas of organ functional reserves, including: Continuous dynamic body measurement data of an individual were acquired simultaneously using a multispectral skin-mucosal imaging system and optical coherence tomography of the tongue surface microstructure. The multispectral skin-mucosal imager can be an imaging device capable of simultaneously acquiring the reflection and absorption characteristics of skin and mucosal regions across multiple spectral bands. It is understood that the multispectral skin-mucosal imager, through a multi-wavelength LED light source and a high-sensitivity spectral sensor array, simultaneously acquires multispectral image sequences from the visible to near-infrared bands, thereby providing high-resolution spatial distribution data on superficial tissue vascular distribution, pigmentation, and microcirculatory oxygenation status. Furthermore, the multispectral skin-mucosal imager can be used in conjunction with optical coherence tomography (OCT) of the tongue's microstructure to form a multimodal acquisition source for continuous dynamic body imaging data, providing superficial vascular structure input for the construction of microcirculatory topology maps. For example, the multispectral skin-mucosal imager can include, but is not limited to, one or more of the following: visible-near-infrared multispectral imager, polarized light skin imager, and fluorescence mucosal imager.

[0048] Optical coherence tomography (OCT) of the tongue surface microstructure can be an optical device that utilizes the principle of low-coherence interference to perform submicron-level tomographic imaging of tongue surface tissue. It can be understood that OCT of the tongue surface microstructure generates backscattered interference signals in the tongue mucosa tissue by scanning a near-infrared laser beam, reconstructing a three-dimensional microstructure tomographic sequence. This allows for the acquisition of high-resolution three-dimensional structural information on the submucosal microvascular network, epithelial thickness variations, and tissue density heterogeneity of the tongue body. Furthermore, OCT of the tongue surface microstructure can be acquired synchronously with a multispectral skin-mucosal imager to provide structural depth information on the tongue's microcirculation, supporting the three-dimensional reconstruction of the microcirculation topology. For example, OCT of the tongue surface microstructure can include, but is not limited to, one or more of the following: a frequency-domain OCT tongue surface imaging system, an ultra-high-speed OCT mucosal scanner, and a Doppler OCT blood flow imaging module.

[0049] Multi-scale vascular network skeleton extraction was performed on individual continuous dynamic body imaging data to construct a microcirculation topology connection map.

[0050] The microcirculation topology connectivity map can be a graph structure representation abstracted from the structure of the tissue microvascular network, containing the topological attributes of nodes and edges. Understandably, the microcirculation topology connectivity map, through skeletonization processing of multi-scale vascular images, extracts the vascular centerline, uses branch points as nodes and vascular segments as edges, and constructs a weighted graph model. This transforms the morphological vascular structure into a computable network topology index for quantifying perfusion connectivity and redundancy. Furthermore, the microcirculation topology connectivity map can serve as a structural input for tissue perfusion-metabolic coupling modeling, jointly constructing an organ-level functional reserve dynamic atlas with infrared thermal metabolic images. For example, the microcirculation topology connectivity map can include, but is not limited to, one or more of the following: vascular branch density map, path connectivity map, and blood flow path entropy map.

[0051] The multi-scale vascular network skeleton extraction module can be an image processing unit used to automatically extract microvascular centerlines and construct topological structures from continuous dynamic volumetric images. Understandably, this module employs Hessian matrix enhancement, morphological skeletonization, and graph-based connectivity analysis algorithms to separate the vascular network skeleton from OCT and multispectral images, thereby generating the structural basis for a microcirculation topological connectivity map. Furthermore, this module can serve as a pre-processing module for the microcirculation topological connectivity map, and its output directly affects the accuracy of subsequent perfusion-metabolism coupling modeling.

[0052] By integrating microcirculation topology maps with infrared thermo-metabolic images, tissue perfusion-metabolic coupling modeling is performed to generate dynamic atlases of organ-level functional reserves.

[0053] Infrared thermometabolic imaging can be thermal imaging data reflecting the distribution of thermal radiation on the tissue surface and the dynamics of local metabolic heat production. It can be understood that infrared thermometabolic imaging acquires a two-dimensional thermal image sequence of the tissue surface temperature field changing over time using a non-contact infrared thermal imager, thereby characterizing the intensity of tissue metabolic activity and local heat dissipation patterns, serving as an indirect indicator of functional requirements. Furthermore, infrared thermometabolic imaging can be spatially aligned with microcirculation topology maps, serving as input for metabolic demand and participating in tissue perfusion-metabolic coupling modeling.

[0054] Tissue perfusion-metabolism coupling modeling can be a mathematical model that establishes a nonlinear dynamic relationship between tissue blood supply structure and local metabolic demand. Understandably, this model employs graph neural networks or nonlinear regression models to jointly map the node degree and path length of the microcirculation topology graph with the heat flux density of infrared thermometabolic images, thereby quantifying the functional matching between blood supply capacity and tissue metabolic demand. Furthermore, tissue perfusion-metabolism coupling modeling can be a core computational module for generating dynamic maps of organ-level functional reserves, and its output directly determines the basis for calculating functional redundancy and compensatory attenuation slope. For example, tissue perfusion-metabolism coupling modeling can include, but is not limited to, one or more of the following: perfusion-metabolism matching index model, thermo-blood flow synergistic attenuation model, and microcirculation supply-demand imbalance scoring model.

[0055] An organ-level functional reserve dynamic atlas can be a dynamic atlas characterizing the evolution of functional reserve capacity over time in the three dimensions of structure, perfusion, and metabolism, with organs as the unit. It can be understood that the organ-level functional reserve dynamic atlas, based on tissue perfusion-metabolism coupling modeling results, aggregates the functional matching degree of local regions to form an organ-scale temporal functional reserve matrix, thereby providing a systematic and quantifiable dynamic representation of organ functional reserves. Furthermore, the organ-level functional reserve dynamic atlas can serve as input to modules calculating functional redundancy and compensation slope, and also as the analysis object for identifying key nodes of functional compensation imbalance. For example, the organ-level functional reserve dynamic atlas may include, but is not limited to, one or more of the following: liver functional reserve dynamic atlas, kidney functional reserve dynamic atlas, and myocardial functional reserve dynamic atlas.

[0056] Based on the dynamic map of organ-level functional reserves, the functional redundancy index and compensatory attenuation slope of each target organ are calculated.

[0057] The functional redundancy index can be a quantitative indicator measuring the additional functional reserve capacity that an organ can mobilize under stress conditions. Understandably, the functional redundancy index is based on an organ-level functional reserve dynamic map, calculating the ratio of the difference between the current state and the theoretical maximum reserve capacity, thus reflecting the size of the organ's current functional buffer space and used to determine whether the compensatory potential is sufficient. Furthermore, the functional redundancy index can be used, along with the compensatory attenuation slope, as an input variable in cross-organ correlation analysis to identify key nodes of functional compensation imbalance. For example, the functional redundancy index may include, but is not limited to, one or more of the following: blood flow redundancy index, metabolic reserve reserve index, and structural redundancy potential index.

[0058] The compensatory decay slope can be a dynamic rate of change index characterizing the rate of decline in organ functional reserve capacity over time. Understandably, the compensatory decay slope is obtained by performing linear or nonlinear regression on the time series of the functional redundancy index to extract the slope parameter of its downward trend, thereby quantifying the acceleration of organ functional decline and identifying organs about to enter the compensatory collapse stage. Furthermore, the compensatory decay slope can be a core two-dimensional input, alongside the functional redundancy index, for identifying pivotal nodes. For example, the compensatory decay slope can include, but is not limited to, one or more of the following: short-term compensatory decay rate, medium-term compensatory decay rate, and long-term compensatory decay rate.

[0059] Cross-organ correlation analysis was conducted on the functional redundancy index and the compensatory attenuation slope to identify the key nodes of functional compensation imbalance.

[0060] In this context, the functional compensation imbalance hub node can be an organ node that plays a key mediating role in a multi-organ functional compensation network and whose compensatory capacity declines first. Understandably, by employing graph neural networks or dynamic Bayesian networks, the functional compensation imbalance hub node analyzes the cross-organ correlation between functional redundancy and compensatory decline slope of each organ, identifying nodes with high centrality and high decline gradients. This allows for the localization of early-stage drivers of systemic imbalances in chronic diseases, revealing the starting point of the compensation chain break. Furthermore, the functional compensation imbalance hub node can serve as an anchor point for a back-mapping module, driving the spatial localization generation of a dynamic map of organ functional reserves. For example, the functional compensation imbalance hub node can include, but is not limited to, one or more of metabolic compensation hubs, blood flow compensation hubs, and structural compensation hubs.

[0061] Using the hub node as the anchor point, the data is back-mapped to the anatomical structure space to generate a dynamic atlas of organ function reserves, which includes a set of spatial positioning labels.

[0062] The spatial location tag set can be a set of anatomical spatial coordinates and regional semantic labels attached to an organ function reserve dynamic atlas. Understandably, the spatial location tag set maps the identification results of functional compensation imbalance hub nodes back to the three-dimensional spatial coordinates of the original multispectral and OCT images through image registration technology, thereby achieving a precise correspondence between functional abnormalities and anatomical locations. Furthermore, the spatial location tag set can serve as an enhanced attribute of the organ function reserve dynamic atlas, directly used for the precise localization of high-risk target organ regions. For example, the spatial location tag set can include, but is not limited to, one or more of organ region localization tags, tissue layer localization tags, and microvascular network localization tags.

[0063] Anatomical structure reverse mapping can be a registration unit used to reverse-locate functionally abnormal nodes back to the anatomical space of the original image. Understandably, this module uses rigid registration and a non-rigid deformation model to map the abstract positions of pivot nodes in the functional atlas back to the three-dimensional coordinate space of multispectral and OCT images, thereby generating a dynamic atlas of organ functional reserves with a set of spatial localization labels. Furthermore, this module can take functionally compensated imbalance pivot nodes as input and output a set of spatial localization labels, completing a closed-loop mapping from function to structure.

[0064] Using hub nodes as anchor points, the data is back-mapped to the anatomical structural space to generate a dynamic atlas of organ function reserves. This can be achieved by receiving and parsing user-submitted data through a user interface. Furthermore, this operation can be performed using non-rigid registration based on feature points to map hub node coordinates to pixel coordinates in OCT and multispectral images, or by employing a deep learning registration network to directly learn the end-to-end mapping relationship between functionally abnormal regions and anatomical regions. This enables precise inversion of functional abnormality location to anatomical sites, making the functional atlas spatially executable. Taking the early compensatory collapse warning of non-alcoholic fatty liver disease as an example, the intelligent body composition analysis and health management method integrating an AI large model in this embodiment can be as follows: An obese individual undergoes simultaneous multispectral skin imaging and tongue OCT scanning. The system extracts the density of sublingual microvascular branches and the connectivity of the skin capillary network to construct a microcirculation topology map; simultaneously, infrared thermograms of the tongue surface are acquired, revealing local metabolic heat accumulation areas; tissue perfusion-metabolic coupling modeling shows a severe imbalance between microcirculation perfusion and metabolic demand in the liver region; dynamic organ-level functional reserve atlas shows that the liver functional redundancy has decreased to 30%, and the compensatory attenuation slope continues to rise; cross-organ correlation analysis identifies the liver as a metabolic compensation hub node, whose... The compensatory pressure is significantly higher than that of the kidneys and pancreas; the anatomical structure reverse mapping module locates this pivotal node to the S5 region of the right lobe of the liver; the system generates a spatial positioning label set "significant compensatory attenuation in the S5 region of the liver" and inputs it into the risk model; based on this, the risk model upgrades the original "mildly abnormal liver function" to an early warning of "the liver compensatory system is about to collapse"; the intervention window prediction module triggers the combined intervention task of "low-fat diet + liver thermotherapy + blood flow enhancement exercise" 72 hours in advance; the dynamic scheduling module matches the nearest community liver rehabilitation station and AI nutritionist service; the feedback data after the intervention shows that the liver area's thermo metabolism has returned to normal and the redundancy has increased; the system updates the baseline trajectory, closes the original compensatory attenuation prediction, and forms a closed loop.

[0065] In one embodiment, multi-scale vascular network skeleton extraction is performed on continuous dynamic body imaging data of an individual to construct a microcirculation topology connectivity map, including: Perform multispectral channel light scattering consistency calibration on continuous dynamic body measurement image data of individuals; Among them, inter-channel light scattering consistency calibration can be a calibration mechanism to eliminate geometric and intensity distortions caused by differences in the scattering characteristics of different wavelengths of light within tissues in multispectral imaging. It can be understood that this calibration mechanism establishes an inter-channel scattering transfer function by acquiring the response of a standard scattering reference plate in each spectral channel, and performs spatial alignment and radiation intensity normalization on the original image. Furthermore, inter-channel light scattering consistency calibration can include, but is not limited to, one or more of the following: radiation calibration mapping, spatial registration compensation, and scattering path normalization. In this embodiment, this calibration mechanism serves as a preprocessing step before pixel-level segmentation of microvascular networks, and its output directly determines the reliability of the enhanced texture image.

[0066] An adaptive local contrast enhancement algorithm is used to sharpen the texture of the calibrated image, generating an enhanced texture image.

[0067] The adaptive local contrast enhancement algorithm can be a nonlinear enhancement method that dynamically adjusts contrast based on the statistical characteristics of the local image. Understandably, this algorithm calculates the entropy and variance of the local gray-level distribution within a sliding window, applies nonlinear stretching to low-contrast regions, and suppresses over-enhancement of high-contrast regions. Furthermore, the adaptive local contrast enhancement algorithm can include, but is not limited to, one or more of adaptive histogram equalization, contrast-limited adaptive histogram equalization, and local gradient enhancement mapping. In this embodiment, the algorithm is applied to an image after multispectral consistency calibration to generate an enhanced texture image, which serves as input for directional texture feature extraction.

[0068] A multi-scale Gabor filter bank was constructed based on enhanced texture images to extract directional texture features of tissue microstructure.

[0069] The enhanced texture image can be high-fidelity image data with clear microvascular structure edges and preserved background noise after multispectral calibration and local contrast enhancement. It can be understood that this image consists of a grayscale matrix sequence output from the pixel-by-pixel processing of the calibration image by an adaptive local contrast enhancement algorithm. Furthermore, the enhanced texture image can be one or more of the following, including but not limited to enhanced vascular edge images, microstructure-preserving texture images, and noise-suppressed enhancement maps. In this embodiment, the enhanced texture image serves as the direct input to a multi-scale Gabor filter bank, determining the accuracy and stability of the directional texture features.

[0070] The multi-scale Gabor filter bank can be a set of Gabor kernels composed of different scales and orientation parameters, used to simulate the sensitivity of human vision to texture orientation and frequency. Understandably, this filter bank constructs a Gabor filter matrix containing multiple spatial frequencies (e.g., 0.05–0.3 cycles / pixel) and orientations (e.g., 0°–165°, stride 15°), which is then convolved layer by layer with the input image. Furthermore, the multi-scale Gabor filter bank can include, but is not limited to, one or more of the following: low-frequency directional response group, mid-frequency edge response group, and high-frequency texture response group. In this embodiment, this filter bank is used to enhance texture images, outputting directional texture features as supervision signals or feature inputs for the lightweight U-Net segmentation model.

[0071] The directional texture features of tissue microstructures can be a multidimensional vector set describing the orientation and arrangement of microvessels, generated by the response of Gabor filter banks. Understandably, this feature extracts the response amplitude and phase of each pixel on all Gabor filters, forming a joint direction-scale feature vector. Furthermore, the directional texture features of tissue microstructures can include, but are not limited to, one or more of the following: direction entropy feature vector, scale response consistency index, and local texture anisotropy. In this embodiment, this feature serves as the input feature or auxiliary supervision signal for the lightweight U-Net segmentation model, improving the segmentation's ability to identify small, broken blood vessels.

[0072] A lightweight U-Net segmentation model is trained using directional texture features to achieve pixel-level segmentation of microvascular networks.

[0073] The lightweight U-Net segmentation model can be a parametrically compressed encoder-decoder neural network specifically designed for real-time microvessel segmentation. Understandably, this model is based on the standard U-Net architecture and employs depthwise separable convolution, channel pruning, and knowledge distillation techniques to compress the model size, reducing computational overhead while maintaining segmentation accuracy. Furthermore, the lightweight U-Net segmentation model can include, but is not limited to, one or more of depthwise separable U-Net, channel-compressed U-Net, and knowledge-distilled U-Net. In this embodiment, the model receives enhanced texture images and directional texture features as input and outputs pixel-level segmentation results of the microvessel network, serving as the core execution unit for structural analysis.

[0074] The pixel-level segmentation result of the microvascular network can be a binary or probabilistic mask output by a lightweight U-Net model, identifying whether each pixel belongs to a microvascular structure. Understandably, this result involves forward inference frame-by-frame on the enhanced texture image to generate a segmentation probability map with the same spatial dimensions as the original image, which is then thresholded to obtain a binary mask. Furthermore, the pixel-level segmentation result of the microvascular network can include, but is not limited to, one or more of the following: high-confidence vessel masks, low-confidence fracture region maps, and edge blurring response maps. In this embodiment, this result is the direct input to the graph theory optimization module, and its completeness and accuracy determine the quality of the final topology connectivity graph.

[0075] Graph theory optimization is performed on the segmentation results to remove pseudo-connected branches and complete the broken blood vessel segments, outputting a microcirculation topology connection graph.

[0076] The graph theory optimization module can be a graph structure processing unit that performs topological repair and purification on the pixel-level segmentation results. It can be understood that this module models the segmentation results as an undirected graph, with nodes representing blood vessel endpoints and branching points, and edges representing continuous blood vessel segments, and corrects them through connectivity analysis and topological constraint rules. Furthermore, the graph theory optimization module can include, but is not limited to, one or more of the following: a connected component purification module, a path completion module, and a topological consistency verification module. In this embodiment, this module receives the pixel-level segmentation results and outputs a micro-circulation topology connection graph, which is the final optimization step in the structure analysis process.

[0077] Falsely connected branches can be non-physiological vascular structural connections formed by noise, imaging artifacts, or misjudgments of tissue shadows. Understandably, this branch is identified through graph theory analysis as an isolated subgraph with abnormal connectivity, excessively short path length, or no functional association with the main vessel. Furthermore, falsely connected branches can include, but are not limited to, one or more of noise-induced false connections, shadowed false vessel segments, and motion artifact connected components. In this embodiment, this branch is the target of elimination by the graph theory optimization module, as it reduces the physiological reliability of the microcirculation topology connectivity graph. Broken vessel segments can be interruptions in the physical continuity of vascular structures caused by low contrast, motion blur, or insufficient imaging resolution. Understandably, this segment is identified by detecting missing connections between nodes, path interruptions, or abnormal degrees (such as too many terminal nodes). Furthermore, broken vessel segments can include, but are not limited to, one or more of local broken segments, edge-blurred interruptions, and missing connections at bifurcation points. In this embodiment, this segment is the target of repair by the graph theory optimization module, and its completion level determines the physiological representativeness of the topology graph.

[0078] A microcirculation topology connectivity graph can be a quantitative graph structure composed of nodes (branching points / endpoints) and edges (vascular segments), containing topological attributes and physical parameters. It can be understood that the segmentation result of this graph structure after processing by a graph theory optimization module is transformed into a graph data structure with attributes such as node coordinates, edge lengths, branch order, and local density. Furthermore, the microcirculation topology connectivity graph can include, but is not limited to, one or more of the following: oriented vascular network graphs, branch density topology graphs, and path connectivity graphs. In this embodiment, this graph structure is the final output of this operation sequence, providing a computable structural basis for subsequent tissue perfusion-metabolic coupling modeling.

[0079] Taking the quantitative detection of early diabetic microvascular complications as an example, the intelligent physical fitness assessment and health management method integrating an AI large model in this embodiment can be as follows: Diabetic patients undergo simultaneous acquisition of multispectral skin imaging and tongue surface OCT. Due to the difference in scattering between near-infrared and visible light bands, the position of the sublingual vessels in the original images is offset. The system first performs light scattering consistency calibration between multispectral channels to align the vessel centerline in multiple bands. Subsequently, an adaptive local contrast enhancement algorithm enhances the grayscale gradient in the vessel edge region to generate enhanced texture images. A multi-scale Gabor filter bank extracts the main vessel orientation and branch angle distribution features to form directional texture vectors. The lightweight U-Net segmentation model combines these features to accurately identify the microvascular network even in low-contrast regions. Even with local ruptures caused by microthrombi, the graph theory optimization module identified and removed three pseudo-connected branches (all noise clumps) formed by imaging artifacts. Simultaneously, using the continuity constraint of vessel diameter, it reconstructed three connection paths between the two ruptured vessel segments that conformed to physiological diameter changes. The final generated microcirculation topology connection graph contained 142 nodes and 218 edges, with an average branch order of 2.7 and a clustering coefficient of 0.41, showing significant differences from the healthy control group. When this graph was input into the tissue perfusion-metabolism coupling modeling module, it was found that the microvascular network density decreased by 28% and the path redundancy decreased, indicating early diabetic microvascular remodeling. Based on this, the system triggered the "microcirculation strengthening exercise + antioxidant intervention" task three months in advance, forming a closed-loop management.

[0080] In one embodiment, a multi-level chronic disease progression risk gradient model is constructed by integrating individualized physiological homeostasis baseline trajectories and dynamic organ function reserve maps, outputting multi-timescale disease activity prediction curves, including: By mapping individualized physiological homeostasis baseline trajectories to the spatiotemporal coordinate system of organ functional reserve dynamic map, a cross-modal homeostasis-function joint representation space is constructed. The cross-modal homeostasis-functional joint representation space can be a high-dimensional joint state space formed by mapping individualized physiological homeostasis baseline trajectories and organ functional reserve dynamic maps to a unified spatiotemporal coordinate system. Understandably, this space uses nonlinear alignment algorithms (such as differentiable time series registration and spatiotemporal attention alignment networks) to perform coordinate mapping and feature fusion of temporal physiological trajectories and spatially distributed functional maps within a unified spatiotemporal grid. Furthermore, as the foundation for constructing a multidimensional risk potential field, the dimension and alignment accuracy of the cross-modal homeostasis-functional joint representation space directly affect the biointerpretability and computational stability of the risk potential field. For example, the cross-modal homeostasis-functional joint representation space can include, but is not limited to, one or more of the following: temporal-spatial alignment embedding space, multimodal dynamic embedding space, and physiological-functional coupled manifold.

[0081] A spatiotemporal coordinate system can serve as a unified four-dimensional reference framework representing the temporal evolution of physiological homeostasis and the spatial distribution of organ function. This coordinate system, with time as the horizontal axis and organ anatomical space as the three-dimensional coordinates, constructs a four-dimensional mapping space, achieving alignment between the two types of data through interpolation and registration. Furthermore, as the geometric basis for fusing the baseline trajectory of physiological homeostasis with the dynamic map of organ functional reserves, the coordinate alignment method of the spatiotemporal coordinate system determines the geometric structure of the joint representation space.

[0082] Within the cross-modal steady-state-functional joint characterization space, a multidimensional risk potential field for the progression of chronic diseases is defined. The potential field is composed of the inflammatory factor fluctuation entropy, the metabolite accumulation gradient, and the amplitude of neuroendocrine axis perturbation. The multidimensional risk potential field can be a scalar potential function space representing the tendency of chronic disease progression, composed of multiple biodynamic variables, within a cross-modal steady-state-functional joint representation space. Understandably, this potential field, based on prior biomedical knowledge, uses three dimensions—inflammatory factor fluctuation entropy, metabolite accumulation gradient, and neuroendocrine axis perturbation amplitude—as independent variables of the potential function, constructing a potential surface through a weighted nonlinear combination. Furthermore, as the driving field of the disease activity evolution manifold, the shape of the multidimensional risk potential field determines the distribution of local minima and critical points along the path. For example, the multidimensional risk potential field can include, but is not limited to, one or more of the following: inflammation-dominated potential field, metabolism-dominated potential field, and neuroendocrine-dominated potential field.

[0083] Inflammatory factor fluctuation entropy can be a quantitative indicator reflecting the intensity of disorder and non-steady-state fluctuations of inflammation-related molecular signals in the body over time. Understandably, this indicator is based on time-series detection data from serum or tissue fluid, using sample entropy or fuzzy entropy algorithms to calculate the complexity of inflammatory marker sequences such as IL-6, TNF-α, and CRP. Furthermore, as one of the input variables in a multidimensional risk potential field, an increase in inflammatory factor fluctuation entropy corresponds to a rise in local energy within the potential field, indicating system instability. For example, inflammatory factor fluctuation entropy can include, but is not limited to, one or more of the following: acute inflammation fluctuation entropy, chronic low-grade inflammation entropy, and entropy related to circadian rhythm disruption.

[0084] The metabolite accumulation gradient can describe the rate of change in spatial concentration of metabolites due to their non-uniform distribution in tissue space. Understandably, this gradient is based on the metabolite distribution mapping in the dynamic map of organ functional reserves, and the local concentration gradient vector field is calculated using the finite difference method or Gaussian process regression. Furthermore, the metabolite accumulation gradient, in conjunction with the inflammatory factor fluctuation entropy, determines the spatial distribution pattern of the potential energy field and the direction of the energy gradient. For example, the metabolite accumulation gradient can include, but is not limited to, one or more of the following: the lipid accumulation gradient in the liver region, the lactate gradient in muscle tissue, and the insulin deposition gradient in the pancreas.

[0085] The amplitude of neuroendocrine axis perturbation can reflect the intensity of disturbances in the hypothalamic-pituitary-adrenal (HPA) neuroendocrine regulatory system that deviate from normal rhythms. Understandably, this amplitude is calculated based on time-series data such as cortisol rhythm, ACTH pulse frequency, and sympathetic / parasympathetic balance index, determining its deviation from a standard rhythm template in terms of magnitude and phase shift. Furthermore, the amplitude of neuroendocrine axis perturbation affects the temporal evolution rate of the potential energy field; an increase in amplitude reduces the system's ability to recover from potential energy depressions. For example, the amplitude of neuroendocrine axis perturbation may include, but is not limited to, one or more of the following: HPA axis phase shift amplitude, autonomic nervous system instability amplitude, and circadian rhythm phase delay amplitude.

[0086] Based on a multidimensional risk potential field, a Lagrange mechanical framework is used to simulate the disease progression path and generate a disease activity evolution manifold. The disease activity evolution manifold can be the set of optimal trajectories for the evolution of disease states over time, derived from the Lagrange mechanical framework within a multidimensional risk potential energy field. Understandably, this manifold treats individual physiological states as moving particles in a potential energy field, obtaining the system's evolutionary path by solving the Euler-Lagrange equations using the principle of least action. Furthermore, the disease activity evolution manifold serves as input to the causal inference enhancement module, and its topological structure determines the identifiability of key inflection points. For example, the disease activity evolution manifold can be, but is not limited to, one or more of the following: a steady-state maintenance manifold, a compensatory transition manifold, and a functional failure manifold.

[0087] The Lagrange mechanical framework can be considered a set of dynamic equations describing the evolution of a system along a path of minimum action in a potential energy field. Understandably, this framework treats the individual's physiological state as a generalized coordinate system, defines the Lagrangian function as the difference between kinetic and potential energy, and derives the state evolution trajectory through the Euler-Lagrange equations. Furthermore, the Lagrange mechanical framework serves as a mechanism for generating the manifold of disease activity evolution, with a multidimensional risk potential energy field as input and the manifold trajectory as output.

[0088] Input a pre-trained multimodal large model, which integrates clinical guideline knowledge graphs, real-world cohort survival data, and individual gene polymorphism information to enhance causal inference of the disease activity evolution manifold; The multimodal large model can be a deep neural network system that integrates heterogeneous medical knowledge from multiple sources, possessing cross-modal reasoning and causal structure learning capabilities. Understandably, this model integrates clinical guideline knowledge graphs, real-world cohort survival data, and gene polymorphism information during the pre-training phase, employing graph attention mechanisms and structured causal learning for joint representation learning. Furthermore, the multimodal large model serves as the core engine of the causal inference enhancement module, taking the disease activity evolution manifold as input and outputting the enhanced prediction curve and key nodes.

[0089] A clinical guideline knowledge graph can be a structured medical knowledge network that expresses the logic of chronic disease diagnosis and treatment, pathological mechanisms, and intervention recommendations. Understandably, this graph extracts entities (such as disease stages, biomarkers, and treatment plans) and relationships (such as "cause," "contraindication," and "recommendation") from authoritative clinical guideline texts to construct a directed knowledge graph. Furthermore, the clinical guideline knowledge graph serves as prior knowledge input for a multimodal large-scale model, constraining the rational path space of the evolutionary manifold. For example, a clinical guideline knowledge graph may include, but is not limited to, one or more of the following: a diabetes progression path graph, a hypertension target organ damage graph, and a liver fibrosis staging diagnosis graph.

[0090] Real-world cohort survival data can be long-term follow-up data from large populations, including individual physiological indicators, clinical outcomes, and time information. Understandably, this data is extracted from electronic health records and regional health databases, containing cohort data of chronic disease patients with complete follow-up records, and preprocessed for survival analysis. Furthermore, real-world cohort survival data provides statistical patterns of disease evolution at the population level, enhancing the model's generalization ability to rare pathways. For example, real-world cohort survival data may include, but is not limited to, one or more of the following: cardiovascular event survival cohorts, diabetic nephropathy progression cohorts, and non-alcoholic fatty liver disease with hepatocellular carcinoma transformation cohorts.

[0091] Individual genetic polymorphism information can be a set of single nucleotide polymorphisms (SNPs) and copy number variations (CNVs) reflecting genetic variations in key metabolic, inflammatory, and endocrine pathways. Understandably, this information is obtained through gene chips or whole-exome sequencing to identify the types of variations at gene loci such as APOE, FTO, IL6R, and ACE. Furthermore, individual genetic polymorphism information serves as a personalized input to a multimodal macro-model, influencing the local resistance and bias of pathways in the potential energy field. For example, individual genetic polymorphism information can include, but is not limited to, one or more of the following: metabolic-related gene polymorphisms, inflammation-regulating gene polymorphisms, and neuroendocrine-regulating gene polymorphisms.

[0092] The output includes disease activity prediction curves covering multiple time granularities, with each curve accompanied by uncertainty quantification indicators and key inflection point warning indicators.

[0093] The multi-level temporal granularity disease activity prediction curve can be a probability sequence of disease progression output at different temporal resolutions. Understandably, this curve generates hourly to monthly prediction sequences by multi-scale sampling and probability density estimation of the enhanced disease activity evolutionary manifold. Furthermore, the multi-level temporal granularity disease activity prediction curve is the final output of the causal inference enhancement module, and its shape is influenced by both the evolutionary manifold and knowledge enhancement. For example, the multi-level temporal granularity disease activity prediction curve can include, but is not limited to, one or more of the following: hourly acute fluctuation prediction curve, daily metabolic cycle prediction curve, and monthly functional decline prediction curve.

[0094] Uncertainty quantification indicators can be mathematical measures describing the confidence level of a disease activity prediction curve. Understandably, this indicator outputs a prediction interval, entropy value, or probability density distribution through Monte Carlo Dropout, quantile regression, or a Bayesian neural network. Furthermore, uncertainty quantification indicators are synchronously output by a multimodal large model during causal inference, reflecting the determinism of path deduction. For example, uncertainty quantification indicators may include, but are not limited to, one or more of the following: prediction interval width, prediction entropy value, and posterior distribution dispersion.

[0095] Key inflection point warning markers can be clinically significant bifurcation points or critical state markers identified in the disease activity evolution manifold. Understandably, these markers are automatically labeled with bifurcations, threshold breakthroughs, and irreversible points by detecting abrupt changes in potential gradients, zero-crossing points of the second derivative, or Lyapunov exponent reversals in the evolutionary path. Furthermore, key inflection point warning markers are generated by a causal inference enhancement module based on joint reasoning using the evolutionary manifold and knowledge graph, and are directly used for intervention window prediction. For example, key inflection point warning markers can include, but are not limited to, one or more of the following: bifurcation point warnings, threshold breakthrough point warnings, and steady-state collapse point warnings.

[0096] Taking the physically interpretable prediction of the progression from non-alcoholic fatty liver disease to liver fibrosis as an example, the intelligent physical fitness management method based on an integrated AI big model in this embodiment can be as follows: A patient's liver region is continuously monitored with ultrasound elastography sequences and time-series data of serum IL-6, cortisol, and fasting insulin. The AI ​​big model maps the patient's physiological homeostatic baseline trajectory (long-term increase in insulin resistance) and liver tissue microstructure texture atlas (increased texture entropy in fibrotic areas) to a spatiotemporal coordinate system, constructing a joint representation space. A multidimensional risk potential field is defined, where the inflammatory factor fluctuation entropy reflects the periodic outbreaks of IL-6, the metabolite accumulation gradient characterizes the spatial heterogeneity of intrahepatic lipid deposition, and the neuroendocrine axis perturbation amplitude shows cortisol rhythm disorder. A Lagrange mechanical framework simulates three potential evolutionary pathways: homeostasis maintenance, compensatory transition, and functional failure. A multimodal large-scale model integrates a knowledge graph from clinical guidelines stating that "the key drivers of liver fibrosis progression are persistent inflammation and metabolic burden," survival data from real-world cohorts showing that "IL-6 > 8 pg / mL for 6 weeks predicts accelerated fibrosis," and patient risk genotype information for PNPLA3rs738409. A causal inference enhancement module identifies "IL-6 fluctuation + lipid gradient" as the key driver pair, excluding non-critical variables. A monthly prediction curve is generated, showing a bifurcation point at week 4 and marking it as a "critical point warning." The prediction curve includes a 95% confidence interval and entropy value, allowing physicians to initiate precise intervention (anti-inflammatory diet + exercise prescription) at week 3 to prevent irreversible progression.

[0097] In one embodiment, high-risk target organ regions are identified based on individual continuous dynamic body imaging; combined with multi-timescale disease activity prediction curves, personalized intervention window periods are extrapolated for each target organ region, generating a set of precise intervention tasks by organ and stage, including: Based on the spatial localization tag set of the dynamic atlas of organ function reserve in continuous dynamic body measurement images, four types of high-risk target organ regions, namely liver, kidney, pancreas and retina, are identified. The spatial location label set of the dynamic organ function reserve atlas can be a structured index set carrying spatial coordinates and functional levels, which can be used to achieve precise mapping of functional states in anatomical space. It can be understood that the spatial location label set of the dynamic organ function reserve atlas can output the functional reserve score of each voxel through a semantic segmentation network and map it to a standard anatomical coordinate system. Furthermore, the spatial location label set of the dynamic organ function reserve atlas can include, but is not limited to, liver segment functional label sets, renal cortex functional label sets, and pancreatic islet functional label sets.

[0098] Pathological imaging feature transfer learning was performed on each target organ region to identify subclinical lesion markers of early fibrosis, microaneurysms, and β-cell apoptosis; Pathological image feature transfer learning can be a deep learning unit that transfers lesion patterns from pathological sections to non-invasive images, and can be used to drive the automatic extraction of subclinical lesion biomarkers. Understandably, a pathological image feature transfer learning module can be fine-tuned on a pathological image database using a pre-trained convolutional network to extract texture and morphological feature codes for fibrosis, microaneurysms, and apoptosis. Furthermore, the pathological image feature transfer learning module may include, but is not limited to, a fibrosis texture transfer submodule, a microvascular morphology transfer submodule, and a β-cell apoptosis and nuclear fragmentation transfer submodule.

[0099] Subclinical lesion biomarkers are early functional decline signals that occur before macroscopic morphological changes appear in tissue structure. They can be used for non-invasive identification of early lesions such as fibrosis, microvascular injury, and apoptosis. Understandably, subclinical lesion biomarkers can be extracted from pre-trained models using transfer learning of pathological image features, extracting texture, morphology, and dynamic response patterns corresponding to histopathological sections. Furthermore, subclinical lesion biomarkers may include, but are not limited to, early fibrosis texture biomarkers, microaneurysm morphological biomarkers, and β-cell apoptosis and nuclear fragmentation biomarkers.

[0100] Based on the detection intensity, spatial distribution density, and deviation from the physiological homeostatic baseline trajectory of subclinical lesion markers, the intervention urgency score of each target organ was calculated. Intervention urgency scores can be quantitative scores based on multidimensional indicators such as the detection intensity of subclinical lesion markers, and can be used to achieve comparability assessment of risks between different organs. Understandably, intervention urgency scores can be fused into a single value by weighted linear combination or nonlinear neural network mapping of standardized indicators. Furthermore, intervention urgency scores can include, but are not limited to, structure-function deviation scores, spatial clustering scores, and signal intensity scores.

[0101] Based on individual daily activity trajectories and environmental exposure data, a spatiotemporal accessibility map of interventionable resources is constructed, and feasible intervention paths that conform to physiological rhythm constraints are extracted. A spatiotemporal reachability map of interventionable resources can be a dynamic network expressing resource availability using graph theory, and can be used to quantify the feasibility of intervention measures in a specific spatiotemporal context. Understandably, a spatiotemporal reachability map of interventionable resources can be constructed by integrating individual daily activity trajectories and environmental exposure data to create a graph structure with spatiotemporal weights. Furthermore, a spatiotemporal reachability map of interventionable resources can include, but is not limited to, community service node reachability maps, remote device response delay maps, and environmental exposure interference maps.

[0102] Feasible intervention pathways can be a set of intervention execution sequences constrained by physiological rhythms, which can be used to ensure that the intervention plan is executed within physiologically appropriate time periods. Understandably, feasible intervention pathways can be generated by using graph search algorithms to eliminate time windows that conflict with physiological rhythms from the resource network. Feasible intervention pathways may include, but are not limited to, morning metabolic window pathways, nighttime liver metabolic activity pathways, and midday activity recovery pathways.

[0103] The intervention urgency score was Pareto optimally matched with feasible intervention pathways to deduce the optimal intervention initiation time window and minimum effective intervention intensity for each target organ. Pareto optimal matching can be a multi-objective optimization process that can be used to achieve a global balance between intervention urgency and resource feasibility. Understandably, Pareto optimal matching can search for undominated solutions in a multi-dimensional space using evolutionary algorithms or constrained optimization methods. Furthermore, Pareto optimal matching can include, but is not limited to, payoff-cost Pareto fronts, rhythm-intensity Pareto fronts, and time-resource Pareto fronts.

[0104] The optimal intervention initiation time window can be a continuous time interval determined by the Pareto optimal solution, which can be used to determine the timing of the intervention task. Understandably, the optimal intervention initiation time window can be derived by jointly inferring the peak of physiological rhythms and the inflection point of risk acceleration. Furthermore, the optimal intervention initiation time window may include, but is not limited to, the peak insulin sensitivity window, the hepatic metabolic clearance window, and the morning autonomic activation window.

[0105] Minimum effective intervention intensity can be defined as the minimum resource input required to ensure that an intervention produces measurable therapeutic effects, and can be used to reduce the burden of over-intervention. Understandably, minimum effective intervention intensity can be determined by modeling dose-response curves using historical intervention response data. Furthermore, minimum effective intervention intensity can include, but is not limited to, low-intensity exercise prescriptions, moderate-dose nutritional adjustments, and short-term remote reminders.

[0106] By integrating the target organ intervention time window, intensity threshold, and resource constraints, a set of precise intervention tasks with execution priority, dose gradient, and efficacy monitoring nodes is generated.

[0107] A precision intervention task set can be a structured set of instructions containing parameters such as execution priority, which can be used to achieve executable outputs of the intervention plan. It is understood that a precision intervention task set can be structured and output by integrating parameters such as the optimal intervention initiation time window for each target organ. Furthermore, a precision intervention task set can include, but is not limited to, a liver metabolic regulation task set, a pancreatic islet function protection task set, and a retinal microcirculation maintenance task set.

[0108] Execution priority can be an intervention sequence determined by risk contribution, used to ensure high-risk organs receive intervention first when resources are limited. It is understood that execution priority can be generated by ranking risk contribution in a Pareto optimal solution. Furthermore, execution priority can include, but is not limited to, high-priority organ tasks, medium-priority organ tasks, and low-priority organ tasks. Dose gradient can be a multi-level intensity classification based on physiological tolerance range, used to achieve personalized adaptation of intervention intensity. It is understood that dose gradient can be defined by deducing from historical individual response data and physiological models. Furthermore, dose gradient can include, but is not limited to, low-intensity dose intervals, medium-intensity dose intervals, and high-intensity dose intervals. Efficacy monitoring nodes can be preset data collection points at specific time points, used to provide structured data anchors for efficacy attribution analysis. It is understood that efficacy monitoring nodes can be biologically meaningful feedback collection windows set by intervention type and organ response dynamics. Furthermore, efficacy monitoring nodes can include, but are not limited to, short-term response monitoring points, medium-term steady-state monitoring points, and long-term adaptation monitoring points.

[0109] In one embodiment, available personalized health intervention resources are identified; based on a precise intervention task set, dynamic adaptation and scheduling of personalized health intervention resources under spatiotemporal constraints are implemented, and intervention response feedback data is collected simultaneously, including: Identify available personalized health intervention resources, including nutrient sustained-release microcapsules, transcutaneous electrical stimulation parameter sets, personalized exercise prescription libraries, and combinations of gut microbiota modulators; The nutrient sustained-release microcapsules can be oral formulation units that release active nutrients at a predetermined rate through material encapsulation technology. Furthermore, the operational principle of nutrient sustained-release microcapsules, based on polymer membrane controlled release, liposome encapsulation, or pH-responsive gel systems, can be explained in context, involving phased release in the gastrointestinal tract triggered by physiological conditions. In this embodiment, the nutrient sustained-release microcapsules, as one of the modeling objects of the digital twin, have their release kinetics coupled with the individual's physiological homeostatic baseline trajectory, influencing the construction of the efficacy-toxicity trade-off matrix. For example, nutrient sustained-release microcapsules can include, but are not limited to, one or more of glucose-regulated sustained-release capsules, fat-soluble vitamin delayed-release capsules, and antioxidant time-delayed-release capsules. The transcutaneous acupoint electrical stimulation parameter set can be a combination of electrophysiological parameters used to control the output of a transcutaneous electrical stimulation device, including frequency, intensity, pulse width, and stimulation timing. Furthermore, the operational principle of the transcutaneous acupoint electrical stimulation parameter set, based on traditional Chinese medicine meridian theory and neurophysiological evidence, can be explained in context, involving setting electrical stimulation patterns for specific acupoints to regulate autonomic nerve activity. In this embodiment, the parameter combinations of the transcutaneous acupoint electrical stimulation parameter set are modeled as input variables of a digital twin, and its physiological response is quantified as the neuromodulation dimension in the efficacy-toxicity trade-off matrix. For example, the transcutaneous acupoint electrical stimulation parameter set can employ a low-frequency / high-frequency alternating stimulation parameter set, a pulse width modulation stimulation parameter set, or a circadian rhythm synchronized stimulation parameter set, etc.

[0110] A personalized exercise prescription library can be a set of preset combinations of exercise types, intensities, durations, and rhythms based on an individual's physiological capabilities and health goals. Furthermore, the operational principle of this personalized exercise prescription library, which generates a multi-dimensional exercise program library based on cardiopulmonary function assessment, joint range of motion detection, and historical exercise adherence data, can be explained in context. In this embodiment, the personalized exercise prescription library serves as the intervention input for a digital twin, and its metabolic response is modeled as a dynamic coupling function with the physiological homeostatic baseline trajectory. A gut microbiota modulator combination can be an oral regulatory preparation composed of various probiotics, prebiotics, or post-biotics in specific proportions. Furthermore, the operational principle of this gut microbiota modulator combination, which selects strains that can target and regulate specific metabolic pathways based on metagenomic analysis results, can be explained in context. In this embodiment, the effect of the gut microbiota modulator combination is delayed and cumulative, and its pharmacokinetic response is modeled as a time-lag coupling relationship with the individual's homeostatic baseline.

[0111] A digital twin is created for individualized health intervention resources, and their pharmacokinetic response is simulated under the individual's physiological homeostatic baseline trajectory to generate a resource efficacy-toxicity trade-off matrix.

[0112] A digital twin can be a virtual replica of a personalized health intervention resource, biophysically modeled to simulate its dynamic response in a real physiological environment. Furthermore, the digital twin can, in context, explain the operational principles of its computational model, which constructs a model encompassing release mechanisms, absorption pathways, and metabolic effects based on the material properties, pharmacokinetic parameters, and clinical research data of the intervention resource. In this embodiment, the digital twin serves as the core mediator, connecting the intervention resource with the efficacy-toxicity trade-off matrix. Its input is the personalized physiological homeostatic baseline trajectory, and its output is the pharmacokinetic response.

[0113] The individualized physiological homeostasis baseline trajectory can reflect the dynamic evolutionary path of an individual's physiological parameters deviating from the normal range over a long timescale. Furthermore, the operational principle of the individualized physiological homeostasis baseline trajectory, which constructs a time-centric non-parametric probability distribution trajectory by periodically decomposing and trend-separating multidimensional physiological time-series data, can be explained in context. In this embodiment, the individualized physiological homeostasis baseline trajectory serves as the core input of the digital twin, coupled with intervention resource modeling, determining the generation of the pharmacodynamic response and trade-off matrix.

[0114] The pharmacokinetic response of a simulated digital twin under an individual's physiological homeostatic baseline trajectory can be achieved by using the individualized physiological homeostatic baseline trajectory as input to drive the operation of the digital twin and outputting the temporal evolution of the intervention response. Furthermore, this operation can be achieved by using blood glucose fluctuation trajectory as input to simulate the delayed release effect of a sustained-release capsule at a hyperglycemic peak, or by using the HRV (Heart Rate of Change) decline trend as input to simulate the recovery rate of parasympathetic tone after electrical stimulation, thereby generating dynamic response curves of the intervention under the individual's real physiological environment.

[0115] The resource efficacy-toxicity trade-off matrix can be a high-dimensional quantitative matrix characterizing the efficacy benefits and potential side effects of various intervention resources under different physiological states. Furthermore, the operational principle of the resource efficacy-toxicity trade-off matrix can be explained in context, as it operates under the drive of a digital twin on an individual's physiological homeostatic baseline trajectory, outputting the joint distribution of efficacy indicators and toxicity probabilities for multiple interventions. In this embodiment, the resource efficacy-toxicity trade-off matrix serves as the foundation for constructing the reward function of a multi-objective reinforcement learning algorithm, determining the optimization direction of intervention scheduling. For example, the resource efficacy-toxicity trade-off matrix may include, but is not limited to, one or more of the following: a metabolic improvement-hypoglycemia risk matrix, a neural regulation-skin irritation risk matrix, and a gut microbiota regulation-diarrhea occurrence probability matrix.

[0116] Generating a resource efficacy-toxicity tradeoff matrix can quantify the pharmacokinetic response of each intervention in two dimensions: efficacy benefits and potential toxicity probabilities. Furthermore, this operation can be achieved by quantifying the improvement in HbA1c by sustained-release capsules and the probability of hypoglycemic events, or by quantifying the correlation function between electrical stimulation and the rate of increase in heart rate variability and the risk of skin burns. This allows for the construction of a multi-dimensional evaluation space for multiple interventions, supporting optimization decision-making.

[0117] Based on the resource efficiency-toxicity trade-off matrix and the precise intervention task set, a multi-objective reinforcement learning algorithm is used for dynamic resource scheduling, outputting an intervention execution instruction sequence with spatiotemporal coordinates, and simultaneously collecting intervention response feedback data.

[0118] A precise intervention task set can be a sequence of combined tasks consisting of multiple organ-specific and stage-adaptive intervention actions. The operational principle of a precise intervention task set, which is generated based on a pre-defined intervention strategy library and matched with personalized intervention windows, can be explained in context. As the state space of a multi-objective reinforcement learning algorithm, the precise intervention task set constrains the priority and combination range of intervention scheduling. For example, a precise intervention task set may include, but is not limited to, one or more of the following: liver metabolic regulation task set, myocardial oxygen supply optimization task set, and insulin sensitivity improvement task set.

[0119] Multi-objective reinforcement learning algorithms can be seen as sequential decision-making models that learn optimal policies under multiple conflicting objectives. Furthermore, multi-objective reinforcement learning algorithms, based on PPO-MO or NSGA-III architectures, use the intervention task set as the state space, spatiotemporal constraints as the action boundaries, and employ a dual-objective optimization principle of maximizing effectiveness and minimizing toxicity. In this embodiment, the multi-objective reinforcement learning algorithm uses the resource effectiveness-toxicity tradeoff matrix as the reward function, the precise intervention task set as the state input, and outputs a sequence of intervention execution instructions.

[0120] In one embodiment, efficacy attribution analysis and threshold drift detection are performed based on intervention response feedback data to drive real-time reconstruction of the physiological homeostasis baseline trajectory, forming a closed-loop dynamic threshold health management strategy, including: Multi-source heterogeneous signal alignment was performed on the intervention response feedback data, including continuous blood glucose fluctuation spectrum, dynamic changes in the urinary microalbumin / creatinine ratio, time-series concentrations of inflammatory factors, and subjective symptom diary text.

[0121] Multi-source heterogeneous signal alignment can be a process of synchronously aligning physiological and behavioral feedback data of different modalities, sampling frequencies, and semantic structures in the time axis and semantic space. This can be understood as acquiring data through methods including aligning continuous physiological signals using timestamp interpolation, extracting the temporal semantics of symptom text using natural language processing techniques, and achieving spatiotemporal consistency mapping of multi-source data through a cross-modal alignment network. Furthermore, multi-source heterogeneous signal alignment can serve as an input preprocessing unit for an intervention efficacy attribution graph neural network, with its output providing a structurally consistent multimodal feedback feature matrix for the attribution network. This can include, but is not limited to, one or more of time-domain alignment, semantic-domain alignment, and modality-aligned encoding.

[0122] Continuous blood glucose fluctuation profiles can be dynamic sequences reflecting the transient fluctuation characteristics of glucose metabolism at homeostasis. They are obtained by collecting subcutaneous tissue fluid glucose concentration data every 5 minutes using a continuous glucose monitoring system. Understandably, their purpose is to provide real-time indicators of insulin sensitivity and hepatic glucose metabolism function. Furthermore, continuous blood glucose fluctuation profiles can be synergistically correlated with the calculation of target organ functional redundancy index.

[0123] The dynamic changes in the urinary microalbumin / creatinine ratio can be a time-series sequence reflecting the subclinical trend of glomerular filtration barrier function. This is obtained by regularly collecting urine samples using a home urine analyzer and calculating the daily ratio. Understandably, its purpose is as a sensitive biomarker of renal functional reserve. Furthermore, the dynamic changes in the urinary microalbumin / creatinine ratio can be correlated with the renal sub-pattern within target organ functional reserve. The time-series concentrations of inflammatory factors can be a dynamic change sequence of the concentrations of key inflammatory markers in the blood. This is obtained by acquiring concentration data through immunoturbidimetry or digital ELISA. Understandably, its purpose is to characterize the dynamic evolution of systemic chronic inflammation. Furthermore, the time-series concentrations of inflammatory factors can be correlated with the liver and vascular sub-patterns within target organ functional reserve. This can include, but is not limited to, one or more of the following: acute stress-induced elevation sequences, sustained low-level elevation sequences, and fluctuating rebound sequences.

[0124] Subjective symptom diary text can be natural language descriptions including dimensions such as fatigue, pain, appetite, and sleep. It is obtained by collecting user text logs through a mobile application interface and extracting symptom type, intensity, and timestamps. Understandably, its purpose is to supplement the subjective perception dimension of objective physiological data. Furthermore, subjective symptom diary text can be semantically encoded and aligned with physiological signals, enhancing the attribution network's ability to model behavior-physiological associations.

[0125] We constructed an attribution graph neural network for intervention efficacy to analyze the contribution weights of each intervention to the functional reserve of different target organs.

[0126] The intervention efficacy attribution graph neural network can be a graph-structured deep learning model used to quantify the contribution weights of interventions to organ function. It is obtained by constructing a heterogeneous graph structure and aggregating multi-source feedback signals through a message passing mechanism. Its purpose is to achieve interpretable causal modeling between intervention measures and organ function responses. Furthermore, the intervention efficacy attribution graph neural network can receive a unified feature matrix after aligning multi-source heterogeneous signals and output the contribution weights of each intervention measure to the target organ's functional reserve. The target organ's functional reserve can be the potential capacity of an organ to maintain normal physiological function even with structural micro-damage, obtained from a dynamic map of organ functional reserves. Its purpose is to serve as the target variable for the intervention efficacy attribution graph neural network. Furthermore, the target organ's functional reserve can be directly correlated with the target organ's functional redundancy index, which is the difference between its current state and the upper limit of reserve. The contribution weights can be edge weights representing the strength of the causal influence of intervention measures on changes in target organ functional reserve, obtained by aggregating multi-source feedback signals through the message passing mechanism of the graph neural network. Its purpose is to quantify the effectiveness of each intervention action. Contribution weights can be output by the intervention efficacy attribution graph neural network and used as the basis for strategy iteration decisions. When the functional redundancy index of any target organ falls below a preset safety threshold and the duration exceeds a preset duration, a threshold drift detection mechanism is triggered.

[0127] The target organ functional redundancy index is a quantitative indicator characterizing an organ's buffering capacity in response to stress disturbances. It is obtained by comparing the predicted value from a dynamic organ functional reserve map with the current observed value. Understandably, its purpose is to serve as a trigger for a threshold drift detection mechanism. Furthermore, the target organ functional redundancy index can be calculated jointly from the dynamic organ functional reserve map and intervention response feedback data.

[0128] A preset safety threshold can be a minimum acceptable level set based on population health benchmarks and individual historical data. It is obtained by determining the percentile of the healthy population distribution and the low-fluctuation range of the individual's baseline stability period. Understandably, its purpose is to serve as a boundary for judging whether organ function has entered a dangerous state. The preset safety threshold can include one or more of the following: organ-specific safety threshold, stage-adaptive safety threshold, and environment-dependent safety threshold. The duration can be the cumulative length of time the target organ's functional redundancy index is below the preset safety threshold, obtained by recording time periods using a sliding window timer. Understandably, its purpose is to distinguish between transient fluctuations and systemic decline.

[0129] The threshold drift detection mechanism can be an online system for detecting systematic shifts in an individual's physiological homeostasis baseline. It is acquired using cumulative sum control charts or Bayesian breakpoint detection algorithms. Understandably, its purpose is to identify the true evolution of the physiological baseline. Furthermore, the threshold drift detection mechanism can receive the comparison results of the target organ functional redundancy index with a preset safety threshold, triggering the retraining of the individualized physiological homeostasis baseline generator.

[0130] Based on the threshold drift detection results, the individualized physiological homeostasis baseline generator is retrained, the physiological homeostasis baseline trajectory and its dynamic confidence interval are updated, and the closed-loop dynamic threshold health management strategy is iterated.

[0131] The individualized physiological homeostasis baseline generator can be a time-series modeler used to generate and update the individual's physiological homeostasis baseline trajectory. It can be obtained using LSTM-AE, a Transformer encoder, or an online Gaussian process regression model. Its purpose is to dynamically generate a homeostasis baseline trajectory that reflects the individual's true physiological state. The individualized physiological homeostasis baseline generator can receive trigger signals from a threshold drift detection mechanism and, after retraining, outputs an updated physiological homeostasis baseline trajectory and its dynamic confidence interval. The dynamic confidence interval can represent the probability range of the predicted uncertainty of the individualized physiological homeostasis baseline trajectory, and it is obtained based on the predicted distribution output by the retrained baseline generator. Its purpose is to provide a quantitative basis for risk prediction uncertainty. The width of the dynamic confidence interval can be dynamically adjusted with the addition of new data.

[0132] The physiological homeostasis baseline trajectory can be seen as the dynamic evolutionary path of an individual's physiological parameters deviating from the normal range over a long timescale. It is obtained from the output of a personalized physiological homeostasis baseline generator. Understandably, its purpose is to serve as input for risk models. Furthermore, the physiological homeostasis baseline trajectory can contain the temporal evolution sequence of multidimensional physiological parameters. The closed-loop dynamic threshold health management strategy can be an adaptive health management framework based on feedback-driven baseline reconstruction and strategy iteration. It is obtained by integrating intervention efficacy attribution, threshold drift detection, baseline retraining, and confidence update processes. Understandably, its purpose is to enable the health management model to have continuous evolutionary capabilities.

[0133] Taking a patient with diabetes and fatty liver as an example, the intelligent health management method based on an integrated AI model in this embodiment can be as follows: the patient undergoes continuous blood glucose monitoring, liver texture analysis using ultrasound, and IL-6 inflammatory factor detection, and subjective symptoms are recorded as "bitter taste in the mouth upon waking" and "drowsiness after meals." A multi-source alignment module uniformly encodes blood glucose fluctuation spectrum, liver texture heterogeneity, IL-6 concentration, and symptom text. Attribution graph network analysis reveals that "high-intensity interval training in the morning" significantly increases liver function redundancy (weight 0.81), while "carbohydrate intake at dinner" is strongly correlated with increased IL-6 (weight -0.69). When liver function redundancy remains below the threshold for 96 hours, the drift detection mechanism is activated, the baseline generator is retrained, and the liver metabolic baseline trajectory is updated. The narrowing of the dynamic confidence interval indicates that the system's judgment of liver function trends is more certain. After the closed-loop strategy iteration, the system automatically adjusts the diet plan to "low-GI breakfast + no-carbohydrate dinner" and increases the frequency of morning exercise reminders. Subsequent monitoring shows a decrease in IL-6 and an improvement in liver texture heterogeneity, verifying that the closed-loop mechanism successfully drives the evolution of the individualized strategy.

[0134] Furthermore, to achieve the above objectives, the present invention also provides an intelligent physical fitness assessment and health management system integrating a large AI model, the system comprising: The data modeling module is used to collect time-series data of individual multidimensional physiological characteristics; perform biorhythm analysis and steady-state shift modeling on the time-series data of individual multidimensional physiological characteristics, and generate individualized physiological steady-state baseline trajectory. The image analysis module is used to acquire continuous dynamic body measurement image data of an individual; based on the continuous dynamic body measurement image data of the individual, it performs tissue microstructure texture characterization extraction and cross-modal functional coupling analysis to generate a dynamic atlas of organ function reserves; The risk modeling module is used to integrate individualized physiological homeostasis baseline trajectories and dynamic maps of organ function reserves to construct a multi-level chronic disease progression risk gradient model and output disease activity prediction curves at multiple time scales. The intervention simulation module is used to identify high-risk target organ regions based on individual continuous dynamic body imaging; combined with multi-timescale disease activity prediction curves, it performs personalized intervention window simulations for each target organ region, generating a set of precise intervention tasks by organ and stage. The resource scheduling module is used to identify available individualized health intervention resources; based on the precise intervention task set, it performs dynamic adaptation scheduling of individualized health intervention resources under spatiotemporal constraints, and simultaneously collects intervention response feedback data; The closed-loop management module is used to perform efficacy attribution analysis and threshold drift detection based on intervention response feedback data, drive real-time reconstruction of the physiological homeostasis baseline trajectory, and form a closed-loop dynamic threshold health management strategy.

[0135] Other embodiments or specific implementations of the intelligent physical fitness management system integrating large AI models described in this invention can be referred to the above-mentioned method embodiments, and will not be repeated here.

[0136] In addition, to achieve the above objectives, the present invention also provides a body composition analyzer, the body composition analyzer comprising: a memory, a processor, and an intelligent body composition analysis and health management program integrating an AI large model stored in the memory and executable on the processor, the intelligent body composition analysis and health management program integrating an AI large model configured to implement the steps of the intelligent body composition analysis and health management method integrating an AI large model as described above.

[0137] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for intelligent physical fitness testing and health management integrating a large AI model, characterized in that, The method includes: Collect time-series data of individual multidimensional physiological characteristics; perform biorhythm analysis and steady-state shift modeling on the time-series data of individual multidimensional physiological characteristics to generate individualized physiological steady-state baseline trajectories; Acquire continuous dynamic body measurement image data of an individual; perform tissue microstructure texture characterization extraction and cross-modal functional coupling analysis based on the continuous dynamic body measurement image data of the individual to generate a dynamic atlas of organ function reserve; By integrating individualized physiological homeostasis baseline trajectory and dynamic atlas of organ function reserve, a multi-level chronic disease progression risk gradient model is constructed, and multi-timescale disease activity prediction curves are output. High-risk target organ regions are identified based on individual continuous dynamic body imaging; combined with multi-timescale disease activity prediction curves, personalized intervention window periods are extrapolated for each target organ region, generating a set of precise intervention tasks by organ and stage. Identify available individualized health intervention resources; based on the precise intervention task set, dynamically adapt and schedule the individualized health intervention resources under spatiotemporal constraints, and simultaneously collect intervention response feedback data; Based on intervention response feedback data, efficacy attribution analysis and threshold drift detection are performed to drive real-time reconstruction of the physiological homeostatic baseline trajectory, forming a closed-loop dynamic threshold health management strategy.

2. The intelligent physical fitness and health management method integrating a large AI model as described in claim 1, characterized in that, The collection of individual multidimensional physiological signs time-series data; The individual's multidimensional physiological characteristic time-series data are analyzed for biological rhythms and modeled for homeostasis shifts to generate an individualized physiological homeostasis baseline trajectory, including: The individual multidimensional physiological signs time series data are collected by the collaborative collection of wearable sensor array and non-contact body measurement terminal. The individual multidimensional physiological signs time series data include at least heart rate variability, skin conductance response, respiratory tidal volume, microcirculatory blood oxygen saturation and sublingual microvascular flow velocity. The individual multidimensional physiological signs time series data are subjected to diurnal rhythm phase decoupling and autonomic nerve tension spectrum decomposition to generate an individualized biological rhythm parameter matrix; Based on the individualized biological rhythm parameter matrix, a physiological homeostatic dynamic equation is constructed, and the topology of the homeostatic attractor is solved. Long-term steady-state shift quantization is performed using steady-state attractor topology to generate a steady-state shift entropy value sequence. By integrating the steady-state offset entropy sequence with the individualized biological rhythm parameter matrix, an individualized physiological homeostasis baseline generator is trained. The individualized physiological homeostasis baseline generator outputs an individualized physiological homeostasis baseline trajectory, which includes a dynamic confidence interval and a pathological perturbation sensitive zone.

3. The intelligent physical fitness testing and health management method integrating a large AI model as described in claim 1, characterized in that, The process includes acquiring continuous dynamic body image data of an individual; performing tissue microstructure texture characterization extraction and cross-modal functional coupling analysis based on the continuous dynamic body image data of the individual, and generating a dynamic atlas of organ functional reserves, including: Continuous dynamic body measurement data of an individual were acquired simultaneously using a multispectral skin-mucosal imaging system and optical coherence tomography of the tongue surface microstructure. Multi-scale vascular network skeleton extraction was performed on the individual's continuous dynamic body imaging data to construct a microcirculation topology connection map; By integrating microcirculation topology maps with infrared thermo-metabolic images, tissue perfusion-metabolic coupling modeling is performed to generate dynamic atlases of organ-level functional reserves. Based on the dynamic map of organ-level functional reserves, the functional redundancy index and compensatory attenuation slope of each target organ are calculated. Cross-organ correlation analysis of functional redundancy index and compensatory attenuation slope was conducted to identify key nodes of functional compensation imbalance. Using the hub node as the anchor point, the data is back-mapped to the anatomical structure space to generate a dynamic atlas of organ function reserves, which includes a set of spatial positioning labels.

4. The intelligent physical fitness and health management method integrating a large AI model as described in claim 3, characterized in that, The step of extracting the multi-scale vascular network skeleton from the individual's continuous dynamic body imaging data and constructing a microcirculation topology map includes: Perform multispectral channel light scattering consistency calibration on continuous dynamic body measurement image data of individuals; An adaptive local contrast enhancement algorithm is used to sharpen the texture of the calibrated image, generating an enhanced texture image. A multi-scale Gabor filter bank was constructed based on enhanced texture images to extract directional texture features of tissue microstructures. A lightweight U-Net segmentation model was trained using directional texture features to achieve pixel-level segmentation of microvascular networks. Graph theory optimization is performed on the segmentation results to remove pseudo-connected branches and complete the broken blood vessel segments, outputting a microcirculation topology connection graph.

5. The intelligent physical fitness testing and health management method integrating a large AI model as described in claim 1, characterized in that, The model integrates individualized physiological homeostasis baseline trajectories and dynamic organ function reserve maps to construct a multi-level chronic disease progression risk gradient model, outputting multi-timescale disease activity prediction curves, including: By mapping individualized physiological homeostasis baseline trajectories to the spatiotemporal coordinate system of organ functional reserve dynamic map, a cross-modal homeostasis-function joint representation space is constructed. Within the cross-modal steady-state-functional joint characterization space, a multidimensional risk potential field for the progression of chronic diseases is defined, which is composed of the inflammatory factor fluctuation entropy, the metabolite accumulation gradient, and the amplitude of neuroendocrine axis perturbation. Based on the multidimensional risk potential field, a Lagrange mechanical framework is used to simulate the disease progression path and generate a disease activity evolution manifold. Input a pre-trained multimodal large model, which integrates clinical guideline knowledge graphs, real-world cohort survival data, and individual gene polymorphism information to enhance causal inference of the disease activity evolution manifold; The output includes disease activity prediction curves covering multiple time granularities, with each curve accompanied by uncertainty quantification indicators and key inflection point warning indicators.

6. The intelligent physical fitness and health management method integrating a large AI model as described in claim 1, characterized in that, The method involves identifying high-risk target organ regions based on individual continuous dynamic body imaging; combining multi-timescale disease activity prediction curves to perform personalized intervention window period simulations for each target organ region, generating a precise intervention task set by organ and stage, including: Based on the spatial localization tag set of the dynamic atlas of organ function reserve in continuous dynamic body measurement images, four types of high-risk target organ regions, namely liver, kidney, pancreas and retina, are identified. Pathological imaging feature transfer learning was performed on each target organ region to identify subclinical lesion markers of early fibrosis, microaneurysms, and β-cell apoptosis; Based on the detection intensity, spatial distribution density, and deviation from the physiological homeostatic baseline trajectory of subclinical lesion markers, the intervention urgency score of each target organ was calculated. Based on individual daily activity trajectories and environmental exposure data, a spatiotemporal accessibility map of interventionable resources is constructed, and feasible intervention paths that conform to physiological rhythm constraints are extracted. Pareto optimal matching was performed between intervention urgency scores and feasible intervention pathways to deduce the optimal intervention initiation time window and minimum effective intervention intensity for each target organ. By integrating the target organ intervention time window, intensity threshold, and resource constraints, a set of precise intervention tasks with execution priority, dose gradient, and efficacy monitoring nodes is generated.

7. The intelligent physical fitness assessment and health management method integrating a large AI model as described in claim 1, characterized in that, The process involves identifying available individualized health intervention resources; dynamically adapting and scheduling these resources under spatiotemporal constraints based on the precise intervention task set; and simultaneously collecting intervention response feedback data, including: Identify available personalized health intervention resources, including nutrient sustained-release microcapsules, transcutaneous electrical stimulation parameter sets, personalized exercise prescription libraries, and combinations of gut microbiota modulators; A digital twin is established for the individualized health intervention resources, and its pharmacokinetic response is simulated under the individual's physiological homeostatic baseline trajectory to generate a resource efficacy-toxicity trade-off matrix. Based on the resource efficiency-toxicity trade-off matrix and the precise intervention task set, a multi-objective reinforcement learning algorithm is used for dynamic resource scheduling, outputting an intervention execution instruction sequence with spatiotemporal coordinates, and simultaneously collecting intervention response feedback data.

8. The intelligent physical fitness testing and health management method integrating a large AI model as described in claim 1, characterized in that, The efficacy attribution analysis and threshold drift detection based on intervention response feedback data drive real-time reconstruction of the physiological homeostasis baseline trajectory, forming a closed-loop dynamic threshold health management strategy, including: Multi-source heterogeneous signal alignment was performed on the intervention response feedback data, including continuous blood glucose fluctuation spectrum, dynamic changes in the urine microalbumin / creatinine ratio, time-series concentrations of inflammatory factors, and subjective symptom diary text; We constructed an attribution graph neural network for intervention efficacy to analyze the contribution weights of each intervention to the functional reserve of different target organs. When the functional redundancy index of any target organ is lower than the preset safety threshold and the duration exceeds the preset duration, the threshold drift detection mechanism is triggered. Based on the threshold drift detection results, the individualized physiological homeostasis baseline generator is retrained, the physiological homeostasis baseline trajectory and its dynamic confidence interval are updated, and the closed-loop dynamic threshold health management strategy is iterated.

9. A smart health management system integrating a large AI model, characterized in that, The system includes: The data modeling module is used to collect time-series data of individual multidimensional physiological characteristics; perform biorhythm analysis and steady-state shift modeling on the time-series data of individual multidimensional physiological characteristics, and generate individualized physiological steady-state baseline trajectory. The image analysis module is used to acquire continuous dynamic body measurement image data of an individual; based on the continuous dynamic body measurement image data of the individual, it performs tissue microstructure texture characterization extraction and cross-modal functional coupling analysis to generate a dynamic atlas of organ function reserves; The risk modeling module is used to integrate individualized physiological homeostasis baseline trajectories and dynamic maps of organ function reserves to construct a multi-level chronic disease progression risk gradient model and output disease activity prediction curves at multiple time scales. The intervention simulation module is used to identify high-risk target organ regions based on individual continuous dynamic body imaging; combined with multi-timescale disease activity prediction curves, it performs personalized intervention window simulations for each target organ region, generating a set of precise intervention tasks by organ and stage. The resource scheduling module is used to identify available individualized health intervention resources; based on the precise intervention task set, it performs dynamic adaptation scheduling of individualized health intervention resources under spatiotemporal constraints, and simultaneously collects intervention response feedback data; The closed-loop management module is used to perform efficacy attribution analysis and threshold drift detection based on intervention response feedback data, drive real-time reconstruction of the physiological homeostasis baseline trajectory, and form a closed-loop dynamic threshold health management strategy.

10. A body composition analyzer, characterized in that, The body composition analyzer includes: a memory, a processor, and an intelligent body composition analysis and health management program with an integrated AI big model stored in the memory and executable on the processor. The intelligent body composition analysis and health management program with an integrated AI big model is configured to implement the steps of the intelligent body composition analysis and health management method with an integrated AI big model as described in any one of claims 1 to 9.

Images