Multimodal-based physiological age calculation and fracture risk data processing method

Abstract

The application provides a multi-modal-based physiological age calculation and fracture risk data processing method, which comprises the following steps: collecting multi-source heterogeneous data of a target individual; obtaining a feature tensor set based on the multi-source heterogeneous data; processing the multi-source heterogeneous data and the feature tensor set to obtain a bone system health state latent space representation; processing the bone system health state latent space representation through a preset machine learning regression model to obtain the physiological age of the target individual; determining the age acceleration state based on the physiological age and a preset threshold; and constructing an osteoporotic fracture risk prediction model based on the above data to output an osteoporotic fracture risk probability and a post-fracture all-cause mortality risk prediction value. Through the above design, the application solves the problems of single evaluation dimension, limited application scene, single data mode and missing intervention strategy in the prior art, and cannot well meet the more accurate evaluation and intervention requirements.

Description

Technical Field

[0001] This invention relates to the field of human health data processing technology, and in particular to a method for calculating physiological age and processing fracture risk data based on multimodal approaches. Background Technology

[0002] Traditional aging assessments rely on chronological age, but fail to reflect individual physiological heterogeneity. Biochronological age (BA), through AI-integrated multi-organ physiological markers, can more accurately capture the rate of aging and mortality risk, surpassing chronological age. However, the association between BA and fractures, and the role of genetic susceptibility, remains poorly understood. Fractures are a key marker of skeletal aging, and current research largely focuses on cardiovascular and cerebrovascular fields, with limited understanding of the synchronicity between skeletal aging and overall aging. Furthermore, osteoporosis has a strong genetic predisposition, with the Genetic Risk Score (GRS) serving as an important innate indicator; however, how the interaction between BA and GRS affects fracture risk remains poorly investigated. To address this issue, some institutions have proposed solutions.

[0003] For example, prior art with publication number CN120388745A discloses a method for accelerating the assessment of healthy lifespan and disease risk based on phenotypic age, including the following steps: S101: Obtaining baseline data of the target population, including chronological age and multi-system clinical chemical biomarker data; S102: Calculating phenotypic age based on the baseline data, and determining the accelerated phenotypic age value through regression analysis between phenotypic age and chronological age; S103: Stratifying the target population by aging degree based on the accelerated phenotypic age value; S104: Evaluating the association between aging degree stratification and disease-free survival years and disease risk through a pre-set survival analysis model; S105: Outputting the assessment results, which characterize the differences in healthy lifespan and changes in disease risk in the target population. The aforementioned disease risks include fractures, osteoporosis, etc. This application reliably assesses the degree of biological aging, providing a scientific basis for personalized health management and the prevention of diseases such as fractures.

[0004] However, this existing technology still has shortcomings. Although it has solved some of the problems of how the interaction between biological age and genetic risk score affects fracture risk, the technology has problems in practical applications, such as a single assessment dimension, limited application scenarios, a single data modality, and a lack of intervention strategies. It cannot meet the needs for more accurate assessment and intervention. Summary of the Invention

[0005] Therefore, it is necessary to provide a multimodal physiological age calculation and fracture risk data processing method to address the problems of existing technologies, such as single assessment dimensions, limited application scenarios, single data modality, and lack of intervention strategies, which cannot well meet the needs of more accurate assessment and intervention.

[0006] This invention provides a method for calculating physiological age and processing fracture risk data based on multimodal approaches, comprising:

[0007] Collect multi-source heterogeneous data of the target individual;

[0008] Based on the aforementioned multi-source heterogeneous data, a set of feature tensors is obtained;

[0009] The multi-source heterogeneous data and feature tensor set are processed to obtain a latent space representation of the health status of the bone system;

[0010] The latent space of the health status of the skeletal system is represented and processed by a preset machine learning regression model to obtain the physiological age of the target individual;

[0011] Based on the physiological age and a preset threshold, an accelerated aging state is determined;

[0012] Based on the latent space representation of the bone system health status, physiological age, aging acceleration status, and multi-source heterogeneous data, an osteoporotic fracture risk prediction model is constructed to output the probability of osteoporotic fracture risk and the predicted value of all-cause mortality risk after fracture.

[0013] In the above scheme, the multi-source heterogeneous data includes at least static bone and joint imaging data, continuous dynamic physiological signal data, temporal gait mechanics data, traditional Chinese medicine constitution identification data, genetic risk data, and multi-dimensional physiological biomarker data.

[0014] The method for obtaining a set of feature tensors based on the multi-source heterogeneous data includes:

[0015] Three-dimensional reconstruction and depth feature extraction are performed on the static image data of the bone joints to obtain the first feature tensor characterizing the morphology and density of bone microstructure.

[0016] The continuous dynamic physiological signal data is subjected to adaptive modal fusion and cross-modal representation learning based on uncertainty perception to obtain a second feature tensor representing cardiovascular and neuromodal functions;

[0017] Nonlinear dynamic analysis is performed on the temporal gait mechanics data to extract posture control disorder indices and obtain the third characteristic tensor.

[0018] In the above scheme, the method for adaptive modal fusion based on uncertainty awareness includes:

[0019] Real-time assessment of the prediction uncertainty of each physiological signal mode, with the goal of minimizing overall decision uncertainty;

[0020] The weights of each mode are dynamically adjusted during the fusion process to address signal quality fluctuations and mode degradation issues.

[0021] The cross-modal representation learning employs a mask fusion strategy and a variational autoencoder architecture to align and uniformly represent different physiological modalities with metadata related to age, gender, and circadian rhythms in a Gaussian latent space.

[0022] In the above scheme, the method for extracting posture control disorder indicators includes:

[0023] Calculate the asymmetric coefficients of the left and right foot drive index and braking index in each gait cycle to obtain the posture control factor;

[0024] Curve fitting was performed on the sequence of posture control factors within a continuous time window to analyze the frequency and slope changes of their numerical distribution, so as to obtain the control disorder coefficient that reflects gait stability and neuromuscular control ability.

[0025] In the above scheme, the method for processing the multi-source heterogeneous data and feature tensor set to obtain the latent space representation of the bone system health status includes:

[0026] The first feature tensor, the second feature tensor, the third feature tensor, TCM constitution identification data, and genetic risk data are input into a multimodal deep fusion network based on tensor decomposition and expert product technology for alignment and fusion to generate a unified high-dimensional latent space representation of the bone system health status.

[0027] The method for aligning and fusing the input into a multimodal deep fusion network based on tensor decomposition and expert product techniques includes:

[0028] The feature tensors extracted from each modality are concatenated to form a higher-order tensor.

[0029] The higher-order tensors are reduced in dimensionality and their core information is extracted using Tucker decomposition or CP decomposition.

[0030] To address potential modality gaps, an expert product technique is employed to combine available modality expert networks to generate a complete latent space representation of the skeletal system's health status.

[0031] In the above scheme, the method for obtaining the physiological age of the target individual includes:

[0032] Based on the multidimensional physiological biomarker data, the physiological age of the target individual is calculated by processing the data using a phenotypic age algorithm and / or the Klemera-Doubal method age algorithm.

[0033] Physiological age is a comprehensive indicator that is independent of actual age and quantifies the degree of biological aging of the skeletal system.

[0034] The preset machine learning regression model is a gradient boosting regression tree or a deep neural network, and it is trained using multi-center, multi-ethnic cohort data.

[0035] The multicenter, multi-ethnic cohort data includes at least a bone age dataset for children in plains and plateau regions, a pelvic CT age estimation dataset for Han Chinese adults in western China, and a multinational dataset integrating the UK Biobank, the US National Health and Nutrition Examination Survey, and the Chinese Osteoporosis Fracture Specialty Cohort, to ensure the model's generalization ability to different ethnic and regional populations.

[0036] In the above scheme, the method for determining the accelerated aging state based on the physiological age and a preset threshold includes:

[0037] Based on the physiological age and the actual age of the target individual, the age acceleration residual is calculated, and the individual is determined to be in an "accelerated aging" or "non-accelerated aging" state based on a preset threshold.

[0038] In the above scheme, the method for calculating physiological age and processing fracture risk data based on multimodality also includes: combining the TCM constitution identification data, and generating a personalized health management strategy that includes integrated TCM and Western medicine intervention measures through a preset rule engine and recommendation algorithm.

[0039] The method for generating the personalized health management strategy includes:

[0040] If the condition is diagnosed as "accelerated aging" and the fracture risk level is "high," the recommended strategy is to prioritize anti-bone resorption drugs, anti-inflammatory treatment, and traditional Chinese medicine plasters and acupoint applications for those with "kidney yang deficiency" or "qi and blood deficiency."

[0041] If the aforementioned control disorder indicators are abnormally elevated, balance training and physical therapy are recommended simultaneously.

[0042] The personalized health management strategy is output in the form of digital reports and visual charts, and is connected to the traditional Chinese medicine health management platform for continuous tracking.

[0043] In the above scheme, the osteoporotic fracture risk prediction model uses an attention mechanism to dynamically weight different dimensions of the latent space representation of the bone system health status, and it also has a built-in mediation effect analysis module to quantify the proportion of the mediating effect of physiological age between fracture and mortality risk.

[0044] In the above scheme, the method for calculating physiological age and processing fracture risk data based on multimodal approaches also includes: implementing a refined operation process for full-cycle health management through an integrated human-computer interaction and data management platform;

[0045] The refined operation process includes:

[0046] Data standardization, collection, and quality control;

[0047] Streamlined monitoring of feature processing and fusion processes;

[0048] Strategic generation and delivery of risk assessments;

[0049] Dynamic feedback on intervention effects and optimization of the intervention plan.

[0050] The above technical solution has the following advantages or beneficial effects: This invention constructs a hierarchical, adaptive multimodal fusion pipeline. At the data layer, it extensively integrates data across all dimensions, from genetics to gait, from CT images to ECG, and from serum biomarkers to traditional Chinese medicine constitution. At the feature layer, it employs the most suitable advanced technologies for different modalities. Deep learning is used to extract deep omics features from images; uncertainty-aware adaptive fusion is used to process physiological signals in noisy environments; nonlinear dynamics is used to analyze gait time-series data; and standardized scales and intelligent devices are used to quantify traditional Chinese medicine constitution. At the fusion and computation layer, tensor decomposition is introduced to process high-dimensional heterogeneous features, and cross-modal representation learning technology is used to achieve semantic alignment in the latent space. Finally, an innovative comprehensive indicator, "physiological age," is output through an ensemble learning model. At the application layer, the calculated physiological age and multimodal features are input into a risk assessment model. This model not only predicts risk but also reveals pathological pathways through mediation analysis. This solves the problems of existing technologies, such as single assessment dimensions, limited application scenarios, single data modalities, and a lack of intervention strategies, which fail to adequately meet the needs for more precise assessment and intervention. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of the process steps of the multimodal physiological age calculation and fracture risk data processing method of the present invention in one embodiment;

[0052] Figure 2 This is a schematic diagram of the process steps of the multimodal physiological age calculation and fracture risk data processing method of the present invention in another embodiment;

[0053] Figure 3 This is a schematic diagram of the process steps of the multimodal physiological age calculation and fracture risk data processing method of the present invention in another embodiment;

[0054] Figure 4This is a schematic diagram of the process steps of the multimodal physiological age calculation and fracture risk data processing method of the present invention in another embodiment;

[0055] Figure 5 This is a schematic diagram of the process steps of the multimodal physiological age calculation and fracture risk data processing method of the present invention in another embodiment;

[0056] Figure 6 This is a schematic diagram of the process steps of the multimodal physiological age calculation and fracture risk data processing method of the present invention in another embodiment;

[0057] Figure 7 This is a schematic diagram of the process steps of the multimodal physiological age calculation and fracture risk data processing method of the present invention in another embodiment. Detailed Implementation

[0058] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the specific details described below are only a part of the embodiments of the present invention, and the present invention can be implemented in many other embodiments different from those described herein. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0059] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0060] This invention provides a method for calculating physiological age and processing fracture risk data based on multimodal approaches, such as... Figure 1 As shown, its execution includes at least the following steps:

[0061] Step S100: Collect multi-source heterogeneous data of the target individual;

[0062] Step S200: Based on the multi-source heterogeneous data, obtain the feature tensor set;

[0063] Step S300: Process the multi-source heterogeneous data and feature tensor set to obtain the latent space representation of the health status of the bone system;

[0064] Step S400: Represent the latent space of the health status of the skeletal system, process it through a preset machine learning regression model, and obtain the physiological age of the target individual;

[0065] Step S500: Based on the physiological age and a preset threshold, determine the state of accelerated aging;

[0066] Step S600: Based on the latent space representation of the bone system health status, physiological age, aging acceleration status, and multi-source heterogeneous data, construct an osteoporotic fracture risk prediction model to output the osteoporotic fracture risk probability and the predicted value of all-cause mortality risk after fracture.

[0067] The aforementioned multi-source heterogeneous data may include static bone and joint imaging data, continuous dynamic physiological signal data, temporal gait biomechanics data, traditional Chinese medicine constitution identification data, genetic risk data, and multi-dimensional physiological biomarker data, etc.

[0068] Specifically, in this embodiment, the present invention constructs a hierarchical, adaptive multimodal fusion pipeline using the above scheme. At the data layer, it extensively integrates data across all dimensions, from genetics to gait, from CT images to ECGs, and from serum biomarkers to traditional Chinese medicine constitution. At the feature layer, it employs the most suitable advanced technologies for different modalities. Deep learning is used to extract deep omics features from images; uncertainty-aware adaptive fusion is used to process physiological signals in noisy environments; nonlinear dynamics is used to analyze gait time-series data; and standardized scales and intelligent devices are used to quantify traditional Chinese medicine constitution. At the fusion and computation layer, tensor decomposition is introduced to process high-dimensional heterogeneous features, and cross-modal representation learning technology is used to achieve semantic alignment in the latent space. Finally, an integrated learning model outputs the innovative comprehensive indicator of "physiological age." At the application layer, the calculated physiological age and multimodal features are input into a risk assessment model. This model not only predicts risk but also reveals pathological pathways through mediation analysis.

[0069] This invention creates a "latent space representation of the health status of the skeletal system." Unlike simply piecing together features, this invention maps multi-source data into a unified, interpretable latent space through tensor decomposition and cross-modal learning, achieving for the first time the digitization of the overall biological state of the skeletal system.

[0070] Uncertainty-aware fusion is introduced into bone health assessment. The MUSE framework is applied to physiological signal fusion, enabling the system to remain robust in real-world environments with noise and signal loss.

[0071] Achieving a three-dimensional assessment of "morphology-function-constitution". This involves cross-disciplinary integration of adult age estimation technology based on CT three-dimensional reconstruction from forensic anthropology, gait analysis technology from biomechanics, and the traditional Chinese medicine "constitution-organs-meridians" identification system, expanding the assessment dimensions from simple anatomical morphology to dynamic function and overall constitution.

[0072] Constructing a closed loop of "risk prediction - mechanism interpretation - intervention generation": The system not only outputs the risk probability, but also reveals the potential biological mechanism through mediation effect analysis and correlation enrichment analysis.

[0073] The multi-source heterogeneous data includes at least: static bone and joint imaging data, continuous dynamic physiological signal data, temporal gait mechanics data, traditional Chinese medicine constitution identification data, and genetic risk data.

[0074] To facilitate understanding and explanation, the implementation and execution of the above steps will be further explained and illustrated below with reference to the embodiments.

[0075] For step S100, the core content during its implementation is the collection and standardized storage of multi-source heterogeneous data.

[0076] The specific equipment and data acquisition methods used in acquiring the still image data are as follows:

[0077] Equipment and Protocol: Dual-source CT was used to perform automatic modulation of the "Low-Dose High-Resolution Spine Scan Protocol". Simultaneously, a dual-energy X-ray absorptiometry (DXA) scanner was used to perform standard DXA scans of the lumbar spine (L1-L4) and left femoral neck.

[0078] Data acquisition: CT DICOM images and DXA PDF reports (including T-value, Z-value, BMD g / cm³). 2 The data is automatically pushed to the "Multimodal Bone Health Data Center" of this invention through the hospital PACS system interface.

[0079] The specific equipment and data transmission methods used in acquiring continuous dynamic physiological signal data are as follows:

[0080] Equipment and Deployment: Wear a wearable patch applied to the suprasternal notch for continuous monitoring for 7 days. The device simultaneously acquires single-lead electrocardiogram (ECG), triaxial acceleration (ACC), and electrical skin activity (EDA).

[0081] Data transmission: The device connects to the patient's mobile app via Bluetooth. The app uploads encrypted data packets to the cloud server every 4 hours. The cloud service analyzes the raw signal and performs preliminary quality labeling.

[0082] The specific methods used to acquire the temporal gait mechanics data are as follows:

[0083] Data collection environment: The data was collected on a standard 6-meter-long electronic walkway in the rehabilitation department.

[0084] Data recording: The system automatically records the spatiotemporal parameters of each gait cycle: gait speed, stride length, stride width, single-foot support time, and the continuous trajectory coordinates of the plantar pressure center.

[0085] The specific methods used to obtain the TCM constitution identification data are as follows:

[0086] Quantitative data collection: Complete the questions based on the "Classification and Judgment Scale of Traditional Chinese Medicine Constitution (Standard of China Association of Traditional Chinese Medicine)" through a self-service terminal, and the system will automatically score the results.

[0087] Tongue diagnosis: High-definition images of the tongue surface are acquired under standard light source, and the instrument's built-in AI algorithm automatically analyzes the tongue color, shape, coating color, and texture.

[0088] The doctor enters key symptoms, such as aversion to cold, cold limbs, soreness and weakness in the lower back and knees, frequent urination at night, and loose stools.

[0089] The system makes a comprehensive judgment based on the scale score, tongue appearance characteristics and symptom keywords, and generates a feature vector containing the weights of each constitution dimension.

[0090] The specific methods used to obtain the genetic risk data are as follows:

[0091] Detection and Calculation: Oral mucosal swabs were collected for testing. Based on the results of the genome-wide association study of fractures published by the GEFOS consortium, 56 independent SNPs that reached genomic significance were selected. The effect size (β) and risk allele frequency (RAF) of each SNP were derived from a GEFOS meta-analysis. The polygenic risk score (PRS) was calculated using the following formula:

[0092] PRS = Σ(β i * G i ), where β i G represents the effect size of the i-th SNP. i Count the risk alleles at this SNP locus.

[0093] In one embodiment, the multi-source heterogeneous data includes at least static bone and joint imaging data, continuous dynamic physiological signal data, temporal gait mechanics data, traditional Chinese medicine constitution identification data, genetic risk data, and multi-dimensional physiological biomarker data.

[0094] The aforementioned static bone and joint imaging data includes at least one of the following: left wrist X-ray, pelvic CT scan, and dual-energy X-ray absorptiometry (DEXA) images of the spine or hip. For X-rays, an improved ResNet or Xception network is used for end-to-end bone age-related feature regression. For CT images, a three-dimensional convolutional neural network is used to segment the reconstructed skeletal model and extract morphological and textural features, including ventral edge, dorsal edge, commissural plane morphology, trabecular bone number, trabecular bone thickness, and trabecular bone separation.

[0095] The biomarkers mentioned include at least one of albumin, creatinine, glucose, C-reactive protein, lymphocyte percentage, mean corpuscular volume, chronological age, alkaline phosphatase, white blood cell count, red blood cell distribution width, glycated hemoglobin, total cholesterol, blood urea nitrogen, and systolic blood pressure.

[0096] The continuous dynamic physiological signal data includes at least two of the following: electrocardiogram (ECG) signal, photoplethysmography (PPG) signal, and arterial blood pressure signal.

[0097] Among them, such as Figure 2 As shown, the step of obtaining the feature tensor set based on the multi-source heterogeneous data includes:

[0098] Step S210: Perform three-dimensional reconstruction and depth feature extraction on the static image data of the bone joint to obtain the first feature tensor characterizing the morphology and density of bone microstructure;

[0099] Step S220: Perform uncertainty-aware adaptive modal fusion and cross-modal representation learning on the continuous dynamic physiological signal data to obtain a second feature tensor representing cardiovascular and neuromodulation functions;

[0100] Step S230: Perform nonlinear dynamic analysis on the time-series gait mechanics data, extract posture control disorder indexes, and obtain the third feature tensor.

[0101] The specific method used to obtain the first feature tensor in step S210 above is as follows:

[0102] First, CT bone microstructure analysis was performed: the CT sequence of the L3 lumbar vertebra was imported into a Python-based 3D Slicer platform, and the trabecular bone region within the vertebral body was automatically segmented using a pre-trained nnU-Net model. The following key parameters were calculated:

[0103] Bone volume fraction (BV / TV): BV / TV = (trabecular bone volume / total volume) × 100%.

[0104] Number of trabeculae (Tb.N): Tb.N = (1 / Tb.Sp)×(BV / TV).

[0105] Trabecular separation (Tb.Sp): Calculated by measuring the average distance between the medullary canals of the trabeculae.

[0106] Structural Model Index (SMI): Measures the trabecular, plate-like or rod-like structure of bones.

[0107] Next, DXA feature extraction was performed: the report was analyzed to obtain the average BMD and T values ​​of the lumbar spine and the BMD and T values ​​of the femoral neck.

[0108] Finally, the first feature tensor is generated: the above 8 image features (BV / TV, Tb.N, Tb.Sp, SMI, lumbar spine BMD, lumbar spine T value, femoral neck BMD, femoral neck T value) are normalized and encoded into a 128-dimensional first feature tensor through a fully connected layer.

[0109] The specific method used to obtain the second feature tensor in step S220 above is as follows:

[0110] First, signal preprocessing is performed: 7 days of ECG and ACC data are aligned, the Pan-Tompkins algorithm is used to detect R waves, and the ACC signal is used to identify and remove periods with severe motion artifacts.

[0111] Next, feature calculations are performed.

[0112] When calculating heart rate variability (HRV), the nighttime resting period is selected, and frequency domain indices are calculated: low-frequency power (LF), high-frequency power (LF), and LF / HF ratio.

[0113] When calculating pulse wave characteristics, the pulse wave conduction time (PTT) is calculated from the R wave and PPG waveform of the ECG (estimated from the EDA signal), and the arteriosclerosis estimation formula PWV is used. est = K×(height / PTT), where K is the calibration factor.

[0114] When calculating sleep quality indicators, the Cole-Kripke algorithm was used to estimate total sleep time (TST) and sleep efficiency (SE) based on ACC data.

[0115] Finally, the second feature tensor is generated: [LF, HF, LF / HF, PWV] est The second feature tensor consists of 15 physiological characteristics, including average nighttime heart rate, TST, and SE.

[0116] The method used to obtain the third feature tensor in step S230 above is as follows:

[0117] First, average walking speed, stride length variation coefficient, and the proportion of double support are obtained directly from the trail data.

[0118] Then, nonlinear dynamic analysis was performed: the displacement time series of the center of plantar pressure (COP) in the medial and lateral (ML) directions for 20 consecutive gait cycles were extracted.

[0119] Calculate the maximum Lyapunov exponent (λ) max Using the Wolf algorithm, with embedding dimension m and time delay τ, λ is calculated. max A positive value indicates that the gait system is sensitive to small perturbations and has poor local stability.

[0120] Calculate the sample entropy (SampEn). A lower value indicates that the gait pattern tends to be simple and rigid, with poor adaptability.

[0121] The formula used to calculate the posture control disorder coefficient is as follows:

[0122]

[0123] Where, λ ref SampEn ref These are reference values ​​for healthy women of the same age, with α and β being weighting coefficients.

[0124] Finally, tensor generation is performed: [step velocity, step size coefficient of variation, proportion of double-support phase, λ] max [SampEn,PCD] constitutes the third feature tensor.

[0125] In one scheme, such as Figure 3 As shown, the steps of the adaptive modal fusion based on uncertainty awareness include:

[0126] Step S221: Real-time assessment of the prediction uncertainty of each physiological signal mode, with the goal of minimizing the overall decision uncertainty;

[0127] Step S222: Dynamically adjust the weights of each mode during the fusion process to address signal quality fluctuations and mode degradation issues;

[0128] The cross-modal representation learning employs a mask fusion strategy and a variational autoencoder architecture to align and uniformly represent different physiological modalities with metadata related to age, gender, and circadian rhythms in a Gaussian latent space.

[0129] Specifically, in the implementation of uncertainty-aware fusion, taking the fusion of ECG and PPG signals to estimate heart rate variability (HRV) as an example:

[0130] The first step is to slice the signal. For example, the nighttime signal can be divided into multiple time windows with a window length of 5 minutes and a 50% overlap.

[0131] The second step is to quantify the uncertainty.

[0132] For the ECG mode: Calculate the confidence level Cecg for R-wave detection within this window. Formula: Cecg = (1 - artifact R-wave proportion) × (Signal-to-noise ratio of R-wave amplitude). The signal-to-noise ratio is calculated as the ratio of the signal power within this window to the baseline noise power.

[0133] For the PPG mode: Calculate the confidence level C of the PPG signal within this window. ecg Formula: C ecg = Perfusion Index (PI) × (1 - Motion Artifact Index). Where PI is calculated from the AC / DC component ratio, and the Motion Artifact Index is determined by the variance of the synchronized ACC signal.

[0134] The third step is adaptive weight calculation.

[0135] A fusion rule based on Dempster-Shafer evidence theory is adopted. The confidence level of each modality is regarded as the "quality" of the evidence provided by that modality. For the i-th window, the quality of evidence provided by ECG and PPG are m respectively. ecg(i) = C ecg(i) m ppg(i) = C ppg(i) The reliability assignment for "heart rate value H" after fusion is as follows:

[0136]

[0137] Where j, k correspond to all conflicting evidence pairs.

[0138] Ultimately, the window's fused heart rate value HR fused(i) It is determined by the highest heart rate zone of Bel(H).

[0139] The fourth step is global feature calculation.

[0140] HR of all windows fused(i) Sequence splicing is performed to calculate the frequency domain HRV index (LF, HF) for the entire night. This fusion method ensures that the PPG signal quality is good during periods of suppression and poor quality (C). ecg (Low), the system mainly relies on ECG evidence (C ecg (High), thus obtaining a more reliable second feature tensor input.

[0141] In the fusion and computation layer of this invention, in order to achieve unified alignment of different modal features at the semantic level and improve the system's robustness to signal quality fluctuations, cross-modal representation learning adopts a multimodal variational autoencoder fusion network, the specific implementation steps of which are as follows:

[0142] The first step involves constructing independent encoders for static image feature tensors, dynamic physiological signal feature tensors, gait mechanics feature tensors, TCM constitution identification vectors, and genetic risk scores (PRS).

[0143] The static image feature tensor is mapped to a 64-dimensional vector h through a three-layer fully connected network (each layer has 128, 64, and 64 neurons respectively, the activation function is ReLU, and batch normalization uses LayerNorm). img .

[0144] The dynamic physiological signal feature tensor is mapped to a 64-dimensional vector h through a one-dimensional convolutional network (three layers, kernel size 5, stride 2, activation function ReLU). phy .

[0145] The gait mechanics feature tensor is mapped to a 64-dimensional vector h via a bidirectional LSTM (32 hidden units). gait .

[0146] Among them, the TCM constitution identification vector is mapped to a 32-dimensional vector h through a fully connected layer. tcm .

[0147] The genetic risk score (PRS) is mapped to a 16-dimensional vector h through an embedding layer (16 embedding dimensions). gen .

[0148] The second step is modal confidence weighting.

[0149] Before the features enter the fusion module, the signal quality of each modality (e.g., ECG signal-to-noise ratio, gait data missing rate, image resolution) is evaluated in real time, modality confidence coefficients α∈[0,1] are generated, and the private representations of each modality are weighted h'. m α m *h m This is to reduce the interference of low-quality modes on the fusion results.

[0150] The third step is the cross-modal attention fusion module.

[0151] The weighted modal vectors are concatenated to form [h] img ][h phy ][h gait ][h tcm ][h gen(Dimension 240), input a multi-head self-attention mechanism (4 attention heads, each with dimension 60), calculate the attention weights of each modality to the fusion result, and generate the fusion feature vector h. fused .

[0152] The fourth step is metadata conditionalization.

[0153] Metadata such as actual age, gender, and circadian rhythm tags are encoded into a conditional vector c (32 dimensions), and then transformed by linear transformation W. c Injection fusion feature: h' fused =h fused +W c *c enables the latent space representation to carry demographic and temporal context information.

[0154] Step 5: Share the latent space encoder.

[0155] The conditionalized fusion features are mapped to a unified Gaussian latent space;

[0156] An intermediate representation is obtained through a fully connected layer (240→128, activation functions ReLU, LayerNorm);

[0157] The two branches generate the mean μ and the log-variance log(σ). 2 (Both are 256-dimensional);

[0158] The latent space vector Z is obtained by sampling using the reparameterization technique:

[0159]

[0160] The vector Z is the latent space representation of the health status of the bone system as described in this invention.

[0161] Step 6: Loss Function Design and Training. During training, the following three types of losses are jointly optimized:

[0162] Reconstruction loss L rec The modal features are decoded and the mean square error is minimized to ensure that the latent space retains the original information.

[0163] KL divergence loss L KL Constraining the latent space to approximate a standard normal distribution improves generalization ability.

[0164] Cross-modal contrast loss L contra It brings the latent space representations of the same object closer together and widens the representations of different objects, thus enhancing semantic alignment.

[0165] The total loss function is: L total =λ rec L rec +λ KL LKL +λ contra L contra The typical weight is set to λ. rec =1.0,λ rec =0.01,λ contra =0.1, and can be adjusted based on the performance on the validation set.

[0166] Through the above structure, the present invention achieves high-fidelity and robust representation of multimodal features in a unified latent space, providing a reliable semantic alignment feature basis for subsequent physiological age calculation and dynamic assessment of fracture risk.

[0167] In summary, the cross-modal representation learning using a multimodal variational autoencoder fusion network can be understood as including:

[0168] A modality-specific encoder is used to map each modality feature into a private space vector;

[0169] The modal confidence weighting module is used to apply weights to each modal vector based on the signal quality assessment results.

[0170] A cross-modal attention fusion module is used to fuse weighted modal vectors based on a multi-head self-attention mechanism;

[0171] The metadata conditionalization module is used to encode actual age, gender, and circadian rhythm labels into conditional vectors and inject them into fusion features;

[0172] A shared latent space encoder is used to map the conditionalized fused features to a unified Gaussian latent space representation;

[0173] The network jointly optimizes the reconstruction loss, KL divergence loss, and cross-modal contrast loss during training.

[0174] In one scheme, such as Figure 4 As shown, the step of extracting posture control disorder indicators includes:

[0175] Step S231: Calculate the asymmetric coefficients of the left and right foot drive index and braking index in each gait cycle to obtain the posture control factor;

[0176] Step S231: Perform curve fitting on the sequence of posture control factors within a continuous time window, analyze the frequency and slope changes of their numerical distribution, and comprehensively obtain the control disorder coefficient that reflects gait stability and neuromuscular control ability.

[0177] In one approach, the method for processing the multi-source heterogeneous data and feature tensor set to obtain a latent space representation of the bone system health status includes:

[0178] The first feature tensor, the second feature tensor, the third feature tensor, TCM constitution identification data, and genetic risk data are input into a multimodal deep fusion network based on tensor decomposition and expert product technology for alignment and fusion to generate a unified high-dimensional latent space representation of the bone system health status.

[0179] Among them, such as Figure 5 As shown, the step of aligning and fusing the input into a multimodal deep fusion network based on tensor decomposition and expert product techniques includes:

[0180] Step S310: Concatenate the feature tensors extracted from each modality to form a higher-order tensor;

[0181] Step S320: Use Tucker decomposition or CP decomposition to reduce the dimensionality of the higher-order tensor and extract core information;

[0182] Step S330: To address potential modality loss, the available modality expert networks are combined using the expert product technique to generate a complete latent space representation of the skeletal system's health status.

[0183] In a specific embodiment, the first feature tensor, the second feature tensor, and the third feature tensor, along with the TCM constitution vector and the PRS scalar, are input into the fusion module. Then, the following steps are performed:

[0184] The first step is tensor construction. Features of different dimensions are unified into a 256-dimensional vector through padding and linear projection, and then stacked into a third-order tensor of "sample x modality x feature".

[0185] The second step is Tucker decomposition. The third-order tensor is decomposed using the Tucker decomposition formula:

[0186]

[0187] in, It is the core tensor, U (1) U (2) U (3) These are the factor matrices for the sample, modality, and feature space, respectively. After decomposition, the core tensor... It compresses cross-modal interaction information.

[0188] The third step is latent space representation. This involves representing the core tensor... The process unfolds and inputs a two-layer perceptron, ultimately outputting a 256-dimensional, unified "latent space representation of bone system health status." One dimension of this representation may correspond to a synergistic aging pattern of "bone metabolism-sympathetic nervous system-gait stability."

[0189] In one embodiment, the method for obtaining the physiological age of the target individual includes:

[0190] Based on the multidimensional physiological biomarker data, the physiological age (BA) of the target individual is calculated by processing the data using the PhenoAge algorithm and / or the Klemera-Doubal method age (KDMAge) algorithm.

[0191] Physiological age is a comprehensive indicator that is independent of actual age and quantifies the degree of biological aging of the skeletal system.

[0192] The preset machine learning regression model is a gradient boosting regression tree or a deep neural network, and it is trained using multi-center, multi-ethnic cohort data.

[0193] The multicenter, multi-ethnic cohort data includes at least a bone age dataset for children in plains and plateau regions, a pelvic CT age estimation dataset for Han Chinese adults in western China, and a multinational dataset integrating the UK Biobank, the US National Health and Nutrition Examination Survey, and the Chinese Osteoporosis Fracture Specialty Cohort, to ensure the model's generalization ability to different ethnic and regional populations.

[0194] Specifically, in this embodiment, the biomarkers include at least one of albumin, creatinine, glucose, C-reactive protein, lymphocyte percentage, mean corpuscular volume, chronological age, alkaline phosphatase, white blood cell count, red blood cell distribution width, glycated hemoglobin, total cholesterol, blood urea nitrogen, and systolic blood pressure. The calculation formula for the above phenotypic age algorithm is as follows:

[0195]

[0196] Among them, Z b = -18.511 - 0.2502×Albumin + 0.0063×Cretin + 0.1741×Glucose + 0.1393×ln(C-reactive protein) - 0.0126×Lymphocyte percentage + 0.0292×Mean red blood cell volume + 0.0777×Live age + 0.0021×Alkaline phosphatase + 0.0547×White blood cell count + 0.2318×Red blood cell distribution width $$ $$White blood cell count + 0.2318×Red blood cell distribution width.

[0197] KDMAge is estimated using eight physiological biomarkers: log-transformed C-reactive protein, serum creatinine, glycated hemoglobin, serum albumin, total cholesterol, blood urea nitrogen, alkaline phosphatase, and systolic blood pressure. In the calculation, i represents the sample index, and j represents the biomarker. Parameters k, q, and s represent the slope, intercept, and root mean square error of each biomarker in the regression analysis of actual age, respectively. The formula for calculating age using the Klemera-Doubal method is as follows:

[0198]

[0199]

[0200]

[0201]

[0202] Where, k j q j s j These represent the slope, intercept, and root mean square error of the regression analysis of the j-th biomarker on actual age, respectively. 2 j G represents the proportion of variance explained by the regression. A This is an adjustment item.

[0203] In one embodiment, the method for determining an accelerated aging state based on the physiological age and a preset threshold includes:

[0204] Based on the physiological age and the actual age of the target individual, the age acceleration residual is calculated, and the individual is determined to be in an "accelerated aging" or "non-accelerated aging" state based on a preset threshold.

[0205] In a specific embodiment, the method for determining accelerated aging is roughly as follows:

[0206] The first step is to establish a reference line.

[0207] For example, using physiological age as the dependent variable and chronological age as the independent variable, a linear regression equation is fitted: Physiological age = 0.96 × Chronological age + 5.2.

[0208] The second step is to calculate the expected value and the residual.

[0209] For example, if the actual age is 68 years old, the expected physiological age = 0.96 × 68 + 5.2 = 70.48 years old. The actual physiological age is 76.8 years old. Therefore, the accelerated age residual = 76.8 - 70.48 = +6.32 years.

[0210] The third step is threshold determination.

[0211] For example, the 90th percentile of the reference population residuals is +4.5 years. If the residuals (+6.32 years) > +4.5 years, it is determined that the individual is in a state of "clearly accelerated aging".

[0212] In one embodiment, the method for calculating physiological age and processing fracture risk data based on multimodal approaches further includes: combining the TCM constitution identification data with a preset rule engine and recommendation algorithm to generate a personalized health management strategy that includes integrated TCM and Western medicine intervention measures.

[0213] The method for generating the personalized health management strategy includes:

[0214] If the condition is diagnosed as "accelerated aging" and the fracture risk level is "high," the recommended strategy is to prioritize anti-bone resorption drugs, anti-inflammatory treatment, and traditional Chinese medicine plasters and acupoint applications for those with "kidney yang deficiency" or "qi and blood deficiency."

[0215] If the aforementioned control disorder indicators are abnormally elevated, balance training and physical therapy are recommended simultaneously.

[0216] The personalized health management strategy is output in the form of digital reports and visual charts, and is connected to the traditional Chinese medicine health management platform for continuous tracking.

[0217] In one approach, the osteoporotic fracture risk prediction model uses an attention mechanism to dynamically weight different dimensions of the latent space representation of the skeletal system's health status, and it also incorporates a mediation effect analysis module to use logistic regression and Cox proportional hazards models to quantify the mediating effect proportion of physiological age between fracture and mortality risk.

[0218] The integrated model architecture is as follows:

[0219] Input layer: Receives a hybrid feature vector consisting of latent space representation (256 dimensions), physiological age, acceleration state flag (0 / 1), PRS value, and key primitive features (such as Tb.Sp, PCD, LF / HF).

[0220] Attention layer: A multi-head self-attention layer used to allow the model to automatically focus on the interactions between different features. For example, it may learn the synergistic enhancement effect on fracture risk when "high PCD" and "low Tb.N" occur simultaneously.

[0221] Parallel prediction head: Head A (fracture risk): A fully connected network connected to a Sigmoid output layer, outputting the fracture probability between 0 and 1. Head B (mortality risk): Connected to a parameter estimation layer of a Cox proportional hazards model, outputting a hazard ratio (HR) related prediction.

[0222] The specific steps in the mediation effect analysis are as follows:

[0223] The first step is to fit a Cox model with "whether a fracture occurred" as the independent variable, for example, with "whether the person died one year after the fracture" as the dependent variable, to obtain the total effect coefficient c of the fracture on mortality.

[0224] The second step is to simultaneously add two independent variables, "whether a fracture has occurred" and "physiological age," to the model, fit the Cox model, and obtain the coefficient c1 for fracture.

[0225] Third step, the mediating effect = c - c1. The mediating effect ratio = (c - c1) / c.

[0226] The fourth step involves sampling the original data 1000 times with replacement, repeating step ac each time, to obtain 1000 estimates of the mediation effect proportions. The 2.5th and 97.5th percentiles of these estimates constitute the 95% confidence interval (2.5%–10.1%).

[0227] In a specific embodiment, latent space representation, physiological age, acceleration state, and PRS are input into the risk prediction model.

[0228] The first step is fracture risk prediction.

[0229] For example, the logistic regression sub-model outputs the probability of developing a major osteoporotic fracture within 5 years. Risk levels are categorized according to pre-defined clinical grading criteria: low risk (<10%), medium risk (10-20%), and high risk (>20%).

[0230] The second step is to predict the risk of death.

[0231] For example, Cox proportional hazards submodel analysis showed that if a hip fracture occurs, the risk of all-cause mortality within one year after surgery is 2.8 times that of individuals of the same age, sex, and non-accelerated aging (hazard ratio HR=2.8, 95% CI: 2.1-3.7).

[0232] The third step is mediation effect analysis (Bootstrap method):

[0233] We analyze the mediating effect of physiological age in the pathway of "accelerated physiological age → fracture → death".

[0234] For example, with Bootstrap repeated 1000 times, the mediating effect was calculated to be 6.3% (95% CI: 2.5% - 10.1%). This means that of the increased risk of death from a fracture, approximately 6.3% can be attributed to the accelerated aging of the skeletal system itself.

[0235] In one embodiment, the method for calculating physiological age and processing fracture risk data based on multimodal approaches further includes: implementing a refined operational process for full-cycle health management through an integrated human-computer interaction and data management platform.

[0236] like Figure 6 , Figure 7 As shown, this refined operation process, namely step S700, includes the following steps during execution:

[0237] Step S710: Data standardization, collection, and quality control;

[0238] Step S720: Streamlined monitoring of feature processing and fusion;

[0239] Step S730: Strategic generation and delivery of risk assessments;

[0240] Step S740: Dynamic feedback on intervention effects and optimization of the plan.

[0241] Specifically, in this embodiment, the data normalization, collection, and quality control steps for step S710 are as follows:

[0242] The platform generates a data acquisition task list for the target individual based on a preset sample of "Multimodal Data Acquisition Operation Specifications". It automatically acquires data from medical image archiving systems, certified wearable device manufacturers' cloud platforms, traditional Chinese medicine constitution identification terminals, and gene testing platforms through interfaces. For each type of data received, it performs rule-based automated quality control, including checking whether the image resolution is lower than a preset threshold (e.g., CT image slice thickness > 1.25 mm), whether the duration of continuous missing physiological signals exceeds the allowable range (e.g., > 5% of the total duration), and the completeness of the traditional Chinese medicine constitution questionnaire. Data that fails the quality control is automatically triggered to trigger a re-acquisition or correction request.

[0243] For step S720, the workflow monitoring steps for feature processing and fusion are as follows:

[0244] The platform will distribute the quality control data to the corresponding feature processing engines according to modality and record the processing status log of each engine. The platform's built-in process monitoring module monitors the nodes of four key sub-processes, namely "3D reconstruction and segmentation", "uncertainty perception fusion", "nonlinear dynamic analysis" and "TCM body quality scoring", according to the "Bone Health Assessment Feature Extraction Flowchart". If any sub-process times out or reports an error, the system administrator will be automatically notified and an attempt will be made to enable a backup computing node.

[0245] For step S730, the strategy-based generation and push steps of the risk assessment are as follows:

[0246] After receiving the output from the risk assessment module, the platform calls the rule engine and, in conjunction with the target individual's TCM constitution identification data, matches and generates a preliminary strategy draft containing specific measures, dosages, and frequencies from the pre-set "Personalized Health Management Strategy Knowledge Base." After the draft undergoes compliance review built into the platform, a final digital health management report is generated and pushed to the target individual's bound mobile application and their designated attending physician via an encrypted channel.

[0247] For step S740, the dynamic feedback and optimization steps of the intervention effect are as follows: During the strategy execution cycle, the platform regularly collects individual compliance feedback, subjective feeling scores, and authorized simplified review data (such as tongue images collected through mobile phone cameras and weekly average gait stability coefficients collected through wearable devices) through mobile application terminals. The platform uses this feedback data to dynamically evaluate the initial strategy through a pre-trained optimization algorithm model. If the evaluation indicators (such as compliance rate <70% or pain score decrease not meeting expectations) do not meet the preset targets, the platform automatically generates strategy adjustment suggestions, which are updated and pushed to the new version of the strategy after online confirmation by the physician.

[0248] This invention also provides an osteoporotic fracture risk assessment system based on multimodal physiological age, comprising:

[0249] The biological age calculation module is used to execute the PhenoAge and KDMAge algorithms in the above embodiments;

[0250] The risk prediction module is used to construct fracture risk models based on biological age, age acceleration index, and PRS.

[0251] The results output module is used to output the probability of fracture risk, mortality risk, and intervention recommendations.

[0252] The causal verification module is used to verify the causal relationship between biological age and fracture using Mendelian randomization methods.

[0253] The mechanism analysis module is used to identify key genes and pathways through enrichment analysis.

[0254] The system will perform the following steps when it is executed:

[0255] Collect multidimensional physiological biomarker data from individuals;

[0256] Based on the biomarker data, the individual's biological age (BA) is calculated using the PhenoAge algorithm and / or the Klemera-Doubal method age (KDMAge) algorithm.

[0257] The residual between the biological age and the actual age is calculated to obtain the age acceleration index (PhenoAgeAccel or KDM-AgeAccel);

[0258] Obtain an individual's polygenic risk score (PRS) for osteoporotic fractures;

[0259] Based on the aforementioned biological age, age acceleration index, and PRS, an osteoporotic fracture risk prediction model is constructed using a logistic regression model, a Cox proportional hazards model, or a machine learning model, outputting fracture risk probability and all-cause mortality assessment results.

[0260] The input features of the aforementioned fracture risk prediction model also include: the mediating effect of biological age on the risk of death after fracture, the mediating effect of accelerated aging, and the dose-response relationship between accelerated aging and PRS.

[0261] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0262] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications, substitutions, and improvements without departing from the concept of the present invention, and these should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of this invention should be determined by the claims.

Claims

1. A method for calculating physiological age and processing fracture risk data based on multimodal approaches, characterized in that, include: Collect multi-source heterogeneous data of the target individual; Based on the aforementioned multi-source heterogeneous data, a set of feature tensors is obtained; The multi-source heterogeneous data and feature tensor set are processed to obtain a latent space representation of the health status of the bone system; The latent space of the health status of the skeletal system is represented and processed by a preset machine learning regression model to obtain the physiological age of the target individual; Based on the physiological age and a preset threshold, an accelerated aging state is determined; Based on the latent space representation of the bone system health status, physiological age, aging acceleration status, and multi-source heterogeneous data, an osteoporotic fracture risk prediction model is constructed to output the probability of osteoporotic fracture risk and the predicted value of all-cause mortality risk after fracture.

2. The method for calculating physiological age and processing fracture risk data based on multimodal data according to claim 1, characterized in that, The multi-source heterogeneous data includes at least static bone and joint imaging data, continuous dynamic physiological signal data, temporal gait mechanics data, traditional Chinese medicine constitution identification data, genetic risk data, and multi-dimensional physiological biomarker data. The method for obtaining a set of feature tensors based on the multi-source heterogeneous data includes: Three-dimensional reconstruction and depth feature extraction are performed on the static image data of the bone joints to obtain the first feature tensor characterizing the morphology and density of bone microstructure. The continuous dynamic physiological signal data is subjected to adaptive modal fusion and cross-modal representation learning based on uncertainty perception to obtain a second feature tensor representing cardiovascular and neuromodal functions; Nonlinear dynamic analysis is performed on the temporal gait mechanics data to extract posture control disorder indices and obtain the third characteristic tensor.

3. The method for calculating physiological age and processing fracture risk data based on multimodal data according to claim 2, characterized in that, The method for adaptive modal fusion based on uncertainty awareness includes: Real-time assessment of the prediction uncertainty of each physiological signal mode, with the goal of minimizing overall decision uncertainty; The weights of each mode are dynamically adjusted during the fusion process to address signal quality fluctuations and mode degradation issues. The cross-modal representation learning employs a mask fusion strategy and a variational autoencoder architecture to align and uniformly represent different physiological modalities with metadata related to age, gender, and circadian rhythms in a Gaussian latent space.

4. The method for calculating physiological age and processing fracture risk data based on multimodal data according to claim 2, characterized in that, The method for extracting posture control disorder indicators includes: Calculate the asymmetric coefficients of the left and right foot drive index and braking index in each gait cycle to obtain the posture control factor; Curve fitting was performed on the sequence of posture control factors within a continuous time window to analyze the frequency and slope changes of their numerical distribution, so as to obtain the control disorder coefficient that reflects gait stability and neuromuscular control ability.

5. The method for calculating physiological age and processing fracture risk data based on multimodal data according to claim 2, characterized in that, The method for processing the multi-source heterogeneous data and feature tensor set to obtain the latent space representation of the bone system health status includes: The first feature tensor, the second feature tensor, the third feature tensor, TCM constitution identification data, and genetic risk data are input into a multimodal deep fusion network based on tensor decomposition and expert product technology for alignment and fusion to generate a unified high-dimensional latent space representation of the bone system health status. The method for aligning and fusing the input into a multimodal deep fusion network based on tensor decomposition and expert product techniques includes: The feature tensors extracted from each modality are concatenated to form a higher-order tensor. The higher-order tensors are reduced in dimensionality and their core information is extracted using Tucker decomposition or CP decomposition. To address potential modality gaps, an expert product technique is employed to combine available modality expert networks to generate a complete latent space representation of the skeletal system's health status.

6. The method for calculating physiological age and processing fracture risk data based on multimodal analysis according to any one of claims 2 to 5, characterized in that, The method for obtaining the physiological age of the target individual includes: Based on the multidimensional physiological biomarker data, the physiological age of the target individual is calculated by processing the data using a phenotypic age algorithm and / or the Klemera-Doubal method age algorithm. Physiological age is a comprehensive indicator that is independent of actual age and quantifies the degree of biological aging of the skeletal system. The preset machine learning regression model is a gradient boosting regression tree or a deep neural network, and it is trained using multi-center, multi-ethnic cohort data. The multi-center, multi-ethnic cohort data includes at least a bone age dataset of children in plains and plateau regions, a pelvic CT age estimation dataset of Han Chinese adults in western China, and a multinational dataset integrating the UK Biobank, the US National Health and Nutrition Examination Survey, and the Chinese Osteoporosis Fracture Specialty Cohort, to ensure the generalization ability of the preset machine learning regression model to different ethnic and regional populations.

7. The method for calculating physiological age and processing fracture risk data based on multimodal data according to any one of claims 2 to 5, characterized in that, The method for determining accelerated aging based on the physiological age and a preset threshold includes: Based on the physiological age and the actual age of the target individual, the age acceleration residual is calculated, and the individual is determined to be in an "accelerated aging" or "non-accelerated aging" state based on a preset threshold.

8. The method for calculating physiological age and processing fracture risk data based on multimodal data according to claim 7, characterized in that, Also includes: Based on the TCM constitution identification data, a personalized health management strategy that includes integrated TCM and Western medicine intervention measures is generated through a preset rule engine and recommendation algorithm. The method for generating the personalized health management strategy includes: If the condition is diagnosed as "accelerated aging" and the fracture risk level is "high," the recommended strategy is to prioritize anti-bone resorption drugs, anti-inflammatory treatment, and traditional Chinese medicine plasters and acupoint applications for those with "kidney yang deficiency" or "qi and blood deficiency." If the aforementioned control disorder indicators are abnormally elevated, balance training and physical therapy are recommended simultaneously. The personalized health management strategy is output in the form of digital reports and visual charts, and is connected to the traditional Chinese medicine health management platform for continuous tracking.

9. The method for calculating physiological age and processing fracture risk data based on multimodal analysis according to any one of claims 1 to 5, characterized in that, The osteoporotic fracture risk prediction model uses an attention mechanism to dynamically weight different dimensions of the latent space representation of the bone system's health status, and it also has a built-in mediation effect analysis module to quantify the proportion of the mediating effect of physiological age between fracture and mortality risk.

10. The method for calculating physiological age and processing fracture risk data based on multimodal analysis according to any one of claims 1 to 5, characterized in that, It also includes a refined operational process for performing full-cycle health management through an integrated human-computer interaction and data management platform, the refined operational process including: Data standardization, collection, and quality control; Streamlined monitoring of feature processing and fusion processes; Strategic generation and delivery of risk assessments; Dynamic feedback on intervention effects and optimization of the intervention plan.

Images