A bone health status personalized tracking and risk alert system

Abstract

The present application belongs to the field of intelligent medical treatment and digital health technology, and provides a bone health state personalized tracking and risk prompting system, which comprises a data sensing layer, a feature fusion layer, an intelligent decision-making layer and an interactive application layer connected in sequence. The data sensing layer collects PPG signals, motion signals and structured health data of users through intelligent wearable devices. The feature fusion layer extracts western medicine physiological feature sequences, traditional Chinese medicine Zangxiang feature vectors and lifestyle feature vectors. The intelligent decision-making layer adopts a double-track driving mechanism of personal dynamic baseline tracking and traditional Chinese and western medicine fusion feature analysis. Early warning and personalized intervention schemes are distributed in the interactive application layer, forming a closed loop of monitoring, evaluation, prompting, intervention and re-evaluation. The present application realizes long-term, dynamic and non-invasive continuous tracking and prospective risk prompting of bone health state, and solves the pain points of traditional imaging examination, such as staticity, fragmentation and inability to be integrated into daily management.

Description

Technical Field

[0001] This invention relates to the field of smart healthcare and digital health technology, and in particular to a personalized tracking and risk alert system for bone health status. Background Technology

[0002] Bone health issues, especially osteoporosis and the resulting fragility fractures, have become an increasingly serious public health challenge worldwide. Osteoporosis, a systemic bone disease characterized by low bone mass and destruction of bone microstructure, has an insidious onset. Patients are often only diagnosed after fragility fractures occur in areas such as the vertebrae and hip, at which point it has already caused serious and irreversible damage to their personal health and quality of life, and imposes a huge socioeconomic burden.

[0003] Currently, the gold standard for assessing bone health and diagnosing osteoporosis in clinical practice is dual-energy X-ray absorptiometry (DXA). DXA can accurately measure bone mineral density in specific areas (such as the lumbar spine and hip) and generate T-scores for diagnosis. However, DXA has significant limitations: First, the equipment is expensive and bulky, requiring operation by technicians in specialized medical institutions, resulting in low accessibility; second, the examination involves ionizing radiation, and although the dose is small, it is still unsuitable as a frequent, long-term monitoring method; third, DXA measurements are essentially static, reflecting only the cumulative bone mineral density at the time of testing, and cannot capture the short-term fluctuations and long-term trends of bone metabolism, a dynamic physiological process, nor can it assess the biomechanical function of bones and fall risk.

[0004] To obtain more accurate volumetric bone mineral density and bone microstructure information, quantitative computed tomography (QCT) has been applied. Although QCT can provide three-dimensional bone mineral density data and distinguish between cortical bone and cancellous bone, its radiation dose is significantly higher than that of DXA, and its cost is higher. It also has the same limitations of static assessment and reliance on large medical equipment, making it difficult to integrate into daily health management.

[0005] In recent years, breakthroughs in artificial intelligence technology have provided new technological pathways for bone health screening. For example, some studies have used deep learning models to automatically analyze vertebral CT values ​​or texture features by mining spinal images included in routine chest CT scans, indirectly assessing bone density and achieving low-cost preliminary screening with a single image serving multiple purposes. Systems such as BoneVoyage use a cascaded lightweight neural network and deep network integration framework to directly analyze bone density images, reportedly achieving high-precision classification. In terms of multimodal data fusion, some works combine X-ray and MRI data and introduce cross-attention mechanisms to improve the accuracy of diagnosis and grading of knee osteoarthritis. The latest risk assessment research uses graph neural networks to quantify the transmission path of fracture risk by simulating the mechanical connections between vertebrae in the spine, thereby enhancing the interpretability of the model.

[0006] However, the aforementioned AI-based technical solutions still fail to overcome a fundamental limitation: the static and fragmented nature of the assessments. Whether relying on single DXA, QCT, or AI analysis of single CT or X-ray images, the output results cannot meet the management needs for long-term, continuous, and personalized tracking of bone health. At the same time, these methods are severely limited by professional medical imaging equipment and cannot implement continuous and imperceptible monitoring in users' daily life scenarios. Therefore, they cannot capture the microscopic and cumulative effects of lifestyle habits, daily activity patterns, and physiological rhythm fluctuations on bone health, thus missing the opportunity for early risk warning.

[0007] Existing publicly available continuous monitoring and early warning systems have not yet had their technological advantages systematically guided and deeply applied to the specific and important vertical field of bone health. Current technologies have not systematically addressed the following key issues: how to specifically design and extract features strongly correlated with bone metabolism, muscle, and skeletal function based on multi-source continuous data collected from wearable devices; how to deeply couple the quantitative features of traditional Chinese medicine pulse diagnosis with Western medical physiological indicators at the algorithmic level to construct a dedicated assessment model for bone health; how to establish a trend tracking and risk prediction mechanism based on a fully individualized dynamic baseline; and how to form a complete management closed loop from data collection, risk identification, personalized intervention to effect feedback. These issues have not yet been systematically elaborated and resolved in existing technologies.

[0008] Therefore, there is a need for a complete system that can utilize convenient wearable devices to integrate continuous physiological data with quantitative characteristics of traditional Chinese medicine, enabling long-term, dynamic, and personalized tracking of individual bone health status and forward-looking intelligent risk alerts. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of the prior art and provide a personalized bone health status tracking and risk warning system. By integrating wearable sensors and multimodal data fusion technology, a layered architecture of data perception, feature fusion, intelligent decision-making, and interactive application is constructed to realize continuous monitoring, dynamic assessment, and graded early warning of bone health status.

[0010] This invention provides a personalized bone health status tracking and risk warning system, comprising a data perception layer, a feature fusion layer, an intelligent decision-making layer, and an interactive application layer, which are connected sequentially. The data perception layer collects photoplethysmography (PPG) signals, triaxial acceleration signals, and gyroscope signals through the smart wearable device worn by the user, and collects structured data including age, gender, diet, exercise, medication, history of fracture, basic bone mineral density classification, and major TCM constitution type through the application APP used by the user, and generates the collected data. The feature fusion layer preprocesses the collected data and extracts Western medicine physiological feature sequences, traditional Chinese medicine Zang-Xiang feature vectors, and lifestyle feature vectors. The intelligent decision-making layer, based on the Western medicine physiological feature sequence, the traditional Chinese medicine Zang-Xiang feature vector, and the lifestyle feature vector, uses the personal bone health dynamic baseline model, the integrated traditional Chinese and Western medicine bone status assessment model, and the time series prediction and risk warning model to conduct personal bone health dynamic baseline tracking and integrated traditional Chinese and Western medicine feature analysis to obtain bone health analysis results. Among them, the personal bone health dynamic baseline model is used to establish a dynamic fluctuation range of personalized health indicators based on the continuous data collected by the user wearing smart wearable devices in the initial stage of data collection, and to continuously evolve the dynamic fluctuation range based on the sliding window and Bayesian update mechanism. A bone health assessment model integrating traditional Chinese and Western medicine employs an architecture including a two-stream encoder and a cross-modal attention fusion module. Based on the fusion features, it outputs a bone health status index and a distribution of TCM syndrome patterns. The two-stream encoder comprises a first encoder for processing Western medicine physiological feature sequences and a second encoder for processing TCM organ-manifestation feature vectors. The cross-modal attention fusion module is configured based on a formula... Calculate attention weights and fuse features based on formulas. Calculation generated; The time-series prediction and risk warning model predicts the future trend of the bone health status index based on historical data, obtains the prediction results, and triggers graded warnings based on the comparison between the prediction results and the dynamic fluctuation range. The interactive application layer distributes bone health analysis results and corresponding personalized intervention plans to user terminals, family member terminals, and doctor workstations to form a closed-loop management process of monitoring, evaluation, prompting, intervention, and re-evaluation.

[0011] Furthermore, the data perception layer integrates a data processing module, which is used to preprocess and initially calculate the user's posture stability index, gait symmetry index, and gait regularity index based on the three-axis acceleration signal and gyroscope signal, and incorporate the posture stability index, gait symmetry index, and gait regularity index as components of the collected data.

[0012] Furthermore, the Western medical pathological feature sequence includes vascular elasticity parameters and microcirculation state parameters; the vascular elasticity parameters and microcirculation state parameters are indirect biochemical correlation features related to bone metabolism estimated by inputting PPG signals into a pre-trained derivative model.

[0013] Furthermore, when establishing the dynamic fluctuation range of personalized health indicators, a robust statistical method based on the median and absolute median difference is used for calculation.

[0014] Furthermore, based on the sliding window and Bayesian update mechanism, the dynamic fluctuation range is continuously evolved, including: introducing a weighted sliding window update strategy to assign differentiated weights to historical data of different time periods; setting up a trend detection unit to use robust statistical methods to determine whether there is a significant monotonic trend in the Western medicine pathological feature sequence, the traditional Chinese medicine Zang-Xiang feature vector, and the lifestyle feature vector; when the trend is confirmed, triggering the fluctuation range drift update rule according to the Bayesian update mechanism to adjust the center position and the slope of the extended trend response; and setting a multi-level threshold mechanism to distinguish between trend drift and acute abnormal events in order to maintain the detection sensitivity of pathological indicator changes that deviate from the normal fluctuation pattern in the short term.

[0015] Furthermore, the extraction of TCM Zang-Xiang feature vectors includes: identifying and extracting quantitative features related to TCM kidney, liver, and spleen functions by performing time-frequency domain joint analysis and waveform morphology modeling on the pulse components in the acquired PPG signals; among which, the quantitative features related to kidney function include at least the pulse depth feature value and the pulse weakness index; the quantitative features related to liver function include at least the pulse stringiness feature value; and the quantitative features related to spleen function include at least the pulse softness and slowness feature value.

[0016] Furthermore, the tiered early warning system includes: When the bone health status index deviates from the dynamic fluctuation range for more than a certain number of days within a consecutive predetermined number of days, or when it is predicted that the bone health status index will enter the preset health attention data range in the future, the first-level warning for pushing lifestyle adjustment reminders will be triggered. When the bone health status index is within the preset sub-health data range, or when the TCM syndrome distribution continues to point to a specific TCM syndrome, a second-level warning is triggered to recommend TCM conditioning and medical examination. When the user's short-term fall risk predicted based on posture stability, gait symmetry, and gait regularity indicators exceeds the high-risk value, or when the bone health status index is about to enter the high-risk data range, a Level 3 warning is triggered to simultaneously notify emergency contacts or family doctors.

[0017] Furthermore, the intelligent decision-making layer is also used for: Based on the collected data, static portrait data is extracted and generated. The static portrait data includes at least age, gender, baseline bone density classification, and major TCM constitution types. Based on static profile data, a set of personalized modulation parameters is generated. The modulation parameters include: baseline offset vector, feature channel weight vector, and risk judgment threshold adjustment coefficient. The baseline offset vector is used to perform initial calibration of the center value of the dynamic fluctuation range based on population subgroups; the feature channel weight vector is used to weight the importance of each dimension of features in the Western medicine physiological feature sequence and the traditional Chinese medicine Zang-Xiang feature vector; the risk judgment threshold adjustment coefficient is used to personalize the scaling of the number of days threshold, high-level risk value, and high-risk data interval boundary in the graded early warning system. When initializing the personal bone health dynamic baseline model, a baseline offset vector is loaded; before the dual-stream encoder in the integrated Chinese and Western medicine bone status assessment model processes the features of each dimension in the Western medicine physiological feature sequence and the Chinese medicine Zang-Xiang feature vector, the features of each dimension in the Western medicine physiological feature sequence and the Chinese medicine Zang-Xiang feature vector are modulated using the feature channel weight vector; when making risk decisions in the time series prediction and risk warning model, a risk judgment threshold adjustment coefficient is applied.

[0018] Furthermore, the intelligent decision-making layer includes a decision routing module; the decision routing module stores a bone health intervention knowledge graph, which is used to generate personalized intervention plans based on bone health analysis results, including nutritional advice, exercise programs, traditional Chinese medicine dietary therapy recipes, acupoint massage guidance, and medical treatment indications.

[0019] Furthermore, the intelligent decision-making layer also includes a model co-optimization engine; the model co-optimization engine is configured as follows: Monitor and record user compliance feedback data, subsequent physiological indicator changes data, and external medical diagnosis results data after each tiered warning is triggered; Based on compliance feedback data, physiological indicator change data, and external medical diagnosis results data, an intervention effect verification dataset was constructed. Using the intervention effect validation dataset, reinforcement learning was performed on the cross-modal attention fusion module to optimize the learnable weight matrix; Based on the intervention effect verification dataset, the verified effective intervention patterns and results are extracted, and the association rules in the bone health intervention knowledge graph are updated with confidence weight or nodes are added.

[0020] Compared with the prior art, the present invention has the following advantages and beneficial effects: (i) A fundamental shift in assessment models and scenarios; bone health assessment will be transformed from static, single-point imaging examinations relying on large hospital equipment to dynamic, continuous tracking based on wearable devices, extending the assessment scenarios to daily life and capturing the microscopic and long-term effects of physiological rhythms and lifestyle habits on bone health; at the same time, wearable devices enable non-invasive, radiation-free, and continuous data collection, overcoming the limitations of traditional examination equipment that is expensive, inconvenient, and has radiation, making long-term, high-frequency bone health management possible and suitable for early screening and long-term follow-up of large populations.

[0021] (II) Multi-dimensional health profile and the synergistic mechanism of traditional Chinese and Western medicine; it breaks through the limitations of a single bone density index, and forms a comprehensive assessment system that reflects bone metabolism function, internal dynamics and external risks by integrating continuous Western medical psychological characteristics (such as nocturnal heart rate variability and vascular elasticity), quantifying traditional Chinese medicine Zang-Xiang characteristics (such as the strength of the cun pulse and the kidney qi index) and user behavior data; through the designed cross-modal attention fusion algorithm, the AI ​​model can understand and quantify the complex relationship between traditional Chinese medicine syndromes (such as kidney deficiency) and specific Western medical psychological abnormalities (such as autonomic nervous system disorders), realizing the deep coupling and mutual verification of traditional Chinese and Western medicine theories at the algorithm level.

[0022] (III) Personalized dynamic baseline assessment and graded early warning; a unique dynamic fluctuation range of bone health is established and continuously evolved for each user. The assessment standard has changed from comparison with the general threshold of the group to deviation analysis of the trend of the individual's historical baseline, which greatly improves the degree of personalization and assessment sensitivity and can effectively distinguish between physiological fluctuations and pathological trends. The system not only assesses the current status, but also uses time series models to predict the bone health trend in the future. Combining trend deviation, status threshold, TCM syndrome type and fall risk and other multi-dimensional information, it triggers graded early warning from life reminders to emergency notifications.

[0023] (iv) Intelligent closed-loop management and self-evolution capability; The system is not only a monitoring tool, but also a management engine. It can automatically generate and push personalized nutrition, exercise, traditional Chinese medicine and medical intervention plans based on risk assessment results, and track user compliance and physiological feedback to form a complete data-driven closed loop of monitoring, assessment, prompting, intervention and reassessment, which significantly improves the compliance and long-term effect of health management. In addition, through the model collaborative optimization engine, the system can continuously learn and optimize assessment and intervention strategies during use, and has the ability to self-evolve.

[0024] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and drawings.

[0025] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0026] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 A schematic diagram of a personalized bone health status tracking and risk alert system; Figure 2 A schematic diagram illustrating the components of a tiered early warning system; Figure 3 This is a schematic diagram illustrating the configuration of the model collaborative optimization engine. Detailed Implementation

[0027] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0028] This invention provides a personalized bone health status tracking and risk alert system, such as... Figure 1 As shown, it includes a data perception layer, a feature fusion layer, an intelligent decision-making layer, and an interactive application layer, which are connected sequentially. The data perception layer collects photoplethysmography (PPG) signals, triaxial acceleration signals, and gyroscope signals through the smart wearable device worn by the user, and collects structured data including age, gender, diet, exercise, medication, history of fracture, basic bone mineral density classification, and major TCM constitution type through the application APP used by the user, and generates the collected data. The feature fusion layer preprocesses the collected data and extracts Western medicine physiological feature sequences, traditional Chinese medicine Zang-Xiang feature vectors, and lifestyle feature vectors. The intelligent decision-making layer, based on the Western medicine physiological feature sequence, the traditional Chinese medicine Zang-Xiang feature vector, and the lifestyle feature vector, uses the personal bone health dynamic baseline model, the integrated traditional Chinese and Western medicine bone status assessment model, and the time series prediction and risk warning model to conduct personal bone health dynamic baseline tracking and integrated traditional Chinese and Western medicine feature analysis to obtain bone health analysis results. Among them, the personal bone health dynamic baseline model is used to establish a dynamic fluctuation range of personalized health indicators based on the continuous data collected by the user wearing smart wearable devices in the initial stage of data collection, and to continuously evolve the dynamic fluctuation range based on the sliding window and Bayesian update mechanism. A bone health assessment model integrating traditional Chinese and Western medicine employs an architecture including a two-stream encoder and a cross-modal attention fusion module. Based on the fusion features, it outputs a bone health status index and a distribution of TCM syndrome patterns. The two-stream encoder comprises a first encoder for processing Western medicine physiological feature sequences and a second encoder for processing TCM organ-manifestation feature vectors. The cross-modal attention fusion module is configured based on a formula... Calculate attention weights and fuse features based on formulas. Calculation generation, For attention weights, This represents the characteristic sequence of Western medical psychology. This represents the characteristic vector of Zang-Xiang in Traditional Chinese Medicine. , , These are the learnable weight matrices for queries, keys, and values, respectively. Scaling factor Let be the dimension of the key vector. The query matrix represents the learnable weight matrix. Feature vector of Zang-Xiang in Traditional Chinese Medicine Perform a linear transformation. The key matrix represents the key matrix, which is then processed by a learnable weight matrix. Western medical physiological characteristic sequence Perform a linear transformation. represents the transpose of the key matrix, used for dot product with the query matrix; Softmax represents normalizing the attention scores to a probability distribution. Represents the characteristics of integration; The integrated traditional Chinese and Western medicine bone status assessment model was trained using an end-to-end supervised learning approach. The training dataset was constructed as follows: synchronous data from several subjects were collected, with each dataset including PPG signals, triaxial acceleration signals, gyroscope signals for at least 7 consecutive days, and structured data collected via an app; an annotation expert group composed of at least three orthopedic surgeons with associate chief physician titles or above and three traditional Chinese medicine physicians annotated the dataset based on the subjects' DXA bone mineral density test results, serum bone metabolism markers (CTX, P1NP) test results, and traditional Chinese medicine... The results of the four diagnostic methods of traditional Chinese medicine were combined and labeled with two tags: (1) Bone Health Index (BHI), with a value range of 0-100, which was comprehensively evaluated by the expert group based on bone mineral density T-value, fracture risk assessment tool (FRAX) score and TCM syndrome score; (2) TCM syndrome distribution, which is the probability distribution vector of seven types of bone health-related TCM syndromes (kidney essence deficiency syndrome, spleen and kidney yang deficiency syndrome, liver and kidney yin deficiency syndrome, blood stasis obstruction syndrome, wind-cold-dampness bi syndrome, qi and blood deficiency syndrome, no obvious syndrome), which was labeled by the expert group based on the syndrome classification standard of "TCM Clinical Diagnosis and Treatment Terminology"; During model training, a joint loss function consisting of mean squared error loss and KL divergence loss is used. The mean squared error loss is used to optimize the regression accuracy of the bone health status index, while the KL divergence loss is used to optimize the fit of the TCM syndrome tendency distribution. The Adam optimizer is used during training, with an initial learning rate of 0.001, a batch size of 64, and 100 training epochs. The early stopping strategy is based on terminating when the validation set loss does not decrease for 10 consecutive epochs. All learnable weight matrices in the dual-stream encoder and cross-modal attention fusion module are optimized end-to-end synchronously using the backpropagation algorithm.

[0029] The time-series prediction and risk warning model predicts the future trend of the bone health status index based on historical data, obtains the prediction results, and triggers graded warnings based on the comparison between the prediction results and the dynamic fluctuation range. The interactive application layer distributes bone health analysis results and corresponding personalized intervention plans to user terminals, family member terminals, and doctor workstations to form a closed-loop management process of monitoring, evaluation, prompting, intervention, and re-evaluation.

[0030] The working principle of the above technical solution is as follows: the core technical architecture of the system consists of a data perception layer, a feature fusion layer, an intelligent decision-making layer and an interactive application layer connected in sequence. Each layer works together to form a complete closed-loop system from physiological signal acquisition to intelligent assessment and intervention feedback. The data perception layer is the system's input foundation, responsible for collecting multi-dimensional health data from users. This layer acquires three key physiological signals in real time through a smart wearable device worn by the user: photoplethysmography (PPG) signals, triaxial acceleration signals, and gyroscope signals. The smart wearable device is a wrist-worn physiological monitoring device equipped with a PPG sensor, a triaxial acceleration sensor, and a gyroscope sensor. The PPG sampling rate is no less than 50Hz, and the acceleration and gyroscope sampling rates are no less than 25Hz. It is worn on the wrist to meet the requirements for daily non-invasive continuous data acquisition. Among them, the PPG signal is used to reflect changes in microcirculation blood flow, indirectly characterizing the perfusion status of the skeletal microenvironment. The triaxial acceleration signal and gyroscope signal... The gyroscope signal is used to capture kinematic parameters such as daily activity patterns, gait stability, and fall risk, providing a basis for bone load and fracture risk modeling. At the same time, the system collects structured user information through a mobile APP, covering static demographic characteristics (age, gender), dynamic behavioral factors (dietary habits, exercise frequency), clinical history (medication status, history of fractures), medical test results (basal bone mineral density classification), and TCM constitution identification results (main TCM constitution types). The above two types of data together constitute the raw data, which has time series and multimodal characteristics, laying the data foundation for subsequent in-depth analysis. The feature fusion layer preprocesses the acquired data, transforming the raw signals into high-order feature representations with clear physiological significance. This process involves three parallel paths: Extraction of Western medical physiological feature sequences: For PPG signals, triaxial acceleration signals and gyroscope signals, after segmentation by sliding window, the time domain (mean, standard deviation, number of peaks), frequency domain (power spectral density main frequency) and nonlinear features (sample entropy, Lyapunov exponent) are calculated to construct Western medical physiological feature sequences that evolve over time. These features can quantify heart rate variability, limb activity rhythm and balance ability, and are associated with neuroendocrine pathways regulating bone metabolism. Traditional Chinese medicine Zang-Xiang feature vector extraction: By performing time-frequency domain joint analysis and waveform morphology modeling on the pulse components in the collected PPG signals, quantitative physiological parameters related to the functional status of the kidney, liver, and spleen in traditional Chinese medicine theory are identified and extracted. Lifestyle feature vectors are constructed by converting user-inputted calcium intake, vitamin D intake, estimated sunshine duration, and resistance exercise frequency into corresponding time-series event labels and intensity indices. All feature vectors have been standardized and time-aligned to ensure that data from different sources and scales can participate in fusion analysis in a unified space. The intelligent decision-making layer is the core reasoning engine of the system, which includes three cascaded models: a personal bone health dynamic baseline model, a traditional Chinese and Western medicine integrated bone status assessment model, and a time series prediction and risk warning model, which respectively complete the establishment of individualized benchmarks, multimodal status assessment, and future trend warning. Personal bone health dynamic baseline model: Using multi-source data collected continuously for 2-4 weeks in the initial stage of the user, a dynamic fluctuation range of user bone health related indicators is established. This model introduces a sliding window mechanism to advance with a fixed-length time window and combines a Bayesian update strategy. A Traditional Chinese Medicine (TCM) and Western Medicine integrated bone condition assessment model: This model employs a dual-stream encoder and a cross-modal attention fusion architecture to achieve information complementarity between Western medicine microscopic indicators and TCM macroscopic syndrome differentiation. Specifically: The first encoder employs an architecture combining a one-dimensional convolutional neural network and a gated recurrent unit to process Western medicine physiological feature sequences and output context-aware hidden state sequences. The second encoder uses a fully connected network architecture to process TCM Zang-Xiang feature vectors and generate fixed-length semantic embeddings. The cross-modal attention fusion module uses TCM Zang-Xiang features as queries and Western medicine feature sequences as keys and values. It can automatically focus on relevant Western medicine indicators (such as decreased nocturnal heart rate variability and reduced stride) based on an individual's TCM constitution type (such as kidney deficiency), thereby generating more interpretable fused features. Finally, the fused features output two results through the classification head: one is a continuous bone health status index (0–100), which comprehensively reflects the current overall functional level of the skeletal system; the other is a discrete TCM syndrome tendency distribution, which gives the probability distribution of each syndrome (such as kidney essence deficiency and spleen and kidney yang deficiency), supporting syndrome differentiation and treatment. Time-series prediction and risk warning model: Based on the historical bone health status index series (such as the daily average of the past 30 days), time series prediction algorithms (including but not limited to LSTM, ARIMA, Prophet, TemporalFusionTransformer, adaptively selected according to data length and noise level to ensure prediction stability and accuracy) are used to model its long-term trend, periodicity and sudden disturbance components, predict the status trend in the next 5-8 weeks, compare the prediction results with the individual dynamic baseline interval, and trigger graded warnings based on the comparison results.

[0031] The interactive application layer is responsible for transforming intelligent decision-making results into actionable health management recommendations and distributing them to multiple terminals to achieve a closed-loop process of monitoring, assessment, alerts, intervention, and reassessment: pushing concise and visual reports to user terminals, sending major early warning notifications (such as level 2 or higher risks) to family terminals, along with emergency contact guidelines and care suggestions to enhance family support capabilities; and uploading complete analysis logs, abnormal event records, and model confidence scores to the doctor's workstation to assist clinical decision-making and support remote follow-up and intervention plan optimization. The system supports feedback from users after implementing intervention measures (such as completing exercise check-ins and symptom improvement), triggering a new round of data collection and reassessment, forming a continuously iterative cycle of personalized health management.

[0032] The beneficial effects of the above technical solution are as follows: The solution provided in this embodiment can achieve personalized, dynamic tracking and precise risk alerts for users' bone health status through multimodal data fusion combined with traditional Chinese and Western medicine theories; it combines continuous physiological signals (PPG, triaxial acceleration, gyroscope) collected by smart wearable devices with structured clinical and lifestyle data input from the APP, breaking through the limitations of traditional bone health monitoring relying on single bone density detection, and providing rich data dimensions for comprehensively depicting bone health status; it constructs an assessment model that integrates Western medicine microphysiological indicators and traditional Chinese medicine macrophysiological characteristics, using a dual-stream encoder to process Western medicine physiological feature sequences and traditional Chinese medicine Zang-Xiang feature vectors respectively, and utilizes a cross-modal attention mechanism to achieve deep interaction and focus on key information between the two, outputting a bone health status index and traditional Chinese medicine... The distribution of syndrome patterns provides a theoretical basis for the integration of traditional Chinese and Western medicine in subsequent interventions. Personalized dynamic fluctuation ranges are established based on continuous initial data collection from users, and continuously evolved through sliding windows and Bayesian update mechanisms. This accurately captures the natural fluctuation patterns of individual health indicators, effectively avoiding misjudgments caused by using group standards. The time-series prediction and risk warning model can predict future bone health trends based on historical data and combine it with individual dynamic baselines for risk assessment and graded early warning, providing users, families, and doctors with ample intervention time windows, helping to control bone health risks in their early stages. The analysis results and intervention plans are distributed to user, family, and doctor terminals through an interactive application layer, and post-intervention feedback data is supported, forming a complete closed-loop management process of monitoring, assessment, alerts, intervention, and reassessment.

[0033] In one embodiment, the data perception layer integrates a data processing module for preprocessing and initially calculating the user's posture stability index, gait symmetry index, and gait regularity index based on triaxial acceleration signals and gyroscope signals. These indices are incorporated as components of the collected data. The posture stability index represents the quantified value of the degree of center of gravity shift and recovery ability during standing or walking, calculated based on low-impact body motion characteristic data, including trunk micro-swing frequency, limb movement amplitude change rate, and movement initiation / termination smoothness. The gait symmetry index includes the left-right step length difference rate, double support phase time consistency, and swing symmetry index. The gait symmetry index is evaluated based on gait symmetry characteristic data using gait cycle segmentation and symmetry analysis algorithms. The gait regularity data includes step frequency fluctuation coefficient, step length variability, and rhythm continuity score; based on this data, the gait regularity index is determined.

[0034] The working principle of the above technical solution is as follows: the data perception layer integrates a data processing module, which receives raw motion signals from the triaxial accelerometer and gyroscope, and extracts time domain and frequency domain features through sensor fusion technology to form a basic body motion data stream; The posture stability index is calculated based on low-impact body motion characteristic data. The frequency of trunk micro-swaying is obtained by Fourier transform analysis of energy concentration in the range of 0.5-3Hz. The rate of change of limb movement amplitude is obtained by normalizing the first difference of the acceleration signal and then calculating the root mean square. The smoothness of the start / end of the movement is quantified by detecting the continuity of the slope of the rising and falling edges of the acceleration curve and calculating its Jerk integral value. Finally, the three are weighted and fused to generate a posture stability score. The lower the value, the stronger the center of gravity control ability. Gait symmetry index relies on precise segmentation of the gait cycle. First, gait event points (such as the moment of heel contact) are detected using the peak angular velocity of a gyroscope to divide the single-step cycle. Then, the left and right step length difference rate (defined as |left step length − right step length| / (left step length + right step length) × 100%), the consistency of the double support phase time (calculated as the average of the absolute values ​​of the time difference between the two feet in the left and right gait cycles), and the swing symmetry index (based on the degree of deviation after taking the logarithmic transformation of the ratio of swing phase duration) are extracted. After dimensionality reduction by principal component analysis, the above parameters are linearly combined to form a comprehensive symmetry score. A score close to zero indicates high symmetry. Gait regularity index is determined by the stability of gait rhythm. The gait frequency fluctuation coefficient is calculated by the sliding window method as the ratio of the standard deviation to the mean of the gait frequency within each 30-second time interval. The gait length variability is calculated by the coefficient of variation (CV) of the gait length sequence of ten consecutive walking cycles. The rhythm continuity score uses autocorrelation analysis to identify the consistency of the position of the maximum lag peak in the gait cycle sequence, and combines it with a hidden Markov model to determine the frequency of gait pattern switching. The three indices are standardized and weighted to generate the final regularity index. The higher the value, the more stable and predictable the walking rhythm.

[0035] The beneficial effects of the above technical solution are as follows: by adopting the solution provided in this embodiment, non-invasive monitoring of bone health risk-related motor abilities can be achieved through the refined quantification of multi-dimensional motion characteristics.

[0036] In one embodiment, the Western medical pathological feature sequence includes vascular elasticity parameters and microcirculation state parameters; the vascular elasticity parameters and microcirculation state parameters are: indirect biochemical correlation features related to bone metabolism estimated by inputting the PPG signal into a pre-trained derivative model; the derivative model refers to a multi-scale feature extraction network built based on a deep learning architecture, which uses the time-domain waveform features, frequency-domain power spectrum features, and nonlinear dynamic features (such as approximate entropy and sample entropy) of the PPG signal as input layer neurons, and adaptively captures subtle fluctuation patterns related to vascular elasticity and microcirculation perfusion in the signal through a hierarchical structure including convolutional layers, recurrent neural network layers, and attention mechanism layers; wherein, convolutional... The layer uses 1D convolutional kernels to extract local features of the PPG signal, such as pulse wave velocity and ascending limb slope. The recurrent neural network layer (LSTM or GRU) is used to model the dynamic changes of the signal during the cardiac cycle. The attention mechanism layer focuses on the characteristic frequency bands related to the secretion cycle of bone metabolism regulating hormones (such as parathyroid hormone and calcitonin). Finally, the fully connected layer outputs parameters such as arterial stiffness index (ASI) and pulse wave velocity (PWV) reflecting vascular elasticity, and microvascular perfusion volume (MFP) and capillary blood flow velocity (CBV) characterizing microcirculation status, realizing end-to-end estimation of indirect correlation features from non-invasive PPG signal to bone metabolism.

[0037] The working principle of the above technical solution is as follows: PPG signal, as an easy-to-collect non-invasive physiological signal, contains information on vascular elasticity and microcirculation status, which has a deep indirect relationship with bone metabolism. Bone metabolism balance depends on the dynamic regulation of osteoblasts and osteoclasts, a process influenced by various hormones (such as parathyroid hormone and calcitonin) and the local microenvironment (such as blood perfusion and nutrient supply). Decreased vascular elasticity (e.g., elevated arterial stiffness index (ASI) and increased pulse wave velocity (PWV)) is usually accompanied by changes in vascular wall structure and function, potentially leading to insufficient bone blood perfusion and affecting osteoblast activity and bone matrix synthesis. Simultaneously, decreased vascular elasticity may indirectly affect calcium and phosphorus metabolism and related hormone secretion through pathways such as the renin-angiotensin system, thereby interfering with bone metabolism balance. Microcirculatory parameters (such as microvascular perfusion volume (MFP) and capillary blood flow velocity (CBV)) directly reflect the blood perfusion status of the local bone and systemic microvascular beds. Sufficient microcirculatory perfusion is fundamental to ensuring bone tissue receives adequate oxygen, nutrients (such as calcium, phosphorus, and vitamin D), and removes metabolic waste. When microcirculatory perfusion is insufficient... When bone mass is insufficient, osteoblast proliferation and differentiation are inhibited, while osteoclast activity may be relatively enhanced, leading to bone loss and decreased bone quality. Derivative models, through deep learning architectures, particularly the synergistic effect of convolutional layers, recurrent neural network layers, and attention mechanisms, can accurately capture subtle patterns related to changes in vascular elasticity and microcirculation from the complex fluctuations of PPG signals. Convolutional layers first perform detailed local feature extraction on the time-domain waveform of PPG signals. For example, pulse wave propagation velocity reflects the speed of pulse wave propagation in the arterial system and is directly related to arterial elasticity; the slope of the ascending branch is related to myocardial contractility and vascular resistance, indirectly affecting hemodynamics. Recurrent neural network layers (LSTM or GRU) utilize their ability to model sequential data to deeply explore the dynamic evolution of PPG signals within one or even multiple cardiac cycles, capturing the dynamic characteristics of vascular elasticity and microcirculation changes with cardiac pumping activity.

[0038] The beneficial effects of the above technical solution are as follows: By adopting the solution provided in this embodiment, through such multi-scale and multi-level feature extraction and fusion, the parameters such as ASI, PWV, MFP, and CBV finally output by the derived model become important biological markers that can indirectly reflect the bone metabolism status, providing valuable input for subsequent bone health status assessment and risk warning.

[0039] In one embodiment, when establishing the dynamic fluctuation range of personalized health indicators, a robust statistical method based on the median and absolute median difference is used for calculation. Specifically, this includes: performing time window sliding processing on the multidimensional bone health-related physiological parameter sequence data of each user, extracting the historical observation set of the corresponding indicators at each time point, calculating the median as the central trend estimate for the historical observation set, and then calculating the median of the absolute values ​​of the differences between each observation and the median, i.e., the absolute median difference, and combining it with a preset scaling factor to determine the upper and lower boundaries of the dynamic fluctuation range.

[0040] The working principle of the above technical solution is as follows: By using time window sliding processing, the user's physiological parameter sequence data can be divided into multiple continuous subsequences according to time order. Each subsequence represents the changes in physiological parameters within a specific time period, thus capturing the fluctuation characteristics of parameters in different periods more precisely. For the set of historical observations within each time window, the median, as an estimate of the central trend, is less susceptible to interference from extreme outliers than the mean, and can robustly reflect the typical level of the data within that window. For example, when a user's ASI parameter shows individual extremely high or low outliers within a certain period, the median can still represent the overall central level of ASI for that period well. The absolute median difference further quantifies the dispersion of data around the median, also exhibiting strong anti-interference capabilities. The preset scaling factor is based on the critical period... The dynamic fluctuation range is determined by the bed requirements, the characteristics of physiological parameters, and the sensitivity requirements for anomaly detection. The lower boundary of the range is the median minus the "scale factor × absolute median difference," and the upper boundary is the median plus the "scale factor × absolute median difference." Together, these constitute the dynamic fluctuation range of the corresponding indicator at that time point. This dynamic fluctuation range is essentially a personalized normal range constructed based on the user's own historical data. It is continuously updated as the time window slides to adapt to the long-term, slow changes in the user's physiological state. For example, with age or changes in lifestyle, the typical fluctuation range of certain bone health-related parameters may drift. Robust statistical methods can capture such changes in a timely manner and adjust the fluctuation range, making the system's assessment of the user's bone health status more individualized and dynamically adaptable.

[0041] The beneficial effects of the above technical solution are as follows: the solution provided in this embodiment can significantly improve the personalization and accuracy of the dynamic fluctuation range of bone health indicators; by combining time window sliding processing with robust statistical methods, the dynamic fluctuation range can closely follow the user's own physiological parameter change pattern, rather than relying on a broad population reference range; the application of median and absolute median difference effectively reduces the impact of extreme outliers (such as accidental measurement errors, short-term physiological stress responses, etc.) on the determination of the interval boundary, ensuring the stability and reliability of the fluctuation range.

[0042] In one embodiment, the dynamic fluctuation range is continuously evolved based on a sliding window and Bayesian update mechanism, including: introducing a weighted sliding window update strategy to assign differentiated weights to historical data at different time periods; setting up a trend detection unit to jointly determine whether the sequence has a significant monotonic trend based on the Theil-Sen estimator and the Mann-Kendall test; when the trend is confirmed, triggering the fluctuation range drift update rule according to the Bayesian update mechanism to adjust the center position and the slope of the extended trend response; and setting a multi-level threshold mechanism to distinguish between trend drift and acute abnormal events in order to maintain the detection sensitivity of pathological indicator changes that deviate from the normal fluctuation pattern in the short term.

[0043] The working principle of the above technical solution is as follows: First, the weighted sliding window update strategy assigns weights to historical data through a time decay function, with recent data being given higher weights (the weight coefficient decays exponentially with time intervals, and the decay factor is optimized to 0.92 through cross-validation), making the data distribution within the window more closely match the user's recent physiological state characteristics. The trend detection unit first uses the Theil-Sen estimator to calculate the robust slope of the bone health index sequence (based on the median slope of all data point pairs within the window), and then verifies the significance of the slope using the Mann-Kendall test (significance level set to 0.05). When the test result rejects the no-trend hypothesis, the system determines that there is a significant monotonic trend. At this time, the Bayesian update mechanism is triggered, using the slope parameter obtained from trend detection as prior information, combined with the likelihood of the latest 10% of data within the window. The function calculates the posterior distribution and dynamically adjusts the center position of the fluctuation range (center offset = posterior expected slope × time step) and the extended trend response slope (slope response coefficient is inversely proportional to the posterior distribution variance). A multi-level threshold mechanism sets a basic fluctuation threshold (±1.5 × dynamic standard deviation), a trend drift threshold (±2.5 × dynamic standard deviation + trend slope × warning delay time), and an acute anomaly threshold (±3.5 × dynamic standard deviation). When the indicator value exceeds the basic threshold but does not reach the trend drift threshold, it is marked as a potential trend drift. When the trend drift threshold is exceeded, the interval drift update process is initiated. When the acute anomaly threshold is exceeded and the rate of change exceeds 3 times the historical maximum daily volatility, it is judged as an acute anomaly event, which independently triggers the risk warning channel, ensuring that while capturing long-term trend changes, no sudden pathological anomalies in the short term are missed. This dynamic baseline update mechanism is specifically designed to address the unique needs of bone health assessment. Changes in bone health indicators tend to accumulate slowly and then accelerate suddenly. Conventional dynamic baseline methods (such as simple moving averages) struggle to effectively distinguish between physiological drift, pathological trends, and measurement noise. This application employs a weighted sliding window to assign higher weights to recent data, ensuring continued sensitivity in tracking slow trends in bone health indicators. It uses a Theil-Sen estimator combined with the Mann-Kendall test to determine trend significance, avoiding misinterpreting short-term fluctuations as pathological trends. Furthermore, a multi-level threshold mechanism separates trend drift from acute abnormal events, ensuring an independent early warning channel for high-risk events such as rapid declines in bone density.

[0044] The beneficial effects of the above technical solution are as follows: By adopting the solution provided in this embodiment, the baseline of the user's bone health status can be accurately tracked through a dynamically weighted historical data window. This avoids the interference of historical data caused by a fixed window, and highlights the representativeness of the recent physiological status through exponential decay weights, making the update of the fluctuation range more timely and personalized.

[0045] In one embodiment, the extraction of TCM Zang-Xiang feature vectors includes: identifying and extracting quantitative features related to TCM kidney function, liver function, and spleen function by performing time-frequency domain joint analysis and waveform morphology modeling on the pulse components in the acquired PPG signal; wherein the quantitative features related to kidney function include at least the pulse depth feature value and the pulse weakness index; the quantitative features related to liver function include at least the pulse stringiness feature value; and the quantitative features related to spleen function include at least the pulse softness and slowness feature value.

[0046] The working principle of the above technical solution is as follows: First, the PPG signal is preprocessed, and an adaptive filtering algorithm is used to remove motion artifacts and baseline drift interference, retaining the pure pulse wave components. For the joint time-frequency domain analysis of the pulse components, the time-domain pulse signal is decomposed into sub-signals of different frequency bands through continuous wavelet transform, and the low-frequency wave characteristics corresponding to the kidney meridian in the α band (0.05-0.15Hz), the mid-frequency characteristics corresponding to the liver meridian in the β band (0.15-0.4Hz), and the high-frequency characteristics corresponding to the spleen meridian in the γ band (0.4-1.0Hz) are extracted. In the waveform morphology modeling stage, a pulse feature space was constructed, including 12 morphological parameters such as the slope of the ascending limb, peak delay, and depth of the descending isthmus. An improved Hidden Markov Model was used to fit the pulse waveform piecewise. The pulse depth feature value was calculated by comprehensively analyzing characteristic parameters reflecting vascular elasticity and peripheral resistance in the PPG signal (such as pulse wave propagation time, ascending limb slope, and descending isthmus depth). The pulse weakness index was defined as the ratio of the energy decay rate of the pulse signal during diastole to the maximum amplitude during systole. For the pulse stringency feature value, a waveform sharpness analysis method based on the second derivative was used to quantify the rate of curvature change of the pulse curve in the ascending limb of the main wave. The pulse softness feature value was calculated by combining the pulse cycle variability coefficient and the trough flatness parameter, where trough flatness is defined as the percentage of the duration of the diastolic waveform plateau segment to the entire cardiac cycle. By normalizing the extracted quantitative features, a 3×N TCM Zang-Xiang feature vector matrix is ​​constructed (N is the feature sub-dimension corresponding to the function of each Zang-Fu organ), providing multimodal feature input for subsequent TCM syndrome differentiation analysis of bone health status. The physiological and traditional Chinese medicine basis of the above feature extraction method is as follows: Traditional Chinese medicine theory believes that the kidney governs bones and produces marrow, and the state of kidney function directly affects bone strength and bone metabolism balance. The cun pulse (cubit pulse) is the manifestation of the kidney, and a weakened pulse with a thin and weak quality is a typical manifestation of insufficient kidney qi. The liver governs tendons and stores blood. Liver dysfunction can lead to malnourishment of tendons and muscles, affecting muscle-bone biomechanical balance. An increased wiry pulse is related to liver qi stagnation and tendon stiffness. The spleen governs muscles and transports and transforms the essence of food and water. Spleen dysfunction leads to muscle weakness and malnourishment of bones. A soft and slow pulse is a typical indicator of spleen deficiency and dampness. Modern physiological research has confirmed that the above pulse characteristics are intrinsically related to physiological parameters such as the state of autonomic nervous system function, peripheral vascular resistance, and microcirculation perfusion. These parameters, in turn, affect bone metabolism through the neural, endocrine, and immune networks.

[0047] The beneficial effects of the above technical solution are as follows: by adopting the solution provided in this embodiment, the abstract organ function state in traditional Chinese medicine theory can be transformed into a quantifiable feature vector through the objective analysis of PPG signals, thereby realizing the modernization and standardization of traditional Chinese medicine pulse diagnosis.

[0048] In one embodiment, such as Figure 2 As shown, the tiered early warning system includes: When the bone health status index deviates from the dynamic fluctuation range for more than a certain number of days within a consecutive predetermined number of days, or when it is predicted that the bone health status index will enter the preset health attention data range in the future, the first-level warning for pushing lifestyle adjustment reminders will be triggered. When the bone health status index is within the preset sub-health data range, or when the TCM syndrome distribution continues to point to a specific TCM syndrome, a second-level warning is triggered to recommend TCM conditioning and medical examination. When the user's short-term fall risk predicted based on posture stability, gait symmetry, and gait regularity indicators exceeds the high-risk value, or when the bone health status index is about to enter the high-risk data range, a Level 3 warning is triggered to simultaneously notify emergency contacts or family doctors.

[0049] The working principle of the above technical solution is as follows: The system continuously monitors the user's bone health status index and constructs a dynamic fluctuation range based on historical data. This range reflects the pattern of bone health changes under normal physiological conditions. When the monitoring finds that the bone health status index deviates from this dynamic fluctuation range for more than a preset threshold number of days within a predetermined number of consecutive days, it indicates that the user's current lifestyle, nutritional intake, or activity patterns may have a continuous negative impact on the skeletal system. Although it has not yet reached the clinical abnormality standard, there is a potential degenerative trend. At the same time, the system combines time series prediction models (such as LSTM or ARIMA) to extrapolate the bone health status in the future. If the prediction results show that the index will enter the preset health concern data range (i.e., the transition area that is close to but has not reached the sub-health threshold), it is judged as an early risk exposure. At this time, the first-level warning is triggered, and personalized lifestyle adjustment reminders are pushed. The content includes dietary optimization suggestions (such as increasing calcium / vitamin D intake), moderate weight-bearing exercise recommendations, and work and rest adjustment plans. The aim is to block the deterioration process through non-medical intervention and achieve proactive health management. When the real-time monitored bone health status index falls into the preset sub-health data range (corresponding to quantifiable indicators such as decreased bone density and abnormal bone metabolism markers), the system determines that bone function has reached an identifiable pre-pathological state. At the same time, the system integrates a TCM syndrome identification module, which calculates the tendency probability distribution of each TCM syndrome (such as liver and kidney deficiency, spleen and kidney yang deficiency, and qi stagnation and blood stasis) based on the user's multidimensional data (such as tongue appearance, pulse appearance, symptom self-assessment scale, and constitution identification questionnaire). If the tendency probability of a specific syndrome (such as liver and kidney deficiency) is consistently higher than the set threshold and shows a stable trend in multiple consecutive assessment cycles, it is considered that there is a clear TCM pathological basis. When any of the above conditions are met, a second-level warning is triggered. Its core goal is upgraded from lifestyle intervention to preparation for medical intervention. The system generates a composite suggestion that includes modern medical examination recommendations (such as DXA bone density test and serum CTX / PTH test) and TCM conditioning plan (such as Chinese herbal prescription recommendations and acupuncture point guidance), and prompts the user to make an appointment with a specialist doctor to achieve early diagnosis and collaborative intervention combining TCM and Western medicine. A fall risk assessment model based on posture stability, gait symmetry, and gait regularity indices is used to predict users' short-term fall risk. When the fall risk score output by the model exceeds the high-risk value (a critical point determined based on large-scale cohort studies, corresponding to a fall probability >30% within the next 72 hours), or when the bone health index is determined to cross the critical threshold and enter the high-risk data range within a very short period of time (e.g., 48 hours) (e.g., T-value < -2.5 with a rapid downward trend), the highest level, Level 3 warning, is triggered. This warning mechanism is designed to focus on the prevention and control of acute events. The system immediately initiates the emergency response process: automatically sending alarm information containing user location, risk type, and suggested treatment measures to the preset emergency contact, while simultaneously notifying the family doctor platform, pushing a structured risk report, and suggesting the initiation of remote consultation or home assessment. This tiered early warning system, with continuous monitoring, dynamic modeling, multi-dimensional analysis, and layered response as its core components, constructs a closed-loop intelligent bone health management framework. First, it collects high-frequency physiological and behavioral data through wearable devices and mobile applications to form individualized baselines. Second, it uses statistical methods (such as moving percentiles) and machine learning algorithms to establish dynamic fluctuation ranges and predictive models, distinguishing between physiological fluctuations and pathological deviations. Third, it integrates modern medical quantitative indicators with traditional Chinese medicine syndrome identification results to achieve cross-validation from both biomedical and traditional medical perspectives. Finally, it categorizes response strategies into three levels based on risk level: Level 1 focuses on preventative guidance, Level 2 promotes diagnostic integration, and Level 3 ensures emergency response coordination, forming a comprehensive technical support system from daily maintenance to crisis intervention.

[0050] The beneficial effects of the above technical solution are as follows: By adopting the solution provided in this embodiment, a hierarchical and progressive early warning mechanism can be constructed to achieve accurate identification and differentiated response to bone health risks, effectively solving the pain points of traditional bone health management, such as emphasizing detection over monitoring, treatment over prevention, and the separation of information between traditional Chinese and Western medicine.

[0051] In one embodiment, the intelligent decision-making layer is also used for: Based on the collected data, static portrait data is extracted and generated. The static portrait data includes at least age, gender, baseline bone density classification, and major TCM constitution types. Based on static profile data, a set of personalized modulation parameters is generated. The modulation parameters include: baseline offset vector, feature channel weight vector, and risk judgment threshold adjustment coefficient. The baseline offset vector is used to perform initial calibration of the center value of the dynamic fluctuation range based on population subgroups; the feature channel weight vector is used to weight the importance of each dimension of features in the Western medicine physiological feature sequence and the traditional Chinese medicine Zang-Xiang feature vector; the risk judgment threshold adjustment coefficient is used to personalize the scaling of the number of days threshold, high-level risk value, and high-risk data interval boundary in the graded early warning system. When initializing the personal bone health dynamic baseline model, a baseline offset vector is loaded; before the dual-stream encoder in the integrated Chinese and Western medicine bone status assessment model processes the features of each dimension in the Western medicine physiological feature sequence and the Chinese medicine Zang-Xiang feature vector, the features of each dimension in the Western medicine physiological feature sequence and the Chinese medicine Zang-Xiang feature vector are modulated using the feature channel weight vector; when making risk decisions in the time series prediction and risk warning model, a risk judgment threshold adjustment coefficient is applied.

[0052] The working principle of the above technical solution is as follows: The intelligent decision-making layer is configured to perform the following operations: First, based on multi-source physiological and constitution-related data collected from the user terminal, feature extraction and classification are performed to generate static portrait data representing the inherent attributes of an individual; the static portrait data includes at least age information, gender information, basic bone density classification results, and the determination results of the main TCM constitution types; the basic bone density classification is determined based on the T-value range obtained by dual-energy X-ray absorptiometry or equivalent assessment methods, and is divided into one of the three categories: normal, osteopenia, and osteoporosis; the main TCM constitution types are determined according to the "Classification and Determination of TCM Constitutions" standard, through standardized questionnaires and identification algorithms, and are comprehensively determined as one or more biased types among balanced constitution, qi deficiency constitution, yang deficiency constitution, yin deficiency constitution, phlegm-dampness constitution, damp-heat constitution, blood stasis constitution, qi stagnation constitution, or special constitution; A set of personalized modulation parameters is generated based on static profile data. The personalized modulation parameters include baseline offset vector, feature channel weight vector, and risk judgment threshold adjustment coefficient. The baseline offset vector is a multi-dimensional offset constructed based on the statistical regularity of subgroups of the population. Its dimension matches the state space of the personal bone health dynamic baseline model. It is used to calibrate the center value of the dynamic fluctuation range during model initialization, so that the initial baseline is closer to the group mean of a specific subgroup (such as the same sex, age group and bone density level), thereby improving the prediction accuracy and stability of the model in the cold start stage. The feature channel weight vector is a set of learnable or rule-based mapping weights aligned with the dimensions of the Western medicine physiological feature sequence and the traditional Chinese medicine Zang-Xiang feature vector. Each element corresponds to a measure of the relative importance of different physiological indicators or TCM syndrome elements. This weight vector is applied to the input features at the processing front end of the dual-stream encoder of the integrated TCM and Western medicine bone status assessment model, realizing adaptive weighted modulation of different feature channels. In the Western medicine physiological feature branch, key parameters such as calcium metabolism indicators, hormone levels, and inflammatory factors are assigned differentiated weights. In the TCM Zang-Xiang feature branch, the weights of liver and kidney function status, Qi and blood circulation status, and tongue and pulse quantitative features are enhanced or suppressed to reflect the unbalanced contribution of individual constitution differences to the path of bone health.

[0053] The risk assessment threshold adjustment coefficient is a scalar or multidimensional scaling factor, generated by jointly mapping age, gender, baseline bone mineral density classification, and TCM constitution type from the static profile. It is used to personalize the graded early warning mechanism in the time-series prediction and risk warning model. In specific applications, the adjustment coefficient acts on multiple decision boundary parameters, including the threshold for the number of consecutive abnormal days required for accelerated bone loss warning, the critical value of the rate of bone mineral density decline corresponding to high-level clinical risk, and the upper and lower boundary values ​​of the high-risk data interval. By introducing this coefficient, the warning threshold can be dynamically tightened for sensitive groups (such as elderly women and those with phlegm-dampness constitution), while the criteria can be appropriately relaxed for low-risk individuals, thereby improving the system's personalized service level while ensuring safety. During the system initialization phase, the baseline offset vector is loaded into the initial state variables of the individual bone health dynamic baseline model, completing the initial setting of the individualized baseline. During the integrated traditional Chinese and Western medicine bone status assessment, the feature channel weight vector performs channel-by-channel multiplication on the original feature vector before the dual-stream encoder receives the input, completing feature modulation. During the risk assessment phase, the risk judgment threshold adjustment coefficient is input into the risk classification module of the time series prediction and risk warning model, which corrects the numerical benchmark of each level of risk judgment in real time, and finally outputs the risk level and intervention suggestions adapted to individual characteristics.

[0054] The beneficial effects of the above technical solution are as follows: By constructing baseline calibration, feature weighting and threshold scaling based on static portrait data, personalized adaptation of the entire process from model initialization, feature expression to risk decision-making is realized, which enhances the universality and accuracy of the system in heterogeneous populations. It is especially suitable for refined modeling of the evolution trend of bone health status and early risk intervention in a long-term monitoring environment.

[0055] In one embodiment, the intelligent decision-making layer includes a decision routing module; the decision routing module stores a bone health intervention knowledge graph, which is used to generate personalized intervention plans based on bone health analysis results, including nutritional advice, exercise programs, traditional Chinese medicine dietary therapy recipes, acupoint massage guidance, and medical indications.

[0056] The working principle of the above technical solution is as follows: The decision routing module integrated in the intelligent decision layer, as the core component for generating personalized intervention plans, pre-stores a bone health intervention knowledge graph. This knowledge graph is a structured medical knowledge network, including multi-dimensional data nodes covering the physiological and pathological characteristics of the skeletal system, the mechanisms of nutrient action, the relationship between exercise load and bone density, traditional Chinese medicine dietary therapy theories, the functions of meridians and acupoints, and their impact on bone metabolism. Logical mapping relationships between nodes are established through semantic association edges. When receiving individualized analysis results from the bone health analysis module, the decision routing module performs reasoning and matching operations based on the bone health intervention knowledge graph, using a graph traversal algorithm to identify... The system identifies knowledge paths that align with the user's current health status and extracts corresponding intervention strategies from the data graph. These intervention strategies include at least: nutritional recommendations based on calcium, vitamin D, and protein intake requirements; progressive exercise plans tailored to the user's age, gender, bone density, and underlying medical conditions; dietary therapy formulas recommended based on traditional Chinese medicine syndrome differentiation; specific acupoint massage guidance designed according to the theory that the kidneys govern bone and marrow production; and tiered medical treatment indications based on the severity of the condition. Finally, the decision routing module integrates these multimodal interventions into a unified, personalized intervention plan and outputs it to the user's terminal or medical service platform through a standardized interface, achieving closed-loop decision support from data analysis to health intervention.

[0057] The beneficial effects of the above technical solution are as follows: by constructing a structured bone health intervention knowledge graph and combining it with intelligent reasoning algorithms, the generation process of personalized intervention plans has a high degree of scientificity and accuracy.

[0058] In one embodiment, such as Figure 3 As shown, the intelligent decision-making layer also includes a model co-optimization engine; the model co-optimization engine is configured as follows: Monitor and record user compliance feedback data, subsequent physiological indicator changes data, and external medical diagnosis results data after each tiered warning is triggered; Based on compliance feedback data, physiological indicator change data, and external medical diagnosis results data, an intervention effect verification dataset was constructed. Using the intervention effect validation dataset, reinforcement learning was performed on the cross-modal attention fusion module to optimize the learnable weight matrix; Based on the intervention effect verification dataset, the verified effective intervention patterns and results are extracted, and the association rules in the bone health intervention knowledge graph are updated with confidence weight or nodes are added.

[0059] The working principle of the above technical solution is as follows: The intelligent decision-making layer further integrates a model co-optimization engine, which is used to realize a closed-loop feedback-driven dynamic model optimization mechanism. Specifically, the model co-optimization engine is configured to continuously monitor and record multi-source feedback information generated after each graded warning is triggered, including user compliance feedback data, subsequent physiological indicator changes collected after intervention, and authoritative diagnostic results provided by external medical systems. Based on the collected multi-dimensional feedback information, a structured intervention effect verification dataset is constructed, where each data sample corresponds to a warning, intervention, and feedback event chain and is labeled with the actual intervention effect label. The intervention effect verification dataset is used as the reward signal source in the reinforcement learning process to perform online policy optimization on the cross-modal attention fusion module. Training involves iteratively updating the learnable weight matrix within the module by maximizing long-term positive intervention rewards, thereby enhancing its adaptive weighting ability for key features in complex and heterogeneous data environments. Simultaneously, based on statistically significant effective intervention patterns and their corresponding clinical outcomes in the intervention effect validation dataset, high-confidence causal association rules are extracted. These rules are then applied to the existing node relationships in the bone health intervention knowledge graph with confidence weighting, or new knowledge nodes and semantic links are added based on newly discovered intervention and response relationships, achieving dynamic evolution of the knowledge graph and accumulation of domain knowledge. This forms a complete technical closed loop of early warning output, feedback collection, effect evaluation, model optimization, and knowledge updating, enabling the system to continuously evolve and improve the accuracy and clinical applicability of personalized early warnings during long-term operation.

[0060] The beneficial effects of the above technical solution are as follows: by adopting the solution provided in this embodiment, the system can achieve self-iteration and continuous evolution in the process of practical application by constructing a closed-loop feedback-driven dynamic model optimization mechanism.

[0061] In a specific implementation example, a user is a 55-year-old woman in the perimenopausal stage. First, the system generates an individualized static profile based on the basic information entered by the user during the first registration and the initial multi-source data collection results: age 55 (perimenopause), gender female, basic bone mineral density classification as osteopenia (portable ultrasound bone mineral density screening T value -1.5 to -2.0), and the main TCM constitution type is identified as kidney deficiency with qi deficiency. Based on this, the system automatically generates a set of personalized modulation parameters: (1) Baseline offset vector: Based on the statistical mean of the subgroup of the same age group, gender, bone density level and TCM syndrome type, the initial center value of the user's personal bone health dynamic baseline model is calibrated so that the initial baseline of the bone health index (BHI) is set at 72 points instead of the general population average of 68 points, thereby improving the accuracy of trend modeling in the cold start stage; (2) Feature channel weight vector: Weight modulation is applied to the front end of the dual-stream encoder of the TCM-Western medicine integration assessment model. In the TCM pathology branch, key calcium metabolism indicators such as serum 25-(OH)D, parathyroid hormone (PTH), and type I collagen cross-linked C-terminal peptide (CTX) are given higher weights (1.35, 1.20, and 1.15, respectively). In the TCM Zangxiang branch, the theoretical path of the kidney governing bone and producing marrow is significantly enhanced by increasing the weight of characteristic dimensions such as liver and kidney function status score, cun pulse strength, and tongue coating thinness (set to 1.40, 1.30, and 1.10 respectively), while suppressing the influence of damp-heat related characteristics (such as yellow greasy coating weight reduced to 0.70); (3) Risk judgment threshold adjustment coefficient: the scaling factor γ=0.85 is generated by static image joint mapping, which is used to tighten multiple criteria in the graded early warning mechanism, shorten the number of consecutive abnormal days required for L1 level trend deviation from the usual 7 days to 6 days, adjust the critical value of bone loss rate recommended for L2 level from a monthly decrease of 0.5% to 0.4%, and simultaneously reduce the upper and lower boundaries of the high-risk interval by 10% to improve the sensitivity of this high-risk population; Next, in the first month without intervention, the system continuously collected data from the user's wearable devices (heart rate variability (HRV), body rhythm, sleep structure), dietary logs (calcium intake, vitamin D source), TCM diagnostic information (pulse waveform parameters), and ambient light intensity. Through a personal bone health dynamic baseline model, the physiological fluctuation range of each core indicator was constructed using the percentile method, and an individualized trend evolution model was established by combining an LSTM time series prediction network. The results showed that the BHI was stable within the range of 72±2, the nighttime HRV low-frequency power (LF) was within the normal fluctuation range (45–60 ms²), and the pulse depth feature value was maintained between 3.8 and 4.2 in standardized units, confirming that the initial baseline was effectively established. Over the next six months, the system's backend model continuously learned and updated trend parameters. Although all individual indicators remained within the traditional medical reference range, the cross-modal attention fusion module detected multi-dimensional synergistic change signals: BHI showed a significant downward trend (slope approximately -0.3 points / month, p<0.05), nighttime HRVLF power decreased by about 10% compared to baseline, and the pulse depth characteristic value also showed a cumulative decrease of more than 10%. This change pattern was determined to be consistent with the evolutionary path of TCM pathogenesis of gradual decline of kidney qi and loss of bone nourishment, and it was highly consistent with the modern medical mechanism of accelerated bone turnover caused by estrogen decline in perimenopausal women. By the end of the 7th month, the system's first-level early warning engine, based on residual analysis of the dynamic fluctuation range, found that the current BHI value had been below the P10 lower limit for 9 consecutive days, and the downward trend was confirmed to be statistically significant by the Mann-Kendall trend test (Z=2.67, p=0.0076). Thus, a first-level trend attention early warning was triggered: The system has detected that your basic bone health status has recently shown a slow downward trend, which may be related to age changes and weakened endocrine regulation. It is recommended to ensure adequate daily intake of calcium (≥800mg) and vitamin D (≥600IU), and appropriately increase outdoor sunlight exposure time (no less than 20 minutes per day). The weighted sliding window decay factor of 0.92, the trend test significance level of 0.05, and the personalized scaling factor of 0.85 were all determined through cross-validation of multi-center cohort data, consistent with the clinical bone health risk grading standard, and all thresholds and parameters were statistically significant and reproducible.

[0062] After receiving the bone health analysis report corresponding to this first-level warning, the decision routing module immediately starts the bone health intervention knowledge graph reasoning process. Based on the user's static profile and current dynamic status, the system matches the following multimodal intervention strategy set: (1) Nutritional advice: It is recommended to consume a combination of calcium-rich foods daily, including tofu (100g), low-fat milk (300ml), dried shrimp (5g) and dark green leafy vegetables (such as kale 150g), and supplement with vitamin D3 preparations (400IU / day); (2) Exercise program: Configure a progressive weight training plan, initially focusing on brisk walking and Baduanjin exercises 3 times a week for 20 minutes each time, with emphasis on strengthening the two hands holding the sky to regulate the three jiaos, shaking the head and wagging the tail to remove heart fire, and the two hands climbing the feet to strengthen the body. (2) Kidney and waist exercises to regulate the kidney meridian, gradually transitioning to light resistance elastic band training; (3) Traditional Chinese medicine diet: Based on the results of the diagnosis of kidney qi deficiency, the classic medicinal diet Eucommia ulmoides and pork kidney soup (10g Eucommia ulmoides and 1 pork kidney) is recommended to be eaten twice a week, and combined with wolfberry and chrysanthemum tea as a daily tea substitute; (4) Acupoint massage guidance: Push the picture and video tutorials to guide them to press the Shenshu acupoint (1.5 cun lateral to the spinous process of the second lumbar vertebra on both sides) and Taixi acupoint (the depression behind the medial malleolus) daily, massage each acupoint for 3 minutes, once in the morning and once in the evening; (5) Medical indications: Clearly inform them that if the BHI continues to drop by more than 5 points or symptoms of back pain occur within the next three months, they should go to a medical institution as soon as possible for DXA bone density testing and endocrinology specialist evaluation; Next, after reviewing the detailed report on the app, the user demonstrated good adherence, proactively adjusting their diet and maintaining their exercise routine. The system recorded their adherence feedback data: a diet check-in rate of 82%, an exercise completion rate of 76%, and an acupressure massage frequency of 1.8 times per day. Three months later, retest data showed that the rate of BHI decline had significantly slowed (approximately -0.08 points / month), nighttime HRVLF power had recovered to over 90% of baseline levels, and the pulse depth characteristic value had increased by nearly 0.4 standardized units from its pre-intervention low. Based on this, the system determined the intervention was effective, automatically lifted the continuous warning status, and provided positive incentive feedback in the latest health report: "Your recent lifestyle adjustments have had a positive impact on bone health; please continue!" The model co-optimization engine incorporated this event chain into the intervention effect verification dataset, labeling it as a positive response sample. Based on this, the engine performed two optimizations: First, using this sample as a positive reward signal, it fine-tuned the cross-modal attention fusion module through reinforcement learning, increasing its attention to the combined changes in the intensity of pulse palpation and HRVLF in a similar perimenopausal group with kidney deficiency; Second, it extracted the combined intervention pattern of Baduanjin (Eight Pieces of Brocade), soy products, and vitamin D supplementation from this case, and statistical tests confirmed that it was significantly associated with BHI stability. Based on this, a high-confidence rule edge was added to the bone health intervention knowledge graph: middle-aged women with kidney deficiency, recommended Baduanjin combined with plant-based calcium intake, and improved bone metabolic stability, realizing the dynamic evolution of the knowledge system.

[0063] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims and their equivalents, this invention is also intended to include these modifications and variations.

Claims

1. A personalized bone health status tracking and risk alert system, characterized in that, It includes a data perception layer, a feature fusion layer, an intelligent decision-making layer, and an interactive application layer, which are connected sequentially. The data perception layer collects photoplethysmography (PPG) signals, triaxial acceleration signals, and gyroscope signals through the smart wearable device worn by the user, and collects structured data including age, gender, diet, exercise, medication, history of fracture, basic bone mineral density classification, and major TCM constitution type through the application APP used by the user, and generates the collected data. The feature fusion layer preprocesses the collected data and extracts Western medicine physiological feature sequences, traditional Chinese medicine Zang-Xiang feature vectors, and lifestyle feature vectors. The intelligent decision-making layer, based on the Western medicine physiological feature sequence, the traditional Chinese medicine Zang-Xiang feature vector, and the lifestyle feature vector, uses the personal bone health dynamic baseline model, the integrated traditional Chinese and Western medicine bone status assessment model, and the time series prediction and risk warning model to conduct personal bone health dynamic baseline tracking and integrated traditional Chinese and Western medicine feature analysis to obtain bone health analysis results. Among them, the personal bone health dynamic baseline model is used to establish a dynamic fluctuation range of personalized health indicators based on the continuous data collected by the user wearing smart wearable devices in the initial stage of data collection, and to continuously evolve the dynamic fluctuation range based on the sliding window and Bayesian update mechanism. A bone health assessment model integrating traditional Chinese and Western medicine employs an architecture including a two-stream encoder and a cross-modal attention fusion module. Based on the fusion features, it outputs a bone health status index and a distribution of TCM syndrome patterns. The two-stream encoder comprises a first encoder for processing Western medicine physiological feature sequences and a second encoder for processing TCM organ-manifestation feature vectors. The cross-modal attention fusion module is configured based on a formula... Calculate attention weights and fuse features based on formulas. Calculation generates; where, For attention weights, This represents the characteristic sequence of Western medical psychology. This represents the characteristic vector of Zang-Xiang in Traditional Chinese Medicine. , , These are the learnable weight matrices for queries, keys, and values, respectively. Scaling factor Let be the dimension of the key vector. Represents the query matrix. Represents the bond matrix. represents the transpose of the key matrix, and Softmax represents normalizing the attention scores into a probability distribution. Represents the characteristics of integration; The time-series prediction and risk warning model predicts the future trend of the bone health status index based on historical data, obtains the prediction results, and triggers graded warnings based on the comparison between the prediction results and the dynamic fluctuation range. The interactive application layer distributes bone health analysis results and corresponding personalized intervention plans to user terminals, family member terminals, and doctor workstations to form a closed-loop management process of monitoring, evaluation, prompting, intervention, and re-evaluation.

2. The personalized bone health status tracking and risk alert system according to claim 1, characterized in that, The data perception layer integrates a data processing module, which is used to preprocess and initially calculate the user's posture stability index, gait symmetry index, and gait regularity index based on the three-axis acceleration signal and gyroscope signal, and incorporate the posture stability index, gait symmetry index, and gait regularity index as components of the collected data.

3. The personalized bone health status tracking and risk alert system according to claim 1, characterized in that, Western medical pathological feature sequences include vascular elasticity parameters and microcirculation state parameters; vascular elasticity parameters and microcirculation state parameters are indirect biochemical correlation features related to bone metabolism estimated by inputting PPG signals into a pre-trained derivative model.

4. The personalized bone health status tracking and risk alert system according to claim 1, characterized in that, When establishing the dynamic fluctuation range of personalized health indicators, a robust statistical method based on the median and absolute median difference is used for calculation.

5. The personalized bone health status tracking and risk alert system according to claim 1, characterized in that, The dynamic fluctuation range is continuously evolved based on the sliding window and Bayesian update mechanism, including: introducing a weighted sliding window update strategy to assign differentiated weights to historical data of different time periods; setting up a trend detection unit to use robust statistical methods to determine whether there is a significant monotonic trend in the Western medicine pathological feature sequence, the traditional Chinese medicine Zang-Xiang feature vector, and the lifestyle feature vector; when the trend is confirmed, the fluctuation range drift update rule is triggered according to the Bayesian update mechanism to adjust the center position and the slope of the extended trend response; and setting a multi-level threshold mechanism to distinguish between trend drift and acute abnormal events in order to maintain the detection sensitivity of pathological indicator changes that deviate from the normal fluctuation pattern in the short term.

6. The personalized bone health status tracking and risk alert system according to claim 1, characterized in that, The extraction of TCM Zang-Xiang feature vectors includes: identifying and extracting quantitative features related to TCM kidney, liver, and spleen functions by performing time-frequency domain joint analysis and waveform morphology modeling on the pulse components in the acquired PPG signals; among them, the quantitative features related to kidney function include at least the pulse depth feature value and the pulse weakness index; the quantitative features related to liver function include at least the pulse stringiness feature value; and the quantitative features related to spleen function include at least the pulse softness and slowness feature value.

7. A personalized bone health status tracking and risk alert system according to claim 2, characterized in that, Tiered early warning includes: When the bone health status index deviates from the dynamic fluctuation range for more than a certain number of days within a consecutive predetermined number of days, or when it is predicted that the bone health status index will enter the preset health attention data range in the future, the first-level warning for pushing lifestyle adjustment reminders will be triggered. When the bone health status index is within the preset sub-health data range, or when the TCM syndrome distribution continues to point to a specific TCM syndrome, a second-level warning is triggered to recommend TCM conditioning and medical examination. When the user's short-term fall risk predicted based on posture stability, gait symmetry, and gait regularity indicators exceeds the high-risk value, or when the bone health status index is about to enter the high-risk data range, a Level 3 warning is triggered to simultaneously notify emergency contacts or family doctors.

8. A personalized bone health status tracking and risk alert system according to claim 7, characterized in that, The intelligent decision-making layer is also used for: Based on the collected data, static portrait data is extracted and generated. The static portrait data includes at least age, gender, baseline bone density classification, and major TCM constitution types. Based on static profile data, a set of personalized modulation parameters is generated. The modulation parameters include: baseline offset vector, feature channel weight vector, and risk judgment threshold adjustment coefficient. The baseline offset vector is used to perform initial calibration of the center value of the dynamic fluctuation range based on population subgroups; the feature channel weight vector is used to weight the importance of each dimension of features in the Western medicine physiological feature sequence and the traditional Chinese medicine Zang-Xiang feature vector; the risk judgment threshold adjustment coefficient is used to personalize the scaling of the number of days threshold, high-level risk value, and high-risk data interval boundary in the graded early warning system. When initializing the personal bone health dynamic baseline model, a baseline offset vector is loaded; before the dual-stream encoder in the integrated Chinese and Western medicine bone status assessment model processes the features of each dimension in the Western medicine physiological feature sequence and the Chinese medicine Zang-Xiang feature vector, the features of each dimension in the Western medicine physiological feature sequence and the Chinese medicine Zang-Xiang feature vector are modulated using the feature channel weight vector; when making risk decisions in the time series prediction and risk warning model, a risk judgment threshold adjustment coefficient is applied.

9. A personalized bone health status tracking and risk alert system according to claim 1, characterized in that, The intelligent decision-making layer includes a decision routing module; the decision routing module stores a bone health intervention knowledge graph, which is used to generate personalized intervention plans based on bone health analysis results, including nutritional advice, exercise programs, traditional Chinese medicine dietary therapy recipes, acupoint massage guidance, and medical indications.

10. A personalized bone health status tracking and risk alert system according to claim 9, characterized in that, The intelligent decision-making layer also includes a model co-optimization engine; The model co-optimization engine is configured as follows: Monitor and record user compliance feedback data, subsequent physiological indicator changes data, and external medical diagnosis results data after each tiered warning is triggered; Based on compliance feedback data, physiological indicator change data, and external medical diagnosis results data, an intervention effect verification dataset was constructed. Using the intervention effect validation dataset, reinforcement learning was performed on the cross-modal attention fusion module to optimize the learnable weight matrix; Based on the intervention effect verification dataset, the verified effective intervention patterns and results are extracted, and the association rules in the bone health intervention knowledge graph are updated with confidence weight or nodes are added.

Images