Early prediction model construction method for senile sarcopenia and metabolic high-risk phenotype

Abstract

The invention discloses an early prediction model construction method for senile sarcopenia and metabolic high-risk phenotypes. The method comprises the steps that 3D scanning data, body composition data and health phenotype data of senile patients are acquired, and data preprocessing is carried out; respectively extracting feature vectors of the three types of data after data preprocessing, and performing alignment and weighted fusion to obtain fusion vectors; inputting the fusion vector into a subtype classification model for training to obtain a preliminary prediction model; in the training process of the subtype classification model, an unsupervised learning method and a supervised learning method are adopted at the same time, and a staged freezing strategy and a five-fold cross validation method are adopted; the preliminary prediction model is verified and finely adjusted by using a verification data set prepared in advance, so that a final early prediction model of senile sarcopenia and metabolic high-risk phenotypes is obtained, and accurate diagnosis of sarcopenia and intelligent identification of metabolic high-risk subtypes can be realized through the constructed prediction model.

Description

Technical Field

[0001] The present application relates to the technical field of sarcopenia in the elderly, and specifically to a method for constructing an early prediction model for sarcopenia in the elderly and metabolic high-risk phenotypes. Background Art

[0002] Sarcopenia refers to the pathological phenomenon of progressive loss of muscle mass, strength, and function with aging, particularly in people aged 65 and older. With the aging of the global population, sarcopenia has become a major health issue for the elderly. It not only leads to frailty, unsteady gait, and an increased risk of falls, but is also closely associated with a variety of chronic diseases, such as diabetes and cardiovascular disease, as well as increased mortality. The pathogenesis of sarcopenia is complex, involving multiple factors, including neurological, endocrine, metabolic, and nutritional factors. Therefore, early identification and effective intervention are crucial.

[0003] The patent document with publication number CN119889712A discloses a method for predicting the risk of sarcopenia in the elderly. The method obtains sampling indicators of the elderly sarcopenia risk prediction object in multiple dimensions; the sampling indicators corresponding to each dimension are input into the corresponding sub-model in the sarcopenia risk prediction model, and the predicted sarcopenia result corresponding to the risk prediction object output by the sarcopenia risk prediction model is obtained. This method is mainly based on sampling indicators of multiple dimensions for prediction, but the data types of these dimensions are relatively simple, mainly concentrated in traditional medical data such as blood markers, intestinal flora and clinical characteristics, and it cannot distinguish between high-risk metabolic subtypes in the elderly population.

[0004] Patent document No. CN117894454A discloses a sarcopenia diagnosis method, device, and electronic device, which obtains target muscle ultrasound scan images and target clinical test indicators; inputs the target muscle ultrasound scan images and target clinical test indicators into a sarcopenia diagnosis model to obtain a diagnostic result output by the sarcopenia diagnosis model; wherein the sarcopenia diagnosis model is obtained based on multimodal data fusion network training. However, this method mainly diagnoses sarcopenia based on two data types: muscle ultrasound scan images and clinical test indicators. The data types are also relatively simple and cannot provide richer information for the prediction of sarcopenia. Moreover, it is not aimed at diagnosing sarcopenia in the elderly, nor can it distinguish between metabolic high-risk subtypes in the elderly population.

[0005] Patent publication CN116864103A discloses a method for diagnosing sarcopenia based on multimodal contrastive learning. This method uses non-local class activation mapping (NLC) enhancement to capture regional information in images, uses visual-text feature fusion (VFF) to mine correlations between visual and textual features, and uses auxiliary contrast representation (ACR) to improve the model's feature representation capabilities. However, this method primarily diagnoses sarcopenia based on hip X-ray images and clinical information (gender, age, BMI, height, and weight). The data is relatively limited and lacks the ability to identify high-risk metabolic subtypes.

[0006] In summary, current sarcopenia diagnosis relies primarily on simple measurements (e.g., grip strength, gait speed) and single imaging data (e.g., dual-energy X-ray DXA). However, these methods lack the ability to accurately distinguish between different subtypes and high-risk phenotypes of metabolic abnormalities. Furthermore, sarcopenia can lead to a variety of metabolic diseases and is also causally related to multiple metabolic diseases. Among sarcopenic individuals, especially those in the elderly, high-risk metabolic subtypes are difficult to distinguish using traditional methods, resulting in less targeted interventions. Currently, there is no system for identifying sarcopenia and its high-risk subtypes in the elderly population using multimodal data. The few risk prediction models available have limitations in generalizability and interpretability across populations.

[0007] Therefore, existing sarcopenia research lacks accurate identification of different subtypes in the elderly population and differentiation of high-risk phenotypes of metabolic abnormalities. Different types of sarcopenia, such as simple muscle loss, sarcopenic obesity, and sarcopenia with osteoporosis, as well as different somatotypes of sarcopenic patients, indicate different metabolic risks. Therefore, it is of great significance to accurately classify sarcopenia phenotypes in the elderly population using multidimensional health phenotypes and somatotype measurement data and identify high-risk representative phenotypes associated with sarcopenia. Summary of the Invention

[0008] To this end, the present application provides a method for constructing an early prediction model for sarcopenia and metabolic high-risk phenotypes in the elderly, so as to solve the problem that existing technologies are unable to accurately identify and differentiate sarcopenia and its high-risk subtypes in the elderly population.

[0009] In order to achieve the above objectives, this application provides the following technical solutions:

[0010] First, a method for constructing an early prediction model for sarcopenia and metabolic high-risk phenotypes in the elderly, comprising:

[0011] Step 1: Obtain 3D scanning data, body composition data, and health phenotype data of elderly patients;

[0012] Step 2: performing data preprocessing on the 3D scanning data, the body composition data, and the health phenotype data;

[0013] Step 3: extracting feature vectors of the 3D scan data, the body composition data, and the health phenotype data after data preprocessing respectively; wherein, a convolutional neural network is used to extract the feature vectors of the 3D scan data;

[0014] Step 4: aligning and weightedly fusing the feature vectors of the 3D scanning data, the body composition data, and the health phenotype data to obtain a fusion vector;

[0015] Step 5: Input the fusion vector into the subtype classification model for training to obtain a preliminary prediction model; the subtype classification model adopts both unsupervised learning and supervised learning methods during the training process, and adopts a staged freezing strategy and a five-fold cross-validation method;

[0016] Step 6: Use the pre-prepared validation dataset to verify and fine-tune the preliminary prediction model to obtain the final early prediction model for sarcopenia and metabolic high-risk phenotypes in the elderly.

[0017] Preferably, in step 1, the 3D scanning data includes muscle morphology data and body shape feature data.

[0018] Preferably, in step 2, when performing data preprocessing on the 3D scan data, the method specifically includes: first performing rigid and non-rigid registration on the 3D scan data, then using a denoising algorithm to remove isolated noise points, and finally normalizing the grid coordinates to the same world coordinate system.

[0019] Preferably, in step 2, when preprocessing the body composition data, it specifically includes: using wavelet transform to filter out high-frequency interference in the body composition data, and dividing the muscle and fat amounts in the body composition data by the overall body mass, converting them into ratios and mapping them between 0 and 1.

[0020] Preferably, in step 2, when performing data preprocessing on the health phenotype data, the data preprocessing specifically includes: applying Min-Max transformation to the continuous indicators in the health phenotype data, and performing one-hot encoding on the categorical variables; for the missing values ​​in the health phenotype data, using a pre-trained GAN generator to synthesize and fill in the missing values ​​based on other sample distributions, and then reducing the hundreds of dimensions of metabolic and inflammatory indicators in the health phenotype data to 10-15 dimensional principal component vectors through principal component analysis.

[0021] Preferably, in step 3, when the convolutional neural network is used to extract the feature vector of the 3D scan data, it specifically includes: first using a multi-scale convolutional layer to extract the muscle morphology and body shape features of the 3D scan data, then inputting the muscle morphology and body shape features into a global average pooling layer for aggregation to obtain morphological embedding vector body composition data, and finally inputting the morphological embedding vector body composition data into two layers of fully connected layers to obtain the feature vector of the 3D scan data.

[0022] Preferably, in step 3, a Transformer encoder with two layers and four attention heads in each layer is used to extract the feature vector of the health phenotype data.

[0023] Preferably, step 4 specifically includes: aligning the feature vectors of the 3D scanning data, the body composition data and the health phenotype data to 256 dimensions through linear mapping, calculating the correlation weights using a cross-modal self-attention layer, and fusing them according to the correlation weights to obtain a fusion vector.

[0024] Preferably, a multi-layer perceptual network is further included, and the multi-layer perceptual network is used to learn the adaptive coefficient of each attention path of the cross-modal self-attention layer.

[0025] The second aspect is an early prediction system for sarcopenia and metabolic high-risk phenotypes in the elderly, including:

[0026] A data acquisition module is used to obtain 3D scanning data, body composition data and health phenotype data of the elderly patient to be examined;

[0027] a data preprocessing module, configured to perform data preprocessing on the 3D scanning data, the body composition data, and the health phenotype data;

[0028] a feature extraction module, configured to extract feature vectors of the 3D scanning data, the body composition data, and the health phenotype data after data preprocessing;

[0029] a feature fusion module, configured to align and weightedly fuse the feature vectors of the 3D scanning data, the body composition data, and the health phenotype data to obtain a fusion vector;

[0030] An inference and prediction module is used to input the fusion vector into a prediction model, where the prediction model is obtained using the method for constructing an early prediction model for sarcopenia and metabolic high-risk phenotypes in the elderly, and the prediction results of sarcopenia and metabolic high-risk phenotypes in the elderly patient to be tested are inferred through the prediction model.

[0031] Compared with the prior art, this application has at least the following beneficial effects:

[0032] 1. This application provides a method for constructing an early prediction model for sarcopenia and metabolic high-risk phenotypes in the elderly, including: obtaining 3D scan data, body composition data, and healthy phenotype data of elderly patients and performing data preprocessing; extracting feature vectors of the preprocessed 3D scan data, body composition data, and healthy phenotype data respectively; aligning and weightedly fusing the feature vectors of the 3D scan data, body composition data, and healthy phenotype data to obtain a fusion vector; inputting the fusion vector into a subtype classification model for training to obtain a preliminary prediction model; the subtype classification model uses both unsupervised learning and supervised learning methods during training, and adopts a staged freezing strategy and a 5-fold cross-validation method; and using a pre-prepared validation dataset to verify and fine-tune the preliminary prediction model, thereby obtaining a final early prediction model for sarcopenia and metabolic high-risk phenotypes in the elderly. By fusing 3D scan data, body composition data, and healthy phenotype data (the three different data sources provide complementary information), this application enables the constructed prediction model to comprehensively assess the health status of elderly individuals from multiple perspectives, thereby more accurately identifying sarcopenia and metabolic high-risk phenotypes in the elderly.

[0033] 2. The early prediction system for sarcopenia and metabolic high-risk phenotypes in the elderly provided in this application can realize the accurate diagnosis of sarcopenia and intelligent identification of metabolic high-risk subtypes, solving the problem that the traditional single data source cannot fully reflect the complex health status of the elderly population, and enhancing the accuracy of diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] To more intuitively illustrate the prior art and the present application, exemplary drawings are provided below. It should be understood that the specific shapes and structures shown in the drawings should not generally be considered as limiting conditions for implementing the present application. For example, based on the technical concepts disclosed in this application and the exemplary drawings, those skilled in the art are capable of easily making routine adjustments or further optimizations to the addition / reduction / attribution division of certain units (components), the specific shapes, positional relationships, connection methods, and dimensional ratios.

[0035] Figure 1 This is a flow chart of a method for constructing an early prediction model for sarcopenia and metabolic high-risk phenotypes in the elderly provided in Example 1 of the present application;

[0036] Figure 2 A schematic diagram of a method for constructing an early prediction model for sarcopenia and metabolic high-risk phenotypes in the elderly provided in Example 1 of the present application;

[0037] Figure 3 This is a schematic diagram of the structure of the multimodal fusion network provided in Example 1 of the present application. DETAILED DESCRIPTION

[0038] The present application will be further described below in detail through specific embodiments in conjunction with the accompanying drawings.

[0039] In the description of this application: unless otherwise specified, "plurality" means two or more. The terms "first," "second," "third," etc. in this application are intended to distinguish the objects referred to and do not have any special technical connotations (for example, they should not be understood as emphasizing importance or order). Expressions such as "including," "comprising," and "having" also mean "not limited to" (certain units, components, materials, steps, etc.).

[0040] The terms such as "upper", "lower", "left", "right", "middle", etc. cited in this application are usually used to indicate the general relative position relationship for the convenience of intuitive understanding by referring to the drawings, and are not absolute limitations on the position relationship in the actual product.

[0041] Example 1

[0042] See also Figure 1 and Figure 2 This embodiment provides a method for constructing an early prediction model for sarcopenia and metabolic high-risk phenotypes in the elderly, comprising:

[0043] S1: Obtain 3D scanning data, body composition data, and health phenotype data of elderly patients;

[0044] In this step, 3D scan data is collected using a 3D optical scanner. Based on anthropometric and ergonomic requirements, the 3D optical scanner automatically identifies feature points throughout the human body and calculates over 160 dimensional data items, including muscle morphology and body shape data, including length, width, thickness, circumference, height, and angles. These dimensions comply with ISO 20685 (GBT 5703) and ISO 8559 (GBT 16160). Body composition data is measured using bioelectrical impedance (BIA) to measure muscle mass, fat mass, bone mass, and the distribution of muscle and fat mass in different left and right regions. Health phenotypic data includes blood metabolic indicators (such as insulin resistance, inflammatory factors, blood sugar, blood lipids, uric acid, and renal function), nutritional intake, and lifestyle.

[0045] Specifically, after initialization is complete, the first step is to accurately collect 3D scanning data and body composition data. For 3D optical scanners, the resolution (e.g., point pitch ≤ 1mm), scanning speed, and ambient lighting conditions must be pre-set to ensure the geometric accuracy and reconstruction consistency of the mesh model; at the same time, the scanning posture standard (e.g., arms outstretched or feet parallel) must be defined, and the internal and external parameters of each scan must be recorded in the software. Body composition data is measured using a 50kHz bioelectrical impedance analyzer. Electrodes are placed on the subject's left and right hands and feet using the quadrupole method to measure muscle mass, fat mass, and bone mass throughout the body and in various regions (upper limbs, trunk, and lower limbs). The effects of surface temperature and humidity on resistance measurements are calibrated in real time. All raw data is written to the database through a unified API interface and automatically timestamped with a device ID.

[0046] S2: Data preprocessing of 3D scanning data, body composition data, and health phenotype data;

[0047] Specifically, after all data collection is complete, data preprocessing is required. This includes: first, performing rigid and non-rigid registration on the 3D scan data. During these registrations, key human points (acromion, acetabulum, knee, and ankle) are used to achieve rigid alignment between different individuals and scanned poses. Next, a Gaussian filter-based denoising algorithm is used to remove isolated noise points. Finally, the mesh coordinates are normalized to the same world coordinate system.

[0048] The data preprocessing of body composition data includes: applying wavelet transform to filter out high-frequency interference, and converting the muscle and fat mass of each partition into a ratio divided by the overall body mass, and mapping it to [0, 1].

[0049] Data preprocessing of healthy phenotypic data includes: normalization: using Min-Max transformation for continuous indicators and one-hot encoding for categorical variables; for missing values, calling a pre-trained GAN generator to synthesize and fill in the missing values ​​based on other sample distributions, and then using principal component analysis to reduce hundreds of dimensions of metabolic and inflammatory indicators to 10-15 dimensional principal component vectors.

[0050] S3: extract the feature vectors of 3D scanning data, body composition data and health phenotype data after data preprocessing respectively;

[0051] This step requires feature extraction of three types of data (3D scanning data, body composition data, and health phenotype data) separately and mapping them into a low-dimensional embedding space for subsequent fusion.

[0052] A convolutional neural network (CNN, for example, based on MobileNetV3) is used to extract muscle morphology feature vectors from 3D scan data. Specifically, the 3D scan data is passed through multi-scale convolutional layers to extract muscle morphology and body shape features, and the information is aggregated at the end of the network through a global average pooling layer to obtain a 512-dimensional morphological embedding vector. The body composition data (BIA ratio vector) is then sent to a two-layer fully connected network, first mapped to a 128-dimensional hidden layer, and then reduced to 64 dimensions to form a component embedding, thereby obtaining the feature vector of the 3D scan data.

[0053] Body composition data and health phenotype data are generally structured data, so fully connected networks or Transformers are sufficient for feature extraction. Because health phenotype data has already been filtered and reduced in dimension during data preprocessing, a two-layer Transformer encoder with four attention heads per layer is used to automatically capture the relationships between different metabolic indicators. This outputs a 128-dimensional phenotype embedding, thereby generating a feature vector for the health phenotype data.

[0054] S4: aligning and weightedly fusing the feature vectors of 3D scanning data, body composition data, and health phenotype data to obtain a fusion vector;

[0055] See also Figure 3 This step mainly integrates multi-source data through a feature fusion layer (e.g., attention mechanism). Specifically, in order to fuse multimodal information in the same space, this embodiment first aligns the three feature vectors to 256 dimensions through a linear mapping, which can be expressed as:

[0056]

[0057] in, , represents the transformed features (i.e. the unified 256-dimensional feature vector), Represents the linear transformation matrix (that is, the original d i dimensional feature mapping to a 256-dimensional weight matrix), represents the original features (i.e. the original feature vectors of each mode), represents the bias vector (i.e., a 256-dimensional bias term used to adjust the transformation baseline), shape, comp, and meta represent three feature vectors.

[0058] Then, the cross-modal self-attention layer calculates the correlation weights between features and generates weighted representations through the Query-Key-Value mechanism:

[0059] .

[0060] To enable the model to dynamically adjust the contribution of each modality based on task requirements, this embodiment also introduces a small MLP (Multi-Layer Perceptron) network (MLP) based on the attention output to learn the adaptive coefficients for each channel and normalize them to ensure their sum is unity. Finally, the three attention outputs are combined using a weighted sum to generate a 64-dimensional fusion vector, which serves as the input for subsequent classification or clustering modules.

[0061] In summary, this step adopts a multimodal data fusion method based on the attention mechanism to dynamically adjust the importance of different data sources in the model, and ultimately optimize the contribution of various data sources to the final model. Specifically, the attention mechanism automatically identifies which features are more critical to the prediction task by calculating the correlation between the features of each data source. During the fusion process, not only the individual features of each data source are considered, but also different weights are adaptively assigned to them according to the characteristics of different input modalities. For example: 3D scanning data may play a greater role in identifying muscle morphology, while metabolic phenotype data is more important for the prediction of subtype classification. Through this dynamic weight adjustment, data from different sources can be processed more accurately, avoiding redundancy or information loss of a certain modality, thereby improving the predictive ability of the final model.

[0062] S5: Input the fusion vector into the subtype classification model for training to obtain a preliminary prediction model; the subtype classification model adopts both unsupervised learning and supervised learning methods during the training process, and adopts a staged freezing strategy and a five-fold cross-validation method;

[0063] During the training process, this embodiment provides a set of joint loss functions that perform a weighted summation of the supervised cross entropy classification loss and the unsupervised clustering or contrastive learning loss, and adds a parameter regularization term to avoid overfitting. Specifically, the hyperparameters 、 and The weights for classification, clustering, and regularization losses are controlled separately and automatically tuned using a validation set. During each training cycle, the model samples both labeled and unlabeled samples in a certain ratio, ensuring that supervised and unsupervised learning proceed in parallel. A phased freezing strategy is employed—initially training only the morphological and component networks, adding the phenotypic subnetwork mid-term, and finally jointly fine-tuning the entire network to stabilize the convergence of different modules.

[0064] In this embodiment, the loss function adopts the form of a weighted sum of three parts: on the one hand, the cross entropy is used to measure the classification performance of known subtype labels, on the other hand, the center loss (or contrast loss) is used to enhance the compactness of unsupervised clustering, and the L2 regularization term is added to control the complexity of the model; the hyperparameter 、 and The optimal combination is determined through grid search on multiple validation sets. Training uses five-fold cross-validation, with the morphological, compositional, and phenotypic sub-networks and the fusion layer updated synchronously within each fold to ensure that all modalities are fully learned.

[0065] In this embodiment, unsupervised learning (such as clustering) can discover potential high-risk metabolic subtypes, while supervised learning (such as SVM, random forest) can predict labels for subtypes.

[0066] S6: Use the pre-prepared validation dataset to verify and fine-tune the preliminary prediction model to obtain the final early prediction model for sarcopenia and metabolic high-risk phenotypes in the elderly.

[0067] After training, this example tests and verifies the model on external public datasets (such as public elderly body shape and metabolic data) to evaluate its generalization ability and accuracy in detecting different subtypes. The fusion strategy and classification threshold are then fine-tuned accordingly to obtain the final early prediction model for sarcopenia and high-risk metabolic phenotypes in the elderly.

[0068] Using the early prediction model of sarcopenia and metabolic high-risk phenotypes in the elderly trained in this embodiment, not only can the risk of sarcopenia in the elderly be scored (for example: 0-100%), but the subtypes can also be classified (for example: "inflammation-dominated type" "metabolic disorder type"), and finally an interpretable report (for example: visualization of key feature contribution) can be obtained. Specifically, this embodiment can provide each patient with personalized intervention recommendations based on the metabolic risks of different subtypes, including personalized nutrition and exercise plans for different sarcopenia subtypes (such as "inflammation-dominated type" and "metabolic disorder type"). This personalized intervention strategy can help patients effectively manage sarcopenia risks in the early stages and slow down the progression of sarcopenia. It can also effectively prevent metabolic diseases related to sarcopenia.

[0069] In addition to providing real-time health assessment and intervention, this embodiment also supports long-term health management of patients. Through regular data collection and evaluation, it helps patients and doctors monitor treatment effects and conduct health status assessment and adjustments.

[0070] This example uses multimodal data fusion technology, combining 3D scanning data with metabolic phenotyping data for the first time. This not only accurately captures the muscle morphology and body shape of older adults through 3D scanning, but also combines body composition data (such as muscle mass, fat mass, and bone mass) measured by bioelectrical impedance (BIA) with blood metabolic indicators (such as insulin resistance and inflammatory factors) to achieve precise diagnosis of sarcopenia and intelligent identification of high-risk metabolic subtypes. This data fusion addresses the problem that traditional single data sources cannot fully reflect the complex health conditions of the elderly population, thereby enhancing diagnostic accuracy.

[0071] The method provided in this embodiment not only fills the gap in the field of early identification of sarcopenia and metabolic diseases in the elderly, but also can effectively combine emerging artificial intelligence technologies with research results in medical imaging, biomedical engineering and other fields, promote the development of primary medical institutions in the direction of intelligence and precision, and provide strong support for the health management of the elderly population.

[0072] Example 2

[0073] This embodiment provides an early prediction system for sarcopenia and metabolic high-risk phenotypes in the elderly, including:

[0074] A data acquisition module is used to obtain 3D scanning data, body composition data and health phenotype data of the elderly patient to be examined;

[0075] a data preprocessing module, configured to perform data preprocessing on the 3D scanning data, the body composition data, and the health phenotype data;

[0076] a feature extraction module, configured to extract feature vectors of the 3D scanning data, the body composition data, and the health phenotype data after data preprocessing;

[0077] a feature fusion module, configured to align and weightedly fuse the feature vectors of the 3D scanning data, the body composition data, and the health phenotype data to obtain a fusion vector;

[0078] An inference and prediction module is used to input the fusion vector into a prediction model, where the prediction model is obtained using the method for constructing an early prediction model for sarcopenia and metabolic high-risk phenotypes in the elderly, and the prediction results of sarcopenia and metabolic high-risk phenotypes in the elderly patient to be tested are inferred through the prediction model.

[0079] Taking into account the diversity of devices and application scenarios, this system has designed a lightweight AI model that can run on low-power mobile devices, such as portable scanners used in community hospitals. This design not only reduces the system's hardware requirements but also greatly improves the model's applicability in primary care settings, enabling early screening for sarcopenia and its high-risk metabolic subtypes in the elderly.

[0080] For the specific implementation content of each module in the early prediction system of sarcopenia and metabolic high-risk phenotype in the elderly, please refer to the above definition of the method for constructing an early prediction model of sarcopenia and metabolic high-risk phenotype in the elderly, which will not be repeated here.

[0081] The technical features of the above embodiments can be combined arbitrarily (as long as there is no contradiction in the combination of these technical features). In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described; these embodiments that are not explicitly written should also be considered to be within the scope of this specification.

Claims

1. A method for constructing an early prediction model for sarcopenia and metabolic high-risk phenotypes in the elderly, characterized by: include: Step 1: Obtain 3D scanning data, body composition data, and health phenotype data of elderly patients; Step 2: performing data preprocessing on the 3D scanning data, the body composition data, and the health phenotype data; Step 3: extracting feature vectors of the 3D scan data, the body composition data, and the health phenotype data after data preprocessing respectively; wherein, a convolutional neural network is used to extract the feature vectors of the 3D scan data; Step 4: aligning and weightedly fusing the feature vectors of the 3D scanning data, the body composition data, and the health phenotype data to obtain a fusion vector; Step 5: Input the fusion vector into the subtype classification model for training to obtain a preliminary prediction model; the subtype classification model adopts both unsupervised learning and supervised learning methods during the training process, and adopts a staged freezing strategy and a five-fold cross-validation method; Step 6: Use the pre-prepared validation dataset to verify and fine-tune the preliminary prediction model to obtain the final early prediction model for sarcopenia and metabolic high-risk phenotypes in the elderly.

2. The method for constructing an early prediction model for sarcopenia and metabolic high-risk phenotype in the elderly according to claim 1, characterized in that: In step 1, the 3D scanning data includes muscle morphology data and body shape feature data.

3. The method for constructing an early prediction model for sarcopenia and metabolic high-risk phenotype in the elderly according to claim 1, characterized in that: In step 2, when performing data preprocessing on the 3D scan data, specifically including: firstly performing rigid and non-rigid registration on the 3D scan data, then using a denoising algorithm to remove isolated noise points, and finally normalizing the grid coordinates to the same world coordinate system.

4. The method for constructing an early prediction model for sarcopenia and metabolic high-risk phenotype in the elderly according to claim 1, characterized in that: In step 2, when the body composition data is preprocessed, it specifically includes: using wavelet transform to filter out high-frequency interference in the body composition data, and dividing the muscle and fat amounts in the body composition data by the total body mass, converting them into ratios and mapping them between 0 and 1.

5. The method for constructing an early prediction model for sarcopenia and metabolic high-risk phenotype in the elderly according to claim 1, characterized in that: In step 2, data preprocessing of the health phenotype data specifically includes: applying Min-Max transformation to the continuous indicators in the health phenotype data and one-hot encoding the categorical variables; using a pre-trained GAN generator to fill in the missing values ​​in the health phenotype data based on other sample distributions, and then using principal component analysis to reduce the hundreds of dimensions of metabolic and inflammatory indicators in the health phenotype data to 10-15 dimensional principal component vectors.

6. The method for constructing an early prediction model for sarcopenia and metabolic high-risk phenotype in the elderly according to claim 1, characterized in that: In step 3, when the convolutional neural network is used to extract the feature vector of the 3D scan data, it specifically includes: first using a multi-scale convolutional layer to extract the muscle morphology and body shape features of the 3D scan data, then inputting the muscle morphology and body shape features into a global average pooling layer for aggregation to obtain morphological embedding vector body composition data, and finally inputting the morphological embedding vector body composition data into two layers of fully connected layers to obtain the feature vector of the 3D scan data.

7. The method for constructing an early prediction model for sarcopenia and metabolic high-risk phenotype in the elderly according to claim 1, characterized in that: In step 3, a two-layer Transformer encoder with a four-head attention mechanism in each layer is used to extract the feature vector of the health phenotype data.

8. The method for constructing an early prediction model for sarcopenia and metabolic high-risk phenotype in the elderly according to claim 1, characterized in that: Step 4 specifically includes: aligning the feature vectors of the 3D scan data, the body composition data, and the health phenotype data to 256 dimensions through linear mapping, calculating the correlation weights using a cross-modal self-attention layer, and fusing them according to the correlation weights to obtain a fusion vector.

9. The method for constructing an early prediction model for sarcopenia and metabolic high-risk phenotype in the elderly according to claim 8, characterized in that: It also includes a multi-layer perceptual network, which is used to learn the adaptive coefficients of each attention path of the cross-modal self-attention layer.

10. An early prediction system for sarcopenia and metabolic high-risk phenotypes in the elderly, characterized by: include: A data acquisition module is used to obtain 3D scanning data, body composition data and health phenotype data of the elderly patient to be examined; a data preprocessing module, configured to perform data preprocessing on the 3D scanning data, the body composition data, and the health phenotype data; a feature extraction module, configured to extract feature vectors of the 3D scanning data, the body composition data, and the health phenotype data after data preprocessing; a feature fusion module, configured to align and weightedly fuse the feature vectors of the 3D scanning data, the body composition data, and the health phenotype data to obtain a fusion vector; An inference and prediction module is used to input the fusion vector into a prediction model, wherein the prediction model is obtained by using the early prediction model construction method for sarcopenia and metabolic high-risk phenotypes in the elderly described in any one of claims 1 to 9, and the prediction result of sarcopenia and metabolic high-risk phenotypes in the elderly patient to be tested is inferred by the prediction model.

Images