Method, system, and medium for visceral fat prediction based on language model augmentation

Abstract

The present application belongs to the technical field of artificial intelligence medical treatment, and provides a visceral fat prediction method and system based on language model enhancement and a medium, comprising: acquiring baseline physiological data and a diet plan; generating semantic vectors for features through a pre-trained language model, and fusing the semantic vectors with numerical data to generate semantic enhancement features; predicting a first prediction result of visceral fat change through a first prediction model; acquiring early physiological response data in the middle of intervention; generating a second prediction result through a second prediction model in combination with the early response data; and generating personalized intervention suggestions based on comparison of the second prediction result with a preset threshold. The present application introduces biomedical semantics of features and constructs a two-stage dynamic prediction architecture, solves the problems of low prediction accuracy and inability to utilize feature biomedical semantics, significantly improves prediction accuracy, and realizes closed-loop personalized intervention guidance.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence medical technology, specifically, it relates to a method, system and medium for predicting visceral fat based on language model enhancement, and in particular, it relates to a system and method for predicting changes in visceral fat under dietary intervention. Background Technology

[0002] Visceral fat is a key indicator for assessing cardiovascular metabolic risk, and its accurate prediction is crucial for personalized health management. Current technologies for predicting the effects of dietary interventions typically employ traditional statistical methods or basic machine learning models, such as random forest or support vector machine models. However, these methods often perform poorly when applied to clinical data with small sample sizes, exhibiting limited predictive accuracy and poor generalization ability.

[0003] To improve predictive accuracy, some existing technologies attempt to employ neural network models and dynamically evaluate data from users during interventions to continuously update the health assessment model and simulate the subsequent effects of different intervention programs. However, these technologies still have significant drawbacks. First, existing models typically treat all physiological characteristics, such as white blood cell count and triglycerides, as indiscriminate numerical values, completely ignoring the rich biomedical semantic information inherent in these characteristic names. This prevents the model from effectively utilizing vast amounts of prior medical knowledge to aid learning, resulting in severely insufficient feature representation in small-sample scenarios such as clinical trials, thus limiting predictive accuracy. Second, the dynamic adjustment mechanisms of existing technologies are usually a general, continuously updated closed-loop process, lacking a structured predictive correction process. They fail to explicitly utilize specific physiological response data from the early stages of intervention to accurately calibrate the mid- to long-term intervention endpoints, potentially missing the optimal time to adjust the intervention program.

[0004] The patent document "Nutritional Health Assessment System for Chronic Heart Failure Based on Machine Learning" (CN120089383A) discloses a system based on machine learning that integrates multiple data modalities and analysis techniques to solve the problem that existing systems cannot fully characterize patients' nutritional behavior and physiological state, and achieves efficient and accurate nutritional assessment and personalized nutritional recommendations. However, it lacks the accuracy and reliability of prediction in scenarios with small sample data.

[0005] Therefore, there is a lack of effective solutions in the current technology for in-depth modeling and accurate prediction of visceral fat as a key risk factor. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the present invention aims to provide a method, system, and medium for predicting visceral fat based on language model enhancement, which addresses the problems of low prediction accuracy on small sample clinical data, inability to utilize characteristic biomedical semantics, and lack of a dynamic correction prediction mechanism based on early intervention feedback.

[0007] A method for predicting visceral fat based on language model enhancement according to the present invention includes:

[0008] Obtain the user's baseline physiological data and at least one dietary intervention to be evaluated; Based on the baseline physiological data, semantically enhanced features are generated; The semantic vector is fused with the numerical data corresponding to the feature to obtain the enhanced feature; The enhanced features are combined with other features in the baseline physiological data that do not generate semantic vectors to obtain the semantic enhancement features; The semantic enhancement features and the scheme features characterizing the at least one dietary intervention scheme are used as input to the first model. Through the first prediction model, a first prediction result is used to predict the change in visceral fat of the user at the end of a preset intervention period under the at least one dietary intervention scheme. At a predetermined midpoint within the predetermined intervention period, the user's early physiological response data is acquired. Combining the baseline physiological data and the early physiological response data, a second prediction result is generated by the second prediction model to show the change in visceral fat of the user at the end of the preset intervention period; Based on the second prediction result, the second prediction result is compared with a preset threshold to determine whether the intervention plan needs to be adjusted, and personalized intervention suggestions are generated accordingly.

[0009] The generation of semantically enhanced features includes generating a semantic vector representing the semantic information of at least one feature in the baseline physiological data through a pre-trained language model.

[0010] Preferably, the generation of personalized intervention recommendations includes: If the second prediction result does not reach the preset threshold, the first prediction model is used to simulate the prediction results of the other alternative dietary interventions by taking the user's semantic enhancement features and the scheme features representing other alternative dietary intervention schemes extracted from a preset scheme library as input, and recommending an alternative intervention scheme with the highest predicted value for visceral fat reduction.

[0011] Preferably, the generated semantically enhanced features include: Prior to the fusion, the semantic vector is subjected to dimensionality reduction processing.

[0012] Preferably, the fusion includes: The reduced semantic vector is weighted and fused with the numerical data corresponding to the feature, or multiplied element-wise or concatenated.

[0013] The weights used in the weighted fusion are preset weights or weights obtained through model training.

[0014] Preferably, the first prediction model and / or the second prediction model is a multilayer perceptron network or a gradient boosting decision tree model.

[0015] Preferably, the early physiological response data includes the change in visceral fat area of ​​the user at a preset intermediate time point.

[0016] Preferably, the early physiological response data also includes the user's weight change and / or fasting blood glucose change.

[0017] Preferably, the preset intermediate time point is the 14th day after the intervention begins.

[0018] The preset intervention period is 4 weeks.

[0019] The present invention provides a visceral fat prediction system based on language model enhancement, comprising: at least one processor and a memory communicating with said at least one processor.

[0020] The memory stores instructions.

[0021] When the instruction is executed by the at least one processor, the system performs the language model-enhanced visceral fat prediction method.

[0022] According to the present invention, a computer-readable storage medium storing a computer program is provided, wherein the computer program, when executed by a processor, implements the steps of the language model-enhanced visceral fat prediction method.

[0023] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention solves the problem of insufficient feature representation ability of traditional models due to the inability to utilize prior knowledge on small sample data by using pre-trained language models to enhance the semantics of numerical features.

[0024] 2. This invention introduces early response data for two-stage dynamic correction, which can significantly reduce prediction errors, achieve accurate prediction of individualized intervention effects, and significantly improve prediction accuracy.

[0025] 3. The "baseline recommendation + dynamic optimization" two-stage architecture constructed in this invention solves the problem that static prediction models cannot adapt to individual dynamic changes. It can promptly identify and correct ineffective solutions in the middle of the intervention, avoid users wasting time and resources, and improve the success rate and efficiency of intervention.

[0026] 4. This invention integrates the biomedical semantic vectors of features into numerical prediction tasks, enabling the model to understand the intrinsic relationships between features rather than just processing isolated numerical values, thereby improving interpretability and providing a new data analysis perspective for revealing the complex biological mechanisms behind visceral fat metabolism. Attached Figure Description

[0027] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 A schematic diagram of the architecture of a language model-enhanced visceral fat prediction system; Figure 2 This is a schematic diagram of the process for a language model-enhanced method for predicting visceral fat. Figure 3 This is a schematic diagram of the semantic enhancement feature generation process in an embodiment of the present invention; Figure 4 This is a schematic diagram of feature filtering in an embodiment of the present invention; Figure 5 This is a timing diagram of the signaling interaction within the system in an embodiment of the present invention.

[0028] Figure 1 The text shows: 10. User terminal; 20. Cloud server; 21. Data acquisition module; 22. Semantic enhancement module; 23. Two-stage prediction module; 231. Baseline prediction sub-model; 232. Early response prediction sub-model; 24. Adaptive decision support module; 30. Pre-trained language model service. Detailed Implementation

[0029] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0030] This invention provides a visceral fat prediction system based on language model enhancement. By utilizing a pre-trained language model for feature semantic enhancement and combining it with a two-stage prediction architecture based on a lightweight multilayer perceptron, it achieves a complete process of accurate prediction and dynamic adjustment of the visceral fat reduction effect on an individual under a specific dietary intervention program. Figure 1 For example, specifically: Deployed on a cloud server 20, it interacts with the user terminal 10 via a network. The user terminal 10 can be any device capable of running applications or browsing web pages, such as a smartphone, personal computer, or tablet. Its main functions are user data input and system result display.

[0031] The cloud server 20 serves as the computing core of the entire system, comprising at least one processor and memory. The memory stores computer programs, which, when executed by the processor, are used to implement a language model-enhanced visceral fat prediction method. Functionally, the cloud server 20 can be divided into multiple logical modules.

[0032] Specifically, the data acquisition module 21 is used to receive and store various types of data from the user terminal 10, including the user's baseline physiological data and early physiological response data collected during the intervention process.

[0033] The semantic enhancement module 22, as one of the core innovations of this application, processes the collected physiological features, communicates with the external pre-trained language model service 30 to generate semantic vectors that can represent the biomedical connotation of the features, and fuses them with the original numerical features to generate semantically enhanced features.

[0034] The two-stage prediction module 23 contains two independent prediction models: baseline prediction sub-model 231 and early response prediction sub-model 232.

[0035] The baseline prediction sub-model 231 makes an initial prediction based on semantically enhanced baseline data, while the early response prediction sub-model 232 combines dynamic data from the middle of the intervention to correct the prediction results.

[0036] The adaptive decision support module 24 parses the output of the two-stage prediction module 23 and generates personalized intervention suggestions based on preset clinical thresholds and decision logic, such as recommending an initial plan, suggesting to continue the current plan, or recommending an alternative plan, and sends such suggestions back to the user terminal 10.

[0037] This invention provides a method for predicting visceral fat based on language model enhancement. Figure 2 For example, the specific execution steps include: Step S101: Collect baseline physiological data and dietary plan.

[0038] The data acquisition module 21 obtains comprehensive baseline information from the user. This information is mainly divided into two parts: the user's baseline physiological data and the dietary intervention plan to be evaluated.

[0039] Specifically, baseline physiological data can include multiple dimensions. For example, body composition indicators, which can be measured using bioelectrical impedance analysis devices (such as the InBody 770 body composition analyzer) under standard conditions (e.g., fasting in the morning, empty bladder, and women avoiding menstruation), include, but are not limited to, "intracellular fluid / total water ratio," "thigh fat-free body weight percentage," "leg body fat mass," and "hip circumference." Blood metabolic indicators, obtained through standard clinical biochemical tests, such as "alkaline phosphatase," "urea," "total bilirubin," "fasting blood glucose," "high-density lipoprotein cholesterol," and "triglycerides." Immune and inflammatory indicators, also obtained through routine blood tests, such as "white blood cell count," "absolute lymphocyte count," "absolute monocyte count," and "lymphocyte percentage." In addition, lifestyle indicators may be included, such as "nap duration" recorded through questionnaires or wearable devices and "respiratory quotient" measured using a metabolic cart.

[0040] In more preferred examples, the above 14 core physiological characteristics were collected.

[0041] Simultaneously, a protocol library containing various standardized dietary intervention protocols is pre-defined. For example, this library may include seven clinically validated dietary protocols, such as a balanced diet, a 16:8 time-restricted eating protocol, an alternate-day fasting protocol, a 5:2 intermittent fasting protocol, a continuous calorie restriction protocol, and a very low-calorie diet protocol. Each protocol is distinguished by a unique protocol identifier.

[0042] In more preferred examples, the preset intermediate time point was adjusted to day 21. The collected early physiological response data were expanded into a multidimensional response vector, which, in addition to the change in visceral fat area ΔVFA_week3, also included the change in body weight ΔWeight_week3 and the change in fasting blood glucose ΔFPG_week3.

[0043] Step S102: Generate semantically enhanced features. This involves transforming the biomedical semantic information of feature names, which traditional models cannot understand, into numerical representations that the model can utilize. Figure 3 For example, this is performed by the semantic enhancement module 22.

[0044] The raw data collected in step S101 is preprocessed. For numerical features, Z-score normalization is used, and the formula is as follows:

[0045] in, It is the original value. It is the mean of this feature in the training set. The standard deviation is used to eliminate dimensional differences between different features. For categorical features (such as diet plan IDs), one-hot encoding is used. For missing values ​​that may exist in the data, multiple imputation can be used to fill in the missing values ​​to ensure data integrity.

[0046] Perform feature filtering. Figure 4 For example, in order to reduce model complexity and avoid collinearity, a screening strategy that combines model importance and feature correlation is adopted.

[0047] Specifically, a random forest model is first trained, and the SHAP method is used to calculate the contribution of each feature to the prediction target (i.e., the SHAP value). The Spearman rank correlation coefficients between all pairs of features are then calculated. When the absolute value of the correlation coefficient between a pair of features exceeds a preset threshold (e.g., 0.80), it indicates that the two features are highly correlated and there is information redundancy. In this case, features with higher SHAP values ​​are retained, while features with lower SHAP values ​​are discarded. Figure 4 In this study, different diets were divided into Group 1 to Group 7, each with m features and n participants, and their importance was ranked.

[0048] This method can screen out a core feature set that has a significant ability to distinguish changes in visceral fat. The screened features can effectively distinguish between the "high visceral fat reduction group" and the "low visceral fat reduction group" in the feature space, and permanently replace the statistical results of multivariate analysis of variance (e.g., P=0.016), thus proving the significance of this distinction.

[0049] The core semantic enhancement step involves constructing a text description for each selected core physiological feature, which briefly explains the biomedical meaning of the feature. For example, for the feature "WBC", the text description could be: "White blood cell count, an indicator reflecting systemic inflammation and immune status."

[0050] exist Figure 3 In this context, Feature k represents the k-th feature. The semantic enhancement module 22 calls an external pre-trained language model service 30, such as a Qwen3-Embedding-8B model (embedding model) trained on massive amounts of text data and possessing powerful semantic understanding capabilities. The text descriptions constructed for each feature are sent as input to this language model service. After processing the text, the language model service generates a high-dimensional floating-point vector for each text description, i.e., a semantic vector (or embedding vector).

[0051] The vector can be viewed as the coordinates of the biomedical semantics of the feature in mathematical space. Features with similar semantics (such as "white blood cell count" and "lymphocyte percentage") are also closer in the vector space.

[0052] Since the original semantic vectors generated by language models are usually very high-dimensional, direct use may lead to the curse of dimensionality and model overfitting. Therefore, it is necessary to perform dimensionality reduction processing on these high-dimensional semantic vectors.

[0053] In more preferred examples, principal component analysis (PCA) is used to project the original high-dimensional vector onto a few principal components. For example, the goal can be to retain more than 95% of the original information variance, thereby reducing the vector dimension from thousands to 11 dimensions.

[0054] The reduced semantic vector is fused with the original value corresponding to the feature after Z-score normalization to generate the final semantically enhanced feature.

[0055] There are various fusion methods. In more preferred examples, a weighted fusion method using element-wise multiplication is adopted, and its mathematical expression can be represented as:

[0056] in, This represents the Hadamard product, which is the product of elements.

[0057] This fusion approach can be understood as using semantic vectors to "modulate" or "weight" the original numerical values, so that the model can not only see the magnitude of the numerical value itself, but also understand the biological context in which the numerical value exists.

[0058] The enhanced representations of all features are combined to form a complete feature vector for the next prediction step.

[0059] In more preferred examples, when selecting a pre-trained language model, a language model that has been specifically pre-trained in biomedical literature, such as the BioBERT model, is chosen.

[0060] Because the BioBERT model contains a large amount of biomedical terminology and contextual relationships in its training corpus, it can theoretically understand and represent the biomedical semantics of physiological characteristics more accurately. Accordingly, the semantic enhancement module 22 sends the feature text description to the BioBERT model service to obtain semantic vectors.

[0061] In the feature fusion stage, vector concatenation is used. After principal component analysis (PCA) is performed on the semantic vectors generated by BioBERT to reduce their dimensionality (e.g., to 11 dimensions), they are no longer multiplied element-wise with the standardized original values. Instead, the two vectors are directly concatenated in terms of dimension. For example, if the reduced semantic vector is 11 dimensions and the original values ​​are 1 dimension, the concatenation will form a new 12-dimensional feature vector. Its mathematical expression can be represented as:

[0062] Because the dimensions of the input features have changed (from 11 to 12), the number of input layer nodes in the baseline prediction sub-model 231 and the early response prediction sub-model 232 in the subsequent two-stage prediction module 23 also needs to be adjusted from 11 to 12 accordingly.

[0063] The structure of other parts of the model, training parameters, and overall prediction process remain consistent.

[0064] By using language models to extract semantic information and fuse it with numerical information, rather than relying on a specific model or fusion algorithm, its universality allows it to be compatible with a variety of different language models and fusion strategies.

[0065] Step S103: Baseline prediction is performed using the first model. The baseline prediction sub-model 231 in the two-stage prediction module 23 is activated. The model receives the semantically enhanced features generated in step S102 and the program features representing different dietary intervention programs (e.g., program IDs after one-heat encoding) as input.

[0066] The baseline prediction sub-model 231 employs a lightweight multilayer perceptron network.

[0067] The multilayer perceptron network structure is as follows: an input layer with 11 nodes (corresponding to 11-dimensional semantic enhancement features), followed by a hidden layer with 64 nodes using the ReLU activation function, then a Dropout layer that randomly deactivates some neurons with a probability of 0.2 to prevent overfitting, followed by a hidden layer with 32 nodes and the ReLU activation function, and finally an output layer that outputs a single value.

[0068] This output value is the model's prediction of the expected reduction in the user's visceral fat area at the end of a preset intervention period (e.g., 4 weeks) (unit: square centimeters).

[0069] The training process employs a leave-one-out cross-validation strategy to make full use of the limited clinical data. Within each fold, the network weights are updated using the AdamW optimizer (learning rate set to 1e-3, weight decay to 1e-4) and the mean squared error loss function.

[0070] For a new user, the semantically enhanced baseline features are combined with each dietary plan (e.g., 7 plans) in the plan library and then input into the trained baseline prediction sub-model 231 in batches to obtain the first prediction result for the user under each plan, i.e. the expected reduction in visceral fat area.

[0071] Step S104: Recommend initial plan. The adaptive decision support module 24 recommends an optimal initial intervention plan for the user based on the multiple prediction results obtained in step S103. The recommendation logic can select the plan with the largest predicted value (i.e., the greatest reduction) based on the predicted reduction in visceral fat area.

[0072] To increase the reliability of recommendations, the concept of confidence intervals is introduced. For example, the system calculates a 95% confidence interval for each predicted value. If the overlap rate between the confidence intervals of the best-performing and second-best-performing solutions is less than a threshold (e.g., 20%), the system will strongly recommend the best solution. If the overlap rate is high, both solutions may be offered to the user as preferred options.

[0073] Step S105: Collect early response data. After the user accepts the recommended initial plan and executes it for a period of time, early physiological response data of the user is collected at preset intermediate time points.

[0074] In more preferred examples, the preset intervention period is 4 weeks (28 days), with the preset midpoint being day 14 after the start of the intervention (allowing for a fluctuation of ±1 day). At this time, the user needs to measure their visceral fat area again using the same body composition analysis device. The data acquisition module 21 receives this data and calculates its change relative to the baseline value, denoted as ΔVFA_week2.

[0075] This change is a key dynamic indicator reflecting an individual's early response to the current intervention program.

[0076] Step S106: Dynamically correct the prediction using the second model. The early response prediction sub-model 232 in the two-stage prediction module 23 is activated, with the goal of using newly acquired early response data to make a more accurate corrected prediction of the final intervention effect.

[0077] Construct a new feature vector. This vector consists of the user's baseline physiological data (which can be raw numerical values ​​or semantically enhanced features), the ID of the currently implemented dietary intervention program, and the key early physiological response data ΔVFA_week2.

[0078] This new feature vector, incorporating dynamic information, is input into the early response prediction sub-model 232. This model can employ a similar structure to the baseline prediction sub-model 231 (e.g., another independently trained lightweight multilayer perceptron), but its input layer node number needs to be adjusted according to the new feature vector dimension. This model is specifically trained to learn how to predict the final 4-week visceral fat area reduction from a combination of baseline state, intervention protocol, and early feedback. Its output is a second prediction, a revised prediction of the visceral fat area reduction.

[0079] In more preferred examples, after incorporating early response data, the model's predictive determination coefficient (R²) can be significantly improved from 0.433 to 0.563 from the baseline prediction, demonstrating the effectiveness of dynamic correction.

[0080] The two-stage architecture of “baseline recommendation + dynamic optimization” can be combined with various machine learning models (such as neural networks and ensemble learning models). In more optimal examples, the baseline prediction sub-model 231 and the early response prediction sub-model 232 are gradient boosting decision tree models (such as XGBoost models).

[0081] In more preferred examples, the input to the early response prediction sub-model 232 is a richer feature vector containing multidimensional early response data: [baseline features, diet plan ID, ΔVFA_week3, ΔWeight_week3, ΔFPG_week3].

[0082] Step S107: Determine whether the prediction result meets the target. The adaptive decision support module 24 compares the second prediction result (corrected reduction in visceral fat area) obtained in step S106 with a preset clinically effective threshold.

[0083] The clinically effective threshold is set based on clinical guidelines or expert experience. For example, a reduction of visceral fat area of ​​-5.0 square centimeters is considered to have significant clinical significance.

[0084] Step S108: Simulate and recommend alternative solutions.

[0085] If the judgment result of step S107 is "no", that is, the corrected prediction result fails to reach the preset clinically effective threshold, it indicates that the currently implemented intervention plan may not be effective for the user. At this time, the adaptive decision support module 24 will start an automatic simulation process.

[0086] The user's baseline physiological data and the early physiological response data ΔVFA_week2 obtained were combined with other alternative dietary interventions (e.g., six other options in the protocol library besides the current one) and the early response prediction submodel 232 (or a similar simulation-specific model) were used to simulate and predict the amount of visceral fat area reduction that the user might achieve by the end of week 4 if they switched to these alternatives from now on.

[0087] After obtaining the simulation prediction results of all alternative options, one or more options with the best expected effect are selected and recommended to the user as alternative intervention options.

[0088] In more preferred examples, when recommending alternatives, in addition to considering the predicted effect, a "compliance difficulty" score is also introduced. When multiple alternatives have similar expected effects, the option with lower compliance difficulty is given priority.

[0089] Step S109: It is recommended to continue with the current plan.

[0090] If the judgment result of step S107 is "yes", that is, the corrected prediction result has reached the preset threshold, it means that the current solution is effective for the user. At this time, the system will send positive feedback to the user, suggesting that they continue to implement the current solution until the end of the 4-week intervention period.

[0091] Figure 5 The timing of signaling interactions between various entities in the system in this embodiment is illustrated. The user submits baseline data, the system performs baseline prediction and returns an initial recommendation; in the middle of the intervention, the user submits early data, the system performs dynamic correction prediction, and returns adjustment suggestions based on the results.

[0092] This two-stage architecture of "baseline recommendation + dynamic optimization" enables precise and closed-loop management of individualized interventions. At the same time, the dynamic correction framework can flexibly adjust the timing and content of data collection and incorporate multi-dimensional considerations into decision-making to adapt to diverse clinical application scenarios.

[0093] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function as logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0094] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for predicting visceral fat based on language model enhancement, characterized in that, include: Obtain the user's baseline physiological data and at least one dietary intervention to be evaluated; Based on the baseline physiological data, semantically enhanced features are generated; The semantic vector is fused with the numerical data corresponding to the feature to obtain the enhanced feature; The enhanced features are combined with other features in the baseline physiological data that do not generate semantic vectors to obtain the semantic enhancement features; The semantic enhancement features and the scheme features characterizing the at least one dietary intervention scheme are used as input to the first model. Through the first prediction model, a first prediction result is used to predict the change in visceral fat of the user at the end of a preset intervention period under the at least one dietary intervention scheme. At a predetermined midpoint within the predetermined intervention period, the user's early physiological response data is acquired. Combining the baseline physiological data and the early physiological response data, a second prediction result is generated by the second prediction model to show the change in visceral fat of the user at the end of the preset intervention period; Based on the second prediction result, the second prediction result is compared with a preset threshold to determine whether the intervention plan needs to be adjusted, and personalized intervention suggestions are generated accordingly. The generation of semantically enhanced features includes generating a semantic vector representing the semantic information of at least one feature in the baseline physiological data through a pre-trained language model.

2. The visceral fat prediction method based on language model enhancement according to claim 1, characterized in that, The generated personalized intervention recommendations include: If the second prediction result does not reach the preset threshold, the first prediction model is used to simulate the prediction results of the other alternative dietary interventions by taking the user's semantic enhancement features and the scheme features representing other alternative dietary intervention schemes extracted from a preset scheme library as input, and recommending an alternative intervention scheme with the highest predicted value for visceral fat reduction.

3. The visceral fat prediction method based on language model enhancement according to claim 1, characterized in that, The generated semantic enhancement features include: Prior to the fusion, the semantic vector is subjected to dimensionality reduction processing.

4. The visceral fat prediction method based on language model enhancement according to claim 3, characterized in that, The fusion includes: The reduced semantic vector is weighted and fused with the numerical data corresponding to the feature, multiplied element-wise, or concatenated. The weights used in the weighted fusion are preset weights or weights obtained through model training.

5. The visceral fat prediction method based on language model enhancement according to claim 1, characterized in that, The first prediction model and / or the second prediction model is a multilayer perceptron network or a gradient boosting decision tree model.

6. The visceral fat prediction method based on language model enhancement according to claim 1, characterized in that, The early physiological response data includes the change in visceral fat area of ​​the user at a preset intermediate time point.

7. The visceral fat prediction method based on language model enhancement according to claim 6, characterized in that, The early physiological response data also includes the user's weight change and / or fasting blood glucose change.

8. The visceral fat prediction method based on language model enhancement according to claim 1, characterized in that, The preset intermediate time point is the 14th day after the intervention begins; The preset intervention period is 4 weeks.

9. A visceral fat prediction system based on language model enhancement, characterized in that, include: At least one processor and a memory communicating with said at least one processor; The memory stores instructions; When the instruction is executed by the at least one processor, the system performs the visceral fat prediction method based on language model enhancement as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the visceral fat prediction method based on language model enhancement as described in any one of claims 1 to 8.

Images