Obesity intervention program generation method and system based on multi-modal large model

Abstract

The application discloses a kind of obesity intervention scheme generation method and system based on multimodal large model, and relates to a kind of obesity intervention scheme generation method and system based on multimodal large model, method includes: collection and to the multimodal data of user is preprocessed;Large language model is adjusted to pre-training, and personalized obesity large model is obtained;The multimodal data after pre-processing is input into personalized obesity large model, and obtains causal association mining result, risk prediction result and attribution analysis result;According to these results, generate personalized obesity intervention scheme.The application generates intervention scheme based on causal association mining result, risk prediction result and attribution analysis result through multimodal data preprocessing and personalized obesity large model, including nutrition plan, exercise prescription, behavior intervention plan and psychological intervention plan, improve precision and user compliance.

Description

Technical Field

[0001] This invention relates to the field of obesity management technology, and in particular to a method and system for generating obesity intervention programs based on a multimodal large model. Background Technology

[0002] The global obesity rate continues to rise, and obesity has become a major risk factor for chronic diseases such as cardiovascular and cerebrovascular diseases, diabetes, and various cancers, placing a heavy burden on public health systems. A key technical challenge in the current field of obesity intervention lies in how to generate truly personalized intervention plans. Existing methods are insufficient in collecting and preprocessing users' multimodal data, making it difficult to effectively integrate multidimensional information such as genomic, physiological, behavioral, psychological, and contextual data. This hinders the construction of accurate personal health profiles and makes it impossible to achieve precise causal relationship mining, risk prediction, and attribution analysis based on multimodal data.

[0003] To address the aforementioned issues, existing technologies primarily employ the following approaches: Traditional weight-loss applications acquire basic user information through simple data collection methods and generate generalized diet and exercise recommendations based on preset rules. Some digital health management systems utilize basic data preprocessing methods to clean and standardize the collected user data. Some systems use traditional machine learning models to adjust pre-trained large language models, attempting to build personalized models, such as using transfer learning methods and instruction sets built based on obesity management cases to fine-tune the basic model. In terms of data analysis, existing technologies mainly obtain causal association mining results through statistical methods, risk prediction results through time series analysis, and attribution analysis results through feature importance analysis. These methods typically employ simple linear models when processing multimodal data.

[0004] However, these existing technologies have significant drawbacks. In the data preprocessing stage, current methods are insufficient in cleaning, standardizing, and aligning multimodal data, resulting in limited data quality. During model building, the methods for adjusting large pre-trained language models are relatively simple and fail to fully capture individual characteristics. After inputting preprocessed multimodal data into the model, the accuracy of causal association mining results is limited, the reliability of risk prediction results is insufficient, and the depth of attribution analysis results is inadequate. When generating personalized obesity intervention plans based on the analysis results, existing technologies often only provide generalized suggestions and cannot achieve truly personalized nutrition plans, exercise prescriptions, behavioral intervention plans, and psychological intervention plans. These limitations severely affect the accuracy and effectiveness of obesity interventions.

[0005] In summary, existing technologies have significant shortcomings in multimodal data preprocessing, personalized model construction, acquisition of in-depth analysis results, and generation of intervention plans. There is an urgent need for an innovative solution that can fully realize the entire process from data acquisition to plan generation. This invention is proposed against this technological background, achieving comprehensive optimization and improvement of obesity intervention methods through systematic technological innovation. Summary of the Invention

[0006] The technical problem to be solved by this invention is to address the shortcomings of existing technologies, specifically by providing a method and system for generating obesity intervention programs based on a multimodal large model, as detailed below: 1) In a first aspect, the present invention provides a method for generating obesity intervention programs based on a multimodal large model, the specific technical solution of which is as follows: Collect and preprocess the user's multimodal data; By adjusting the pre-trained large language model, a personalized obesity model can be obtained. The preprocessed multimodal data is input into a personalized obesity model to obtain causal association mining results, risk prediction results, and attribution analysis results. Based on the results of causal association mining, risk prediction, and attribution analysis, a personalized obesity intervention plan is generated.

[0007] The beneficial effects of the obesity intervention program generation method based on a multimodal large model provided by this invention are as follows: By collecting and preprocessing multimodal data from users, this approach effectively integrates multidimensional information such as genomic, physiological, behavioral, psychological, and contextual data, ensuring data quality and consistency and providing a reliable foundation for subsequent analysis. Adjustments are made to a pre-trained large-scale language model to obtain a personalized obesity model that fully utilizes vast amounts of medical knowledge and adapts to individual characteristics, enhancing the model's professionalism and accuracy in obesity management. After inputting the preprocessed multimodal data into the personalized obesity model, it can efficiently perform data fusion and inference, obtaining accurate causal association mining results, reliable risk prediction results, and in-depth attribution analysis results, thereby identifying the root causes and future trends of individual obesity. Based on the causal association mining results, risk prediction results, and attribution analysis results, personalized obesity intervention plans are generated, including nutrition plans, exercise prescriptions, behavioral intervention plans, and psychological intervention plans, achieving highly customized and executable intervention strategies. The entire technical solution forms a systematic process from data collection to plan generation, overcoming the data silos and generalization limitations of existing technologies, improving the accuracy and practicality of interventions, and helping to improve user compliance and long-term success rates.

[0008] Based on the above scheme, the method for generating an obesity intervention scheme based on a multimodal large model of the present invention can be further improved as follows.

[0009] Furthermore, the pre-trained large-scale language model is adjusted to obtain a personalized obesity model, including: The preprocessed multimodal data is encoded to obtain vector data representing the preprocessed multimodal data. The vector data is used as personalized contextual information, and a pre-trained large language model is trained using an instruction set built based on obesity management cases to obtain a personalized obesity model.

[0010] The beneficial effects of adopting the above-mentioned further approach are as follows: Currently, in the field of obesity intervention, existing technologies for adjusting pre-trained large-scale language models are relatively simple and struggle to fully integrate individual characteristics from multimodal data, resulting in intervention plans generated by the model being generalized and lacking precision. This invention encodes pre-processed multimodal data to obtain vector data representing the pre-processed multimodal data, and uses this vector data as personalized contextual information. Simultaneously, it trains the pre-trained large-scale language model using an instruction set constructed based on obesity management cases, resulting in a personalized obesity model. This method effectively overcomes the problems of data silos and insufficient model adaptability. The technical effects are: the vector data encoding process transforms multi-dimensional information such as genomic data, physiological data, behavioral data, psychological data, and contextual data into a unified numerical representation, which is embedded as personalized contextual information into the model input, ensuring that inference can dynamically reference the user's specific background; the instruction set training based on obesity management cases enables the model to learn real-world intervention patterns, improving its professional understanding of the obesity management field. Ultimately, the personalized obesity model can more accurately capture individual differences and generate highly customized intervention plans, thereby improving the applicability and user compliance of the plans.

[0011] Furthermore, it also includes: Monitor users' implementation of personalized obesity intervention programs and feedback data, and optimize the personalized obesity model based on the feedback data.

[0012] The beneficial effects of adopting the above-mentioned further solutions are as follows: A prominent problem in the current field of obesity intervention is that existing solutions are mostly static plans, unable to be dynamically adjusted based on the user's actual implementation and feedback. Existing technologies lack continuous monitoring of the implementation of personalized obesity intervention programs and fail to effectively utilize feedback data to optimize models, resulting in intervention programs that are difficult to adapt to users' changing needs and states. This invention effectively solves the above problems by monitoring users' implementation of personalized obesity intervention programs and feedback data, and optimizing the personalized obesity model based on the feedback data. The specific technical effects are reflected in the following ways: The system continuously tracks the user's adherence to nutrition plans, exercise prescriptions, behavioral intervention plans, and psychological intervention plans, collecting objective data such as diet check-ins, exercise completion rates, and weight changes, while also acquiring subjective feedback data such as user satisfaction and difficulties in implementation. This multi-dimensional data is systematically used to optimize the personalized obesity model, continuously adjusting model parameters through incremental learning to enable the model to learn the user's latest behavioral patterns and physiological responses. Continuous optimization based on feedback data enables the personalized obesity model to generate intervention plans that are more tailored to the user's current state and needs, improving the practicality and operability of the plans. This dynamic optimization process forms a virtuous cycle of learning, allowing the intervention plan to adaptively adjust as the user's state changes, thereby improving the user's long-term compliance and intervention success rate.

[0013] Furthermore, multimodal data includes static and dynamic data. Static data includes genomic data, medical history, and anthropometry data, while dynamic data includes physiological data, behavioral data, psychological data, and contextual data. Personalized obesity intervention programs include nutrition plans, exercise prescriptions, behavioral intervention plans, and psychological intervention plans.

[0014] The beneficial effects of adopting the above-mentioned further solutions are as follows: Current obesity intervention methods suffer from technical problems such as limited data dimensions and one-sided intervention measures. Existing methods often focus only on certain data types, making it difficult to form a comprehensive health profile, and the generated intervention plans lack systematicity. This invention effectively solves these problems by clearly defining multimodal data as including static and dynamic data. Static data encompasses genomic data, medical history, and anthropometry data, while dynamic data includes physiological, behavioral, psychological, and contextual data. Furthermore, it clarifies that personalized obesity intervention plans include nutrition plans, exercise prescriptions, behavioral intervention plans, and psychological intervention plans. The technological benefits are reflected in the following aspects: the complete definition of multimodal data ensures comprehensive information collection from genetic characteristics to real-time environmental factors, laying the foundation for building accurate personal health profiles; static data provides stable references for individual characteristics, while dynamic data reflects real-time status changes, and the combination of the two forms a three-dimensional health assessment system; the comprehensive coverage of intervention programs ensures multi-dimensional coordinated intervention from nutritional intake, physical activity to behavioral habits and psychological state; this structured data definition and program design breaks through the limitations of traditional single-dimensional intervention, enabling the generated intervention programs to consider both inherent individual characteristics and dynamic changing factors, greatly improving the scientific nature and completeness of the programs, and providing systematic support for effectively achieving personalized obesity management.

[0015] 2) In a second aspect, the present invention also provides a system for generating obesity intervention plans based on a multimodal large model, the specific technical solution of which is as follows: It includes: a data acquisition and preprocessing module, a model adjustment module, a model application module, and an intervention plan generation module; The data acquisition and preprocessing module is used to: acquire and preprocess the user's multimodal data; The model tuning module is used to tune a pre-trained large language model to obtain a personalized obesity model. The model application module is used to: input preprocessed multimodal data into a personalized obesity model to obtain causal association mining results, risk prediction results, and attribution analysis results; The intervention plan generation module is used to generate personalized obesity intervention plans based on the results of causal association mining, risk prediction, and attribution analysis.

[0016] Based on the above scheme, the obesity intervention scheme generation system based on a multimodal large model of the present invention can be further improved as follows.

[0017] Furthermore, the model adjustment module is specifically used to: encode the preprocessed multimodal data to obtain vector data representing the preprocessed multimodal data, use the vector data as personalized contextual information, and simultaneously train the pre-trained large language model using an instruction set built based on obesity management cases to obtain a personalized obesity model.

[0018] Furthermore, it also includes a model optimization module, which is used to monitor the user's implementation of the personalized obesity intervention plan and feedback data, and optimize the personalized obesity model based on the feedback data.

[0019] Furthermore, multimodal data includes static and dynamic data. Static data includes genomic data, medical history, and anthropometry data, while dynamic data includes physiological data, behavioral data, psychological data, and contextual data. Personalized obesity intervention programs include nutrition plans, exercise prescriptions, behavioral intervention plans, and psychological intervention plans.

[0020] 3) In a third aspect, the present invention also provides an electronic device, the electronic device including a processor coupled to a memory, the memory storing at least one computer program, the at least one computer program being loaded and executed by the processor, so that the electronic device implements any of the above-mentioned methods for generating obesity intervention programs based on a multimodal large model.

[0021] 4) In a fourth aspect, the present invention also provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements any of the above-mentioned methods for generating obesity intervention programs based on a multimodal large model.

[0022] It should be noted that the beneficial effects of the technical solutions of the second to fourth aspects of the present invention and their corresponding possible implementations can be found in the above description of the technical effects of the first aspect and its corresponding possible implementations, and will not be repeated here. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments of the present invention will be briefly introduced below: Figure 1 This is a flowchart illustrating a method for generating obesity intervention plans based on a multimodal large model according to an embodiment of the present invention. Figure 2 This is a schematic diagram of a multimodal large model-based obesity intervention scheme generation system according to an embodiment of the present invention. Detailed Implementation

[0024] The principles and features of the present invention are described below. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0025] The technical solution of the present invention and how the technical solution of the present invention solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present invention will now be described with reference to the accompanying drawings.

[0026] like Figure 1 As shown in the figure, an embodiment of the present invention provides a method for generating an obesity intervention program based on a multimodal large model, which includes the following steps: S1. Collect and preprocess the user's multimodal data; where multimodal data includes static data and dynamic data, static data includes genomic data, medical history, and anthropometry data, and dynamic data includes physiological data, behavioral data, psychological data, and contextual data; The data acquisition phase is achieved through various technological channels and devices. Genomic data is obtained from raw gene sequence files through application programming interfaces (APIs) provided by commercial genetic testing services; these files are typically stored in FASTQ or VCF formats. Medical history data is extracted from medical institutions' electronic health record systems via standard medical data exchange protocols such as HL7 or FHIR, supplemented by digital health questionnaires completed by users, covering medical history, surgical records, medication use, and family medical history. Anthropometry data is automatically uploaded using IoT devices such as smart scales and body fat scales, which transmit data via Bluetooth or Wi-Fi, supplemented by manually entered measurements such as waist and hip circumference. Physiological data is continuously recorded through sensors in wearable devices, including photoplethysmography sensors on smartwatches to monitor heart rate, electrode sensors on continuous glucose monitors to measure interstitial fluid glucose concentration, and accelerometers on sleep trackers to monitor sleep stages. Behavioral data is acquired through multimodal input from smartphone applications, including images of food taken by cameras, geolocation information recorded by GPS, and 3D acceleration data from motion sensors to record activity intensity. Psychological and contextual data are collected through a natural language user interface, including text input from an emotion diary, a slider rating of stress levels, and environmental parameters such as time and weather conditions inferred by the context-aware module from device sensors.

[0027] The data cleaning stage primarily addresses missing values, outliers, and noise in the data. For missing gene loci in the genomic data, population frequency imputation is used. Specifically, the allele frequency of that locus in the reference population is first calculated, and then... Indicates the frequency of minor alleles; missing values ​​are expressed as probabilities. Perform random interpolation. For outliers in physiological data, such as sudden extremely high or low values ​​in heart rate data, use a sliding window detection algorithm to identify outliers. Define time series data points. in Indicates a time index. Indicates time Physiological data values. First, calculate the mean and standard deviation of the data within the window, let... This represents the mean of the data within the window. This represents the standard deviation of the data within the window. If... ,but These are marked as outliers. Outlier correction employs a time-series-based linear interpolation method. Let... This represents the value at the previous time point. This represents the value at a later point in time, the correction value. Calculated as For image noise in behavioral data, digital image processing techniques are used for noise reduction, such as applying a Gaussian filter. The filter kernel function is expressed as follows: ,in and Represents pixel coordinates, This represents the standard deviation parameter. For text input in psychological data, natural language processing techniques are used to remove stop words and punctuation marks, while retaining key sentiment words.

[0028] The data standardization phase transforms data from different sources and units into a unified format and scale. For numerical data such as anthropometric data, a min-max normalization method is used to map the original values ​​to a range of zero to one. Let... Represents the original value of a feature. This represents the minimum value of the feature in the training dataset. This represents the maximum value of the feature in the training dataset, after normalization. Calculated as For categorical data, such as disease classifications in medical history, one-hot encoding is used to convert them into binary vector representations. For example, if there are three disease types, type one is encoded as vector [1,0,0], type two as [0,1,0], and type three as [0,0,1]. For textual psychological data, word segmentation is performed first, and then word embedding techniques such as Word2Vec are used to map words to a high-dimensional vector space. Let... To represent a word, The word vectors are represented by the vector average method to obtain the overall representation vector for a text. ,in Indicates the number of words. Indicates the first For time-series data such as physiological data, resampling is also performed to standardize the sampling frequency, for example, unifying data from different devices to one sampling point per minute.

[0029] The data alignment phase integrates all data into a unified timeframe. A personal health time-series database is established based on user unique identifiers and device timestamps. For data with different sampling frequencies, interpolation methods are used for time alignment. For example, for high-frequency physiological data collected at one sampling point per minute and low-frequency anthropometric data collected at one sampling point per week, time point matching is performed on the low-frequency data during time alignment. Indicates the target time point. and Indicates adjacent actual sampling time points, and The corresponding values ​​are respectively and Calculate using linear interpolation methods. The value of time For missing time points, null values ​​are retained and handled in subsequent analysis. The resulting personal health time-series database after data alignment contains complete time-series records, each containing fields such as timestamp, user identifier, data type, and data value.

[0030] The entire preprocessing process also includes a data quality verification step, which uses statistical tests to evaluate the distribution characteristics and consistency of the preprocessed data. For example, the Kolmogorov-Smirnov test is used to compare the difference between the data distribution and the expected distribution to ensure that the data meets the requirements of subsequent analysis. The preprocessing module generates a detailed data quality report, recording the parameters and results of each processing step, ensuring the traceability and repeatability of the preprocessing process.

[0031] Genomic data refers to the complete set of genetic information extracted from an individual's deoxyribonucleic acid (DNA) sequence, including genotype data, single nucleotide polymorphisms (SNPs), copy number variations, and epigenetic markers. This data, obtained through high-throughput sequencing or microarray analysis, provides information on genetic factors related to the development of obesity, such as variations in genes regulating adipocyte differentiation, changes in the expression levels of genes related to energy metabolism pathways, and genetic polymorphisms of appetite-regulating hormone receptors. Genomic data is typically stored in variant call format files or genotype matrices and requires specialized bioinformatics workflows for quality control, annotation, and interpretation.

[0032] Medical history refers to a detailed record of all health events and medical interventions an individual has experienced throughout their life, including a list of diagnosed diseases, surgical treatment history, long-term medication records, allergy information, family history of genetic diseases, and routine physical examination results. Medical history data primarily originates from the electronic health record system of medical institutions and is transmitted through standardized medical data exchange protocols such as HL7 or FHIR, supplemented and improved by structured health questionnaires completed by the patient. Medical history information can provide important clues about obesity-related comorbidities and risk factors, such as diagnoses of endocrine and metabolic diseases, a history of cardiovascular events, and the impact of mental health status on weight management.

[0033] Anthropometry data refers to quantitative indicators of an individual's body shape and composition obtained through standardized measurement methods. These include parameters such as height, weight, body fat percentage, waist circumference, hip circumference, neck circumference, limb circumference, and skinfold thickness. This data is acquired using calibrated measuring equipment, such as digital scales to measure weight, bioelectrical impedance analysis to measure body fat composition, measuring tapes to measure circumference, and calipers to measure skinfold thickness. Anthropometry data provides direct evidence for assessing obesity levels and body fat distribution patterns, and is the foundation for calculating body mass index, waist-to-hip ratio, and body fat distribution type.

[0034] Physiological data refers to real-time biosignal measurements reflecting an individual's internal organ function and metabolic state. These include parameters such as resting heart rate, ambulatory blood pressure, blood oxygen saturation, core body temperature, respiratory rate, heart rate variability, skin conductance, and continuous glucose monitoring. This data is continuously collected using medical-grade or consumer-grade wearable biosensors, such as photoplethysmography sensors to monitor blood flow changes, electrode sensors to measure electrophysiological signals, and chemical sensors to detect interstitial fluid glucose concentration. Physiological data can reveal the dynamic characteristics of energy metabolism, the homeostasis of the autonomic nervous system, and stress response patterns, providing crucial information for understanding the physiological mechanisms of weight regulation.

[0035] Behavioral data refers to objective observational information recording an individual's daily life activities and habit patterns. Specifically, it includes the type and quantity of food intake, the type and intensity of physical activity, the duration and frequency of sedentary behavior, the cycle and quality of sleep and wakefulness, and the trajectory and patterns of geographical movement. This data is collected through multiple channels, including smartphone built-in sensors, wearable fitness trackers, diet tracking applications, and environmental sensing devices. For example, accelerometers measure exercise intensity, GPS records location changes, and image recognition algorithms analyze food composition. Behavioral data can establish behavioral patterns of energy intake and expenditure, identifying high-risk behavioral habits and lifestyle factors for weight gain.

[0036] Psychological data refers to self-reported or objectively assessed information reflecting an individual's internal emotional state, cognitive processes, and psychological characteristics. Specifically, it includes dimensions such as perceived stress level, level of emotional pleasure, intensity of food cravings, body satisfaction, self-efficacy, and motivation level. This data is acquired through digital psychological assessment scales, ecological instantaneous assessment methods, natural language interfaces, and emotion recognition technologies. For example, Likert scales measure subjective feelings, text sentiment analysis identifies emotional tendencies, and voice feature analysis infers psychological states. Psychological data can reveal the intrinsic links between psychological factors and eating behavior, exercise adherence, and weight changes, providing an important perspective for understanding the psychological and behavioral mechanisms of obesity.

[0037] Contextual data refers to multi-dimensional information describing the characteristics and background conditions of an individual's external environment, specifically including time-cycle characteristics, spatial location attributes, social environment composition, weather and climate conditions, food accessibility, and work-life arrangements. This data is automatically collected through smartphone sensors, calendar applications, social media connections, and environmental monitoring services; for example, clock functions record time information, GPS determines geographical location, and social graph analysis reflects interpersonal networks. Contextual data can provide environmental context information about the occurrence of behaviors and identify the moderating effects of external environmental factors on dietary choices, physical activity levels, and weight management effectiveness.

[0038] S2. Adjust the pre-trained large language model to obtain a personalized obesity model. Specifically, encode the pre-processed multimodal data to obtain vector data representing the pre-processed multimodal data. Use the vector data as personalized contextual information and simultaneously train the pre-trained large language model using an instruction set built based on obesity management cases to obtain a personalized obesity model. Pre-trained large-scale language models refer to language models pre-trained on large-scale text datasets through self-supervised learning, possessing powerful natural language understanding and generation capabilities. Based on a deep neural network architecture, these models learn the statistical patterns and semantic knowledge of language from massive amounts of text, enabling them to handle various language tasks. In personalized obesity intervention scenarios, pre-trained large-scale language models serve as the foundation, fine-tuned to adapt to the professional needs of obesity management, becoming the core engine of personalized obesity models. For example, the GPT-4 model can be chosen, possessing powerful generation capabilities and extensive medical knowledge; the BERT model can be selected, which performs excellently on language understanding tasks; the T5 model can also be chosen, with its unified text-to-text framework suitable for various generation tasks; the Bloom model, with its multilingual capabilities, is suitable for international application scenarios; and models specifically trained for the biomedical field, such as ClinicalBERT, have a better understanding of medical terminology and concepts.

[0039] The specific process for obtaining vector data is as follows: For genomic data, the encoding process employs a site embedding method. Genomic data contains gene sequences and variation information. First, each single nucleotide polymorphism (SNP) site is represented as a discrete categorical variable, and then mapped to a continuous vector space through an embedding layer. Specifically, let... Indicates the first The original coding values ​​of each single nucleotide polymorphism site, among which The value range is from 1 to , This represents the total number of single nucleotide polymorphism sites in the genomic data. It is achieved through a trainable embedding matrix. , each Convert to vector The formula is ,in, Indicates to One-hot encoding operation. Aggregate representation of all site vectors. Obtained through average pooling, i.e. This process compresses high-dimensional, sparse genomic data into low-dimensional, dense vectors.

[0040] For medical history data, the encoding process combines category embedding and text embedding techniques. Medical history data includes structured information such as disease diagnoses, surgical records, and medication histories. First, each medical history entry is converted into a standardized code, for example, using the International Classification of Diseases (ICD) code. Indicates the first The code value of each medical history entry, among which The value range is from 1 to , This represents the total number of medical history entries. (Used via an embedding matrix.) Each Mapped to vector The formula is For text-based descriptions of medical history, pre-trained language models such as BERT are used to generate sentence-level embeddings. Let... Indicates the first The text description is processed by the BERT model to obtain vectors. Overall medical history data vector The weights are calculated by taking a weighted average of all item vectors and text vectors, with the weights assigned based on the clinical importance of the items.

[0041] For anthropometry data, a multilayer perceptron was used for feature extraction during the encoding process. Anthropometry data includes numerical indicators such as height, weight, and body fat percentage, which have already been normalized. Indicates the first Values ​​of individual anthropometric characteristics, among which The value range is from 1 to , This represents the number of anthropometric features. All features constitute the input vector. Through a fully connected neural network layer Mapping to vector The formula is ,in Represents the weight matrix. This represents the bias vector. This indicates an activation function such as ReLU. This encoding process captures the nonlinear relationships between anthropometric features.

[0042] For physiological data, the encoding process uses a time-series model to process time-series information. Physiological data includes signals that change over time, such as heart rate and blood sugar levels, and this data has been aligned to a unified timestamp. Let... Indicates a point in time Physiological data vectors, where The value range is from 1 to , This represents the length of the time series. The entire sequence is processed using a Long Short-Term Memory (LSTM) network model, including the hidden states. The updated formula is Final physiological data vector Take the hidden state at the last time step, i.e. This method effectively captures long-term dependencies and dynamic patterns in physiological data.

[0043] For behavioral data, the encoding process integrates multimodal features from image, text, and numerical data. Behavioral data includes images of food consumption, motion records, and geolocation information. For food consumption images, convolutional neural networks are used to extract visual features. The input image is processed through convolutional and pooling layers to obtain feature maps, which are then converted into vectors by global average pooling. For motion recording numerical data, let Indicates the first Each activity metric, such as steps, is represented by a vector. Mapped to through a fully connected layer For geolocation text data, a word embedding model is used to convert location descriptions into vectors. Overall behavioral data vector It is obtained by concatenating all sub-vectors and then performing a dimensionality reduction layer, i.e. .

[0044] For both psychological and situational data, the encoding process employs embedding and aggregation strategies. Psychological data includes self-reported information such as emotional state and stress levels; each psychological assessment item is first converted into a numerical score. Indicates the first Scores for each psychological assessment item, among which The value range is from 1 to , This represents the number of psychological assessment items. Each score is mapped to a vector through an embedding layer. Contextual data includes contextual information such as time and environment, making... Indicates the first Each contextual element, such as hour or weather code, is used to obtain a vector through one-hot encoding and embedding. The overall vector of psychological data and situational data. Calculated by averaging the vectors of all items, i.e. ,in Indicates the number of contextual elements.

[0045] After independently encoding all data types, the vectors are integrated into a unified vector data. This integration process is achieved through concatenation operations, making... ,in This represents the final multimodal vector data. This vector data is stored in a vector database as personalized contextual information for real-time retrieval and matching during subsequent model inference. The entire encoding process ensures efficient data representation and semantic integrity, providing reliable input for deep analysis of personalized obesity models.

[0046] The specific implementation process for obtaining the personalized obesity model is as follows: In processing vector data as personalized contextual information, a pre-built vector database is used to store the user's multimodal vector data. The vector database employs an approximate nearest neighbor search algorithm, such as a hierarchical navigable small-world graph, to achieve efficient retrieval. When the model needs to process a new user query, it first retrieves the most relevant historical vector data for that user from the vector database. Let... This represents the vector representation of the current user's query. Represents the first in the vector database 1 stored vector, where The value range is from 1 to , This represents the total number of vectors in the vector database. Similarity is calculated using the cosine similarity function, with the formula: ,in, This represents the dot product operation. This represents the Euclidean norm of the vector. The top-K most similar vectors are selected, where K is a preset integer, e.g., K=5. These vectors form a personalized contextual information set. Then, the personalized contextual information is either converted into a text description or directly concatenated as a numerical vector into the model's input sequence. For the text description approach, a template is used to map each vector element to a natural language statement, such as "User genome data shows that specific gene variations are associated with metabolic rates." For the numerical vector approach, the personalized contextual information is directly concatenated as an additional input vector with the original query vector to form an enhanced input. This process ensures that the model can reference the user's unique health profile when generating a response.

[0047] When training a large pre-trained language model using an instruction set built based on obesity management cases, a large pre-trained language model is first loaded as a base, such as a model based on the Transformer architecture. This represents a pre-trained large language model with the following parameters: The instruction set contains multiple instruction samples, each consisting of input text and expected output text. The input text includes a description or vector representation of the user's multimodal data, as well as possible personalized contextual information; the expected output text is the corresponding obesity intervention plan. This represents a set of instructions built based on obesity management cases, where It is the total number of instruction samples. It is the first The input text for each sample, It is the first The expected output text for each sample. The training process employs supervised fine-tuning to minimize the difference between the model's output and the expected output. The loss function used is cross-entropy loss, formulated as follows: ,in It is the first The length of the output text for each sample. It is the first The first sample output text One token, Indicates the first The first sample output text One token, It is the probability distribution predicted by the model. These are the model parameters. During training, personalized contextual information is dynamically integrated into the input through a retrieval mechanism. In this process, the model learns to incorporate real-time retrieved user vector data when generating responses. The optimizer uses AdamW with a learning rate set to a decreasing schedule, for example, from... Initially, the loss per epoch is reduced to 0.95 times that of the previous epoch. Training is iterated for multiple epochs until the loss on the validation set no longer decreases significantly, thus obtaining a personalized obesity model. .

[0048] The entire implementation process also includes validation and optimization steps. The performance of the personalized obesity model is evaluated using a reserved test set, with metrics including the relevance and accuracy of the generated protocols. Through iterative adjustments to retrieval parameters and training hyperparameters, the model gradually adapts to the specific needs of obesity management, ultimately enabling the generation of highly personalized intervention protocols based on multimodal vector data.

[0049] The instruction set, built upon obesity management case studies, is a dataset specifically created for training a large-scale personalized obesity model. It contains numerous input-output pairs from real-world or simulated obesity management scenarios. Each input typically includes descriptions of various user data types, such as textual or vectorized representations of genomic data, medical history, anthropometry, physiological data, behavioral data, psychological data, and contextual data. The output is a detailed textual description of the corresponding obesity intervention plan, such as a nutrition plan, exercise prescription, behavioral intervention plan, and psychological intervention plan. This instruction set, built by collecting clinical records, expert consultations, and user feedback, ensures coverage of diverse obesity types and intervention strategies. It is used to supervise and fine-tune a large-scale pre-trained language model, enabling it to learn to generate scientific and personalized recommendations given user data.

[0050] S3. Input the preprocessed multimodal data into the personalized obesity model to obtain causal association mining results, risk prediction results, and attribution analysis results. Specifically: In generating the causal association mining results, the personalized obesity big data model first analyzes the time series data and association patterns in the input data. The model uses Transformer-based attention weights to identify potential causal relationships between variables. For example, for continuous glucose monitoring values ​​in physiological data and dietary records in behavioral data, the model calculates conditional probabilities to assess the impact of specific food types on blood glucose fluctuations. This indicates a potential causal variable such as high sugar food intake. For example, when the outcome variable is elevated blood sugar, the model calculates the probability. and To assess the strength of causality. Among them, Indicates the cause variable Outcome variables under the condition of occurrence The probability of occurrence Indicates the cause variable Outcome variables under conditions that do not occur The probability of occurrence. The output of causal association mining is a causal graph or text description, listing the identified causal pairs and their confidence scores; for example, the association strength of sleep deprivation leading to increased appetite the next day is 0.85. The model also uses counterfactual reasoning to simulate different scenarios to verify the robustness of causal relationships.

[0051] In generating risk prediction results, the personalized obesity big data model applies time series prediction and classification techniques. The model treats the input data as a feature vector, including historical weight trends, physiological indicators, and behavioral patterns. The feature vector is encoded from multimodal data. ,in The dimension of the feature vector. Indicates the first The model processes these features through fully connected layers and recurrent neural network layers to output future time points. The risk score is calculated. Risk prediction results include predicted weight change and the probability of metabolic complications. Weight change prediction uses a regression model, with the formula: ,in, This represents the predicted weight change. Represents the weight vector. This represents the bias term. The probability of metabolic complications is calculated using logistic regression, with the formula: ,in Indicates the probability of complications occurring. Represents the weight vector. This indicates the bias term. Risk prediction results are output in numerical and descriptive form, such as a 70% probability that a user will gain 2 kg in the next month, or a 15% risk of developing diabetes.

[0052] In generating the attribution analysis results, the personalized obesity big model focuses on identifying the root causes of poor current weight management outcomes. The model uses feature importance analysis and gradient-based methods to evaluate the contribution of each input variable to the output outcome, such as weight rebound. The loss function, such as weight prediction error, is calculated by the model for each feature. gradient To measure its importance, the attribution analysis results output a list of importance scores. For example, sedentary time in behavioral data contributes 40% to weight gain, while stress levels in psychological data contribute 30%. The model also integrates contextual information, such as user historical vector data retrieved by the personalized obesity big data model, to identify key factors by comparing feature changes over different time periods. The attribution analysis results are presented in a structured report, listing the main attributable factors and supporting evidence.

[0053] The entire implementation process relies on the end-to-end processing capabilities of the personalized obesity big data model. The obesity-related knowledge learned by the model during the training phase ensures the scientific rigor and personalization of the output. Finally, the causal relationship mining results, risk prediction results, and attribution analysis results are integrated into a comprehensive analysis report for use in generating subsequent intervention plans.

[0054] Causal association mining results refer to the ordered set of causal relationships between variables in multimodal data identified through statistical and machine learning methods. These results include a list of causal pairs, association strength indices, and statistical significance levels, used to reveal, for example, the unique effects of specific food types on individual blood glucose levels or the intrinsic link between sleep quality and appetite regulation. Causal association mining results are generated based on conditional probability and hypothesis testing, helping to pinpoint obesity-related driving factors.

[0055] The risk prediction results refer to quantitative estimates of the probability of future obesity-related events based on historical and multimodal data. These results include predictions of weight change trends, risk scores for metabolic complications such as diabetes or cardiovascular disease, and uncertainty intervals. The risk prediction results are calculated using time series models and classification algorithms, providing prospective insights to support early intervention strategies.

[0056] The attribution analysis results refer to a systematic analysis of the main reasons behind poor current weight management or weight rebound. These results include feature importance ranking, contribution scores, and key factor descriptions, used to explain the extent to which, for example, low diet adherence or insufficient exercise affects weight control. The attribution analysis results are derived through gradient analysis and contextual comparison, providing a basis for optimizing intervention programs.

[0057] S4. Based on the causal relationship mining results, risk prediction results, and attribution analysis results, generate a personalized obesity intervention plan, which includes a nutrition plan, exercise prescription, behavioral intervention plan, and psychological intervention plan.

[0058] In generating the nutrition plan, the personalized obesity big data model first calculates daily energy requirements based on weight change trends in risk prediction results and diet-related issues in attribution analysis. Energy requirements are based on basal metabolic rate (BMR) and physical activity level, with BMR calculated using the Mifflin-St Jeor formula: in, This represents the basal metabolic rate (unit: kcal / day). Indicates body weight (unit: kilograms). Height (unit: centimeters) Indicates age (unit: years). Represents a gender constant (male takes...). female selection Total energy demand Calculated using the formula: in, This indicates the total daily energy requirement (unit: kilocalories). This represents the physical activity level coefficient (determined based on the activity intensity in user behavior data; for example, 1.2 for sedentary behavior and 1.5 for moderate activity). This indicates the energy consumed during exercise (unit: kilocalories, estimated from exercise records).

[0059] Then, based on the food and metabolic responses identified in the causal relationship mining results, such as specific foods causing blood sugar fluctuations, nutrient allocation is adjusted. The nutrition plan is specified down to food types, portions, and meal times, for example, recommending low-glycemic index foods to replace high-glycemic index foods, and taking into account the user's dietary preferences and allergy history. The plan output includes daily meal examples, nutritional goals (such as the proportion of carbohydrates, protein, and fat), and meal schedules, ensuring that the plan aligns with individual metabolic characteristics and health goals.

[0060] In the generation of exercise prescriptions, a personalized obesity big data model designs safe and effective exercise programs based on the insufficient exercise problem identified in attribution analysis and the health risks in risk prediction results. Exercise intensity is planned using a percentage of maximum heart rate. The calculation formula is: in, This indicates the maximum heart rate (unit: beats / minute). Age (in years). Exercise intensity. Calculated using the formula: in, Indicates exercise intensity (percentage). This represents heart rate during exercise (in beats per minute). The model recommends exercise type, frequency, duration, and intensity based on the user's fitness level (inferred from anthropometric and physiological data) and health status (obtained from medical history). For example, low-impact exercises such as swimming are recommended for users with joint problems; high-intensity exercises are limited for users with cardiovascular risk. Exercise prescriptions also integrate activity patterns from behavioral data, such as combining exercise time with the user's schedule, to improve the feasibility of the plan. Outputs include specific exercise programs, weekly frequency, duration per session, and progress adjustment rules.

[0061] In generating behavioral intervention plans, the personalized obesity big data model utilizes specific behavioral problems (such as sedentary behavior or emotional eating) identified in attribution analysis and the behavioral-consequence relationships revealed in causal association mining to design behavioral modification strategies. The model is based on behavioral economics theory, setting behavioral goals and incentive mechanisms. For example, for sedentary behavior, it employs timed standing reminders with specific intervals. Calculated dynamically based on daily work patterns: in, Indicates the reminder interval (unit: minutes). This indicates the total working time (in minutes). This indicates the recommended number of rest periods (determined from attribution analysis results, e.g., once per hour). The behavioral intervention plan includes specific behavior substitution techniques (such as using walking instead of elevators), environmental adjustment suggestions (such as reducing the visibility of high-calorie foods), and habit formation methods (such as setting daily goals). The model references the user's willingness to perform from psychological data and patterns from historical behavioral data to ensure the plan is feasible. The output is presented in the form of a step-by-step list, including behavioral goals, implementation methods, and monitoring indicators.

[0062] In generating the psychological intervention plan, the personalized obesity model provides emotional support and cognitive adjustment programs based on psychological factors (such as stress or low motivation) identified in the attribution analysis and psychological risks predicted in the risk forecast. The model uses cognitive behavioral therapy principles to design interventions targeting stress management, emotion regulation, and motivation maintenance. For example, regarding emotional eating, the model generates mindfulness eating training tasks with varying frequency. Adjust dynamically based on pressure level: in, This indicates the weekly training frequency (unit: times / week). This represents a stress score (obtained from psychological data, ranging from 0 to 100). This represents a rounding function. The psychological intervention program integrates the user's contextual data, such as scheduling relaxation exercises during high-stress periods, to provide personalized coping strategies. Outputs include specific exercise descriptions (such as deep breathing techniques), emotion recording templates, and motivation-enhancing activities to ensure the program supports long-term adherence.

[0063] Ultimately, the personalized obesity big data model integrates nutrition plans, exercise prescriptions, behavioral intervention plans, and psychological intervention plans into a unified, comprehensive intervention program. The program output uses structured text and visual charts, including specific implementation steps, timelines, and quantifiable goals, and is presented to users through a human-computer interaction module to facilitate understanding and implementation.

[0064] A nutrition plan refers to a structured dietary arrangement based on an individual's energy needs and metabolic characteristics. This includes food selection, portion control, nutrient ratios, and meal timing. Based on energy balance and nutritional principles, a nutrition plan considers an individual's eating habits, metabolic response, and health goals, aiming to achieve weight control through scientific dietary management. A nutrition plan typically includes specific meal plan examples, food substitution suggestions, and nutrition education content.

[0065] An exercise prescription refers to a systematic physical activity program designed for an individual's health condition and fitness level, specifying in detail the type, intensity, frequency, duration, and schedule of exercise. Based on the principles of exercise physiology and medicine, and considering an individual's exercise capacity, health risks, and preferences, exercise prescriptions aim to promote energy expenditure and improve health through regular exercise. Exercise prescriptions include various forms such as aerobic exercise, strength training, and flexibility exercises.

[0066] Among them, the behavior intervention program refers to a set of behavior modification strategies designed based on behavioral science theory, aiming to help individuals change their daily behaviors and habits related to weight management. The behavior intervention program includes techniques such as goal setting, self-monitoring, environmental control, stimulus control, and reinforcement incentives, providing systematic solutions for specific problems such as eating behavior, physical activity behavior, and sedentary behavior.

[0067] The psychological intervention program refers to a systematic approach that applies psychological methods to improve an individual's mental state and cognitive patterns, focusing on addressing psychological barriers and emotional issues related to weight management. The program includes techniques such as cognitive restructuring, emotion regulation, stress management, motivation enhancement, and mindfulness training, aiming to improve an individual's self-efficacy and long-term commitment, providing psychological support for weight management.

[0068] Optionally, the above technical solution also includes: S5. Monitor users' implementation of the personalized obesity intervention plan and their feedback data, and optimize the personalized obesity model based on the feedback data. The specific implementation process is as follows: User behavior and physiological responses are tracked in real time through a human-computer interaction module and multi-source sensors. The implementation status of the personalized obesity intervention program is acquired through digital logs and automated sensor recordings; for example, users record their food intake and exercise completion status via a mobile application, while wearable devices monitor heart rate, sleep, and activity levels. Feedback data is collected through structured questionnaires and natural language input, such as user ratings of program difficulty, descriptions of their acceptance of suggestions, and self-reports of emotional states. The implementation data and feedback data are integrated into a time-series dataset and stored in a personal health database.

[0069] After data integration, quantitative indicators of performance are calculated, such as protocol compliance scores. The adherence score for a personalized obesity intervention program is calculated using the following formula: in, This indicates the total number of tasks in the plan. Indicates the first The completion status of each task (1 for completed, 0 for incomplete). Indicates the first The weights of each task are assigned based on its importance; for example, eating suggestions in a nutrition plan have a higher weight, while reminders in a behavior intervention plan have a lower weight. Feedback data is then converted into numerical or vector forms, such as user satisfaction ratings. Text feedback is directly recorded as integers from 0 to 100, and then converted into sentiment polarity scores through a sentiment analysis model. The value ranges from -1 to 1.

[0070] When optimizing the personalized obesity model, the collected performance data and feedback data are used as new training samples to incrementally fine-tune the model. This represents the current personalized obesity model, with the following parameters: Construct an optimized dataset. ,in Indicates the number of new samples. Indicates the first The input for each sample includes user multimodal data and performance metrics. Indicates the first The expected output for each sample is adjusted based on feedback data; for example, if users report that the nutrition plan is difficult to implement, then... The nutritional recommendations have been simplified. The optimization process minimizes the loss function. The formula is: in, The cross-entropy loss represents the prediction of task completion. This represents the mean squared error loss of the feedback prediction. The model represents the first The predicted output for each sample, This represents the user satisfaction predicted by the model. and This represents the weight hyperparameters used to balance the task loss and feedback loss. The optimization algorithm uses stochastic gradient descent, and the learning rate is set to... Iteratively update model parameters This continues until the loss function converges.

[0071] The performance is used as a reward signal to further adjust the model. For example, a reward function is defined. The model is updated using a policy gradient method to maximize the expected reward. The optimized personalized obesity model... Redeploy to generate more adaptive intervention plans. The entire process is performed periodically, such as weekly, to ensure the model continuously adapts to users' changing needs and behavioral patterns.

[0072] The implementation status of personalized obesity intervention programs refers to the quantitative record of the degree and accuracy with which users actually implement each task within the personalized obesity intervention program. These records include task completion status, execution time deviation, completion quality scores, and adherence indicators, such as the compliance rate of dietary recommendations in the nutrition plan, the achievement rate of exercise intensity in the exercise prescription, and the number of days of habit formation in the behavioral intervention plan. Implementation data is obtained through automatic sensing and user manual logs, used to evaluate the actual effectiveness of the program and user participation, providing an objective basis for model optimization.

[0073] Feedback data refers to users' subjective evaluations and experience reports on personalized obesity intervention programs, including satisfaction, difficulties encountered, suggestions for improvement, and emotional responses. This data is collected through structured questionnaires, rating scales, free text input, or voice interaction. Examples include user ratings of their preferences for the taste of the nutrition plan, descriptions of their adaptation to the intensity of the exercise prescription, and feedback on the effectiveness of the psychological intervention program. Feedback data helps identify personalized barriers and preferences within the program, providing a direct user perspective for model adjustments.

[0074] The technical solution of the present invention will be further described through another embodiment.

[0075] The purpose of this invention is to overcome the shortcomings of existing technologies and achieve deep fusion and understanding of multimodal human data to construct personalized digital twin metabolic models; accurately identify the core factors and key intervention targets leading to individual obesity; generate highly personalized, dynamically adaptive, and executable comprehensive intervention plans covering multiple dimensions such as nutrition, exercise, behavior, and psychology; and achieve real-time optimization of the intervention plans through continuous human-computer interaction and feedback, thereby improving user compliance and long-term success rates. Specifically, the invention includes the following steps: 1) Multimodal Data Acquisition and Preprocessing: This involves collecting static user data, including genomic data, medical history, and anthropometry data; and dynamic user data, including continuous collection of physiological and behavioral data via wearable devices and smartphones, as well as psychological and contextual data obtained through natural language interaction. The resulting multimodal and heterogeneous data is cleaned, standardized, and aligned to construct a unified-format personal health time-series database. The cleaning process includes handling missing and outlier values; for example, for heart rate values ​​in physiological data, interpolation methods are used to correct values ​​outside a reasonable range. The standardization process converts data of different dimensions to a unified scale; for example, for numerical anthropometry data, a min-max normalization method is used. The alignment process, based on timestamps and user identifiers, integrates the data into the personal health time-series database, ensuring data consistency and traceability.

[0076] 2) Construction and Fine-tuning of a Personalized Obesity Model: A large-scale language model pre-trained with extensive medical literature, clinical guidelines, and nutritional and exercise science knowledge serves as the base model. Model fine-tuning utilizes collected multimodal data, achieved through the following methods: The pre-processed multimodal data is encoded to obtain vector data representing the pre-processed multimodal data. This vector data is used as personalized contextual information, and simultaneously, an instruction set based on obesity management cases is used to train the pre-trained large-scale language model, resulting in a personalized obesity model. The encoding process involves converting different types of data into vector form. For example, for genomic data, an embedding matrix is ​​used to map gene loci to vectors; for text data, a word embedding model is used to generate vector representations. The training process employs supervised fine-tuning to minimize the loss function.

[0077] 3) Multimodal Data Fusion and Deep Analysis: The preprocessed multimodal data is input into a fine-tuned personalized obesity model. The model leverages its powerful reasoning capabilities to perform causal association mining, risk prediction, and attribution analysis. Causal association mining identifies potential causal relationships between different factors, such as the association between a high-GI food and a sudden spike in blood sugar, using conditional probability to calculate the strength of the association. Risk prediction forecasts the user's weight change trend and the risk of developing metabolic complications over a future period, for example, using a regression model to calculate weight change, expressed as: in, This represents the predicted weight change. Represents the weight vector. Represents the eigenvector. This represents the bias term. Attribution analysis identifies the main reasons for poor current weight management results or weight rebound. For example, it uses gradient-based methods to calculate feature importance, expressed by the formula: in, Indicates the first The importance score of each feature Represents the loss function. Indicates the first Each feature value.

[0078] 4) Personalized Obesity Intervention Program Generation: Based on in-depth analysis, the model generates a structured and quantifiable comprehensive intervention program. This program includes a personalized nutrition plan, a customized exercise prescription, a behavioral intervention plan, and a psychological intervention plan. The personalized nutrition plan provides precise recommendations on specific food types, portions, and meal times, taking into account the user's dietary preferences, allergy history, and metabolic response. For example, it calculates daily intake based on energy requirements, expressed by the formula: in, This indicates the total daily energy requirement. This represents the basal metabolic rate. This represents the coefficient of physical activity level. This indicates the energy consumed during exercise. Customized exercise prescriptions recommend exercise types, intensity, frequency, and duration suitable for the user's fitness level, health condition, and interests, such as using heart rate percentage to plan intensity. Behavioral intervention programs offer cognitive behavioral therapy techniques, mindfulness-based eating training, etc. Psychological intervention programs offer stress management strategies, etc. Real-time prompts and feedback push personalized encouragement or warning messages via the app at key times.

[0079] 5) Dynamic Feedback and Program Optimization: Continuously monitor user implementation and feedback data of the personalized obesity intervention program. This includes recording diet check-ins, exercise completion rates, and weight changes via mobile applications, and collecting user satisfaction data through questionnaires. This new feedback data is then input into the personalized obesity model to initiate a new round of analysis and program generation, thereby achieving closed-loop optimization and dynamic adaptation of the intervention program. The optimization process uses incremental learning to update model parameters, for example, employing a stochastic gradient descent algorithm with a learning rate set to... To minimize feedback loss.

[0080] Assume user A, female, 35 years old, with a tendency towards insulin resistance. Multimodal data of user A is collected through a data acquisition module, including static data such as insulin resistance tendency in medical history and anthropometric data such as height, weight, and body fat percentage; dynamic data such as continuous glucose monitoring values ​​from wearable devices, dietary images and exercise records from smartphones; and psychological and contextual data such as emotional state and stress levels obtained through natural language interaction. This multimodal data undergoes cleaning, standardization, and alignment through a data preprocessing and storage module. For example, outliers in continuous glucose monitoring values ​​are corrected using interpolation methods, and emotional state text data is converted into vector representations using word embedding technology, constructing a unified format personal health time-series database. The personalized obesity big data engine inputs the preprocessed multimodal data into a fine-tuned personalized obesity big data model for multimodal data fusion and deep analysis. The model identifies a potential causal relationship between user A's consumption of oatmeal after breakfast and a rapid rise in blood sugar through causal association mining, and calculates the association strength using conditional probability, expressed by the formula: ,in, Represents conditional probability. Represents the outcome variable. and This represents the causal variable. Simultaneously, the model uses attribution analysis to pinpoint the root cause of user A's increased stress levels and snacking urges on Wednesday afternoons, and calculates feature importance using a gradient-based method, expressed by the formula: in, The importance score represents the stress characteristic. Represents the loss function. This represents the stress characteristic value. The model also performs risk prediction, forecasting user A's future weight change trend and risk of metabolic complications, for example, by using a regression model to calculate weight change values.

[0081] Based on in-depth analysis, the model generates a personalized obesity intervention plan. This plan includes a personalized nutrition plan, a customized exercise prescription, a behavioral intervention plan, and a psychological intervention plan. The personalized nutrition plan suggests replacing instant oatmeal with steel-cut oatmeal for breakfast, or opting for a combination of whole-wheat bread, eggs, and avocado, taking into account User A's dietary preferences and metabolic response; a low-sugar Greek yogurt or a small handful of nuts can be prepared in advance on Wednesday afternoon to cope with stress eating. The customized exercise prescription, based on User A's sleep data, adjusts the intensity of Wednesday morning exercise from high-intensity interval training to moderate-intensity yoga or brisk walking, using a percentage of maximum heart rate for intensity planning, expressed by the formula: in, Indicates exercise intensity. Indicates heart rate during exercise. This indicates maximum heart rate. The behavioral intervention program detected Wednesday afternoons as a peak time for stress eating and recommends users practice 5 minutes of deep breathing before meetings and set a healthy snack reminder in their schedules. The psychological intervention program provides stress management strategies, such as mindful eating training and mood regulation techniques.

[0082] The human-computer interaction and feedback module continuously monitors user A's implementation of the personalized obesity intervention plan and the feedback data, such as recording diet check-ins, exercise completion, weight changes, and user satisfaction scores. This new feedback data is then input into the personalized obesity model to initiate a new round of analysis and plan generation, thereby achieving dynamic adaptation and optimization of the intervention plan. The optimization process uses incremental learning to update model parameters, for example, employing a stochastic gradient descent algorithm with a learning rate set to... To minimize feedback loss, the intervention plan was fine-tuned one week later based on new data to ensure it adapted to user A's real-time status and needs.

[0083] Compared with the prior art, the present invention has the following advantages: 1) True Personalization: By collecting and preprocessing users' multimodal data and adjusting a pre-trained large-scale language model, a personalized obesity model is obtained, overcoming the limitations of generalized solutions. This personalized obesity model can deeply interpret multimodal data, generating personalized obesity intervention plans based on causal association mining results, risk prediction results, and attribution analysis results. It provides targeted and precise interventions that directly address the root causes of individual obesity. This personalization is reflected in the plan's full consideration of users' static characteristics such as genomic data, medical history, and anthropometric data, as well as dynamic changes in physiological, behavioral, psychological, and contextual data.

[0084] 2) Multi-dimensional fusion: For the first time, static data such as genomic data, medical history, and anthropometry data are effectively fused with dynamic data such as physiological data, behavioral data, psychological data, and contextual data to construct a comprehensive individual health profile. By inputting preprocessed multimodal data into a personalized obesity model for in-depth analysis, the generated personalized obesity intervention plans, including nutrition plans, exercise prescriptions, behavioral intervention plans, and psychological intervention plans, possess higher scientific rigor and systematicity. Multi-dimensional data fusion provides the model with rich feature information; 3) Dynamic Adaptability and High Adherence: By monitoring user implementation and feedback data of personalized obesity intervention programs, and optimizing the personalized obesity model based on the feedback data, a continuous optimization cycle is formed. This dynamic adaptive mechanism enables the intervention program to be adjusted in real time according to changes in user status, improving user stickiness and adherence through intelligent interaction. The system continuously collects new multimodal data and updates the personalized obesity model to ensure that the intervention program remains synchronized with the user's current situation.

[0085] 4) Strong predictive and preventative capabilities: The personalized obesity big data model has a forward-looking risk prediction capability, enabling early warning and intervention before problems occur. By inputting preprocessed multimodal data into the model to obtain risk prediction results, the system can predict the user's future weight change trend and the risk of metabolic complications, realizing the transformation from passive treatment to proactive prevention.

[0086] 5) High scalability: This personalized obesity model framework adopts a modular design, allowing for the continuous integration of new medical discoveries and data types, such as gut microbiome data. By continuously updating the instruction set built based on obesity management cases and adjusting model parameters, the system can continuously evolve and maintain its technological edge. The model's scalability ensures that it can adapt to advances in medical research and the needs of new health data types.

[0087] In the above embodiments, although the steps are numbered S1, S2, etc., they are only specific embodiments given by the present invention. Those skilled in the art can adjust the execution order of S1, S2, etc. according to the actual situation. The scheme after adjusting the order is also within the protection scope of the present invention. It can be understood that in some embodiments, some or all of the above embodiments may be included.

[0088] like Figure 2 As shown, an embodiment of the present invention provides an obesity intervention program generation system 200 based on a multimodal large model, comprising: a data acquisition and preprocessing module 201, a model adjustment module 202, a model application module 203, and an intervention program generation module 204. The data acquisition and preprocessing module 201 is used to: acquire and preprocess the user's multimodal data; The model adjustment module 202 is used to: adjust the pre-trained large language model to obtain a personalized obesity model; Model application module 203 is used to: input preprocessed multimodal data into a personalized obesity big model to obtain causal association mining results, risk prediction results, and attribution analysis results; The intervention plan generation module 204 is used to generate personalized obesity intervention plans based on the causal association mining results, risk prediction results, and attribution analysis results.

[0089] Optionally, in the above technical solution, the model adjustment module 202 is specifically used to: encode the preprocessed multimodal data to obtain vector data representing the preprocessed multimodal data, use the vector data as personalized contextual information, and simultaneously use the instruction set constructed based on obesity management cases to train the pre-trained large language model to obtain a personalized obesity large model.

[0090] Optionally, the above technical solution also includes a model optimization module, which is used to: monitor the user's implementation of the personalized obesity intervention plan and feedback data, and optimize the personalized obesity model based on the feedback data.

[0091] Optionally, in the above technical solutions, multimodal data includes static data and dynamic data. Static data includes genomic data, medical history, and anthropometry data, while dynamic data includes physiological data, behavioral data, psychological data, and contextual data. Personalized obesity intervention programs include nutrition plans, exercise prescriptions, behavioral intervention plans, and psychological intervention plans.

[0092] In another embodiment, a multimodal large-scale model-based obesity intervention program generation system includes: a data acquisition module, a data preprocessing and storage module, a personalized obesity large-scale model engine, a program generation and output module, and a human-computer interaction and feedback module. The data acquisition module is used to collect users' multimodal health data from various terminals and devices; the data preprocessing and storage module is used to clean and standardize the raw data and store it in a personal health database and a vector database; the personalized obesity large-scale model engine is the core module, containing a finely tuned base large-scale model, responsible for data fusion, deep analysis, and program generation; the program generation and output module transforms the structured programs output by the model into user-friendly interfaces, such as apps or reports; the human-computer interaction and feedback module provides an interface for users to interact with the system, used to receive user feedback, perform check-in and other operations.

[0093] It should be noted that the beneficial effects of the obesity intervention plan generation system 200 based on a multimodal large model provided in the above embodiments are the same as those of the obesity intervention plan generation method based on a multimodal large model described above, and will not be repeated here. Furthermore, the system provided in the above embodiments is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the system can be divided into different functional modules according to the actual situation to complete all or part of the functions described above. In addition, the system and method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process is detailed in the method embodiments, and will not be repeated here.

[0094] An electronic device according to an embodiment of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the above-mentioned methods for generating obesity intervention programs based on a multimodal large model.

[0095] Among them, electronic devices can also be terminal devices, which can be any device that can install applications, including at least one of smartphones, tablets, laptops, desktop computers, smart speakers, smartwatches, smart TVs, and smart in-vehicle devices.

[0096] An embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the above-mentioned methods for generating obesity intervention programs based on a multimodal large model.

[0097] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.

[0098] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for generating obesity intervention programs based on a multimodal large model, characterized in that, include: Collect and preprocess the user's multimodal data; By adjusting the pre-trained large language model, a personalized obesity model can be obtained. The preprocessed multimodal data is input into the personalized obesity model to obtain causal association mining results, risk prediction results, and attribution analysis results. Based on the causal association mining results, risk prediction results, and attribution analysis results, a personalized obesity intervention plan is generated.

2. The method for generating an obesity intervention program based on a multimodal large model according to claim 1, characterized in that, By fine-tuning a pre-trained large language model, a personalized obesity model is obtained, including: The preprocessed multimodal data is encoded to obtain vector data representing the preprocessed multimodal data. The vector data is used as personalized context information, and the pre-trained large language model is trained using an instruction set built based on obesity management cases to obtain a personalized obesity model.

3. The method for generating an obesity intervention program based on a multimodal large model according to claim 1, characterized in that, Also includes: Monitor the user's implementation of the personalized obesity intervention plan and the feedback data, and optimize the personalized obesity model based on the feedback data.

4. A method for generating obesity intervention programs based on a multimodal large model according to any one of claims 1 to 3, characterized in that, The multimodal data includes static data and dynamic data. The static data includes genomic data, medical history, and anthropometry data. The dynamic data includes physiological data, behavioral data, psychological data, and contextual data. The personalized obesity intervention program includes a nutrition plan, an exercise prescription, a behavioral intervention plan, and a psychological intervention plan.

5. A system for generating obesity intervention plans based on a multimodal large model, characterized in that, include: The system includes a data acquisition and preprocessing module, a model adjustment module, a model application module, and an intervention plan generation module. The data acquisition and preprocessing module is used to: acquire and preprocess the user's multimodal data; The model adjustment module is used to: adjust the pre-trained large language model to obtain a personalized obesity model; The model application module is used to: input the preprocessed multimodal data into the personalized obesity model to obtain causal association mining results, risk prediction results, and attribution analysis results; The intervention plan generation module is used to generate personalized obesity intervention plans based on the causal association mining results, risk prediction results, and attribution analysis results.

6. The obesity intervention program generation system based on a multimodal large model according to claim 5, characterized in that, The model adjustment module is specifically used to: encode the preprocessed multimodal data to obtain vector data representing the preprocessed multimodal data, use the vector data as personalized context information, and simultaneously train the pre-trained large language model using an instruction set built based on obesity management cases to obtain a personalized obesity large model.

7. The obesity intervention program generation system based on a multimodal large model according to claim 6, characterized in that, It also includes a model optimization module, which is used to: monitor the user's implementation of the personalized obesity intervention plan and feedback data, and optimize the personalized obesity model based on the feedback data.

8. A system for generating obesity intervention programs based on a multimodal large model according to any one of claims 5 to 7, characterized in that, The multimodal data includes static data and dynamic data. The static data includes genomic data, medical history, and anthropometry data. The dynamic data includes physiological data, behavioral data, psychological data, and contextual data. The personalized obesity intervention program includes a nutrition plan, an exercise prescription, a behavioral intervention plan, and a psychological intervention plan.

9. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the method for generating an obesity intervention program based on a multimodal large model as described in any one of claims 1 to 4.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method for generating an obesity intervention program based on a multimodal large model as claimed in the claim.

Images