Obesity comprehensive evaluation model based on user health data and application thereof

Abstract

The invention relates to the technical field of intelligent medical treatment, and discloses a comprehensive obesity evaluation model based on user health data and application thereof, and the model comprises a data acquisition module, a data preprocessing module, a feature engineering module, a model training module and an evaluation output module. According to the method, multi-source heterogeneous health data are integrated, the limitation of traditional single data assessment is broken through, multi-dimensional factors influencing obesity can be comprehensively captured, data preprocessing and feature engineering optimization are combined, the data and feature quality is effectively improved, and a foundation is laid for accurate assessment; secondly, by adopting a deep learning and causal reasoning hybrid architecture, not only is feature extraction and evaluation precision guaranteed, but also the function logic of the features on obesity evaluation is clarified through interpretability analysis, and the evaluation credibility is improved; besides, the dynamic obesity risk score updated along with the real-time health data of the user is generated, the risk change can be reflected in real time, and accurate and efficient technical support is provided for personalized health management and obesity intervention.

Description

Technical Field

[0001] This invention relates to the field of smart healthcare technology, specifically to a comprehensive obesity assessment model based on user health data and its application. Background Technology

[0002] Current obesity assessment technologies face three main bottlenecks: insufficient utilization of multi-source data, difficulty in balancing model accuracy and interpretability, and a lack of dynamic adaptability in assessment results. At the data level, existing solutions largely rely on single static physiological indicators such as body mass index (BMI) and body fat percentage, severely limiting the quality of model input. At the model architecture level, traditional solutions either employ deep learning architectures to pursue accuracy but lose interpretability due to their black-box nature, failing to clearly define the influence weights and association rules of various health characteristics on obesity risk; or they use traditional statistical models to ensure interpretability but struggle to fit complex, multi-dimensional features, limiting accuracy. This results in low alignment between assessment results and actual health needs, insufficient reliability, and ultimately, limited application value in obesity risk prediction, making it difficult to support personalized obesity intervention guidance. Summary of the Invention

[0003] The purpose of this invention is to provide a comprehensive obesity assessment model based on user health data and its application, so as to solve the technical problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention discloses the following technical solutions: In a first aspect, the present invention discloses a comprehensive obesity assessment model based on user health data, comprising: The data acquisition module is used to acquire multi-source heterogeneous health data of users, including physiological indicator data, wearable device dynamic monitoring data, dietary record data and environmental parameter data. The data preprocessing module is used to clean, standardize, and perform cross-modal fusion processing on the multi-source heterogeneous health data, extract multi-dimensional initial features from the processed data, and output standardized multi-dimensional initial feature data. The feature engineering module is used to receive the output of the data preprocessing module and perform feature filtering, feature transformation and feature fusion operations, remove low-value features and map multi-dimensional data to a unified feature space, and output an engineered feature set for model training. The model training module, based on a hybrid architecture of deep learning and causal inference, includes: The base layer is used to receive the engineered feature set, extract features of different types of data in the engineered feature set through a deep learning network, and calculate and output the corresponding feature weights based on the preliminary correlation between each feature and obesity assessment. The optimization layer receives the output of the base layer and dynamically adjusts the feature weights using a weight optimization algorithm, outputting the adjusted feature weights and corresponding feature data. The interpretation layer receives the output of the optimization layer and analyzes the influence weight of each feature on the obesity assessment result through an interpretability analysis algorithm, generates an interpretable rule set, and feeds back the analyzed influence weights to the optimization layer. An evaluation output module is used to receive the output of the model training module and generate a dynamic obesity risk score based on the feature data output by the optimization layer, the influence weights, and the interpretable rule set, wherein the dynamic obesity risk score is updated with the user's real-time health data.

[0005] Preferably, the process of cleaning, standardizing, and performing cross-modal fusion processing on the multi-source heterogeneous health data, and outputting standardized data in a unified format, includes: The physiological indicator data and the wearable device dynamic monitoring data are initially filtered out using the first-stage outlier filtering rule, and regional anomaly detection is performed on the dietary image data in the dietary record data using a convolutional network to retain valid original data. The physiological index data and the environmental parameter data are converted to the [-1,1] value range by the tan-h normalization algorithm, and the time series values ​​of the wearable device dynamic monitoring data are uniformly mapped to the [0,1] range by the Min-Max normalization algorithm to eliminate the difference in dimensions. Based on the data collection timestamp, the physiological indicator data, the wearable device dynamic monitoring data, the dietary record data, and the environmental parameter data are spatiotemporally aligned. Multi-dimensional initial features are extracted from the aligned data. The multi-dimensional initial features include body fat percentage, hourly exercise intensity, food type code, and time period identifier, forming and outputting standardized multi-dimensional initial feature data.

[0006] Preferably, the feature engineering module includes: The feature filtering unit is configured to: calculate the variance contribution of each initial feature in the standardized multi-dimensional initial feature data and the significance of its association with the obesity assessment label, remove low-value initial features with low variance contribution and low significance of association, and form a filtered feature set; The feature transformation unit is configured to perform transformation operations on the temporal features, text features, and static numerical features in the filtered feature set respectively, and generate transformed features suitable for model training. The feature fusion unit is configured to concatenate and weight the transformed features to generate an engineered feature set.

[0007] Preferably, the calculation of the variance contribution includes the following steps: Detection and removal of outlier features based on box plot rules; The remaining effective feature values ​​are calculated based on the variance contribution of principal component analysis, and features with variance contribution below a preset threshold are removed. Then, the redundancy of the remaining features is detected by the Pearson correlation coefficient. If the absolute value of the Pearson correlation coefficient between two features is greater than the preset threshold, features with low correlation to the obesity assessment label are removed, and features with higher correlation are retained.

[0008] Preferably, the conversion operation includes the following steps: Derived temporal features were extracted from temporal features using a fixed time interval sliding window method. The derived temporal features included the average exercise intensity and the number of nighttime sleep fragmentation within the fixed time interval. The bag-of-words model is used to segment and encode the text features, and text transformation features of food categories and cooking methods are extracted. The static numerical features are mapped to the [0,1] numerical range using a normalization algorithm to generate static transformed features.

[0009] Preferably, the splicing and weighted fusion includes the following steps: The derived temporal features, the text transformation features, and the static transformation features are concatenated into an initial feature matrix; Weights are assigned based on feature type identifiers and initial correlation, with sub-weights allocated to features within the same type according to the proportion of the absolute value of the initial correlation, ultimately generating an engineered feature set.

[0010] Preferably, the acquisition of the preliminary correlation degree includes the following steps in sequence: Determine the type of each feature in the engineering feature set, where physiological index features and environmental parameter features are defined as continuous numerical features, and dietary category features and time period identifier features are defined as discrete categorical features. For the continuous numerical features, the linear association strength between them and the obesity assessment label is calculated using the Pearson correlation coefficient; for the discrete categorical features, the categorical association strength between them and the obesity assessment label is calculated using the chi-square test. The calculated results of the linear correlation strength and the category correlation strength are normalized to obtain the preliminary correlation between each feature and obesity assessment.

[0011] Preferably, the step of analyzing the influence weights of each feature on the obesity assessment results using an interpretability analysis algorithm to generate an interpretable rule set includes the following steps: Based on the Shapley additive interpretable value algorithm, all possible feature subsets in the engineered feature set are generated. For each feature subset, the difference in obesity assessment results output by the model when the feature subset contains the target feature and when it does not is calculated as the marginal contribution of the target feature in that feature subset. Then, according to the size of each feature subset and the probability of uniform distribution, the marginal contribution of the target feature is weighted and averaged to obtain the contribution degree of the target feature to the obesity assessment result, and the absolute value of the contribution degree is used as the influence weight. Features with high weighting were selected as key features. The obesity risk values ​​corresponding to different numerical ranges of each key feature were statistically analyzed to determine the association rule that the obesity risk increases when the value of a key feature exceeds a specific range. For physiological indicators and exercise intensity features, the threshold rules for the normal range of body fat percentage and the normal range of daily exercise intensity for different genders were determined by referring to health and medical standards and combining the numerical ranges of healthy people in the feature value distribution. The key features were combined for testing, and the obesity risk value when the key features were combined was calculated and compared with the risk value when the single feature was combined to determine the synergistic effect rule of the increase in obesity risk under the combined feature effect.

[0012] Preferably, the dynamic obesity risk score is generated through the following steps: Extract the feature data from the output of the optimization layer, multiply each feature data by its corresponding influence weight, and then sum all the multiplication results to obtain the basic obesity risk score. The baseline obesity risk score is adjusted according to the interpretable rule set. If a key feature exceeds the specific range defined in the association rule, the baseline obesity risk score is increased according to the risk increase determined by the association rule. If a key feature falls within the normal range defined by the threshold rule, the baseline obesity risk score is maintained. If there is a combination of key features that meets the synergistic effect rule, the baseline obesity risk score is further adjusted according to the risk increase determined by the synergistic effect rule. When a user's real-time health data is updated, the updated data is sequentially acquired by the data acquisition module, processed by the data preprocessing module, and converted into an engineered feature set by the feature engineering module. After being processed by the model training module, the influence weights are obtained and fed back to the optimization layer. Based on the updated key feature data and the corresponding influence weights, the basic risk score and score adjustment are repeatedly calculated to obtain the updated dynamic obesity risk score.

[0013] Secondly, this embodiment discloses the application of the obesity comprehensive assessment model based on user health data, as described above, in predicting user obesity risk.

[0014] Beneficial Effects: The comprehensive obesity assessment model based on user health data and its application in this application break through the limitations of traditional single-data assessment by integrating multi-source heterogeneous health data. It can comprehensively capture multi-dimensional factors affecting obesity. Combined with data preprocessing and feature engineering optimization, it effectively improves the quality of data and features, laying the foundation for accurate assessment. Secondly, by adopting a hybrid architecture of deep learning and causal reasoning, it not only ensures the accuracy of feature extraction and assessment, but also clarifies the logic of the role of features in obesity assessment through interpretability analysis, thereby improving the credibility of the assessment. In addition, it generates a dynamic obesity risk score that is updated with the user's real-time health data, which can reflect changes in risk in real time and provide accurate and efficient technical support for personalized health management and obesity intervention. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 The diagram below shows the structure of a comprehensive obesity assessment model based on user health data, as provided in an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] In this document, the term "comprising" is intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0019] This embodiment provides a first aspect such as Figure 1 The obesity comprehensive assessment model shown is based on user health data. This model can be integrated into personal health management terminals (such as smart bracelet apps, home health monitoring devices) or health assessment systems of medical institutions to assess users' obesity risk in real time and accurately and provide support for personalized intervention.

[0020] Specifically, a comprehensive obesity assessment model based on user health data includes: The data acquisition module is used to acquire users' multi-source heterogeneous health data, which includes physiological indicator data, wearable device dynamic monitoring data, dietary record data, and environmental parameter data. The data preprocessing module is used to clean, standardize, and perform cross-modal fusion processing on multi-source heterogeneous health data. It extracts multi-dimensional initial features from the processed data and outputs standardized multi-dimensional initial feature data. The feature engineering module receives the output of the data preprocessing module and performs feature filtering, feature transformation and feature fusion operations. It removes low-value features and maps multi-dimensional data to a unified feature space, outputting an engineered feature set for model training. The model training module, based on a hybrid architecture of deep learning and causal inference, includes: The base layer receives the engineered feature set, extracts features of different types of data from the engineered feature set through a deep learning network, and calculates and outputs the corresponding feature weights based on the preliminary correlation between each feature and obesity assessment. The optimization layer receives the output of the base layer and dynamically adjusts the feature weights using a weight optimization algorithm, outputting the adjusted feature weights and the corresponding feature data. The interpretation layer receives the output of the optimization layer and analyzes the influence weight of each feature on the obesity assessment result through the interpretability analysis algorithm, generates an interpretable rule set, and feeds back the analyzed influence weights to the optimization layer. The evaluation output module receives the output of the model training module and generates a dynamic obesity risk score based on the feature data, influence weights, and interpretable rule set output by the optimization layer. The dynamic obesity risk score is updated with the user's real-time health data.

[0021] In this embodiment, the data acquisition module is implemented as follows: physiological indicator data is collected through devices such as a smart body fat scale (to obtain body fat percentage, BMI, and muscle mass), an electronic blood pressure monitor (to obtain blood pressure values), and a blood glucose meter (to obtain blood glucose values); wearable device dynamic monitoring data is collected through fitness trackers (to obtain steps, exercise intensity, and sleep duration) and heart rate belts (to obtain real-time heart rate); dietary record data is obtained through text records manually entered by the user in the health APP (such as "Breakfast: 200ml milk + 1 slice of bread") or uploaded images of food (such as photos of lunch plates); environmental parameter data is collected through temperature and humidity sensors (to obtain temperature and humidity) and air quality monitoring devices (to obtain PM2.5 concentration) in the user's environment. All data carries a collection timestamp and a unique user identifier to ensure data traceability.

[0022] Secondly, in this embodiment, the hybrid architecture of deep learning and causal reasoning is specifically manifested as follows: the base layer adopts a combination architecture of "convolutional neural network (CNN) + long short-term memory network (LSTM)," where CNN is used to extract visual features of dietary image data (such as food type and portion size), and LSTM is used to extract dynamic features of wearable device time-series data (such as the trend of exercise intensity change and sleep cycle); causal reasoning is achieved by constructing a causal graph model between features, for example, based on expert knowledge to pre-set causal relationships such as "dietary calorie intake → change in body fat percentage" and "exercise intensity → energy consumption," and the optimization layer refers to this causal graph when adjusting feature weights to avoid non-causal features interfering with weight allocation; the weight optimization algorithm adopts the Adam optimizer, with a learning rate set to 0.001 and an iteration count set to 100 rounds to ensure that the weight adjustment converges to the optimal solution.

[0023] Feasible methods include cleaning, standardizing, and performing cross-modal fusion processing on multi-source heterogeneous health data, and outputting standardized data in a unified format, including: The first-stage outlier filtering algorithm is used to remove outliers from physiological indicator data and wearable device dynamic monitoring data. Then, a convolutional network is used to detect regional anomalies in dietary image data in dietary records, thus preserving valid original data. The physiological index data and environmental parameter data are transformed to the [-1,1] value range by using the tan-h normalization algorithm, and the time series values ​​of wearable device dynamic monitoring data are uniformly mapped to the [0,1] range by using the Min-Max normalization algorithm, thus eliminating the difference in dimensions. Based on the data collection timestamp, physiological indicator data, wearable device dynamic monitoring data, dietary record data and environmental parameter data are spatiotemporally aligned. Multidimensional initial features are extracted from the aligned data, including body fat percentage, hourly exercise intensity, food type code and time period identifier, forming and outputting standardized multidimensional initial feature data.

[0024] In this embodiment, the specific calculation method of the first-stage outlier filtering rule is as follows: using the box plot rule, the quartiles (Q1, Q3) of physiological indicator data and wearable device dynamic monitoring data are calculated, and the outlier judgment range is determined to be [Q1-1.5×IQR, Q3+1.5×IQR] (IQR=Q3-Q1). Data exceeding this range is judged as outliers and removed. For example, in body fat percentage data, if a user's single body fat percentage detection value is 50% (far exceeding the normal range and falling outside the abnormal range of the box plot), it is judged as an outlier and removed. The convolutional network adopts a lightweight architecture of 2 convolutional layers (convolutional kernel size 3×3, number 16 and 32 respectively) + 1 max pooling layer (pooling kernel size 2×2) to detect regional anomalies in dietary images (such as images containing debris unrelated to food, or images that are blurry and cannot identify food). For image data judged as abnormal, the user is triggered to re-upload or manually enter dietary records.

[0025] In this embodiment, the specific process of spatiotemporal alignment is as follows: Using 15-minute time granularities, all data is mapped to the corresponding 15-minute time window according to the collection timestamp. For example, a user's exercise intensity data from 8:00 to 8:15, body fat percentage data collected at 8:10, and breakfast image data uploaded at 8:12 are all aligned to the "8:00-8:15" time window. Food type coding adopts GB / T23784-2009 "Food Classification and Codes," for example, "rice" is coded as "01". "01.01", "pork" is coded as "02.01.01"; the time period identifier is divided into 6 time periods according to the 24 hours of a day, namely "00:00-06:00 (early morning)", "06:00-09:00 (breakfast)", "09:00-12:00 (morning)", "12:00-14:00 (lunch)", "14:00-18:00 (afternoon)" and "18:00-24:00 (dinner and night)", and the corresponding time period identifier is matched according to the data collection timestamp.

[0026] A feasible feature engineering module includes: The feature filtering unit is configured to: calculate the variance contribution of each initial feature in the standardized multidimensional initial feature data and the significance of its association with the obesity assessment label, remove low-value initial features with low variance contribution and low significance of association, and form a filtered feature set. The feature transformation unit is configured to perform transformation operations on the time-series features, text features, and static numerical features in the filtered feature set, respectively, to generate transformed features suitable for model training. The feature fusion unit is configured to concatenate and weightedly fuse the transformed features to generate an engineered feature set.

[0027] Furthermore, the calculation of the variance contribution includes the following steps: Detection and removal of outlier features based on box plot rules; The remaining effective feature values ​​are calculated based on the variance contribution of principal component analysis, and features with variance contribution below a preset threshold are removed. Then, the redundancy of the remaining features is detected by the Pearson correlation coefficient. If the absolute value of the Pearson correlation coefficient between two features is greater than the preset threshold, features with low correlation to the obesity assessment label are removed, and features with higher correlation are retained.

[0028] In this embodiment, the specific values ​​of the preset thresholds are as follows: The preset threshold for variance contribution is set to 5%, that is, only initial features with a variance contribution of ≥5% are retained. For example, the variance contribution of the feature "environmental humidity" is 2.3% (below the threshold), which is judged as a low-value feature and removed. The preset threshold for Pearson correlation coefficient is set to 0.8. For example, the absolute value of the Pearson correlation coefficient between the features "BMI" and "body fat percentage" is 0.85 (greater than the threshold). The correlation between the two and the obesity assessment label (label is set as "0=normal, 1=overweight, 2=obese") is further calculated. If the correlation between "body fat percentage" and the label (0.72) is higher than that between "BMI" and "BMI" (0.65), the "BMI" feature is removed. The significance of the association is calculated by t-test, and the significance level is set to 0.05. If the t-test p value of a feature and the obesity assessment label is >0.05, it is judged as not significantly associated and removed.

[0029] Furthermore, the execution of the conversion operation includes the following steps: Derived temporal features were extracted from temporal features using a fixed time interval sliding window method. The derived temporal features included the average exercise intensity and the number of nighttime sleep fragmentation within the fixed time interval. The bag-of-words model is used to segment and encode the text features, and text transformation features of food categories and cooking methods are extracted. The static numerical features are mapped to the [0,1] numerical range using a normalization algorithm to generate static transformed features.

[0030] In this embodiment, the specific parameters of the fixed time interval sliding window method are as follows: For the temporal characteristics of exercise intensity, a 30-minute sliding window is used to extract the "30-minute average exercise intensity", and a 24-hour sliding window is used to extract the "daily average exercise intensity"; for nighttime sleep data, a 5-minute sliding window is used to detect sleep-wake states. The criteria for determining the number of nighttime sleep fragmentation are: wake-up time during sleep > 5 minutes, and the interval between two wake-ups < 30 minutes, which is counted as 1 fragmentation. The "total number of sleep fragmentation per night" is counted as a derived temporal characteristic.

[0031] In this embodiment, the word segmentation and encoding process of the bag-of-words model is as follows: For food text records (such as "Dinner: Braised Pork Ribs + Stir-fried Vegetables + 1 Bowl of Rice"), the Jieba word segmentation tool is used to segment the text, obtaining words such as "Dinner," "Braised Pork Ribs," "Stir-fried Vegetables," "Rice," and "1 Bowl." Then, based on a preset dictionary (containing 500 common foods and 20 cooking methods), encoding is performed. For example, "Braised Pork Ribs" is encoded as "Food Category: Meat - Livestock, Cooking Method: Braised," and "Stir-fried Vegetables" is encoded as "Food Category: Vegetables, Cooking Method: Stir-fried." Finally, a text transformation feature vector with a dimension of 64 is generated. The normalization algorithm uses Min-Max normalization, and its formula and parameter definitions are as follows: The Min-Max normalization formula is: ,in, The raw static numerical feature values ​​to be normalized (such as the raw detection values ​​of a user's age, height, basal metabolic rate, and other static physiological indicators). This is the minimum value of the static numerical feature in the model training sample set (obtained from statistics of the sample set, which needs to cover user data of different age groups and body types to ensure universality). This is the maximum value of the static numerical feature in the model training sample set (same as \(x_{min}\, determined based on sample set statistics); This is the static transformed feature value mapped to the [0,1] interval after normalization (to eliminate the differences in the scale of different static features and adapt to the input requirements for model training).

[0032] To further clarify the application of the parameters, the following example is provided: Using "age" as a static numerical feature, the minimum age value in the model training sample set. Age (minimum age threshold for adult users), maximum age Age (the upper age limit commonly seen for middle-aged and elderly users). If a user's original age... Age, substituting into the formula, yields: That is, the static conversion characteristic value of the user's age is 0.36.

[0033] Furthermore, the splicing and weighted fusion process includes the following steps: The derived temporal features, text transformation features, and static transformation features are concatenated into an initial feature matrix; Weights are assigned based on feature type identifiers and initial correlation, with sub-weights allocated to features within the same type according to the proportion of the absolute value of the initial correlation, ultimately generating an engineered feature set.

[0034] Secondly, obtaining the initial correlation involves the following steps: Determine the type of each feature in the engineering feature set, where physiological index features and environmental parameter features are defined as continuous numerical features, and dietary category features and time period identifier features are defined as discrete categorical features. For continuous numerical features, the linear association strength between them and the obesity assessment label is calculated using the Pearson correlation coefficient; for discrete categorical features, the categorical association strength between them and the obesity assessment label is calculated using the chi-square test. The calculated results of linear association strength and category association strength were normalized to obtain the preliminary association degree between each feature and obesity assessment.

[0035] In this embodiment, the composition and dimension definition of the initial feature matrix are first clarified: the dimensions of the derived temporal features, text transformation features, and static numerical features are respectively set as follows: (Including two features: "average exercise intensity over 30 minutes" and "number of sleep fragments per night") (Based on encoding vectors of 500 common foods and 20 cooking methods) (Including static physiological indicators such as age, height, and basal metabolic rate), the initial feature matrix formed by splicing the three data points has a dimension of [missing value]. The matrix form is (Feature matrix for a single user's single evaluation, rows represent samples, columns represent features), elements in the matrix This represents the normalized value of the j-th transformed feature for a single user (range [0,1]).

[0036] In this embodiment, the normalization formula for the linear correlation strength of continuous numerical features and the complete parameter definitions are as follows: To determine the initial linear correlation strength (Pearson correlation coefficient r) between continuous numerical features (such as body fat percentage and average daily exercise intensity) and obesity assessment labels, linear normalization was applied to map them to the [0,1] interval to obtain the preliminary correlation degree of the continuous features. The formula is: , in, The original Pearson correlation coefficient between the continuous numerical feature and the obesity assessment label ranges from -1 to 1. It is calculated based on the linear correlation between the feature and the label in the model training sample set (sample size n=5000, covering users aged 18-65 years of different genders and BMI ranges). A positive correlation indicates that the higher the feature value, the higher the risk of obesity (e.g., r=0.7 for body fat percentage), while a negative correlation indicates that the higher the feature value, the lower the risk of obesity (e.g., r=-0.5 for average daily exercise intensity). This represents the initial correlation degree after normalization of continuous numerical features, with a value range of [0,1]. It serves as the core basis for subsequent sub-weight allocation; the larger this value, the higher the correlation between the normalized feature and obesity assessment; in the formula... The purpose of this is to shift the range of the original Pearson correlation coefficient from [-1,1] to [0,2], and then compress it to [0,1] by dividing by 2, so as to ensure that the correlation strength of negatively correlated features (such as exercise intensity) can be effectively quantified (avoiding negative values ​​from affecting weight allocation).

[0037] To illustrate parameter application with an example: If the original Pearson correlation coefficient r = 0.7 between "body fat percentage" (a continuous feature) and obesity assessment labels (labels defined as "0 = normal, 1 = overweight, 2 = obese", converted to numerical labels during calculation), substituting this into the formula yields the preliminary correlation. That is, the correlation degree of this feature after normalization is 0.85, which belongs to the high correlation feature.

[0038] In this embodiment, the normalization formula for the category association strength of discrete categorical features and the complete parameter definitions are as follows: The chi-square test value is used to determine the correlation strength between discrete categorical features (such as the frequency of high-oil food intake and time period markers) and the original category of obesity assessment labels. The maximum value is normalized and mapped to the [0,1] interval to obtain the preliminary correlation degree of discrete features. The formula is: , in, The original chi-square test value is the sum of the discrete categorical features and the obesity assessment label, with a range of values. The calculation is based on the difference between the observed frequency and the expected frequency (the theoretical frequency assuming the feature and label are independent) of the feature in the training sample set; the larger the value, the stronger the category association between the feature and the label (e.g., "frequency of high-oil food intake"). ); The maximum value of the original chi-square test for all discrete categorical features (32 items in total, including food category, time period identifier, cooking method, etc.) is obtained from the training sample set. In this embodiment, (The chi-square test value corresponding to "daily high sugar food intake frequency"), this value is a fixed constant used to unify the dimensions of the association strength of all discrete features; This represents the initial correlation degree after normalization of discrete categorical features, with a value range of [0,1], and is related to the correlation degree of continuous features. Having the same units of measurement, they can participate together in the allocation of sub-weights.

[0039] Illustrate parameter application with examples: The original chi-square test value of "Frequency of High-Oil Food Intake" (a discrete feature, categorized as "0 times / day", "1-2 times / day", and "3 times or more / day") Substituting into the formula yields the preliminary correlation degree. That is, the correlation degree of this feature after normalization is 0.83, which belongs to the high correlation feature.

[0040] In this embodiment, the sub-weight allocation formula and complete parameter definitions are as follows: Based on the total weight of feature types and the preliminary correlation degree of individual features, the sub-weights of each feature are calculated. The formula is: , in, The final sub-weight of a certain feature, with a value range of (0,1], is used for weighted calculation during feature fusion (the sub-weight is multiplied by the feature value and then used to construct the engineered feature set). The total weight of the feature type is determined by pre-setting the importance of the feature to obesity assessment and optimizing it through 5-fold cross-validation. In this embodiment, the total weight of the derived time-series feature type is... (Dynamic features have the greatest impact on real-time risk assessment), total weight of text conversion feature types (Dietary characteristics are the core influencing factor), total weight of static numerical characteristic types (Static features have a relatively stable impact), and (The total weight of all types is 1, ensuring that the weight allocation is reasonable.) The initial correlation of this feature (for continuous features) Discrete characteristics are ), such as "30-minute average exercise intensity" "Number of sleep fragments per night" ; This is the sum of the preliminary correlation degrees of all m features within the category to which this feature belongs. For example, if m=2 within the derived time-series feature category, the sum is 0.7 + 0.3 = 1.0; the proportional term in the formula... This is used to ensure that highly correlated features within the same type receive higher sub-weights and low-correlation features receive lower sub-weights, thus achieving differentiated weight allocation within the type.

[0041] A feasible approach is to analyze the influence weights of each feature on the obesity assessment results using an interpretability analysis algorithm, and generate an interpretable rule set, including the following steps: Based on the Shapley additive interpretable value algorithm, all possible feature subsets in the engineered feature set are generated. For each feature subset, the difference in the obesity assessment results output by the model when the feature subset contains the target feature and when it does not is calculated as the marginal contribution of the target feature in that feature subset. Then, according to the size of each feature subset and the probability of uniform distribution, the marginal contribution of the target feature is weighted and averaged to obtain the contribution of the target feature to the obesity assessment results, and the absolute value of the contribution is used as the influence weight. Features with high weighting were selected as key features. The obesity risk values ​​corresponding to different numerical ranges of each key feature were statistically analyzed to determine the association rule that the obesity risk increases when the value of a key feature exceeds a specific range. For physiological indicators and exercise intensity features, the threshold rules for the normal range of body fat percentage and the normal range of daily exercise intensity for different genders were determined by referring to health and medical standards and combining the numerical ranges of healthy people in the feature value distribution. The key features were combined for testing, and the obesity risk value when the key features were combined was calculated and compared with the risk value when the single feature was combined to determine the synergistic effect rule of the increase in obesity risk under the combined feature effect.

[0042] In this embodiment, the specific calculation process of the Shapley addendum interpretation value algorithm is as follows: If the engineered feature set contains 10 features, generate all... Let there be n non-empty feature subsets, and the probability of each subset appearing is denoted as n. (Uniform distribution); Taking "average daily exercise intensity" as the target feature, calculate the difference between the model evaluation results (e.g., risk value 0.6) of the subset containing this feature (e.g., {average daily exercise intensity, body fat percentage}) and the evaluation results (e.g., risk value 0.75) of the subset not containing this feature (e.g., {body fat percentage}), which is taken as the marginal contribution of the subset; the marginal contributions of all subsets containing the target feature are weighted by probability to obtain the contribution of the target feature (e.g., -0.12), and its absolute value of 0.12 is the influence weight.

[0043] In this embodiment, the selection criteria and rule set for key features are as follows: Features with the top 20% influence weight are selected as key features, such as "body fat percentage", "average daily exercise intensity", "frequency of high-fat food intake", and "nighttime sleep duration"; association rules include "if the frequency of high-fat food intake is >3 times / day, the obesity risk value increases by 0.2-0.3"; threshold rules refer to the "Guidelines for the Prevention and Control of Overweight and Obesity in Chinese Adults", determining that the normal range for body fat percentage is 15%-20% for men and 20%-25% for women, and the normal range for daily exercise intensity is moderate-intensity exercise ≥30 minutes / day; synergistic influence rules include "if the average daily exercise intensity is <20 minutes / day and the frequency of high-fat food intake is >3 times / day, the obesity risk value increases by an additional 0.15 compared to the effect of a single feature".

[0044] A feasible approach is to generate a dynamic obesity risk score through the following steps: Extract the feature data from the output of the optimization layer, multiply each feature data by its corresponding influence weight, and then sum all the multiplication results to obtain the basic obesity risk score. The baseline obesity risk score is adjusted according to the interpretable rule set. If a key feature exceeds the specific range defined in the association rule, the baseline obesity risk score is increased according to the risk increase determined by the association rule. If a key feature falls within the normal range defined by the threshold rule, the baseline obesity risk score is maintained. If there is a combination of key features that meets the synergistic effect rule, the baseline obesity risk score is further adjusted according to the risk increase determined by the synergistic effect rule. When a user's real-time health data is updated, the updated data is successively acquired by the data acquisition module, processed by the data preprocessing module, and converted into an engineered feature set by the feature engineering module. After being processed by the model training module, the influence weights are obtained and fed back to the optimization layer. Based on the updated key feature data and the corresponding influence weights, the basic risk score and score adjustment are repeatedly calculated to obtain the updated dynamic obesity risk score.

[0045] In this embodiment, the calculation and adjustment of the baseline obesity risk score is illustrated below: The key feature data output by the optimization layer are "body fat percentage 0.6 (after normalization, corresponding to the original value of 28%), average daily exercise intensity 0.3 (corresponding to the original value of 15 minutes / day), and frequency of high-fat food intake 0.8 (corresponding to the original value of 4 times / day)," with corresponding influence weights of 0.3, 0.2, and 0.25, respectively. The baseline obesity risk score = 0.6 × 0.3 + 0.3 × 0.2 + 0.8 × 0.25 = 0.18 + 0.06 + 0.2 = 0.44; (Refer to the rule set) The score for "body fat percentage 28% (out of normal range)" was increased by 0.15 according to the association rule, the score for "high-fat food intake 4 times / day (out of specific range)" was increased by 0.25, and the score for "average daily exercise intensity 15 minutes / day (below normal range)" was increased by 0.1. In addition, the score for "low exercise intensity + high fat intake" was increased by an additional 0.15 according to the synergy rule. The final adjusted score is 0.44 + 0.15 + 0.25 + 0.1 + 0.15 = 1.09 (score range 0-2, 0-0.6 is low risk, 0.6-1.2 is medium risk, and 1.2-2 is high risk).

[0046] In this embodiment, the triggering conditions for real-time data updates are as follows: when the wearable device generates new exercise data (such as updating exercise intensity every 10 minutes), the user uploads new dietary records (such as uploading immediately after a meal), or the physiological indicator detection device acquires new data (such as detecting body fat percentage every morning), the data update process is automatically triggered. The model completes data processing and score updates within 5 seconds and pushes the updated score to the user terminal.

[0047] This embodiment provides, in a second aspect, the application of the comprehensive obesity assessment model based on user health data, as described above, in predicting user obesity risk. In this embodiment, the specific application scenario and implementation method are as follows: Personal health management APP application: The model is integrated into the smart health APP. Users can bind wearable devices and physiological indicator monitoring devices through the APP. The APP obtains multi-source health data in real time and calls the model to generate a dynamic obesity risk score. When the score is in the medium to high risk range, the APP pushes personalized intervention suggestions (such as "You have consumed high-oil foods 3 times today. It is recommended to choose steamed fish and cold vegetables for dinner" or "Your exercise intensity is insufficient today. It is recommended to do a 30-minute brisk walk from 19:00 to 19:30"). At the same time, it generates weekly / monthly risk change trend charts to help users intuitively understand their own obesity risk changes.

[0048] Community healthcare intervention application: Community healthcare centers connect to user health data through the system. Doctors use models to obtain users' dynamic obesity risk scores and interpretable rule sets (such as "the user's high risk is mainly due to body fat percentage exceeding the standard by 20% + insufficient daily exercise"). Combined with clinical experience, they develop personalized intervention plans (such as "three fat loss training sessions per week + low GI diet plan" for users with excessive body fat percentage). The intervention effect is tracked in real time through the model. If the user's risk score continues to decrease, the existing plan is maintained; if the score does not change, the intervention strategy is adjusted in a timely manner.

[0049] Applications in corporate employee health management: Companies equip their employees with health monitoring devices and generate employee obesity risk assessment reports regularly through models. They can organize health lectures, weight loss training camps, and other activities for high-risk employees, while incorporating risk data into employee health records to provide customized health benefits (such as providing exercise subsidies for employees who persist in exercising), helping companies reduce the incidence of obesity-related diseases among their employees.

[0050] In the embodiments provided by this invention, it should be understood that the embodiments described herein can be implemented in hardware, software, firmware, middleware, code, or any suitable combination thereof. For hardware implementation, the processor can be implemented in one or more of the following: application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), processors, controllers, microcontrollers, microprocessors, other electronic units designed to implement the functions described herein, or combinations thereof. For software implementation, some or all of the processes of the embodiments can be performed by a computer program instructing the associated hardware. During implementation, the program can be stored in a computer-readable storage medium or transmitted as one or more instructions or code on a computer-readable storage medium. Computer-readable storage media include computer storage media and communication media, wherein communication media include any medium that facilitates the transmission of a computer program from one place to another. Storage media can be any available medium accessible to a computer. Computer-readable storage media can include, but are not limited to, RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code having the form of instructions or data structures and accessible to a computer.

[0051] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A comprehensive obesity assessment model based on user health data, characterized in that, include: The data acquisition module is used to acquire multi-source heterogeneous health data of users, including physiological indicator data, wearable device dynamic monitoring data, dietary record data and environmental parameter data. The data preprocessing module is used to clean, standardize, and perform cross-modal fusion processing on the multi-source heterogeneous health data, extract multi-dimensional initial features from the processed data, and output standardized multi-dimensional initial feature data. The feature engineering module is used to receive the output of the data preprocessing module and perform feature filtering, feature transformation and feature fusion operations, remove low-value features and map multi-dimensional data to a unified feature space, and output an engineered feature set for model training. The model training module, based on a hybrid architecture of deep learning and causal inference, includes: The base layer is used to receive the engineered feature set, extract features of different types of data in the engineered feature set through a deep learning network, and calculate and output the corresponding feature weights based on the preliminary correlation between each feature and obesity assessment. The optimization layer receives the output of the base layer and dynamically adjusts the feature weights using a weight optimization algorithm, outputting the adjusted feature weights and corresponding feature data. The interpretation layer receives the output of the optimization layer and analyzes the influence weight of each feature on the obesity assessment result through an interpretability analysis algorithm, generates an interpretable rule set, and feeds back the analyzed influence weights to the optimization layer. An evaluation output module is used to receive the output of the model training module and generate a dynamic obesity risk score based on the feature data output by the optimization layer, the influence weights, and the interpretable rule set, wherein the dynamic obesity risk score is updated with the user's real-time health data.

2. The obesity comprehensive assessment model based on user health data according to claim 1, characterized in that, The process of cleaning, standardizing, and performing cross-modal fusion processing on the multi-source heterogeneous health data, and outputting standardized data in a unified format, includes: The physiological indicator data and the wearable device dynamic monitoring data are initially filtered out using the first-stage outlier filtering rule, and regional anomaly detection is performed on the dietary image data in the dietary record data using a convolutional network to retain valid original data. The physiological index data and the environmental parameter data are converted to the [-1,1] value range by the tan-h normalization algorithm, and the time series values ​​of the wearable device dynamic monitoring data are uniformly mapped to the [0,1] range by the Min-Max normalization algorithm to eliminate the difference in dimensions. Based on the data collection timestamp, the physiological indicator data, the wearable device dynamic monitoring data, the dietary record data, and the environmental parameter data are spatiotemporally aligned. Multi-dimensional initial features are extracted from the aligned data. The multi-dimensional initial features include body fat percentage, hourly exercise intensity, food type code, and time period identifier, forming and outputting standardized multi-dimensional initial feature data.

3. The obesity comprehensive assessment model based on user health data according to claim 1, characterized in that, The feature engineering module includes: The feature filtering unit is configured to: calculate the variance contribution of each initial feature in the standardized multi-dimensional initial feature data and the significance of its association with the obesity assessment label, remove low-value initial features with low variance contribution and low significance of association, and form a filtered feature set; The feature transformation unit is configured to perform transformation operations on the temporal features, text features, and static numerical features in the filtered feature set respectively, and generate transformed features suitable for model training. The feature fusion unit is configured to concatenate and weight the transformed features to generate an engineered feature set.

4. The obesity comprehensive assessment model based on user health data according to claim 3, characterized in that, The calculation of the variance contribution includes the following steps: Detection and removal of outlier features based on box plot rules; The remaining effective feature values ​​are calculated based on the variance contribution of principal component analysis, and features with variance contribution below a preset threshold are removed. Then, the redundancy of the remaining features is detected by the Pearson correlation coefficient. If the absolute value of the Pearson correlation coefficient between two features is greater than the preset threshold, features with low correlation to the obesity assessment label are removed, and features with higher correlation are retained.

5. The obesity comprehensive assessment model based on user health data according to claim 3, characterized in that, The conversion operation includes the following steps: Derived temporal features were extracted from temporal features using a fixed time interval sliding window method. The derived temporal features included the average exercise intensity and the number of nighttime sleep fragmentation within the fixed time interval. The bag-of-words model is used to segment and encode the text features, and text transformation features of food categories and cooking methods are extracted. The static numerical features are mapped to the [0,1] numerical range using a normalization algorithm to generate static transformed features.

6. The obesity comprehensive assessment model based on user health data according to claim 5, characterized in that, The splicing and weighted fusion process includes the following steps: The derived temporal features, the text transformation features, and the static transformation features are concatenated into an initial feature matrix; Weights are assigned based on feature type identifiers and initial correlation, with sub-weights allocated to features within the same type according to the proportion of the absolute value of the initial correlation, ultimately generating an engineered feature set.

7. The obesity comprehensive assessment model based on user health data according to claim 6, characterized in that, The acquisition of the initial correlation degree includes the following steps in sequence: Determine the type of each feature in the engineering feature set, where physiological index features and environmental parameter features are defined as continuous numerical features, and dietary category features and time period identifier features are defined as discrete categorical features. For the continuous numerical features, the linear association strength between them and the obesity assessment label is calculated using the Pearson correlation coefficient; for the discrete categorical features, the categorical association strength between them and the obesity assessment label is calculated using the chi-square test. The calculated results of the linear correlation strength and the category correlation strength are normalized to obtain the preliminary correlation between each feature and obesity assessment.

8. The obesity comprehensive assessment model based on user health data according to claim 1, characterized in that, The method of analyzing the influence weights of each feature on the obesity assessment results using an interpretability analysis algorithm to generate an interpretable rule set includes the following steps: Based on the Shapley additive interpretable value algorithm, all possible feature subsets in the engineered feature set are generated. For each feature subset, the difference in obesity assessment results output by the model when the feature subset contains the target feature and when it does not is calculated as the marginal contribution of the target feature in that feature subset. Then, according to the size of each feature subset and the probability of uniform distribution, the marginal contribution of the target feature is weighted and averaged to obtain the contribution degree of the target feature to the obesity assessment result, and the absolute value of the contribution degree is used as the influence weight. Features with high weighting were selected as key features. The obesity risk values ​​corresponding to different numerical ranges of each key feature were statistically analyzed to determine the association rule that the obesity risk increases when the value of a key feature exceeds a specific range. For physiological indicators and exercise intensity features, the threshold rules for the normal range of body fat percentage and the normal range of daily exercise intensity for different genders were determined by referring to health and medical standards and combining the numerical ranges of healthy people in the feature value distribution. The key features were combined for testing, and the obesity risk value when the key features were combined was calculated and compared with the risk value when the single feature was combined to determine the synergistic effect rule of the increase in obesity risk under the combined feature effect.

9. The obesity comprehensive assessment model based on user health data according to claim 8, characterized in that, The dynamic obesity risk score is generated through the following steps: Extract the feature data from the output of the optimization layer, multiply each feature data by its corresponding influence weight, and then sum all the multiplication results to obtain the basic obesity risk score. The baseline obesity risk score is adjusted according to the interpretable rule set. If a key feature exceeds a specific range defined in the association rule, the baseline obesity risk score is increased according to the risk increase determined by the association rule; if a key feature falls within the normal range defined by the threshold rule, the baseline obesity risk score is maintained. If a combination of key features that meets the synergistic effect rule exists, the baseline obesity risk score will be further adjusted according to the risk increase determined by the synergistic effect rule. When a user's real-time health data is updated, the updated data is sequentially acquired by the data acquisition module, processed by the data preprocessing module, and converted into an engineered feature set by the feature engineering module. After being processed by the model training module, the influence weights are obtained and fed back to the optimization layer. Based on the updated key feature data and the corresponding influence weights, the basic risk score and score adjustment are repeatedly calculated to obtain the updated dynamic obesity risk score.

10. The application of the comprehensive obesity assessment model based on user health data according to any one of claims 1-9 in predicting user obesity risk.

Images