Human basal metabolic rate prediction method based on bioelectrical impedance and multi-modal fusion

Abstract

The application relates to the technical field of health monitoring, and provides a human basal metabolic rate prediction method based on bioelectrical impedance and multi-modal fusion. The method applies an alternating excitation signal of multiple frequencies to a human body through a multi-frequency bioelectrical impedance acquisition module, collects impedance data of the human body at multiple frequencies and extracts impedance features, simultaneously collects weight parameters, height parameters, age parameters, gender parameters, daily activity parameters, heart rate parameters and sleep parameters of the user and extracts physiological behavior features, inputs the impedance features and the physiological behavior features into a multi-modal fusion model to output a fusion feature vector through attention mechanism weighted fusion, and inputs the fusion feature vector into a pre-trained basal metabolic rate prediction model to calculate the real basal metabolic rate of the user, and simultaneously identifies a metabolic abnormality risk according to a deviation between the real basal metabolic rate and a standard basal metabolic rate. The method has the beneficial effects of high basal metabolic rate prediction accuracy, strong individual adaptability and automatic metabolic abnormality identification.

Description

Technical Field

[0001] This application relates to the field of health monitoring technology, and in particular to a method for predicting human basal metabolic rate based on bioelectrical impedance and multimodal fusion. Background Technology

[0002] The current method for calculating basal metabolic rate mainly uses empirical formulas, which calculate the estimated value of basal metabolic rate by substituting the user's weight, height, age and gender into a preset formula.

[0003] The above-mentioned empirical formula method has the following problems: it relies only on four static parameters—weight, height, age, and gender—and does not consider differences in body composition such as muscle mass and body fat percentage, thus failing to reflect an individual's true metabolic level; it does not take into account dynamic physiological behavioral characteristics such as daily activity levels, heart rate, and sleep patterns, resulting in a large deviation between the calculated results and the actual metabolic level; there are significant differences in metabolism among different individuals, and the fixed formula cannot adapt to individual differences, leading to insufficient calculation accuracy; and it lacks the ability to automatically identify metabolic abnormalities, preventing users from understanding their own metabolic health status in a timely manner.

[0004] Therefore, a method is urgently needed to solve at least one of the above problems. Summary of the Invention

[0005] This application provides a method and system for predicting human basal metabolic rate based on bioelectrical impedance and multimodal fusion, aiming to solve a series of problems in the existing basal metabolic rate calculation method, which mainly uses empirical formulas and calculates the estimated value of basal metabolic rate by substituting the user's weight, height, age and gender parameters into a preset formula.

[0006] In a first aspect, embodiments of this application provide a method for predicting human basal metabolic rate based on bioelectrical impedance and multimodal fusion, applied to a health management device including a multi-frequency bioelectrical impedance acquisition module, a multimodal data acquisition module, and a basal metabolic rate prediction module, comprising: The multi-frequency bioelectrical impedance acquisition module applies AC excitation signals of multiple frequencies to the human body, collects impedance data of the human body at multiple frequencies, and extracts impedance features; the multi-modal data acquisition module collects the user's weight parameters, height parameters, age parameters, gender parameters, daily activity parameters, heart rate parameters, and sleep parameters, and extracts physiological and behavioral features. Impedance features and physiological behavior features are input into a multimodal fusion model. The multimodal fusion model performs weighted fusion of impedance features and physiological behavior features through an attention mechanism and outputs a fused feature vector. The fused feature vector is input into the pre-trained basal metabolic rate prediction model to calculate the user's true basal metabolic rate. At the same time, the deviation between the true basal metabolic rate and the standard basal metabolic rate is used to identify the risk of metabolic abnormalities.

[0007] In some embodiments, the step of applying AC excitation signals of multiple frequencies to the human body through a multi-frequency bioelectrical impedance acquisition module and acquiring impedance data of the human body at multiple frequencies includes: placing electrode pads at multiple measurement locations on the user's body surface; applying AC excitation currents of different frequencies to the human body through the electrode pads, wherein the frequency of the AC excitation currents covers both low-frequency and high-frequency bands; acquiring the voltage response signals across the electrode pads at each frequency; calculating the human body impedance value at each frequency based on the AC excitation current and voltage response signals; and obtaining impedance data corresponding to multiple frequencies.

[0008] In some embodiments, the extraction of impedance features includes: performing curve fitting on impedance data corresponding to multiple frequencies to obtain impedance frequency characteristic curves; extracting feature parameters from the impedance frequency characteristic curves, the feature parameters including impedance amplitude, impedance phase angle, real part of impedance in the low-frequency band and real part of impedance in the high-frequency band corresponding to the feature frequency point; and organizing the feature parameters into an impedance feature vector as input to the subsequent multimodal fusion model.

[0009] In some embodiments, the step of collecting user's weight, height, age, gender, daily activity level, heart rate, and sleep parameters through a multimodal data acquisition module and extracting physiological behavioral features includes: collecting user's weight parameters through a scale; collecting user's height parameters through a height measurement module; collecting user's age and gender parameters through a user input module; collecting user's daily activity level and heart rate parameters through a smartwatch; and collecting user's sleep parameters through a sleep monitoring module; and organizing the weight, height, age, gender, daily activity level, heart rate, and sleep parameters into a physiological behavioral feature vector.

[0010] In some embodiments, the step of inputting impedance features and physiological behavior features into a multimodal fusion model, and the multimodal fusion model performing weighted fusion of impedance features and physiological behavior features through an attention mechanism to output a fused feature vector, includes: inputting the impedance feature vector into a first branch of the multimodal fusion model, and inputting the physiological behavior feature vector into a second branch of the multimodal fusion model; the multimodal fusion model calculating the attention weights of the impedance feature vector and the physiological behavior feature vector respectively through an attention mechanism; and performing a weighted summation of the impedance feature vector and the physiological behavior feature vector according to the attention weights to output a fused feature vector.

[0011] In some embodiments, the step of inputting the fused feature vector into a pre-trained basal metabolic rate prediction model to calculate the user's true basal metabolic rate includes: inputting the fused feature vector into the pre-trained basal metabolic rate prediction model; the basal metabolic rate prediction model performing regression calculation based on the fused feature vector to output the user's true basal metabolic rate value; cross-validating the true basal metabolic rate value with the user's input weight, height, and age parameters; re-performing the prediction calculation when the verification deviation exceeds a preset allowable range; and outputting the final true basal metabolic rate value when the verification deviation falls within the preset allowable range.

[0012] In some embodiments, identifying metabolic abnormality risk based on the deviation between the actual basal metabolic rate and the standard basal metabolic rate includes: querying the corresponding standard basal metabolic rate from a standard basal metabolic rate comparison table based on the user's weight, height, age, and gender parameters; calculating the deviation rate between the actual basal metabolic rate and the standard basal metabolic rate; determining a low metabolic risk when the deviation rate is lower than a preset low threshold, a high metabolic risk when the deviation rate is higher than a preset high threshold, and a normal metabolic risk when the deviation rate is within a preset normal range.

[0013] In some embodiments, before inputting impedance features and physiological behavior features into the multimodal fusion model, the method further includes: collecting impedance features and physiological behavior features of historical user samples and corresponding actual basal metabolic rate labels; inputting the impedance features of historical user samples into a first feature extraction network and the physiological behavior features of historical user samples into a second feature extraction network; performing weighted fusion of the outputs of the first feature extraction network and the second feature extraction network through an attention mechanism to train the network parameters of the multimodal fusion model; and storing the network parameters of the trained multimodal fusion model into the running memory of the health management device.

[0014] In some embodiments, before inputting the fused feature vector into the pre-trained basal metabolic rate prediction model, the method further includes: collecting fused feature vectors and corresponding actual basal metabolic rate labels from historical user samples to construct a training dataset; dividing the training dataset into a training subset and a validation subset according to a preset ratio; performing model training on the training subset and model validation on the validation subset using a regression algorithm; stopping training when the prediction error of the validation set falls within a preset error range to obtain the pre-trained basal metabolic rate prediction model; and storing the model parameters of the pre-trained basal metabolic rate prediction model in the running memory of the health management device.

[0015] In some embodiments, the method further includes: collecting the true basal metabolic rate for each monitoring cycle within multiple consecutive monitoring cycles; generating a basal metabolic rate change curve based on the true basal metabolic rate for multiple consecutive monitoring cycles; inputting the basal metabolic rate change curve into a trend analysis model; the trend analysis model identifies metabolic trend types based on the basal metabolic rate change curve; metabolic trend types include a continuous upward trend, a continuous downward trend, and a fluctuating trend; generating personalized health recommendations based on the metabolic trend type and the user's physiological behavioral characteristics; and pushing the personalized health recommendations to the user's terminal.

[0016] The method provided in this application has the following beneficial effects: 1. By collecting human body impedance data through multi-frequency AC excitation and extracting impedance frequency characteristics, the data reflects information on human body composition and muscle mass, breaking through the limitations of traditional calculations that rely solely on weight and height.

[0017] 2. By integrating impedance characteristics with physiological and behavioral characteristics such as weight, height, age, gender, daily activity level, heart rate, and sleep through an attention mechanism, the system comprehensively reflects the factors influencing an individual's metabolism.

[0018] 3. The pre-trained basal metabolic rate prediction model calculates the true basal metabolic rate based on the fused feature vector, and the prediction accuracy far exceeds that of traditional empirical formulas.

[0019] 4. Automatically identify the risk of low or high metabolism based on the deviation between the actual basal metabolic rate and the standard basal metabolic rate, helping users to understand their metabolic health status in a timely manner.

[0020] 5. Identify metabolic trend types by analyzing basal metabolic rate change curves over multiple consecutive monitoring periods, and generate personalized health recommendations.

[0021] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a schematic flowchart illustrating the steps of a method for predicting human basal metabolic rate based on bioelectrical impedance and multimodal fusion, provided in an embodiment of this application. Figure 2This is a schematic diagram illustrating the principle of a method for predicting human basal metabolic rate based on bioelectrical impedance and multimodal fusion, provided in an embodiment of this application. Figure 3 This is a schematic block diagram of a human basal metabolic rate prediction system based on bioelectrical impedance and multimodal fusion provided in an embodiment of this application; Figure 4 This is a schematic block diagram of the structure of a health management device provided in one embodiment of this application.

[0024] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Detailed Implementation

[0025] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content and operations / steps, nor does it necessarily have to be performed in the order described. For example, some operations / steps can be broken down, combined, or partially merged, so the actual execution order may change depending on the actual situation.

[0027] It should be understood that, in order to clearly describe the technical solutions of the embodiments of the present invention, the terms "first" and "second" are used in the embodiments of the present invention to distinguish identical or similar items with essentially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and the terms "first" and "second" are not necessarily different.

[0028] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of the application. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0029] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0030] The current method for calculating basal metabolic rate mainly uses empirical formulas, which calculate the estimated value of basal metabolic rate by substituting the user's weight, height, age and gender into a preset formula.

[0031] The above-mentioned empirical formula method has the following problems: it relies only on four static parameters—weight, height, age, and gender—and does not consider differences in body composition such as muscle mass and body fat percentage, thus failing to reflect an individual's true metabolic level; it does not take into account dynamic physiological behavioral characteristics such as daily activity levels, heart rate, and sleep patterns, resulting in a large deviation between the calculated results and the actual metabolic level; there are significant differences in metabolism among different individuals, and the fixed formula cannot adapt to individual differences, leading to insufficient calculation accuracy; and it lacks the ability to automatically identify metabolic abnormalities, preventing users from understanding their own metabolic health status in a timely manner.

[0032] Therefore, a method is urgently needed to solve at least one of the above problems.

[0033] Please refer to Figure 1 This application provides a method for predicting human basal metabolic rate based on bioelectrical impedance and multimodal fusion, applicable to a health management device including a multi-frequency bioelectrical impedance acquisition module, a multimodal data acquisition module, and a basal metabolic rate prediction module. It should be noted that all information involved in the method provided in this application is extracted with the authorization of the relevant user and in accordance with relevant regulations, and will not infringe on user privacy.

[0034] Please refer to Figure 2 This application provides a human basal metabolic rate prediction system based on bioelectrical impedance and multimodal fusion, which is applied to health management devices (such as intelligent height and weight measuring instruments), including a multi-frequency bioelectrical impedance acquisition module, a multimodal data acquisition module, a multimodal fusion model and a basal metabolic rate prediction model.

[0035] The multi-frequency bioelectrical impedance acquisition module employs a four-electrode method, placing electrode pads (both wrist and ankle electrodes) at multiple measurement locations on the user's body. Weak alternating currents of different frequencies (typically 50μA to 500μA, covering low-frequency bands such as 5kHz to 50kHz and high-frequency bands such as 50kHz to 500kHz) are applied to the body through these electrode pads. The voltage response signal across the electrode pads at each frequency is acquired, and the human body impedance data corresponding to multiple frequencies is calculated. Low-frequency currents primarily flow through extracellular fluid, while high-frequency currents can penetrate cell membranes and flow through intracellular fluid. The multi-frequency impedance data reflects the electrical characteristics of different tissue components in the human body and can be used to estimate body composition indicators such as muscle mass and body fat percentage.

[0036] The multimodal data acquisition module includes a weight scale (for collecting weight parameters), a height measurement column (for collecting height parameters), a user input module (for collecting age and gender parameters), and a smart wearable device that can be linked with the body composition analyzer (a smartwatch / bracelet for collecting daily activity and heart rate parameters, and a sleep monitoring module for collecting sleep parameters). The weight scale uses a pressure sensor array with an accuracy of 0.1 kg; the height measurement column uses ultrasonic or infrared ranging with an accuracy of 0.1 cm.

[0037] The multimodal fusion model is a deep learning model that runs on the device's main control chip or a cloud server. It performs weighted fusion of impedance features and physiological behavior features through an attention mechanism, automatically learns the contribution weights of different features to the prediction of basal metabolic rate, and outputs a fused feature vector.

[0038] The basal metabolic rate prediction model is a pre-trained regression model that takes a fused feature vector as input and outputs the user's true basal metabolic rate (in kcal / day). Based on the deviation between the true value and the standard value, it identifies the risk status of low, normal or high metabolism and pushes the results to the user's terminal (mobile app or device display screen).

[0039] The overall workflow of the above system is as follows: The user stands barefoot on the scale, holds the electrode pads with both hands or attaches the electrodes to the wrists. The body composition analyzer starts multi-frequency bioelectrical impedance measurement to collect weight and height, and at the same time synchronizes daily activity, heart rate and sleep data from the smartwatch. After the impedance characteristics and physiological behavior characteristics are fused through the attention mechanism, they are input into the pre-trained basal metabolic rate prediction model to calculate the true basal metabolic rate, and compare it with the standard value to identify the risk of metabolic abnormalities. The results are then output to the user terminal.

[0040] The provided method for predicting human basal metabolic rate based on bioelectrical impedance and multimodal fusion includes steps S101 to S103. Details are as follows: Step S101. Apply AC excitation signals of multiple frequencies to the human body through the multi-frequency bioelectrical impedance acquisition module, collect impedance data of the human body at multiple frequencies, and extract impedance features; collect the user's weight parameters, height parameters, age parameters, gender parameters, daily activity parameters, heart rate parameters, and sleep parameters through the multi-modal data acquisition module, and extract physiological and behavioral features.

[0041] Specifically, the core function of step S101 is to simultaneously collect two types of input data that reflect the user's metabolic level: one is multi-frequency bioelectrical impedance characteristics that reflect differences in body composition, and the other is multi-modal physiological behavior characteristics that reflect individual basic attributes and dynamic physiological behaviors, providing complete and complementary inputs for subsequent multimodal fusion and basal metabolic rate prediction.

[0042] The multi-frequency bioelectrical impedance acquisition module applies AC excitation signals of multiple frequencies to the human body, acquiring impedance data at these frequencies and extracting impedance characteristics. The user stands barefoot on the scale, with each foot contacting one set of electrode pads on the scale surface (one on the left heel and one on the forefoot). The user holds the handheld electrodes or wears electrode rings on their wrists, forming a current loop. The signal source of the multi-frequency bioelectrical impedance acquisition module applies a weak, frequency-adjustable AC excitation current (100μA, meeting safety standards and causing no stinging sensation) to the human body, covering both low and high frequency bands: typical low-frequency band frequencies are 5kHz and 50kHz, and typical high-frequency band frequencies are 250kHz and 500kHz, totaling 4 to 8 frequency points. At each frequency, the measurement circuit acquires the voltage response signal across the electrode pads and calculates the human body impedance value at that frequency according to Ohm's law (impedance = voltage divided by current), including the impedance amplitude (modulus) and impedance phase angle (the angle corresponding to the ratio of the real part to the imaginary part). Low-frequency currents primarily flow through extracellular fluid, reflecting body fluid distribution and the electrical properties of adipose tissue; high-frequency currents penetrate cell membranes and flow through intracellular fluid, reflecting muscle tissue content. After filtering and analog-to-digital conversion, the multi-frequency impedance data is fitted using curve fitting (e.g., Cole-Cole circle model fitting) to obtain impedance-frequency characteristic curves. Feature parameters are extracted from these curves: impedance amplitude at the characteristic frequency point (the characteristic frequency fc corresponding to the intersection of the impedance trajectory and the real axis), phase angle at the characteristic frequency point, the real part R0 of the low-frequency band (5kHz) impedance (primarily reflecting extracellular fluid impedance), the real part R∞ of the high-frequency band (500kHz) impedance (reflecting the total impedance of intracellular and extracellular fluids), and the difference between the two (R0-R∞, related to muscle water content and muscle mass). These feature parameters are organized into an impedance feature vector, which serves as the input to the subsequent multimodal fusion model.

[0043] The system collects user parameters such as weight, height, age, gender, daily activity level, heart rate, and sleep parameters through a multimodal data acquisition module, extracting physiological and behavioral characteristics. Weight parameters are collected via four-point pressure sensors on the scale, with an accuracy of 0.1 kg; height parameters are collected via an ultrasonic ranging module on the height measuring column, with an accuracy of 0.1 cm; age and gender parameters are input by the user via the device's touchscreen or a mobile app (gender is coded as male=0, female=1). Daily activity level parameters are collected via a smartwatch linked to the body composition analyzer, including average daily steps, average daily energy expenditure (kcal / day), duration of moderate-to-vigorous intensity exercise (min / day), and exercise frequency (times / week); heart rate parameters are collected via the smartwatch's photoelectric pulse sensor, including resting heart rate (beats / min) and average heart rate; sleep parameters are collected via the smartwatch's accelerometer and heart rate sensor, including average daily sleep duration (h), percentage of deep sleep (%), percentage of light sleep (%), sleep efficiency (%), and number of awakenings after falling asleep. Parameters such as weight, height, age, gender, daily activity level, heart rate, and sleep are organized into physiological behavioral feature vectors, which are then input into the subsequent multimodal fusion model along with impedance feature vectors. The two types of features are aligned on the timestamps to reflect the user's body composition and physiological behavioral state at the same point in time.

[0044] Step S102. Input the impedance features and physiological behavior features into the multimodal fusion model. The multimodal fusion model performs weighted fusion of the impedance features and physiological behavior features through an attention mechanism and outputs a fused feature vector.

[0045] Specifically, the core function of step S102 is to effectively fuse impedance features with different modalities, dimensions, and physical meanings with physiological behavioral features through a multimodal fusion model, automatically learn the contribution weights of the two types of features to the prediction of basal metabolic rate using an attention mechanism, and output a unified fused feature vector for use by the basal metabolic rate prediction model.

[0046] The multimodal fusion model employs a dual-branch network structure. Impedance feature vectors are input to the first branch (impedance feature branch), which is a multi-layer fully connected network (e.g., two fully connected layers, 64 to 32 nodes). This branch performs a nonlinear transformation on the impedance feature vectors to extract higher-order impedance semantic features. Physiological behavior feature vectors are input to the second branch (physiological behavior feature branch), which is also a multi-layer fully connected network (e.g., two fully connected layers, 64 to 32 nodes). This branch performs nonlinear transformations on parameters such as weight, height, age, gender, activity level, heart rate, and sleep to extract higher-order physiological behavior semantic features. After the feature dimensions output from both branches are unified, they enter the attention fusion module.

[0047] The attention mechanism calculates attention weights for impedance features and physiological behavior features separately. The attention fusion module takes the concatenated feature vector from the two branches as input and calculates the attention weight coefficients α1 and α2 for each modality (impedance modality and physiological behavior modality) using a small multilayer perceptron (MLP) or softmax function, where α1 + α2 = 1. These weight coefficients are automatically learned from training data: for example, for users with abnormal body fat percentages, where body composition varies greatly, impedance features contribute more significantly to metabolic rate prediction, and the model automatically increases the attention weight of the impedance modality; for users with high daily activity levels and high heart rate variability, behavioral features contribute more significantly, and the model automatically increases the weight of the behavioral modality. The attention weights are not fixed values ​​but are dynamically calculated based on the current input feature vector, achieving adaptive weighting "tailored to the sample."

[0048] The impedance feature vector and the physiological behavior feature vector are weighted and summed based on attention weights: fused feature vector = α1 × impedance branch output feature + α2 multiplied by physiological behavior branch output feature. The weighted fused feature vector retains both body composition information (from impedance) and individual attribute and behavioral information (from physiological behavior), and the contribution weights of each part are automatically learned and optimized by the model. The fused feature vector is typically 32-dimensional or 64-dimensional and is output to the basal metabolic rate prediction model. Compared to simple feature concatenation (directly concatenating two types of features into a long vector input to the regression model), attention-weighted fusion can suppress the interference of noisy features, amplify the contribution of key features, and improve prediction accuracy. At the same time, the weights are interpretable, making it easy to analyze which features are more important for metabolic rate prediction.

[0049] Step S103. Input the fused feature vector into the pre-trained basal metabolic rate prediction model to calculate the user's true basal metabolic rate, and identify the risk of metabolic abnormalities based on the deviation between the true basal metabolic rate and the standard basal metabolic rate.

[0050] Specifically, the core function of step S103 is to input the fused feature vector into the pre-trained basal metabolic rate prediction model, perform regression calculation to obtain the user's true basal metabolic rate value, and compare the predicted value with the standard reference value of the same age, gender, weight and height to identify the risk status of low, normal or high metabolism, and provide the user with a quantitative metabolic health assessment.

[0051] The fused feature vector is input into a pre-trained basal metabolic rate (BMR) prediction model. This BMR prediction model is a pre-trained regression model, which can employ algorithms such as Gradient Boosting Regression Tree (GBRT), Multilayer Perceptron, or Support Vector Regression. The model takes the fused feature vector as input and outputs a scalar value (kcal / day) through regression calculation: the user's true BMR. During the training phase, the model undergoes supervised learning using measured metabolic rates (e.g., through indirect calorimetry or the gold standard measurement from a metabolic chamber) as labels, learning the non-linear mapping relationship from fused features to metabolic rate. During prediction, forward computation is performed on the current user's fused feature vector, outputting the metabolic rate value. To ensure the reliability of the results, the predicted true basal metabolic rate is cross-validated with the estimated value obtained by substituting the user's input weight, height, age, and gender into a classic empirical formula (such as the Mifflin-StJeor formula). If the deviation between the predicted value and the formula's estimated value is within a preset allowable range (e.g., within ±15%), the result is considered reasonable, and the true basal metabolic rate is output. If the deviation is too large (e.g., exceeding ±20%), it indicates that there may be abnormal sensor data or feature extraction errors, triggering data re-collection or re-prediction to avoid outputting erroneous results.

[0052] Metabolic abnormality risk is identified by the deviation between the actual basal metabolic rate and the standard basal metabolic rate (BMR). Based on the user's weight, height, age, and gender parameters, the corresponding standard basal metabolic rate (BMR_std) is retrieved from a standard basal metabolic rate reference table (or a standard empirical formula). The standard reference value is calculated based on gender, age range, weight, and height, representing the average level of a population with similar conditions. The deviation rate is calculated as (actual BMR - BMR_std) divided by BMR_std multiplied by 100%. The metabolic risk level is determined based on the deviation rate: When the deviation rate is below the preset low threshold (e.g., less than -10%), it is considered a low metabolic risk, indicating that the user's basal metabolic rate is lower than the average level of people of the same age and body type, which may indicate insufficient muscle mass, hypothyroidism, or sedentary lifestyle. When the deviation rate is above the preset high threshold (e.g., greater than +10%), it is considered a high metabolic risk, indicating that the user's metabolic rate is high (which may be related to high muscle mass, hyperthyroidism, or recent excessive exercise, and needs to be judged in conjunction with other indicators). When the deviation rate is within the preset normal range (-10% to +10%), it is considered normal metabolism. The risk level, along with the actual basal metabolic rate value, is output to the user's terminal (mobile app or device screen), along with textual interpretation and lifestyle suggestions. Compared to the traditional empirical formula method, which only outputs a static estimate, this step uses a fusion model to obtain a metabolic rate that is closer to the individual's actual level and automatically provides risk stratification and health tips to help users understand their own metabolic health status in a timely manner.

[0053] In some embodiments, the step of applying AC excitation signals of multiple frequencies to the human body through a multi-frequency bioelectrical impedance acquisition module and acquiring impedance data of the human body at multiple frequencies includes: placing electrode pads at multiple measurement locations on the user's body surface; applying AC excitation currents of different frequencies to the human body through the electrode pads, wherein the frequency of the AC excitation currents covers both low-frequency and high-frequency bands; acquiring the voltage response signals across the electrode pads at each frequency; calculating the human body impedance value at each frequency based on the AC excitation current and voltage response signals; and obtaining impedance data corresponding to multiple frequencies.

[0054] Specifically, the multi-frequency bioelectrical impedance analysis (BIA) uses a four-electrode method. The electrode positions are: right heel electrode, right forefoot electrode, left wrist electrode, and right wrist electrode (both feet are on the scale surface electrodes, and both hands are holding the handle electrodes or wearing the electrodes on the wrists). Two of the four electrodes are current electrodes, and two are voltage electrodes, eliminating the influence of electrode contact impedance. Before data acquisition, the user is required to fast, empty their bladder, and stand still for 2 to 5 minutes to avoid fluctuations in body fluid distribution affecting the measurement. The signal source of the multi-frequency bioelectrical impedance analysis module is a high-precision sinusoidal current source, outputting AC excitation current at multiple frequency points. The typical frequency points are 5kHz, 50kHz, 100kHz, 250kHz, and 500kHz, a total of 5 frequencies, with a current intensity of 100μA (compliant with IEC60601 safety standards, harmless and imperceptible to the human body). At each frequency, the voltage measurement circuit (instrumentation amplifier + synchronous detector) acquires the voltage response signal between the two voltage electrodes, measuring the voltage amplitude and phase. According to Ohm's law, at this frequency, the human body impedance value equals the voltage amplitude divided by the current amplitude, and the impedance phase angle equals the voltage phase minus the current phase. The impedance is decomposed into a real part (resistive component R) and an imaginary part (reactant component Xc, capacitive). By sequentially switching frequency points and repeating the measurement, impedance amplitude, phase angle, and real and imaginary part data for multiple frequencies are obtained. The measurement process is completed within 30 to 60 seconds. The data is digitally filtered to remove power frequency interference (50Hz) before being stored as the raw data for subsequent feature extraction.

[0055] In some embodiments, the extraction of impedance features includes: performing curve fitting on impedance data corresponding to multiple frequencies to obtain impedance frequency characteristic curves; extracting feature parameters from the impedance frequency characteristic curves, the feature parameters including impedance amplitude, impedance phase angle, real part of impedance in the low-frequency band and real part of impedance in the high-frequency band corresponding to the feature frequency point; and organizing the feature parameters into an impedance feature vector as input to the subsequent multimodal fusion model.

[0056] Specifically, the Cole-Cole semicircle model is fitted to the impedance data corresponding to multiple frequencies: On the complex impedance plane (the horizontal axis is the real part of resistance R, and the vertical axis is the capacitive reactance -Xc), the human body impedance trajectory is approximately an arc. The arc is fitted by the nonlinear least squares method to obtain the model parameters: the real part of the impedance R0 in the low-frequency band (the resistance when the frequency approaches 0, mainly representing the extracellular fluid resistance), the real part of the impedance R∞ in the high-frequency band (the resistance when the frequency approaches infinity, representing the parallel resistance of the extracellular fluid and intracellular fluid), the characteristic frequency fc (the frequency corresponding to the apex of the arc, i.e. the frequency point with the strongest capacitance, which is related to the cell membrane capacitance and body fluid resistance), the impedance amplitude Zc and the impedance phase angle φmax at the characteristic frequency. The characteristic parameters extracted from the impedance frequency response curve include: R0 (real part of impedance in the low-frequency band), R∞ (real part of impedance in the high-frequency band), the difference between R0 and R∞ (R0-R∞, positively correlated with intracellular fluid volume, which can be used to estimate muscle water content and muscle mass), characteristic frequency fc, impedance amplitude Zc at the characteristic frequency point, and maximum phase angle φmax. These parameters reflect the composition of human tissue from different perspectives: the lower the R0, the more extracellular fluid or the lower the resistance of adipose tissue; the larger the R0-R∞, the more muscle mass and the richer the intracellular fluid; and the lower the fc, the larger the cell membrane capacitance (high proportion of muscle tissue). The above 6 to 8 characteristic parameters are organized into an impedance characteristic vector (e.g., [R0,R∞,R0-R∞,fc,Zc,φmax]), standardized (Z-score standardization, subtracting the mean and dividing by the standard deviation) to eliminate the influence of dimensions, and then used as the input of the first branch of the subsequent multimodal fusion model.

[0057] In some embodiments, the step of collecting user's weight, height, age, gender, daily activity level, heart rate, and sleep parameters through a multimodal data acquisition module and extracting physiological behavioral features includes: collecting user's weight parameters through a scale; collecting user's height parameters through a height measurement module; collecting user's age and gender parameters through a user input module; collecting user's daily activity level and heart rate parameters through a smartwatch; and collecting user's sleep parameters through a sleep monitoring module; and organizing the weight, height, age, gender, daily activity level, heart rate, and sleep parameters into a physiological behavioral feature vector.

[0058] Specifically, weight parameters are collected through four strain gauge pressure sensors inside the scale base. When the sensors are subjected to force, the elastic body deforms, causing a change in resistance. This change is converted into a voltage signal via a Wheatstone bridge, amplified, sampled by an A / D converter, and then weighted at four points to obtain the weight value (unit: kg, accuracy: 0.1 kg). Height parameters are collected through an ultrasonic ranging module at the top of the height measuring column. The ultrasonic transmitter emits ultrasonic waves upwards, receives the reflected echoes, and calculates the height (unit: cm, accuracy: 0.1 cm) based on the flight time, or by using an infrared through-beam sensor array for segmented measurement. Age and gender parameters are entered by the user upon first use via the device's touchscreen or mobile app and stored in the user's account. Gender is coded as Male = 0, Female = 1. Daily activity level and heart rate parameters were collected via a smartwatch and synchronized with the fitness tracker via Bluetooth or Wi-Fi. Daily activity level parameters included average daily steps (steps / day), average daily energy expenditure (kcal / day), duration of moderate-to-vigorous intensity activity (min / day), and estimated resting energy expenditure. Heart rate parameters included resting heart rate (beats / min, morning resting state) and the average heart rate over the past 7 days. Sleep parameters were collected via a sleep monitoring module (smartwatch accelerometer + heart rate sensor), including average total daily sleep duration (h), percentage of deep sleep (%), percentage of light sleep (%), percentage of REM sleep (%), sleep efficiency (%, actual sleep time / bedtime), and number of nighttime awakenings. These parameters were organized into a physiological behavior feature vector, including approximately 12 to 15 dimensions such as weight, height, age, gender coding, average daily steps, energy expenditure, duration of moderate-to-vigorous intensity activity, resting heart rate, average heart rate, sleep duration, percentage of deep sleep, and sleep efficiency. After standardization, these parameters served as input to the second branch of the multimodal fusion model. Data collection adheres to user authorization and privacy protection requirements, and is only collected and synchronized with the user's consent.

[0059] In some embodiments, the step of inputting impedance features and physiological behavior features into a multimodal fusion model, and the multimodal fusion model performing weighted fusion of impedance features and physiological behavior features through an attention mechanism to output a fused feature vector, includes: inputting the impedance feature vector into a first branch of the multimodal fusion model, and inputting the physiological behavior feature vector into a second branch of the multimodal fusion model; the multimodal fusion model calculating the attention weights of the impedance feature vector and the physiological behavior feature vector respectively through an attention mechanism; and performing a weighted summation of the impedance feature vector and the physiological behavior feature vector according to the attention weights to output a fused feature vector.

[0060] Specifically, the multimodal fusion model is deployed on the device's main control microprocessor or a cloud server. Impedance feature vectors (6 to 8 dimensions) are input to the first branch network, which consists of two fully connected layers (32 nodes in the first layer, ReLU activation function; 16 nodes in the second layer, ReLU activation function), outputting 16-dimensional higher-order impedance features. Physiological behavior feature vectors (12 to 15 dimensions) are input to the second branch network, with a structure symmetrical to the first branch (two fully connected layers, 32 to 16 nodes), outputting 16-dimensional higher-order physiological behavior features. The attention fusion module receives the two 16-dimensional features, concatenates them into a 32-dimensional vector, and then calculates the attention weights α1 (impedance modality weight) and α2 (physiological behavior modality weight) for the two modalities through a single fully connected network (outputting 2 nodes, activated by Softmax), where α1 + α2 = 1. Attention weights dynamically change based on the current input. For example, α1 automatically increases when there are abnormalities in the impedance features (R0-R∞, indicating significant differences in muscle mass compared to peers), and α2 automatically increases when there is high discriminative power between daily activity levels and heart rate features. Weighted fusion: Fusion feature = α1 multiplied by the 16-dimensional output of the impedance branch + α2 multiplied by the 16-dimensional output of the physiological behavior branch, resulting in a 16-dimensional fusion feature vector. This fusion method preserves the complementary information of body composition and behavioral features, and automatically suppresses noise and highlights key modalities through attention weights, improving feature quality compared to direct concatenation. The fusion feature vector is output to the basal metabolic rate prediction model. After model training, the parameters are permanently stored in the device's memory. During inference, a single forward computation takes less than 100ms, meeting real-time measurement requirements.

[0061] In some embodiments, the step of inputting the fused feature vector into a pre-trained basal metabolic rate prediction model to calculate the user's true basal metabolic rate includes: inputting the fused feature vector into the pre-trained basal metabolic rate prediction model; the basal metabolic rate prediction model performing regression calculation based on the fused feature vector to output the user's true basal metabolic rate value; cross-validating the true basal metabolic rate value with the user's input weight, height, and age parameters; re-performing the prediction calculation when the verification deviation exceeds a preset allowable range; and outputting the final true basal metabolic rate value when the verification deviation falls within the preset allowable range.

[0062] Specifically, the 16-dimensional fused feature vector is input into a pre-trained basal metabolic rate (BMR) prediction model. The model employs a GBRT regressor (100 decision trees, maximum depth 4) or a small fully connected regression network. Using the fused features as input, the model calculates a scalar value after regression: the true BMR value (in kcal / day). The prediction result represents an estimate of the user's daily energy expenditure at rest, typically approximately 1500-2000 kcal / day for men and 1200-1600 kcal / day for women. To ensure the reliability of the results, cross-validation is performed: the predicted value is compared with the estimated value using the classic Mifflin-StJeor formula: for men, BMR = 10 x weight + 6.25 x height - 5 x age + 5; for women, BMR = 10 x weight + 6.25 x height - 5 x age - 161. The deviation rate is calculated as (model predicted value - formula estimated value) divided by the formula estimated value multiplied by 100%. When the deviation rate falls within the preset allowable range (e.g., within ±15%), the result is deemed reasonable, and the final true basal metabolic rate value is output. When the deviation rate exceeds the allowable range (e.g., greater than ±20%), sensor data anomalies are suspected (e.g., poor electrode contact, weight measurement error, feature extraction error). The control system prompts for remeasurement or re-execution of feature extraction and prediction until the result is reasonable, avoiding the output of abnormal erroneous values. The cross-validation mechanism uses mature empirical formulas as a reasonableness reference to perform fallback verification on the model prediction results, improving the robustness of the system.

[0063] In some embodiments, identifying metabolic abnormality risk based on the deviation between the actual basal metabolic rate and the standard basal metabolic rate includes: querying the corresponding standard basal metabolic rate from a standard basal metabolic rate comparison table based on the user's weight, height, age, and gender parameters; calculating the deviation rate between the actual basal metabolic rate and the standard basal metabolic rate; determining a low metabolic risk when the deviation rate is lower than a preset low threshold, a high metabolic risk when the deviation rate is higher than a preset high threshold, and a normal metabolic risk when the deviation rate is within a preset normal range.

[0064] Specifically, based on the user's weight, height, age, and gender, the standard basal metabolic rate (BMR_std) is calculated using the Mifflin-StJeor formula or the Harris-Benedict formula, serving as an average reference value for a population with similar conditions. The deviation rate Δ is calculated as (BMR_true - BMR_std) divided by BMR_std multiplied by 100%. The risk level is categorized by Δ: When Δ is less than -10% (i.e., the true metabolic rate is more than 10% lower than the standard value), it is considered a low metabolic risk, indicating a low metabolic rate. Common causes include insufficient muscle mass, sedentary lifestyle, low calorie intake, and hypothyroidism. It is recommended to increase strength training, moderately increase calorie intake, and pay attention to thyroid health. When Δ is greater than +10% (i.e., the true metabolic rate is more than 10% higher than the standard value), it is considered a high metabolic risk, indicating a high metabolic rate. This may be related to high muscle mass, high exercise intensity, and hyperthyroidism, and requires comprehensive judgment in conjunction with body composition and other indicators. When Δ is in the range of -10% to +10%, it is considered a normal metabolic rate. The risk level, along with the BMR value, a bar chart comparing it with the standard value, and personalized interpretation text, are displayed to the user through the device screen or mobile app to help the user intuitively understand their own metabolic level and relative position.

[0065] In some embodiments, before inputting impedance features and physiological behavior features into the multimodal fusion model, the method further includes: collecting impedance features and physiological behavior features of historical user samples and corresponding actual basal metabolic rate labels; inputting the impedance features of historical user samples into a first feature extraction network and the physiological behavior features of historical user samples into a second feature extraction network; performing weighted fusion of the outputs of the first feature extraction network and the second feature extraction network through an attention mechanism to train the network parameters of the multimodal fusion model; and storing the network parameters of the trained multimodal fusion model into the running memory of the health management device.

[0066] Specifically, model training is completed offline in the cloud or on a high-performance server. A large-scale historical user sample dataset is collected, with each sample containing the same user's impedance feature vector (R0, R∞, R0-R∞, fc, Zc, φmax, etc.), physiological behavior feature vector (weight, height, age, gender, activity level, heart rate, sleep, etc.), and the corresponding actual basal metabolic rate gold standard label: the label is accurately determined by indirect calorimetry (measuring oxygen consumption and carbon dioxide production with a breathing mask to estimate metabolic rate) or metabolic chamber method, covering people of different genders, age groups, weight ranges, and body fat percentage levels, with a sample size typically ranging from several thousand to tens of thousands of cases. During training, impedance features are input into the first feature extraction network (two fully connected layers), and physiological behavior features are input into the second feature extraction network (two fully connected layers). The attention module calculates the weights of the two modalities and performs weighted fusion to obtain the fused feature. The fused feature is then passed to the subsequent regression network to output the metabolic rate prediction value. The mean squared error between the predicted value and the gold standard label is used as the loss function. The parameters of the two branch networks, the attention module, and the regression network are iteratively updated using the backpropagation algorithm. The Adam optimizer with a learning rate of 0.001 and a batch size of 32 is used for training until the loss converges. After training, the network parameters (weight matrices and bias vectors of each layer) of the multimodal fusion model are exported in a lightweight format (such as ONNX or TFLite) and deployed and stored in the runtime memory of the health management device (such as the Flash memory of the body composition analyzer or a cloud inference service). During device-side inference, the parameters are directly loaded to perform forward calculations, eliminating the need to access the server each time, thus balancing prediction accuracy and response speed.

[0067] In some embodiments, before inputting the fused feature vector into the pre-trained basal metabolic rate prediction model, the method further includes: collecting fused feature vectors and corresponding actual basal metabolic rate labels from historical user samples to construct a training dataset; dividing the training dataset into a training subset and a validation subset according to a preset ratio; performing model training on the training subset and model validation on the validation subset using a regression algorithm; stopping training when the prediction error of the validation set falls within a preset error range to obtain the pre-trained basal metabolic rate prediction model; and storing the model parameters of the pre-trained basal metabolic rate prediction model in the running memory of the health management device.

[0068] Specifically, the training of the basal metabolic rate (BMR) prediction model is completed jointly or independently with the fusion model. A training dataset is constructed by collecting fusion feature vectors (extracted from the trained fusion model) from historical user samples and their corresponding gold standard labels for actual BMR, randomly dividing the dataset into an 80% training subset and a 20% validation subset. The regression algorithm can be GBRT, random forest, or multilayer perceptron regressor: on the training subset, the model is trained by minimizing the mean squared error using fusion features as input and the gold standard metabolic rate as the label; on the validation subset, the mean absolute error (MAE) and mean absolute percentage error (MAPE) between the predicted and actual values ​​are calculated. When the validation set MAPE falls within a preset error range (e.g., less than 8% to 10%, far superior to the typical error of 15% to 20% for traditional empirical formulas), training stops, resulting in a pre-trained BMR prediction model. The model parameters (the tree structure and splitting threshold of GBRT, or the weights of the regression network) are stored in the device's memory. The model can be periodically retrained and updated with newly accumulated user data to continuously improve prediction accuracy. Compared to empirical formulas that use only four linear variables—weight, height, age, and gender—the fusion model significantly reduces prediction error through a nonlinear mapping of body composition impedance characteristics and dynamic behavioral characteristics.

[0069] In some embodiments, the method further includes: collecting the true basal metabolic rate for each monitoring cycle within multiple consecutive monitoring cycles; generating a basal metabolic rate change curve based on the true basal metabolic rate for multiple consecutive monitoring cycles; inputting the basal metabolic rate change curve into a trend analysis model; the trend analysis model identifies metabolic trend types based on the basal metabolic rate change curve; metabolic trend types include a continuous upward trend, a continuous downward trend, and a fluctuating trend; generating personalized health recommendations based on the metabolic trend type and the user's physiological behavioral characteristics; and pushing the personalized health recommendations to the user's terminal.

[0070] Specifically, over multiple consecutive monitoring periods (e.g., daily morning measurements, accumulating to 4 to 12 weeks), the system establishes a metabolic rate profile for the same user. Each monitoring period records the actual basal metabolic rate (BMR), weight, estimated muscle mass, activity level, and other data for that day, arranging them chronologically to generate a basal metabolic rate change curve. This curve is then input into a trend analysis model: a linear regression is performed on the time series, calculating the slope and significance: a significantly positive slope with consecutive weeks of increase is identified as a sustained upward trend (possibly related to muscle-building training or increased activity); a significantly negative slope with consecutive weeks of decrease is identified as a sustained downward trend (possibly related to muscle loss, sedentary lifestyle, dieting, or hypothyroidism, requiring special attention); and a slope close to zero but with large inter-weekly fluctuations is identified as a fluctuating trend (possibly related to irregular measurements, disrupted sleep patterns, or weight fluctuations). Personalized health recommendations are generated based on trend type and user's physiological behavior characteristics: For users with a continuously declining trend, based on their activity data, if their average daily steps are low, it is recommended to increase strength training three times a week and ensure adequate protein intake; if they are sleep-deprived, it is recommended to maintain a regular sleep schedule; for users with a continuously rising trend, based on their body composition data, it is recommended to maintain training intensity and pay attention to calorie intake balance; for users with fluctuating trends, it is recommended to fix measurement conditions (fasting in the morning, measurement at the same time) to reduce fluctuations. The recommendations are displayed via push notifications through the mobile app, forming a closed loop for long-term metabolic health management, helping users track metabolic changes and intervene in adverse trends in a timely manner.

[0071] In some embodiments, the complete workflow of the human basal metabolic rate prediction system based on bioelectrical impedance and multimodal fusion is as follows: Upon first use, the user inputs their age, gender, and height (or automatically measured via a height column) on the touchscreen of the body composition analyzer, stands barefoot on the scale electrodes, holds the electrode handles with both hands, and remains still. The multi-frequency bioelectrical impedance acquisition module applies a 100μA AC current sequentially at five frequencies: 5kHz, 50kHz, 100kHz, 250kHz, and 500kHz, acquires the voltage response at each frequency, calculates the impedance, performs Cole-Cole fitting, and extracts R0, R∞, R0-R∞, fc, Zc, and φmax to form an impedance feature vector. The scale simultaneously acquires the user's weight, and the smartwatch synchronizes the user's average daily steps, activity expenditure, resting heart rate, sleep duration, and deep sleep ratio over the past 7 days via Bluetooth, forming a physiological behavior feature vector. After standardization, the two types of features are input into the multimodal fusion model: the impedance feature is processed through a first branch network, and the physiological behavior feature is processed through a second branch network to extract higher-order features. The attention module calculates the weights of the two modalities and weights them to fuse them into a 16-dimensional fusion feature vector. The system uses a pre-trained GBRT regression model with integrated feature inputs to calculate and output the true basal metabolic rate (e.g., 1720 kcal / day for a 30-year-old male user weighing 75 kg and 175 cm tall). This value is cross-validated with a Mifflin formula estimate of 1650 kcal / day; a deviation rate of 4.2% is within the allowable range of ±15%, confirming the result's validity. Further comparison with a standard value under the same conditions shows a deviation rate of +4.2%, falling within the range of -10% to +10%, indicating normal metabolism. The device screen and mobile app display the true basal metabolic rate value, comparison with peers, and metabolic status rating. After 8 weeks of continuous monitoring, the system generates a metabolic rate change curve, identifies the trend type, and provides personalized suggestions. The entire measurement process takes less than one minute. By integrating an attention mechanism and a pre-trained model, the system reduces the prediction error from 15% to 20% to less than 8% compared to traditional empirical formula methods, while simultaneously achieving automatic identification of metabolic abnormality risks and long-term trend tracking.

[0072] In some embodiments, the multi-device collaborative metabolic rate model personalized optimization method based on federated learning, under the premise of privacy protection by not centrally uploading users' original vital signs data and impedance data, collaboratively trains the basal metabolic rate prediction model through multiple physical testing devices and health management terminals, improves the model's generalization ability by utilizing diverse data across regions and populations, and generates personalized model parameters adapted to each individual user.

[0073] Specifically, each user's body composition analyzer and smartwatch collects impedance and physiological behavior data locally, with the raw data remaining on-premises. Only the locally trained model parameter updates (gradients) are uploaded to the coordination server via an encrypted channel. The coordination server uses a federated averaging (FedAvg) algorithm to aggregate the model parameter updates uploaded by each client, calculates a weighted average to obtain the global model parameters, and then sends them back to each client. Each client initializes with the global parameters and then fine-tunes the model based on its local user data. Through multiple rounds of federated training, the global model incorporates metabolic data patterns from users of different regions, ages, and body types, addressing the model bias caused by limited local data samples and insufficient population coverage on single devices, and improving the model's prediction accuracy for niche groups (such as the elderly, fitness enthusiasts, and different ethnicities). Simultaneously, each client performs local fine-tuning based on the global model using its own historical measurement data, adapting the model to the user's individual characteristics (such as specific body types and exercise habits), achieving a two-layer optimization of "global commonality + individual individuality." Data transmission employs homomorphic encryption and differential privacy technologies. Noise is added and parameters are encrypted during gradient aggregation, preventing any participant from retrieving other users' original vital signs data from the aggregated parameters. This complies with personal information protection requirements and addresses privacy concerns in multi-center data co-construction and sharing. The federated learning-optimized model can further improve prediction accuracy by 3% to 5% for niche populations, while user privacy data remains locally at all times.

[0074] In some embodiments, based on the method of early warning and health event correlation analysis of abnormal time-series metabolic rate, the user's continuous basal metabolic rate time series is correlated with time-series data such as heart rate variability, sleep structure, and activity level collected by smartwatches. The time-series anomaly detection algorithm is used to identify abnormal signals before clinically significant changes in metabolic rate occur (such as early thyroid dysfunction, early muscle loss, and early metabolic adaptation due to dieting), and push health warnings in advance, realizing the upgrade from "numerical measurement" to "risk warning".

[0075] Specifically, the system establishes a long-term health time-series database for each user, storing daily data such as BMR, weight, muscle mass estimation, heart rate, heart rate variability (HRV), sleep indicators, and activity levels, with a daily time resolution. Temporal anomaly detection algorithms (such as IsolationForest, LSTM-AE, or BOCPD) are used to model the BMR sequence and its joint distribution with other physiological indicators: Under normal circumstances, BMR is relatively stable over several weeks (fluctuations less than 5%). When there is no significant weight change but BMR decreases by more than 8% for more than 2 weeks, accompanied by a combination of signals including decreased HRV, increased resting heart rate, and reduced deep sleep, the algorithm identifies it as an early event of metabolic abnormality, associated with health risk directions (such as hypothyroidism tendency, excessive metabolic adaptation / dieting, early muscle loss); when abnormally elevated BMR is accompanied by increased resting heart rate, poor sleep, and weight loss, it suggests a risk of hyperthyroidism or overtraining. The system combines users' recent dietary records (manually entered in the app or synchronized with a nutrition app) and training plan data to distinguish between normal physiological changes (such as a reasonable increase in BMR due to muscle gain, and metabolic adaptation in the later stages of weight loss) and abnormal changes requiring medical examination. It only sends alerts for highly suspicious abnormal events to avoid false alarms and user anxiety. The alerts are pushed via the mobile app, accompanied by BMR change curves, graphs of related physiological indicators, and recommended measures (such as "It is recommended to have a follow-up examination in two weeks, and check thyroid function if necessary," and "It is recommended to resume calorie intake to avoid metabolic damage"), enabling early detection and proactive management of metabolic health.

[0076] In some embodiments, a digital twin-based body composition-metabolic dynamic simulation and intervention program simulation method is used to construct an individual digital twin model of the user. This model integrates muscle mass, body fat percentage, water distribution, and basal metabolic rate calculated using multi-frequency impedance analysis. In a virtual environment, the model simulates the trends of weight, body fat, and metabolic rate changes over the next 4 to 12 weeks under different exercise, diet, and lifestyle intervention programs. This allows the system to recommend the optimal intervention program for the user and continuously correct the twin model with real measurement data during actual implementation, thus achieving a closed-loop health management system of "prediction-intervention-correction".

[0077] Specifically, based on the user's current impedance characteristics (calculation of muscle mass and body fat percentage using R0 and R0-R∞), actual BMR, activity level, heart rate, and sleep data, an individual metabolic digital twin model is constructed for the user. This model includes an energy balance equation (calorie intake - calorie expenditure = weight change), a metabolic adaptation sub-model (adaptive decrease in metabolic rate during long-term calorie restriction), and an exercise response sub-model (the effect of strength training on muscle mass and BMR). Given the user's goals (e.g., losing 5kg of fat and maintaining metabolic rate), multiple intervention plans are simulated in the twin model: Plan A is a daily calorie deficit of 500kcal + 3 strength training sessions per week; Plan B is a daily calorie deficit of 300kcal + 5 aerobic sessions per week; Plan C is a high-protein diet + HIIT training. The model predicts changes in weight, body fat percentage, BMR, and muscle retention after 8 weeks for each plan, outputting the predicted curve, success probability, and side effect risks for each plan (e.g., Plan A may lead to a 3% decrease in metabolic rate, while Plan C offers the best muscle retention), allowing the user and coach to choose the optimal plan. In practice, the system uses newly measured weight, impedance, and BMR data from the body composition analyzer weekly to calibrate the twin model parameters (Bayesian update), correcting model prediction biases and making subsequent predictions increasingly closer to the user's actual response (e.g., if a user's muscle-building efficiency during strength training is above average, the model automatically adjusts their training response coefficient). The entire process is completed on the device or in the cloud. Users view the prediction curve and recommended plan through the app, and the system pushes weekly progress comparisons (actual changes vs. prediction curves), forming a closed loop of precise sports nutrition intervention based on individual digital twins, improving intervention effectiveness and adherence compared to general recommendations.

[0078] Please see Figure 3 As shown, Figure 3 This is a schematic diagram of the structure of a human basal metabolic rate prediction system 200 based on bioelectrical impedance and multimodal fusion provided in this application embodiment. This human basal metabolic rate prediction system 200 based on bioelectrical impedance and multimodal fusion is used to execute the steps of the human basal metabolic rate prediction method based on bioelectrical impedance and multimodal fusion shown in the above embodiments. The human basal metabolic rate prediction system 200 based on bioelectrical impedance and multimodal fusion can be a single server or a server cluster, or it can be a terminal, such as a handheld terminal, a laptop computer, a wearable device, or a robot.

[0079] like Figure 3 As shown, the human basal metabolic rate prediction system 200 based on bioelectrical impedance and multimodal fusion includes: The signal application unit 201 is used to apply AC excitation signals of multiple frequencies to the human body through the multi-frequency bioelectrical impedance acquisition module, acquire impedance data of the human body at multiple frequencies, and extract impedance features; and to acquire the user's weight parameters, height parameters, age parameters, gender parameters, daily activity parameters, heart rate parameters, and sleep parameters through the multi-modal data acquisition module, and extract physiological and behavioral features. The model input unit 202 is used to input impedance features and physiological behavior features into the multimodal fusion model. The multimodal fusion model performs weighted fusion of impedance features and physiological behavior features through an attention mechanism and outputs a fused feature vector. The risk prediction unit 203 is used to input the fused feature vector into the pre-trained basal metabolic rate prediction model to calculate the user's true basal metabolic rate, and at the same time identify the risk of metabolic abnormalities based on the deviation between the true basal metabolic rate and the standard basal metabolic rate.

[0080] In some embodiments, the step of applying AC excitation signals of multiple frequencies to the human body through a multi-frequency bioelectrical impedance acquisition module and acquiring impedance data of the human body at multiple frequencies includes: placing electrode pads at multiple measurement locations on the user's body surface; applying AC excitation currents of different frequencies to the human body through the electrode pads, wherein the frequency of the AC excitation currents covers both low-frequency and high-frequency bands; acquiring the voltage response signals across the electrode pads at each frequency; calculating the human body impedance value at each frequency based on the AC excitation current and voltage response signals; and obtaining impedance data corresponding to multiple frequencies.

[0081] In some embodiments, the extraction of impedance features includes: performing curve fitting on impedance data corresponding to multiple frequencies to obtain impedance frequency characteristic curves; extracting feature parameters from the impedance frequency characteristic curves, the feature parameters including impedance amplitude, impedance phase angle, real part of impedance in the low-frequency band and real part of impedance in the high-frequency band corresponding to the feature frequency point; and organizing the feature parameters into an impedance feature vector as input to the subsequent multimodal fusion model.

[0082] In some embodiments, the step of collecting user's weight, height, age, gender, daily activity level, heart rate, and sleep parameters through a multimodal data acquisition module and extracting physiological behavioral features includes: collecting user's weight parameters through a scale; collecting user's height parameters through a height measurement module; collecting user's age and gender parameters through a user input module; collecting user's daily activity level and heart rate parameters through a smartwatch; and collecting user's sleep parameters through a sleep monitoring module; and organizing the weight, height, age, gender, daily activity level, heart rate, and sleep parameters into a physiological behavioral feature vector.

[0083] In some embodiments, the step of inputting impedance features and physiological behavior features into a multimodal fusion model, and the multimodal fusion model performing weighted fusion of impedance features and physiological behavior features through an attention mechanism to output a fused feature vector, includes: inputting the impedance feature vector into a first branch of the multimodal fusion model, and inputting the physiological behavior feature vector into a second branch of the multimodal fusion model; the multimodal fusion model calculating the attention weights of the impedance feature vector and the physiological behavior feature vector respectively through an attention mechanism; and performing a weighted summation of the impedance feature vector and the physiological behavior feature vector according to the attention weights to output a fused feature vector.

[0084] In some embodiments, the step of inputting the fused feature vector into a pre-trained basal metabolic rate prediction model to calculate the user's true basal metabolic rate includes: inputting the fused feature vector into the pre-trained basal metabolic rate prediction model; the basal metabolic rate prediction model performing regression calculation based on the fused feature vector to output the user's true basal metabolic rate value; cross-validating the true basal metabolic rate value with the user's input weight, height, and age parameters; re-performing the prediction calculation when the verification deviation exceeds a preset allowable range; and outputting the final true basal metabolic rate value when the verification deviation falls within the preset allowable range.

[0085] In some embodiments, identifying metabolic abnormality risk based on the deviation between the actual basal metabolic rate and the standard basal metabolic rate includes: querying the corresponding standard basal metabolic rate from a standard basal metabolic rate comparison table based on the user's weight, height, age, and gender parameters; calculating the deviation rate between the actual basal metabolic rate and the standard basal metabolic rate; determining a low metabolic risk when the deviation rate is lower than a preset low threshold, a high metabolic risk when the deviation rate is higher than a preset high threshold, and a normal metabolic risk when the deviation rate is within a preset normal range.

[0086] In some embodiments, before inputting impedance features and physiological behavior features into the multimodal fusion model, the method further includes: collecting impedance features and physiological behavior features of historical user samples and corresponding actual basal metabolic rate labels; inputting the impedance features of historical user samples into a first feature extraction network and the physiological behavior features of historical user samples into a second feature extraction network; performing weighted fusion of the outputs of the first feature extraction network and the second feature extraction network through an attention mechanism to train the network parameters of the multimodal fusion model; and storing the network parameters of the trained multimodal fusion model into the running memory of the health management device.

[0087] In some embodiments, before inputting the fused feature vector into the pre-trained basal metabolic rate prediction model, the method further includes: collecting fused feature vectors and corresponding actual basal metabolic rate labels from historical user samples to construct a training dataset; dividing the training dataset into a training subset and a validation subset according to a preset ratio; performing model training on the training subset and model validation on the validation subset using a regression algorithm; stopping training when the prediction error of the validation set falls within a preset error range to obtain the pre-trained basal metabolic rate prediction model; and storing the model parameters of the pre-trained basal metabolic rate prediction model in the running memory of the health management device.

[0088] In some embodiments, the method further includes: collecting the true basal metabolic rate for each monitoring cycle within multiple consecutive monitoring cycles; generating a basal metabolic rate change curve based on the true basal metabolic rate for multiple consecutive monitoring cycles; inputting the basal metabolic rate change curve into a trend analysis model; the trend analysis model identifies metabolic trend types based on the basal metabolic rate change curve; metabolic trend types include a continuous upward trend, a continuous downward trend, and a fluctuating trend; generating personalized health recommendations based on the metabolic trend type and the user's physiological behavioral characteristics; and pushing the personalized health recommendations to the user's terminal.

[0089] It should be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the human basal metabolic rate prediction system and its modules based on bioelectrical impedance and multimodal fusion described above can be found in the corresponding contents of the various embodiments of the human basal metabolic rate prediction method based on bioelectrical impedance and multimodal fusion, and will not be repeated here.

[0090] The aforementioned method for predicting human basal metabolic rate based on bioelectrical impedance and multimodal fusion can be implemented as a computer program, which can be used in, for example... Figure 3 It runs on the system shown.

[0091] Please see Figure 4 , Figure 4 This is a schematic block diagram of the structure of a health management device provided in an embodiment of this application. The health management device includes a processor, a memory, and a network interface connected via a device bus, wherein the memory may include a storage medium and internal memory.

[0092] The storage medium can store operating devices and computer programs. The computer program includes program instructions that, when executed, cause the processor to perform any method for predicting human basal metabolic rate based on bioelectrical impedance and multimodal fusion.

[0093] The processor provides computing and control capabilities to support the operation of the entire health management device.

[0094] Internal memory provides an environment for the execution of computer programs in non-volatile storage media. When executed by a processor, the computer program enables the processor to execute any method for predicting human basal metabolic rate based on bioelectrical impedance and multimodal fusion.

[0095] This network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the terminal to which the present application is applied. Specific health management devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0096] It should be understood that the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0097] In one embodiment, the processor is configured to run a computer program stored in memory to perform the following steps: The multi-frequency bioelectrical impedance acquisition module applies AC excitation signals of multiple frequencies to the human body, collects impedance data of the human body at multiple frequencies, and extracts impedance features; the multi-modal data acquisition module collects the user's weight parameters, height parameters, age parameters, gender parameters, daily activity parameters, heart rate parameters, and sleep parameters, and extracts physiological and behavioral features. Impedance features and physiological behavior features are input into a multimodal fusion model. The multimodal fusion model performs weighted fusion of impedance features and physiological behavior features through an attention mechanism and outputs a fused feature vector. The fused feature vector is input into the pre-trained basal metabolic rate prediction model to calculate the user's true basal metabolic rate. At the same time, the deviation between the true basal metabolic rate and the standard basal metabolic rate is used to identify the risk of metabolic abnormalities.

[0098] In some embodiments, the step of applying AC excitation signals of multiple frequencies to the human body through a multi-frequency bioelectrical impedance acquisition module and acquiring impedance data of the human body at multiple frequencies includes: placing electrode pads at multiple measurement locations on the user's body surface; applying AC excitation currents of different frequencies to the human body through the electrode pads, wherein the frequency of the AC excitation currents covers both low-frequency and high-frequency bands; acquiring the voltage response signals across the electrode pads at each frequency; calculating the human body impedance value at each frequency based on the AC excitation current and voltage response signals; and obtaining impedance data corresponding to multiple frequencies.

[0099] In some embodiments, the extraction of impedance features includes: performing curve fitting on impedance data corresponding to multiple frequencies to obtain impedance frequency characteristic curves; extracting feature parameters from the impedance frequency characteristic curves, the feature parameters including impedance amplitude, impedance phase angle, real part of impedance in the low-frequency band and real part of impedance in the high-frequency band corresponding to the feature frequency point; and organizing the feature parameters into an impedance feature vector as input to the subsequent multimodal fusion model.

[0100] In some embodiments, the step of collecting user's weight, height, age, gender, daily activity level, heart rate, and sleep parameters through a multimodal data acquisition module and extracting physiological behavioral features includes: collecting user's weight parameters through a scale; collecting user's height parameters through a height measurement module; collecting user's age and gender parameters through a user input module; collecting user's daily activity level and heart rate parameters through a smartwatch; and collecting user's sleep parameters through a sleep monitoring module; and organizing the weight, height, age, gender, daily activity level, heart rate, and sleep parameters into a physiological behavioral feature vector.

[0101] In some embodiments, the step of inputting impedance features and physiological behavior features into a multimodal fusion model, and the multimodal fusion model performing weighted fusion of impedance features and physiological behavior features through an attention mechanism to output a fused feature vector, includes: inputting the impedance feature vector into a first branch of the multimodal fusion model, and inputting the physiological behavior feature vector into a second branch of the multimodal fusion model; the multimodal fusion model calculating the attention weights of the impedance feature vector and the physiological behavior feature vector respectively through an attention mechanism; and performing a weighted summation of the impedance feature vector and the physiological behavior feature vector according to the attention weights to output a fused feature vector.

[0102] In some embodiments, the step of inputting the fused feature vector into a pre-trained basal metabolic rate prediction model to calculate the user's true basal metabolic rate includes: inputting the fused feature vector into the pre-trained basal metabolic rate prediction model; the basal metabolic rate prediction model performing regression calculation based on the fused feature vector to output the user's true basal metabolic rate value; cross-validating the true basal metabolic rate value with the user's input weight, height, and age parameters; re-performing the prediction calculation when the verification deviation exceeds a preset allowable range; and outputting the final true basal metabolic rate value when the verification deviation falls within the preset allowable range.

[0103] In some embodiments, identifying metabolic abnormality risk based on the deviation between the actual basal metabolic rate and the standard basal metabolic rate includes: querying the corresponding standard basal metabolic rate from a standard basal metabolic rate comparison table based on the user's weight, height, age, and gender parameters; calculating the deviation rate between the actual basal metabolic rate and the standard basal metabolic rate; determining a low metabolic risk when the deviation rate is lower than a preset low threshold, a high metabolic risk when the deviation rate is higher than a preset high threshold, and a normal metabolic risk when the deviation rate is within a preset normal range.

[0104] In some embodiments, before inputting impedance features and physiological behavior features into the multimodal fusion model, the method further includes: collecting impedance features and physiological behavior features of historical user samples and corresponding actual basal metabolic rate labels; inputting the impedance features of historical user samples into a first feature extraction network and the physiological behavior features of historical user samples into a second feature extraction network; performing weighted fusion of the outputs of the first feature extraction network and the second feature extraction network through an attention mechanism to train the network parameters of the multimodal fusion model; and storing the network parameters of the trained multimodal fusion model into the running memory of the health management device.

[0105] In some embodiments, before inputting the fused feature vector into the pre-trained basal metabolic rate prediction model, the method further includes: collecting fused feature vectors and corresponding actual basal metabolic rate labels from historical user samples to construct a training dataset; dividing the training dataset into a training subset and a validation subset according to a preset ratio; performing model training on the training subset and model validation on the validation subset using a regression algorithm; stopping training when the prediction error of the validation set falls within a preset error range to obtain the pre-trained basal metabolic rate prediction model; and storing the model parameters of the pre-trained basal metabolic rate prediction model in the running memory of the health management device.

[0106] In some embodiments, the method further includes: collecting the true basal metabolic rate for each monitoring cycle within multiple consecutive monitoring cycles; generating a basal metabolic rate change curve based on the true basal metabolic rate for multiple consecutive monitoring cycles; inputting the basal metabolic rate change curve into a trend analysis model; the trend analysis model identifies metabolic trend types based on the basal metabolic rate change curve; metabolic trend types include a continuous upward trend, a continuous downward trend, and a fluctuating trend; generating personalized health recommendations based on the metabolic trend type and the user's physiological behavioral characteristics; and pushing the personalized health recommendations to the user's terminal.

[0107] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to implement the steps of the human basal metabolic rate prediction method based on bioelectrical impedance and multimodal fusion as provided in any embodiment of this application.

[0108] The computer-readable storage medium can be an internal storage unit of the health management device described in the foregoing embodiments, such as the hard drive or memory of the health management device. Alternatively, the computer-readable storage medium can be an external storage device of the health management device, such as a plug-in hard drive, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the health management device.

[0109] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting human basal metabolic rate based on bioelectrical impedance and multimodal fusion, applied to a health management device including a multi-frequency bioelectrical impedance acquisition module, a multimodal data acquisition module, and a basal metabolic rate prediction module, characterized in that, include: The multi-frequency bioelectrical impedance acquisition module applies AC excitation signals of multiple frequencies to the human body, acquires impedance data of the human body at multiple frequencies, and extracts impedance characteristics. The multimodal data acquisition module collects users' weight, height, age, gender, daily activity level, heart rate, and sleep parameters to extract physiological and behavioral characteristics. Impedance features and physiological behavior features are input into a multimodal fusion model. The multimodal fusion model performs weighted fusion of impedance features and physiological behavior features through an attention mechanism and outputs a fused feature vector. The fused feature vector is input into the pre-trained basal metabolic rate prediction model to calculate the user's true basal metabolic rate. At the same time, the deviation between the true basal metabolic rate and the standard basal metabolic rate is used to identify the risk of metabolic abnormalities.

2. The method according to claim 1, characterized in that, The process involves applying AC excitation signals of multiple frequencies to the human body via a multi-frequency bioelectrical impedance acquisition module, and acquiring impedance data of the human body at multiple frequencies, including: Electrode pads are placed at multiple measurement locations on the user's body surface, and alternating currents of different frequencies are applied to the human body through the electrode pads. The frequency of the alternating current covers both low-frequency and high-frequency bands. The voltage response signal across the electrode pads is collected at each frequency. The human body impedance value at each frequency is calculated based on the AC excitation current and voltage response signal, resulting in impedance data corresponding to multiple frequencies.

3. The method according to claim 2, characterized in that, The extracted impedance features include: Curve fitting is performed on the impedance data corresponding to multiple frequencies to obtain the impedance frequency response curve; Characteristic parameters are extracted from the impedance frequency response curve. These characteristic parameters include the impedance amplitude, impedance phase angle, real part of impedance in the low-frequency band, and real part of impedance in the high-frequency band at the characteristic frequency point. The feature parameters are organized into impedance feature vectors, which are then used as inputs to the subsequent multimodal fusion model.

4. The method according to claim 1, characterized in that, The process involves collecting user parameters such as weight, height, age, gender, daily activity level, heart rate, and sleep parameters through a multimodal data acquisition module, and extracting physiological and behavioral characteristics, including: The system collects the user's weight parameters through a weighing scale, the user's height parameters through a height measurement module, and the user's age and gender parameters through a user input module. The system collects users' daily activity and heart rate parameters through a smartwatch, and collects users' sleep parameters through a sleep monitoring module. Weight parameters, height parameters, age parameters, gender parameters, daily activity level parameters, heart rate parameters, and sleep parameters are organized into a physiological behavior feature vector.

5. The method according to claim 1, characterized in that, The impedance features and physiological behavior features are input into a multimodal fusion model. The multimodal fusion model performs weighted fusion of the impedance features and physiological behavior features through an attention mechanism, and outputs a fused feature vector, including: The impedance feature vector is input into the first branch of the multimodal fusion model, and the physiological behavior feature vector is input into the second branch of the multimodal fusion model. The multimodal fusion model calculates the attention weights of the impedance feature vector and the physiological behavior feature vector separately through an attention mechanism. The impedance feature vector and the physiological behavior feature vector are weighted and summed according to the attention weights to output a fused feature vector.

6. The method according to claim 1, characterized in that, The step of inputting the fused feature vector into the pre-trained basal metabolic rate prediction model to calculate the user's true basal metabolic rate includes: The fused feature vector is input into the pre-trained basal metabolic rate prediction model. The basal metabolic rate prediction model performs regression calculations based on the fused feature vector and outputs the user's true basal metabolic rate value. The actual basal metabolic rate is cross-validated with the user's input weight, height, and age parameters. If the cross-validation deviation exceeds the preset allowable range, the prediction calculation is re-executed. If the cross-validation deviation falls within the preset allowable range, the final actual basal metabolic rate value is output.

7. The method according to claim 1, characterized in that, The method of identifying metabolic abnormality risk based on the deviation between the true basal metabolic rate and the standard basal metabolic rate includes: Based on the user's weight, height, age, and gender parameters, the corresponding standard basal metabolic rate is retrieved from the standard basal metabolic rate comparison table; The deviation rate between the actual basal metabolic rate and the standard basal metabolic rate is calculated. When the deviation rate is lower than the preset low threshold, it is judged as low metabolic risk. When the deviation rate is higher than the preset high threshold, it is judged as high metabolic risk. When the deviation rate is within the preset normal range, it is judged as normal metabolism.

8. The method according to claim 1, characterized in that, Before inputting impedance features and physiological behavioral features into the multimodal fusion model, the following is also included: Collect impedance characteristics and physiological behavior characteristics of historical user samples, as well as corresponding actual basal metabolic rate labels; The impedance features of historical user samples are input into the first feature extraction network, and the physiological and behavioral features of historical user samples are input into the second feature extraction network. The network parameters of the multimodal fusion model are obtained by training the outputs of the first feature extraction network and the second feature extraction network through a weighted fusion of the outputs of the attention mechanism. The network parameters of the trained multimodal fusion model are stored in the running memory of the health management device.

9. The method according to claim 1, characterized in that, Before inputting the fused feature vector into the pre-trained basal metabolic rate prediction model, the following is also included: Collect fusion feature vectors and corresponding actual basal metabolic rate labels from historical user samples to construct a training dataset; divide the training dataset into a training subset and a validation subset according to a preset ratio; The model is trained on the training subset and validated on the validation subset using a regression algorithm. Training stops when the prediction error on the validation set falls within a preset error range, thus obtaining a pre-trained basal metabolic rate prediction model. The model parameters of the pre-trained basal metabolic rate prediction model are stored in the running memory of the health management device.

10. The method according to claim 1, characterized in that, The method further includes: The true basal metabolic rate is collected for each monitoring period over multiple consecutive monitoring periods, and a basal metabolic rate change curve is generated based on the true basal metabolic rate over multiple consecutive monitoring periods. Input the basal metabolic rate change curve into the trend analysis model. The trend analysis model identifies the metabolic trend type based on the basal metabolic rate change curve. The metabolic trend types include continuous upward trend, continuous downward trend, and fluctuating trend. Personalized health recommendations are generated based on metabolic trend types and user physiological behavior characteristics, and then pushed to the user's device.

Images