Dynamic metabolic rate prediction method and system based on multi-modal biological data

Abstract

The invention relates to a dynamic metabolic rate prediction method and system based on multi-modal biological data, and belongs to the technical field of health monitoring management. The prediction method comprises the following steps: acquiring and processing a multi-modal biological data stream of a user, and generating physiological signal time domain feature data and nutrition quantification data; inputting the body composition data, the physiological signal time domain feature data and the nutrition quantification data into a space-time alignment engine to generate a timestamp-aligned multi-modal fusion data set; calculating and generating a real-time metabolic rate sequence based on the multi-modal fusion data set; constructing a metabolic state vector based on the real-time metabolic rate sequence and the nutrition quantification data; according to the metabolic state vector and preset activity plan data, generating a nutritional requirement ratio through a reverse solution algorithm; and dynamically correcting the nutrition demand ratio according to continuously monitored blood sugar time sequence data, and outputting a final meal replacement scheme. According to the method, accurate prediction of the individual dynamic metabolic rate and intelligent recommendation of the personalized nutrition ratio are realized.

Description

Technical Field

[0001] This application relates to the field of health monitoring and management technology, and in particular to a method and system for predicting dynamic metabolic rate based on multimodal biological data. Background Technology

[0002] With the fast pace of modern life and changes in dietary structure, people's demand for personalized health management is growing, especially in the prevention of metabolic diseases, sports nutrition support and chronic disease management. How to achieve accurate monitoring and scientific guidance of the human body's energy metabolism status has become a key direction for current technological development.

[0003] Traditional methods for measuring metabolic rate primarily rely on indirect calorimetry in a laboratory setting. While this method offers high accuracy, the equipment is expensive, the operation is complex, and it lacks continuous dynamic monitoring capabilities, making it impractical for everyday life. Existing technologies include some metabolic estimation methods based on single data sources, such as estimating exercise energy expenditure through heart rate monitoring or assessing nutritional intake based on questionnaires and food diaries. However, these methods generally suffer from limitations such as limited data dimensions, poor timeliness, and high subjectivity.

[0004] In recent years, with the popularization of IoT technology and smart wearable devices, the collection of multi-source heterogeneous biological data has become possible. However, effectively integrating data streams from different sensing devices and achieving accurate time synchronization and feature fusion still faces significant challenges. Especially in home and community application scenarios, users' physiological states, dietary behaviors, and exercise habits exhibit high individual variability and time-varying characteristics. Existing technologies often struggle to establish unified mathematical models to accurately describe this complex dynamic metabolic process and lack the ability to perceive and dynamically regulate individual real-time physiological states. Particularly in practical application environments, due to significant differences in metabolic characteristics among individuals, common metabolic prediction models often fail to accurately reflect the personalized needs of specific users. This makes the construction of a dynamic metabolic assessment system with adaptive learning capabilities a pressing technical challenge. Summary of the Invention

[0005] To address the aforementioned technical issues, this application provides a method and system for predicting dynamic metabolic rate based on multimodal biological data.

[0006] Firstly, this application provides a dynamic metabolic rate prediction method based on multimodal biological data, employing the following technical solution: A dynamic metabolic rate prediction method based on multimodal biological data, the dynamic metabolic rate prediction method comprising: Acquire user's body composition data, physiological signal data, nutrient intake image data, and blood glucose time series data to generate a multimodal biological data stream; Perform time-domain feature extraction on the physiological signal data to generate physiological signal time-domain feature data; The nutrient intake image data is processed by image recognition to generate nutrient quantification data; The body composition data, physiological signal temporal feature data, and nutritional quantification data are input into the spatiotemporal alignment engine, and a timestamp-aligned multimodal fusion dataset is generated through a dynamic time warping algorithm. Based on the multimodal fusion dataset, the baseline value of basal metabolic rate, the exercise metabolic compensation value, and the food thermic effect compensation value are calculated and superimposed to generate a real-time metabolic rate sequence. Based on the real-time metabolic rate sequence and the nutrient quantification data, a metabolic state vector is constructed; Based on the metabolic state vector and the preset activity plan data, a nutritional requirement ratio is generated through a reverse algorithm. The nutritional requirements are dynamically corrected based on the continuously monitored blood glucose time-series data, and a final meal replacement plan is output.

[0007] By adopting the above technical solution, a leap from static metabolic assessment to real-time dynamic metabolic tracking has been achieved. Combining deep learning, time-series analysis, and inverse optimization algorithms, this not only improves the accuracy and individual adaptability of metabolic prediction but also allows for adaptive correction based on the user's actual physiological feedback (such as blood glucose fluctuations), generating personalized nutritional recommendations. This technical solution is of great significance in health management, chronic disease prevention, and sports nutrition support, providing users with scientific, accurate, and real-time dietary guidance, and promoting the development of digital health and smart healthcare.

[0008] Secondly, this application provides a dynamic metabolic rate prediction system based on multimodal biological data, employing the following technical solution: A dynamic metabolic rate prediction system based on multimodal biological data, the prediction system comprising: The data acquisition module is used to acquire the user's body composition data, physiological signal data, nutritional intake image data, and blood glucose time series data, and generate a multimodal biological data stream; The time-domain feature extraction module is used to perform time-domain feature extraction on the physiological signal data to generate physiological signal time-domain feature data; The image recognition processing module is used to perform image recognition processing on the nutrient intake image data to generate nutrient quantification data; The spatiotemporal alignment and fusion module is used to input the body composition data, physiological signal time-domain feature data and nutritional quantification data into the spatiotemporal alignment engine, and generate a timestamp-aligned multimodal fusion dataset through a dynamic time warping algorithm; The metabolic rate calculation module is used to calculate the baseline value of basal metabolic rate, the exercise metabolic compensation value, and the food thermic effect compensation value based on the multimodal fusion dataset, and then superimpose them to generate a real-time metabolic rate sequence. A metabolic state construction module is used to construct a metabolic state vector based on the real-time metabolic rate sequence and the nutrient quantification data; The reverse solving module is used to generate the nutritional requirement ratio based on the metabolic state vector and the preset activity plan data through a reverse solving algorithm; The dynamic correction output module is used to dynamically correct the nutritional requirement ratio based on the continuously monitored blood glucose time series data and output the final meal replacement plan.

[0009] Thirdly, this application provides a computer-readable storage medium, which adopts the following technical solution: A computer-readable storage medium storing a computer program that can be loaded by a processor and executed as in any of the methods in the first aspect.

[0010] In summary, this application includes at least one of the following beneficial technical effects: By integrating multimodal biological data, it achieves accurate dynamic prediction of users' metabolic rate and personalized nutritional intervention. The system can monitor and analyze users' body composition, physiological signals, nutrient intake, and blood glucose changes in real time. It solves the time synchronization problem of different data sources through spatiotemporal alignment fusion technology, accurately calculating the complete metabolic rate sequence including basal metabolism, exercise expenditure, and thermic effect of food. Based on metabolic state vectors and preset activity plans, the system can inversely solve for the optimal nutrient ratio and perform dynamic correction using continuous blood glucose monitoring data, ultimately outputting a personalized meal replacement nutrition plan. This technical solution improves the accuracy of metabolic rate prediction and the timeliness of nutritional intervention, providing users with scientific and precise health management services. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the first process of a dynamic metabolic rate prediction method based on multimodal biological data according to one embodiment of this application.

[0012] Figure 2 This is a schematic diagram of the second process of a dynamic metabolic rate prediction method based on multimodal biological data according to one embodiment of this application.

[0013] Figure 3 This is a schematic diagram of the third process of a dynamic metabolic rate prediction method based on multimodal biological data according to one embodiment of this application.

[0014] Figure 4 This is a schematic diagram of the fourth process of a dynamic metabolic rate prediction method based on multimodal biological data according to one embodiment of this application.

[0015] Figure 5 This is a schematic diagram of the fifth step of a dynamic metabolic rate prediction method based on multimodal biological data according to one embodiment of this application.

[0016] Figure 6 This is a schematic diagram of the sixth process of a dynamic metabolic rate prediction method based on multimodal biological data according to one embodiment of this application.

[0017] Figure 7 This is a schematic diagram of the seventh process of a dynamic metabolic rate prediction method based on multimodal biological data according to one embodiment of this application.

[0018] Figure 8 This is a schematic diagram of the eighth process of a dynamic metabolic rate prediction method based on multimodal biological data according to one embodiment of this application. Detailed Implementation

[0019] To make the purpose, technical solution, and advantages of this application clearer, the following description is provided in conjunction with the appendix. Figures 1-8 The present application will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the application.

[0020] This application discloses a method for predicting dynamic metabolic rate based on multimodal biological data.

[0021] Reference Figure 1 A dynamic metabolic rate prediction method based on multimodal biological data, specifically including: Step S101: Acquire the user's body composition data, physiological signal data, nutrient intake image data, and blood glucose time series data to generate a multimodal biological data stream; Specifically, during the data acquisition phase, the system collects various biologically relevant signals from users' daily lives through different types of electronic devices, including body composition data (such as muscle mass and fat content), physiological signal data (such as heart rate and body movement acceleration), dietary image data (used to record the types and quantities of food consumed daily), and time-series changes in blood glucose concentration.

[0022] These heterogeneous data constitute the basic material library for subsequent analysis, and come from weight measurement devices (such as bioelectrical impedance analyzers), wearable motion monitoring devices (such as smartwatches or fitness trackers), image capture terminals (mobile phone cameras with image recognition apps), and continuous glucose monitoring systems (CGM).

[0023] Step S102: Perform time-domain feature extraction on the physiological signal data to generate physiological signal time-domain feature data; Among them, the time-domain feature data of physiological signals includes heart rate time-series data and acceleration time-series data; Specifically, mathematical transformations are performed on continuous signals collected from heart rate and triaxial accelerometer sensors of motion monitoring devices to extract statistically significant key indicators. By slicing the original signals into time windows and applying methods such as moving average, peak detection, and standard deviation calculation, a set of numerical values ​​reflecting the real-time physiological response characteristics of the human body is obtained—the so-called time-domain features. These features are easier to process in real-time and retain sufficient dynamic response characteristics compared to frequency domain or other high-dimensional representations, making them particularly suitable for lightweight edge computing task deployments in resource-constrained mobile environments. Furthermore, since these features often intuitively reflect an individual's current physical load and their degree of adaptation to environmental stimuli, they are crucial for subsequently deriving accurate energy consumption estimates.

[0024] Step S103: Perform image recognition processing on the nutrient intake image data to generate nutrient quantification data; The nutritional data includes protein intake, carbohydrate intake, and fat intake. Specifically, in the image recognition processing module, a deep learning network architecture (such as a convolutional neural network CNN) is used to automatically perform semantic segmentation and category discrimination on food image samples ingested by the test subject, thereby converting them into structured quantitative descriptors. This process involves not only learning and training the mapping relationship from pixel-level visual representation to the proportion of macro-nutrient composition, but also considering potential errors caused by factors such as lighting differences and shooting angle shifts. Successfully converting this type of unstructured multimedia data means that sensory information that is difficult to directly participate in computation can be transformed into a set of digital variables that can be further processed, thus bridging the gap between behavioral observation and biochemical metabolic prediction.

[0025] Step S104: Input body composition data, physiological signal temporal feature data and nutritional quantification data into the spatiotemporal alignment engine, and generate a timestamp-aligned multimodal fusion dataset through a dynamic time warping algorithm; Specifically, considering the significant differences in sampling periods and clock synchronization deviations among the aforementioned sensor sources, failure to properly address these issues can easily lead to misalignment or even failure in subsequent joint modeling. Therefore, a specially designed unified space-time coordinate framework is introduced to coordinate the timestamp label distribution of input items from all parties. The classic DTW (Dynamic Time Warping) algorithm is employed as one of the main methods, capable of finding the optimal matching path to bridge phase lag or lead caused by rate fluctuations without requiring strict correspondence. After this normalization operation, all observations are projected into a common reference frame, making subsequent cross-comparison analyses based on these frames reasonable and feasible.

[0026] Step S105: Based on the multimodal fusion dataset, calculate the baseline value of basal metabolic rate, the exercise metabolic compensation value, and the food thermic effect compensation value, and superimpose them to generate a real-time metabolic rate sequence; Specifically, the process begins by using lean body mass (LBM) as an indicator and classical empirical rules such as the Weir formula or the Harris-Benedict equation to deduce the minimum calorie requirement per unit time at rest, i.e., the baseline basal metabolic rate. Next, the specific magnitude of the additional energy expenditure caused by physical activity is calculated using a specific functional relationship, i.e., the exercise-metabolic compensation value. Third, the thermic effect of food itself, which induces a certain degree of increased energy expenditure (TEF), must also be considered; this part also requires corresponding adjustments based on dietary composition details, i.e., the thermic effect of food compensation value. Finally, these three independent but interconnected parts are summed to obtain a new time series that comprehensively depicts the evolution of energy metabolism intensity throughout the day, i.e., the real-time metabolic rate curve.

[0027] Step S106: Construct a metabolic state vector based on real-time metabolic rate sequences and nutrient quantification data; The metabolic state vector includes liver glycogen levels, lipolysis rate, and protein turnover rate. Specifically, the integral difference between carbohydrate intake and real-time metabolic rate data is used as an indicator of liver glycogen level; the ratio of the rate of change in fat intake to real-time metabolic rate data is used as an indicator of lipolysis rate; and the convolution result of protein intake and real-time metabolic rate data is used as an indicator of protein turnover rate.

[0028] Understandably, this step essentially involves reorganizing the various functional indicators that have already been acquired and preliminarily processed into a more compact and interpretable state expression.

[0029] Step S107: Based on the metabolic state vector and preset activity plan data, generate the nutritional requirement ratio through a reverse algorithm; This approach employs inverse optimization to deduce the optimal macronutrient allocation ratios for individual needs under ideal conditions. Unlike traditional methods that typically break down target intake levels from top to bottom, this method first defines the expected performance indicators (e.g., total calorie budget in an exercise plan, expected recovery speed, etc.), and then uses operations research techniques to find a set of solutions that minimizes overall cost while satisfying several boundary constraints. The cost function typically considers factors such as maintaining the stability of vital reserves and avoiding severe imbalances. The final output is a draft of an idealized nutritional guideline that is both scientifically rigorous and practically applicable.

[0030] Step S108: Dynamically correct the nutritional requirements ratio based on the continuously monitored blood glucose time series data and output the final meal replacement plan.

[0031] The system utilizes a continuous blood glucose monitoring feedback mechanism to make timely adjustments to the nutritional requirements until a final meal replacement plan is formed. This application embodiment specifically includes a closed-loop monitoring subsystem responsible for long-term tracking of the subject's glucose fluctuation pattern evolution and periodically updating the recommended menu content to maintain optimal applicability. When the coefficient of variation exceeds the normal range, an alarm mechanism is triggered to initiate a new round of parameter iteration and optimization until the parameters converge back to the safe range before ceasing adjustments.

[0032] The above implementation achieves a leap from static metabolic assessment to real-time dynamic metabolic tracking. By combining deep learning, time-series analysis, and inverse optimization algorithms, it not only improves the accuracy and individual adaptability of metabolic prediction but also adaptively corrects based on the user's actual physiological feedback (such as blood glucose fluctuations), generating personalized nutritional recommendations. This technical solution is of great significance in health management, chronic disease prevention, and sports nutrition support, providing users with scientific, accurate, and real-time dietary guidance, and promoting the development of digital health and smart healthcare.

[0033] Reference Figure 2 As one implementation of step S104, the step of inputting body composition data, physiological signal temporal feature data, and nutritional quantification data into a spatiotemporal alignment engine, and generating a timestamp-aligned multimodal fusion dataset through a dynamic time warping algorithm includes: Step S201: Obtain body composition data sequence, including timestamps and corresponding lean body mass values; Among these methods, body composition analyzers can be used to collect composition data sequences, and parameters such as muscle mass and fat content in an individual can be determined based on electrical impedance or dual-energy X-ray absorptiometry (DEXA). Because these devices have long operating cycles and require active user intervention for each measurement, their output is often a sparse and non-continuous discrete time series, meaning that there may be a long time interval between two adjacent data points.

[0034] It should be noted that, considering that the focus in clinical practice is on lean body mass, a core indicator reflecting the total amount of lean tissue, this variable is specifically extracted here as a representative feature for subsequent procedures.

[0035] Step S202: Obtain the time-domain feature data sequence of physiological signals, including timestamps and corresponding heart rate and acceleration values; This data originates from wearable smart devices, such as photoplethysmography (PPG) and triaxial microelectromechanical systems (MEMS) accelerometer modules in smartwatches or fitness trackers. These devices can provide real-time feedback at frequencies of one second or even higher, covering changes in cardiovascular responses and limb movement patterns during daily activities. It's important to note that heart rate is not only regulated by the autonomic nervous system but also indirectly reflects energy expenditure levels; while acceleration can be used to estimate steps, exercise intensity, and even posture type. Together, they constitute crucial indicators of bodily activity levels.

[0036] Step S203: Obtain the nutritional quantitative data sequence, including timestamps and corresponding protein, carbohydrate, and fat intake; This portion of the data originates from image recognition-assisted food diary recording tools. Typical applications include using deep learning networks to automatically analyze food photos taken with a mobile phone and estimate the proportion of macronutrients. Because dietary behavior is highly subjective and random, and because uploading photos is not a mandatory, continuous process, the relevant entries exhibit a significantly sparse distribution. Therefore, to improve subsequent matching accuracy, it is necessary to further optimize this portion of the original input to make it as close as possible to the actual food intake curve.

[0037] Step S204: Using the sampling frequency of the physiological signal time-domain feature data as the reference time axis, generate a standardized time point sequence with equal intervals, and output the heart rate value sequence and acceleration value sequence. Compared to daily weight checks or irregular photo submissions, heart rate and trunk movement are clearly more suitable as stable reference systems for anchoring global coordinates. The standard time node array constructed in this way is essentially a virtualized public coordinate system, around which all other types of data to be fused will be mapped and transformed to ensure that their relative positional relationships are accurately preserved.

[0038] Step S205: Perform linear interpolation calculation on the body composition data sequence to generate a lean body mass interpolation sequence corresponding to the standardized time point sequence; Linear interpolation is applicable to the approximate function approximation process between any two points. In this step, its main function is to fill in the unknown numerical estimates in the gap period caused by the interval between the use of the weighing scale.

[0039] Step S206: Perform sliding window mean processing on the nutrient quantification data sequence to generate a nutrient smoothing sequence corresponding to the standardized time point sequence; The essence of moving average filtering is to weaken the transient disturbances caused by sudden shocks, making the originally jagged and volatile dietary intake records smoother. Specifically, a fixed-length time window is typically set, and within this window, the arithmetic mean of all elements in the current segment is repeatedly calculated while progressing along the time axis. This arithmetic mean is then used as the sampling result for the corresponding position in the new sequence. This not only effectively suppresses spikes caused by random factors but also better reflects the average eating habits over a period of time, thus enhancing the effectiveness of cross-modal data association mining.

[0040] Step S207: Input the heart rate value sequence, acceleration value sequence, lean body mass interpolation sequence, and nutritional smoothing sequence into the dynamic time warping algorithm to calculate the time offset of each sequence relative to the reference time axis. Specifically, a joint feature vector of heart rate and acceleration sequences is constructed, and the Euclidean distance matrix between the joint feature vector and the fat-free body weight interpolation sequence / nutrition smoothing sequence is calculated. The minimum cumulative distance path is solved by dynamic programming, and the time difference corresponding to the inflection point in the path is used as the time offset.

[0041] Dynamic Time Warping (DTW) is a classic time series similarity measurement method widely used in fields such as speech recognition and gesture recognition. Its basic idea is to find an optimal path that minimizes the cumulative cost without requiring a strict one-to-one correspondence between the two sets of sequences.

[0042] In this embodiment, potential delay differences between each subsequence and its shared time axis are identified, and compensation corrections are made accordingly. It should be noted that traditional DTW focuses more on the stretching and compression relationships between single-dimensional signals, while in complex multi-dimensional data environments, it is also necessary to consider how to construct reasonable comprehensive evaluation criteria.

[0043] Step S208: Perform time-domain translation correction on the lean body mass interpolation sequence and the nutritional smoothing sequence according to the time offset; Specifically, the steps of time-domain translation correction include: identifying the postprandial metabolic response time window in the nutritional smoothing sequence, using quadratic polynomial interpolation instead of linear interpolation within the response time window, and verifying the derivative relationship between the rate of weight change and energy intake after translation.

[0044] Translation correction involves rigidly shifting the original sequence along the time axis according to a pre-calculated offset vector, thereby repositioning events that did not occur at the same time to the correct time points. This adjustment is applied only to weight and nutrition variables, excluding heart rate and acceleration, because the former are identified as primarily affected by lag effects, while the latter, being under high-frequency sampling, are theoretically already close to an ideal time axis and can maintain high consistency without additional intervention.

[0045] Step S209: Merge the corrected heart rate value sequence, acceleration value sequence, lean body mass interpolation sequence, and nutritional smoothing sequence to generate a timestamp-aligned multimodal fusion dataset.

[0046] This involves aggregating all data sequences into a structured database with identical timestamps and complete attribute dimensions. This database not only simplifies the tedious data cleaning process in downstream machine learning modeling but also provides a solid and reliable foundation for in-depth research into a series of profound scientific questions, such as how dietary habits affect muscle growth rate and whether physical labor intensity regulates the appetite center's response mechanism.

[0047] In the above implementation, by cleverly selecting an appropriate time base, making reasonable use of mathematical modeling techniques, and introducing advanced dynamic programming concepts, a high degree of synergistic integration of key physiological and biochemical indicators of the human body is achieved.

[0048] Reference Figure 3 As one implementation of step S105, the step of calculating the baseline value of basal metabolic rate, the exercise metabolic compensation value, and the food thermic effect compensation value based on the multimodal fusion dataset and superimposing them to generate a real-time metabolic rate sequence includes: Step S301: Extract the timestamp-aligned heart rate value sequence, acceleration value sequence, lean body mass interpolation sequence, and nutritional smoothing sequence from the multimodal fusion dataset; The core task of this step is to achieve spatial unification and temporal calibration of the output information from different sensors or data acquisition devices, enabling all variables to be jointly modeled within the same temporal framework. Timestamp alignment means that each record must be assigned a time stamp accurate to the second or even millisecond, and delay biases caused by inconsistent sampling frequencies must be eliminated through interpolation, resampling, or other synchronization strategies. For example, some wearable devices may collect acceleration signals at a frequency of 1Hz, while the heart rate monitoring module of a smartwatch collects signals every 5 seconds. Interpolation algorithms are needed to unify these signals onto a common time axis, ensuring that the values ​​of each variable at the same point in time can effectively participate in subsequent calculations. This cross-modal data integration not only improves the operability of the system but also provides a solid data foundation for complex metabolic models.

[0049] Step S302: Perform sliding window statistical processing on the lean body mass interpolation sequence to generate a lean body mass mean sequence; Among them, for the obtained fat-free body mass interpolation sequence, a sliding window statistical processing is performed to generate a more stable LBM mean sequence. The significance of this step is to reduce the influence of random noise in the original LBM measurement and improve the stability of basal metabolic estimation.

[0050] In practical applications, due to potential measurement errors or short-term fluctuations in body composition analyzers or bioelectrical impedance analysis (BIA) devices, directly using instantaneous readings as input parameters may lead to unstable metabolic predictions. Therefore, the moving average method is a typical noise reduction technique. A fixed-length time window (e.g., 3 minutes) is set, and at each time point, all LBM observations within that window are backtracked and their arithmetic mean is calculated, resulting in a smoother trend curve. This step is essentially an application of a low-pass filter, helping to filter out high-frequency disturbances caused by factors such as respiratory rhythm and muscle tension, making the obtained LBM mean more representative of the individual's true physiological structure. Furthermore, the choice of window width must balance sensitivity and robustness; too wide a window weakens the model's responsiveness to external interventions, while too narrow a window makes it difficult to suppress noise interference.

[0051] Step S303: Calculate the baseline basal metabolic rate sequence based on the fat-free body mass mean sequence; Basal Metabolic Rate (BMR) refers to the minimum amount of energy required for the human body to maintain basic life activities at rest. It is usually significantly affected by factors such as age, sex, body type, and especially lean body mass.

[0052] In one embodiment of this application, a pre-defined empirical regression formula is introduced: BMR = A × LBM + B, where A and B are pre-calibrated constant coefficients that can be derived from population samples. However, considering that individuals in reality experience varying degrees of edema or dehydration, i.e., changes in the extracellular fluid (ECF) ratio can have a non-linear impact on metabolism, an ECF correction mechanism is also implemented: when an abnormally high ECF level is detected, an amplification factor (1 + k × ECF / LBM) is applied to the original formula to compensate for the additional energy consumption caused by water retention. This adjustment reflects the model's high sensitivity and adaptability to changes in homeostasis, making it more biologically reasonable than traditional static models that rely solely on the LBM factor.

[0053] Step S304: Classify the motion intensity based on the acceleration value sequence and generate an intensity level coding sequence; Acceleration signals reflect changes in the speed and direction of body movement. After appropriate transformation, they can be used to distinguish different types of physical activity, such as walking, running, and cycling. To facilitate computer understanding and quantitative analysis, a discretized classification method is used to map continuously changing acceleration amplitudes into several finite levels (e.g., mild, moderate, and vigorous levels), each corresponding to a specific range of acceleration magnitude thresholds. This encoding process not only simplifies the subsequent modeling process but also enhances the system's ability to generalize to different types of movement patterns. More importantly, it provides a feasible way to quantify the degree of human physical exertion, serving as a bridge between external motor performance and internal metabolic regulation.

[0054] Step S305: Combine the intensity level coding sequence and the heart rate value sequence to calculate the exercise metabolic compensation value sequence; Specifically, heart rate, as an important physiological feedback signal, can directly reflect the increased energy demand brought about by enhanced sympathetic nerve excitability. In this process, the resting period (i.e., the period without significant physical activity and where the heart rate is in a low, stable range) is first identified, and a baseline heart rate value is extracted from it. Then, during other active periods, the difference between the current heart rate and the baseline heart rate value is compared, and a weighted calculation considering the time decay effect and exercise intensity weight is performed to obtain the final compensation increment. The formula is: Compensation value = Decay factor × (Current heart rate - Baseline heart rate) × Intensity weight coefficient.

[0055] The decay factor follows an exponential decreasing pattern, simulating the body's gradual return to calm after exercise. The intensity weighting coefficient, on the other hand, increases proportionally to the magnitude of acceleration, reflecting that higher-intensity exercise leads to a more prolonged accumulation of oxygen debt, known as excess post-exercise oxygen consumption (EPOC). The synergistic effect of these two factors ensures that the model not only captures immediate energy bursts but also reasonably estimates the sustained metabolic boost over a period after exercise.

[0056] Step S306: Perform macronutrient classification integration on the nutrient smoothing sequence to generate a food heat effect compensation value sequence. This method assigns a specific thermic effect coefficient to each of the three main energy-providing substances in the daily diet (protein, carbohydrates, and fat), and then accumulates these coefficients to derive the total thermic effect contribution. The thermic effect of food (TEF) refers to the extra energy loss caused by the digestion, absorption, and conversion processes after eating, and is generally considered to decrease in the order of protein > carbohydrates > fat.

[0057] In this embodiment, an integral accumulation method can be adopted, which involves multiplying the actual intake rate of various nutrients per unit time by their corresponding TEF coefficients and then summing the results. This preserves the uniqueness of individual dietary habits while reflecting the differences in the gastrointestinal transit time and metabolic pathways of different nutrients. For example, if a user consumes a large amount of whey protein powder for breakfast, their TEF curve will show a small peak within the next two hours. Conversely, if dinner is mainly composed of fats, the peak will appear later but last longer. Such fine-grained time control greatly improves the model's responsiveness and accuracy to changes in daily dietary behavior.

[0058] Step S307: Overlay the baseline basal metabolic rate sequence, the exercise metabolic compensation sequence, and the food thermic effect compensation sequence to generate a real-time metabolic rate sequence.

[0059] The process involves adding the baseline basal metabolic rate, exercise metabolic compensation value, and heat effect coefficient at the same time point; performing median filtering on the superposition results of three consecutive time points; and finally outputting the filtered real-time metabolic rate sequence.

[0060] Specifically, algebraic superposition is performed within a unified time coordinate system to obtain a complete real-time metabolic rate sequence. A median filter is incorporated into this aggregation process to remove extreme abrupt changes or abnormal peaks, preventing global distortion caused by individual erroneous heartbeat detections or false alarms. The optimized output not only possesses excellent smoothness and visibility but can also directly serve various applications within health management platforms, such as calorie budgeting, weight loss plan recommendations, and even clinical nutritional support.

[0061] The above implementation establishes a highly integrated dynamic metabolic modeling framework. Its innovation lies not only in the ingenious design of each individual technical detail, but also in the organic linkage mechanism formed between them, realizing a fundamental shift from static parameter estimation to dynamic process tracking.

[0062] Reference Figure 4 As one implementation of step S107, the step of generating the nutritional requirement ratio by reverse algorithm based on the metabolic state vector and preset activity plan data includes: Step S401: Obtain metabolic state vector data, including liver glycogen level indicators, lipolysis rate indicators, and protein turnover rate indicators. Among these parameters, liver glycogen, as one of the main sources of rapid short-term energy supply for muscles and the liver, directly affects the priority of carbohydrate demand due to its level; lipolysis rate reflects the mobilization efficiency of adipose tissue, determining the speed and scale of fatty acids entering the bloodstream; and protein turnover rate characterizes the dynamic balance between protein synthesis and degradation in the body, which is particularly crucial for high-intensity trainees or patients with chronic diseases. These parameters together constitute a state-space expression describing an individual's current energy reserves and material conversion capacity, namely, a metabolic state vector.

[0063] Step S402: Receive and parse the preset activity plan data to extract the movement type code and duration for the future target time period; This step embodies a forward-looking judgment on an individual's future behavioral patterns. The activity plan not only covers physical exercise arrangements but also includes changes in work intensity and adjustments to sleep patterns. By structurally analyzing this information, it can be abstracted into a set of standardized coding sequences, such as defining aerobic running as "AEROBIC_RUN_30MIN," which facilitates database indexing and energy consumption model calls.

[0064] Step S403: Based on the exercise type code and duration, match the exercise energy consumption coefficient database to generate predicted energy consumption values ​​for future periods; This step essentially involves performing a multi-dimensional interpolation calculation. Since different types of physical activity correspond to different energy expenditure rates (METs) per unit time, a three-dimensional time-intensity-energy mapping library covering common activity categories needs to be established. For example, jogging burns approximately 5 kilocalories per minute, while strength training may burn 8-10 kilocalories. By querying this database and substituting the user's activity schedule into the corresponding function expression, the estimated total energy expenditure over a future period can be obtained.

[0065] Step S404: With maintaining liver glycogen level homeostasis as the first constraint, calculate the lower limit of carbohydrate requirement to meet the liver glycogen level index. This step focuses on maintaining liver glycogen homeostasis, using a reverse calculation method to determine the minimum supplemental dose required to fill the existing glycogen deficit. Specifically, when a user's current liver glycogen concentration is found to be below the normal range, it must be restored as soon as possible to prevent affecting nerve conduction speed or even triggering hypoglycemia symptoms. To this end, a fixed proportionality factor, the bioavailability coefficient, is introduced to measure the actual conversion efficiency of ingested glucose into usable glycogen (generally considered to be approximately 0.6). The estimation can be completed using the following formula: Lower limit of carbohydrate requirement = (Target homeostasis threshold - Current liver glycogen level) ÷ Bioavailability coefficient, where the bioavailability coefficient is a preset constant.

[0066] Step S405: Using the maintenance of nitrogen balance as the second constraint, calculate the lower limit of protein requirement to meet the protein turnover index. Specifically, this step addresses the long metabolic cycle of proteins by employing a decay function to simulate their gradual change over time. Considering that protein breakdown is a gradual process rather than an instantaneous reaction, an exponential or power-law curve is used to fit it, and then multiplied by a safety margin coefficient greater than 1 to ensure no net loss occurs. This method effectively avoids the bias problems caused by previous rough estimations based solely on body weight, improving the accuracy of nitrogen balance control. The specific formula is: Lower limit of protein requirement = Protein breakdown rate × Safety margin coefficient, where the safety margin coefficient > 1.

[0067] Step S406: Using the maintenance of lipolysis steady state as the third constraint, calculate the lower limit of fat demand that satisfies the lipolysis rate index. The study quantifies a user's current fat metabolism state using the lipolysis rate index and translates this state into a safe boundary constraint for fat intake. The lipolysis rate reflects the efficiency of breaking down adipose tissue into free fatty acids per unit time. When this index is consistently high, it indicates that the body is in an active lipomobilization state (such as fasting, after exercise, or during periods of insulin resistance). In this case, excessive fat intake may exacerbate dyslipidemia or the risk of fat accumulation. Conversely, if the lipolysis rate is low (such as after a high-carbohydrate diet), a minimum fat intake must be ensured to maintain basic physiological functions such as cell membrane synthesis and hormone secretion.

[0068] Specifically, a pre-defined lipolysis-demand mapping model (e.g., a piecewise function based on clinical cohort calibration) can dynamically map the lipolysis rate to a lower limit of fat demand. This model follows the biological principle that "lipolysis rate and lower limit of fat demand are inversely proportional": when the lipolysis rate exceeds the healthy threshold range, the model outputs a lower lower limit of fat demand (inhibiting exogenous fat intake and prioritizing the use of endogenous fat for energy); when the lipolysis rate is below the threshold, the model automatically raises the lower limit of fat demand (avoiding essential fatty acid deficiency). Essentially, this mechanism transforms the organism's metabolic feedback loop (with lipolysis rate as the feedback signal) into a rigid constraint for nutritional decisions, thereby embedding the logic of protecting physiological homeostasis into the mathematical optimization framework. This ensures that the final fat intake value satisfies both the mathematical optimality of energy conservation and the biological feasibility of metabolic health.

[0069] Step S407: Based on the predicted energy consumption value, lower limit value of carbohydrate demand, lower limit value of protein demand, and lower limit value of fat demand for the future period, the optimal ratio of protein demand, carbohydrate demand, and fat demand is generated by using a linear programming algorithm. The objective function of the linear programming algorithm is to minimize the absolute value of the difference between the actual total energy supply and the predicted energy consumption. Constraints include: carbohydrate intake must not be lower than the lower limit of carbohydrate requirement, protein intake must not be lower than the lower limit of protein requirement, and fat intake must not be lower than the lower limit of fat requirement. Furthermore, fat intake is derived as a function of the proportion of energy supplied by fat, based on the principle of energy conservation: fat intake = (predicted energy consumption × proportion of energy supplied by fat) / energy density (approximately 9 kcal / g). After satisfying all lower limit constraints, the optimization algorithm derives the final value through the objective function and the principle of energy conservation, balancing individual needs with energy balance.

[0070] Step S408: Determine whether a user's special physiological identifier has been detected; if yes, proceed to step S409; if no, proceed to step S410. Step S409: Adjust and optimize the proportion of macronutrients in the formula according to the specific physiological marker type to obtain the nutritional requirement formula. Step S410: The optimized ratio is directly used as the nutritional requirement ratio.

[0071] Specifically, if the patient is identified as having diabetes, the carbohydrate requirement is reduced and the fat requirement is increased proportionally; if the patient is identified as having metabolic syndrome, the protein requirement is increased and an upper limit for carbohydrate requirement is set.

[0072] Specifically, specific physiological markers refer to conditions such as diabetes and metabolic syndrome that exhibit significantly different metabolic characteristics compared to healthy individuals. Once such markers are detected, corresponding weighting mechanisms are activated to address them. For example, in diabetic patients with severe insulin resistance, the proportion of carbohydrates is typically reduced and the proportion of fat is increased to compensate, thereby mitigating postprandial blood glucose spikes. For individuals with central obesity and a tendency towards dyslipidemia, the proportion of high-quality protein intake is appropriately increased, while a strict cap is imposed on total carbohydrate intake to help improve pancreatic sensitivity and inflammatory factor levels. It should be noted that this step is a necessary modification to improve the universality and safety of the treatment plan.

[0073] In the above implementation, based on a rigorous data-driven architecture and a multi-level feedback closed-loop mechanism, the integration from macro-lifestyle guidance to micro-molecular metabolic regulation has been successfully achieved. This overcomes the limitations of traditional nutrition research that relies too much on static statistical models and also makes up for the lack of physiological credibility support in pure artificial intelligence black-box reasoning.

[0074] Reference Figure 5 As one implementation of step S108, the step of dynamically correcting the nutritional requirement ratio based on continuously monitored blood glucose time-series data and outputting the final meal replacement plan includes: Step S501: Acquire continuously monitored blood glucose time-series data and perform sliding window segmentation to generate blood glucose subsequences; This step aims to introduce a steady-state analysis perspective within a local time period. The sliding window refers to the process of moving point by point along a fixed-length time interval to extract segments of the original time series; for example, setting the window width to 30 minutes will yield a series of subsets such as [0:30], [1:31], and so on, until the end. This approach can decompose the originally lengthy and complex overall blood glucose trajectory into several relatively independent small units to examine their inherent patterns separately, helping to reveal the microscopic structural fluctuations hidden beneath the macroscopic trends.

[0075] It is important to note that choosing an appropriate window size is crucial. A window that is too short will increase noise interference and reduce statistical power, while a window that is too large may mask the details of short-term, dramatic fluctuations. In practice, the optimal parameter settings are often determined by combining physiological theories (such as a gastric emptying cycle of about 2 hours), clinical observation experience (most people experience postprandial blood glucose peaks 30–90 minutes after eating), and machine learning performance test results.

[0076] Step S502: Calculate the coefficient of variation and fluctuation entropy value of each blood glucose subsequence to generate a blood glucose stability index. The coefficient of variation (CV) is defined as the result of dividing the standard deviation of blood glucose by the mean blood glucose. It is a dimensionless measure of dispersion and is particularly suitable for comparing the variability between data sets with different average levels. Permutation entropy, on the other hand, is a complexity measure function within the category of nonlinear dynamics. By calculating the probability distribution of the permutation and combination patterns of adjacent values ​​within a given time series, it can effectively identify whether there is a potential order structure or the existence of random fluctuations. The specific formula is: Permutation entropy = -Σ(p_i × log p_i), where p_i is the probability distribution of the amplitude of blood glucose changes.

[0077] By combining these two approaches to create a composite glycemic stability index, we not only address the traditional concern about deviations from the central position but also expand our understanding to include the diversity and unpredictability of its evolutionary pathways. This comprehensive evaluation system is more robust and discriminative than traditional methods that rely solely on peak, trough, or average blood glucose levels, making it particularly suitable for managing the needs of chronic disease populations with significant inter-individual heterogeneity.

[0078] Step S503: When the blood glucose stability index exceeds the preset threshold, extract the carbohydrate ratio and protein ratio parameters from the nutritional requirement ratio. This step focuses solely on the ratio of carbohydrates to proteins, the two main energy-providing substances, because numerous studies have confirmed their most direct and significant mechanisms of action in regulating glycemic response. Specifically, carbohydrate intake rapidly raises blood sugar and controls it by stimulating pancreatic β-cells to secrete more insulin; conversely, appropriately increasing high-quality protein sources can not only slow down gastrointestinal digestion and absorption, thus alleviating the problem of rapid blood sugar spikes, but also promote muscle anabolic processes to some extent, which is equally significant for improving long-term glycemic control.

[0079] Step S504: Construct a spatiotemporal correlation matrix between blood glucose fluctuation patterns and macronutrient intake; Specifically, the process of constructing the spatiotemporal correlation matrix includes: identifying nutritional intake events within the postprandial blood glucose response time window; extracting carbohydrate and protein intake within the response time window; calculating the Pearson correlation coefficient between blood glucose fluctuation amplitude and carbohydrate intake; calculating the mutual information entropy between blood glucose fluctuation frequency and protein intake; and generating a two-dimensional correlation matrix: matrix element = [correlation coefficient, mutual information entropy].

[0080] The spatiotemporal correlation matrix is ​​an abstract two-dimensional tabular expression. Rows represent different types of blood glucose fluctuations (e.g., rapid rise and fall, slow rise and stable), while columns correspond to specific amounts of certain nutrients (mainly carbohydrates and proteins). Each cell contains a numerical value representing the strength of the interaction between the two factors, calculated using statistical methods. Common values ​​include Pearson correlation coefficient (suitable for describing linear relationships), Spearman rank correlation (focusing more on monotonically increasing / decreasing trends), and mutual information entropy.

[0081] For example, if a subject's area under the blood glucose curve is significantly lower than that of a control group with the same calories but lacking dietary fiber after each consumption of fiber-rich foods, then it can be inferred that fiber content plays a mediating role in buffering the blood glucose shock effect, and its weight assignment within the entire model framework can be updated accordingly.

[0082] Step S505: Adjust the weight allocation of the carbohydrate percentage parameter and the protein percentage parameter according to the spatiotemporal correlation matrix; The steps for weight adjustment include: when the correlation coefficient is greater than the first critical value, reducing the weight of the carbohydrate percentage parameter; when the mutual information entropy is less than the second critical value, increasing the weight of the protein percentage parameter; the formula for the adjustment magnitude is: new weight = original weight × (1 + k × correlation matrix element), where k is the sensitivity coefficient.

[0083] Specifically, this step introduces a clear set of rules to determine how to modify the proportions of various components in the original preset plan. Generally, the following principles are adopted: if a strong positive correlation is found between carbohydrate intake and blood sugar increase (i.e., the correlation coefficient is higher than a pre-defined first threshold), it indicates that the user is more sensitive to this type of food, and its proportion in the total calorie source should be appropriately reduced; conversely, if it is observed that more protein consumption leads to smaller blood sugar fluctuations (manifested as mutual information entropy lower than a second threshold), its contribution should be appropriately increased.

[0084] In addition, to ensure that each fine-tuning is kept within a reasonable and controllable range and does not cause over-correction, an additional scaling factor k less than or equal to 1 is added for smooth transition. That is, new weight = original weight × (1 + k × elements of the correlation matrix). This preserves the basic outline of the original architecture while achieving the goal of flexibly adapting to changes in the external environment.

[0085] Step S506: Recalculate the nutritional requirement ratio based on the adjusted weight allocation, and output the final meal replacement plan with timestamps.

[0086] The final meal replacement program includes: corrected protein, carbohydrate, and fat requirements; a schedule for the intake of each nutrient; and a warning code for contraindicated ingredients based on the user's pathological markers.

[0087] Specifically, the focus of this step has shifted entirely to accurately matching an individual's actual metabolic capacity with their ideal nutritional needs. To this end, the API interfaces provided by mature nutritional modeling software platforms (such as the MyFitnessPal API and the Nutrium database) can be used to quickly retrieve and match the optimal combinations of various food ingredients that meet the latest weight constraints. Furthermore, prior to this, the objective function needs to be further refined, such as explicitly limiting daily total calorie intake to no more than 2000 kcal, and controlling the proportions of the three macronutrients within the range of 50%-60% carbohydrates, 15%-25% protein, and 20%-30% fat. Only in this way can the subsequent outputs be guaranteed to meet both basic survival needs and specific health management requirements.

[0088] Subsequently, the final meal replacement plan with timestamps is output. This data package is not just a simple menu list; it may also include a series of operation guide documents with precise time markers, specifying in detail what to eat, how much to eat, and when to eat each meal.

[0089] More importantly, it also incorporates a module that can automatically generate a list of prohibited foods based on an individual's past medical records. This module can automatically filter out all dangerous ingredients that may worsen the condition by parsing the relevant disease codes registered in the user's profile (in accordance with the ICD-10 internationally accepted classification standard), greatly enhancing the level of safety assurance.

[0090] In the above embodiments, a dynamic metabolic rate prediction model is constructed by combining continuous blood glucose monitoring data with multimodal bioinformatics, and the nutritional ratio is corrected in real time based on blood glucose stability indicators, thereby outputting a personalized meal replacement plan. This technical solution can effectively improve the accuracy and individual adaptability of blood glucose control, especially in dealing with postprandial blood glucose fluctuations and optimizing the macronutrient intake structure, which shows significant advantages and helps to achieve the long-term health management goals of patients with chronic diseases.

[0091] Reference Figure 6 As a further implementation of the dynamic metabolic rate prediction method for multimodal biological data, after outputting the final meal replacement plan, it also includes: Step S601: Execute the final meal replacement plan and obtain the new biological data stream after the execution of the final meal replacement plan; This process goes beyond simply collecting routine physiological parameters. It integrates multiple high-dimensional sensory channels to form a data stream structure with temporal continuity and multimodal characteristics. This data encompasses changes in body composition (e.g., extracellular fluid / muscle mass), physiological signals (e.g., heart rate variability (HRV) and acceleration amplitude), nutritional behavior records (e.g., meal supplement information derived from image recognition), and continuous monitoring results of blood glucose levels. The value of this dataset lies in its ability to capture the microscopic metabolic response trajectory of the human body under specific nutritional interventions, thus providing the initial basis for subsequent model calibration. For example, if a diabetic patient's blood glucose curve is significantly higher than expected within two hours after consuming a standardized meal replacement, it indicates a possible phase of deterioration in insulin sensitivity. This phenomenon needs to be promptly included in the evaluation to trigger the corresponding model update mechanism.

[0092] Step S602: Perform spatiotemporal alignment processing on the newly added biological data stream to generate a multimodal fusion dataset for validation; This step is designed to overcome issues such as inconsistent timestamps and differences in sampling frequencies between sensors from different sources. Because the operating rhythms of various sensing devices are not uniform—for example, blood glucose tests typically output values ​​every five minutes, while heart rate may refresh every second—failure to perform precise synchronization of these data will lead to serious timing mismatch risks during subsequent data analysis.

[0093] In some embodiments, an incremental alignment method based on an improved Dynamic Time Warping (DTW) algorithm can be employed. This method no longer relies on a fixed reference template for global alignment, but instead allows the length of the baseline axis to be continuously expanded as new data arrives, and reconstructs the mapping relationship on the nonlinear time axis by searching for locally optimal paths. Furthermore, considering real-time performance requirements, an accelerated version, ADTW (Accelerated DTW), can be introduced to effectively compress computational resource consumption by limiting the direction and span of the warping path.

[0094] Step S603: Extract real-time metabolic rate verification sequences and nutrient quantification verification data from the multimodal fusion dataset for verification. Specifically, the key task of this step is to extract core indicators from a normalized multivariate time-series database that can be used to measure the current metabolic state. For the basal metabolic rate, the baseline level of energy expenditure is derived primarily by using the heart rate cycle fluctuation pattern at rest. As for the exercise-induced additional heat release (EPOC effect), the rate of oxygen debt repayment during the recovery period can be estimated by decomposing the amplitude spectrum of movement collected by accelerometers using Fourier transform. The short-term thermic effect (TEF) caused by dietary factors is determined based on the trend of body temperature rise before and after the protein intake event.

[0095] On the other hand, the verification work regarding nutrient intake involves a cross-media data verification process: food images captured by cameras are converted into macronutrient content estimates after being recognized by a convolutional neural network, and then regression fitting is performed with the actual blood glucose change curves fed back by continuous glucose monitoring devices during the same time period. If the deviation between the two exceeds the preset tolerance limit (such as 15%), a higher-precision visual enhancement module will be activated to re-analyze the food details to ensure that subsequent metabolic predictions will not lead to a chain of errors due to front-end misjudgment.

[0096] Step S604: Calculate the error matrix between the real-time metabolic rate verification sequence and the real-time metabolic rate sequence; The error matrix is ​​a composite vector space organized in three dimensions, corresponding to the residual distribution under three typical scenarios: pre-meal, post-meal, and exercise. Each dimension is further subdivided into multiple sub-items to characterize different biological meanings. For example, the pre-meal RMSE mainly reflects the liver's ability to maintain endogenous glucose production, the post-meal AUC difference rate can indirectly evaluate the pancreas's sensitivity to insulin response in peripheral tissues, and the EPOC duration shift helps determine whether mitochondrial oxidative phosphorylation efficiency is within the normal range.

[0097] More importantly, to prevent early errors from dominating the selection of the latest optimization direction, a time weight adjustment mechanism with a decay factor is specially added. This makes the influence of historical errors closer to the current time point in the overall loss function greater, thereby ensuring that the entire learning process always follows the latest metabolic adaptation trend of the subjects.

[0098] Step S605: Update the calculation coefficients of the baseline value of basal metabolic rate, exercise metabolic compensation value, and food thermic effect compensation value according to the error matrix; In order to determine which aspects most urgently need to be corrected first, principal component analysis can be used to discover the main characteristic directions of the error covariance matrix, thereby identifying the categories of error sources with the greatest impact, and then developing targeted improvement strategies for each.

[0099] For example, when the first principal component shows a significant cumulative bias caused by miscalculation of exercise intensity, the mapping ratio coefficient between acceleration and metabolic equivalent should be adjusted accordingly; conversely, if the second principal component points to a problem caused by inaccurate classification of food types, the relevant parameters for protein digestion and absorption kinetics should be optimized.

[0100] In addition, two physical boundary constraints must be added during the entire parameter adjustment process. First, the daily total basal metabolic rate must not fluctuate by more than 5% to avoid drastic shocks that could disrupt the original stable equilibrium point. Second, the rate of decay of excess energy consumption after any exercise must follow an exponential decreasing law, and the half-life constant should fluctuate within a reasonable range (30-120 minutes). Only in this way can we ensure that the proposed mathematical expression conforms to rational physiological mechanisms and maintains good engineering controllability.

[0101] Step S606: When the Euclidean norm of the error matrix exceeds a preset threshold for a consecutive preset number of times, reconstruct the dimension weight allocation of the metabolic state vector. When the system repeatedly observes that actual measurement results deviate from the ideal predicted value beyond a certain tolerance limit, it will automatically initiate a dimensional weight reset procedure, treating different aspects based on the importance of specific physiological nodes. Particularly, those exceeding the critical threshold (e.g., 60%) will receive higher attention, correspondingly increasing the weight of their associated variables in the final score synthesis stage, while moderately weakening the influence of other secondary factors. This approach highlights key problem areas and helps guide subsequent interventions to focus on core issues. For example, for individuals with a history of severe fatty liver disease, liver glycogen reserve capacity and its conversion rate are likely among the most critical factors determining overall glucose and lipid metabolism stability. In such cases, adhering to an egalitarian weighting principle would undoubtedly lead to significant cognitive biases.

[0102] Step S607: Feedback the updated calculated coefficients and dimensional weights to the real-time metabolic rate sequence calculation and metabolic state vector construction steps.

[0103] This step means that each new combination of model parameters, derived after thorough validation, will seamlessly integrate back into the original prediction engine to participate in the task scheduling for the next stage.

[0104] Furthermore, this application embodiment can also be used in conjunction with a metabolic feature fingerprint database built using a graph neural network (GNN) to support more efficient cross-user knowledge transfer services. As long as there is a sufficiently high static attribute similarity and dynamic response behavior consistency between two individuals, they can quickly borrow the personalized model parameters that the other party has already trained as their own starting point to carry out transfer learning, which greatly shortens the threshold of the number of effective samples required for initial deployment.

[0105] The above implementation achieves high-precision dynamic tracking and adaptive regulation of individual metabolic status. This system not only integrates multi-source heterogeneous data from physiological sensors, nutritional behavior recognition, and biochemical indicator monitoring in real time, but also possesses powerful temporal alignment capabilities and an intelligent error correction mechanism, thereby significantly improving the individualized matching and long-term stability of meal replacement intervention effects.

[0106] Reference Figure 7 As a further implementation of the dynamic metabolic rate prediction method for multimodal biological data, the construction of the error matrix includes: Step S701: Divide the time windows into three metabolic scenarios: before meals, after meals, and after exercise; Specifically, the time windows for three metabolic scenarios—before meals, after meals, and after exercise—are divided. The core of this step lies in decoupling the continuous metabolic process in stages through a physiological rhythm-driven approach, thereby achieving refined error modeling.

[0107] Understandably, human metabolic activity is regulated by multiple factors, including circadian rhythms, dietary behavior, and physical exertion, exhibiting distinct phased fluctuations. Traditional modeling methods using a uniform time period struggle to accurately capture the changing patterns of the body's response mechanisms under different metabolic states. Therefore, setting different time window boundaries based on the characteristics of biological rhythms can more effectively match the functional performance of specific metabolic pathways.

[0108] For example, during the pre-meal fasting period, the focus is on the basal glucose metabolism regulation mechanism, where liver glycogenolysis is the primary source of blood glucose. In the post-meal phase, the efficiency of insulin-mediated glucose transport and utilization is highlighted. As for the recovery phase after exercise, the emphasis is on the function of the mitochondrial respiratory chain and its regulatory role in energy supply and demand balance. This physiologically function-oriented time window division not only enhances the model's sensitivity to key metabolic nodes but also improves the spatial resolution and biological interpretability of subsequent error analysis.

[0109] Step S702: Calculate the root mean square error between the real-time metabolic rate verification sequence and the real-time metabolic rate sequence within the time window of each metabolic scenario. The root mean square error (RMSE) is calculated for each scenario. This step aims to quantitatively characterize the deviation between the output of the current metabolic model and its actual measurements under various physiological conditions. By taking the square root of the squared residuals, RMSE assigns a higher weighting coefficient to larger deviations, ensuring that sudden metabolic disturbances (such as a sharp rise in postprandial blood glucose or excessive oxygen consumption response triggered by strenuous exercise) are fully reflected in the error evaluation system.

[0110] Furthermore, the key variables of interest differ depending on the metabolic scenario. For example, in the pre-meal state, the focus is on the estimation accuracy of the resting metabolic rate (RMR), using the actual thermal effect obtained through indirect calorimetry as a reference benchmark; while in the post-meal environment, the emphasis is on the simulation accuracy of the area under the blood glucose curve (AUC) to reveal whether there are lags or insufficient compensation issues in insulin secretion kinetics; similarly, during exercise recovery, the time-cumulative error in the integral form of cardiopulmonary gas exchange parameters is used to measure whether the operation of the mitochondrial oxidative phosphorylation system in muscle cells meets the expected level.

[0111] Step S703: Combine the root mean square error of the time window for each type of metabolic scenario into a three-dimensional error vector; The essence of this operation is to spatially integrate the originally scattered individual error indicators into a multi-dimensional comprehensive deviation descriptor, thereby forming a directional and structured overall evaluation framework. The three dimensions in this vector correspond to the performance deviations of three typical metabolic pathways: the carbohydrate metabolism dimension (representing the liver's gluconeogenesis and glycogenolysis capacity), the insulin response dimension (reflecting the pancreatic β-cell secretory regulation efficiency), and the mitochondrial function dimension (characterizing the oxidative energy supply potential of skeletal muscle tissue).

[0112] Specifically, these three dimensions are closely functionally interconnected: an abnormally high value in one dimension often suggests a potential pathological change in the corresponding pathway, and can even provide early warning information before clinical symptoms appear. To overcome the scale inconsistency caused by differences in basal metabolic levels among individuals, the error vector needs to be Z-score standardized. This involves subtracting the mean of all samples from the original error in each dimension and dividing by its standard deviation, allowing vectors from different users to be compared and clustered on the same coordinate system. The resulting error vector retains the information content of each dimension while possessing good cross-individual transferability and discriminative performance.

[0113] Step S704: Add timestamps and scene identifiers to generate a structured error matrix.

[0114] The purpose of this step is to establish a complete spatiotemporal indexing system for massive historical error records, facilitating efficient retrieval, trend tracking, and source auditing in the later stages. Structured formatting means that all error entries are encapsulated in a two-dimensional or multi-dimensional array structure according to a unified format, where each element contains four basic attribute fields: <error value, timestamp, scene code, data source hash>.

[0115] In the above implementation, a time window segmentation mechanism driven by biological rhythms is constructed, coupled with a deviation evaluation system with multi-dimensional error vectors as the core, and supplemented by a highly structured spatiotemporal index management module, realizing the transformation from simple numerical fitting to a metabolic modeling paradigm with deep mechanistic insights.

[0116] Reference Figure 8 As a further implementation of the dynamic metabolic rate prediction method for multimodal biological data, after the step of constructing the metabolic state vector and before the step of executing the inverse solution algorithm, the method further includes: Step S801: Extract core metabolic indicators, including liver glycogen level, lipolysis rate and protein turnover rate, from the metabolic state vector. Step S802: Obtain macronutrient absorption efficiency parameters from the user's historical data repository; Among them, the parameters of macronutrient absorption efficiency include carbohydrate bioavailability, net protein utilization, and fat absorption rate. Specifically, carbohydrate bioavailability describes the proportion of food that can be effectively absorbed and converted into usable glucose. Its calculation relies on a dose-response model fitting between the historical area under the blood glucose curve (AUC) and the actual total carbohydrate intake, indirectly reflecting the activity of α-glucosidase on the brush border of the small intestine. Net protein utilization is determined by combining nitrogen balance experimental data with historical records of 3-methylhistidine excretion in urine, evaluating whether nitrogen truly participates in the synthesis of functional proteins. Fat absorption rate is mainly derived from the difference between the content of unabsorbed lipids in fecal samples and the total known intake of fats, reflecting the functional performance of exocrine glands, particularly pancreatic lipases.

[0117] Step S803: Based on core metabolic indicators and macronutrient absorption efficiency parameters, calculate the synergistic gain factor of carbohydrates and proteins, and the antagonistic inhibitory factor of fats and proteins through a dynamic coupling function. In this embodiment, the dynamic coupling function is a mathematical model used to quantify the interactions between carbohydrates, proteins, and fats during metabolism in vivo. Its core lies in capturing the biochemical interactions between nutrients through nonlinear mapping. Based on real-time acquired core metabolic indicators (such as liver glycogen levels, lipolysis rate, and protein turnover) and macronutrient absorption efficiency parameters extracted from individual historical data (including carbohydrate bioavailability, net protein utilization, and fat absorption rate), this function constructs a time-dependent regulatory factor system that reflects synergistic and antagonistic inhibitory mechanisms.

[0118] Among them, the synergistic gain factor mainly characterizes the promoting effect of insulin rise induced by carbohydrate intake on protein synthesis pathways (such as mTOR activation), and is weighted and adjusted in combination with liver glycogen reserve status; while the antagonistic inhibitory factor reflects the negative regulatory effect of fatty acid derivatives on the PI3K / Akt signal axis under high-fat environment, thereby affecting muscle protein deposition efficiency. This function not only integrates multi-source physiological and biochemical data, but also embeds a time decay mechanism and a dual-temporal feedback structure, enabling it to intelligently adjust the intensity of each factor at different metabolic stages, thereby accurately simulating the complex nutritional metabolic network of the human body and providing biologically reasonable dynamic constraints for subsequent inverse solving algorithms.

[0119] Step S804: Based on the slope of postprandial blood glucose change in the real-time monitored blood glucose time series data, perform time decay dynamic correction on the synergistic gain factor and the antagonistic inhibitory factor. The time decay mechanism doesn't simply decrease a value by a fixed percentage. Instead, it intelligently adjusts blood glucose levels based on the slope of blood glucose fluctuations within a specific time period, while fully respecting the body's own metabolic rhythm. For example, if a rapid rise in blood glucose is observed in the early postprandial period (e.g., exceeding 2 mmol / min), it may indicate overactive intestinal glucose transport or delayed insulin response. In this case, the synergistic effect between carbohydrates and protein should be appropriately reduced to mitigate the resulting blood glucose fluctuations. Conversely, if blood glucose rises slowly or even stagnates, it suggests impaired gastric emptying or insufficient glucagon secretion. In this case, it's necessary to extend the duration of fat energy supply and moderately upregulate antagonistic inhibitory factors to avoid the potential risk of hypoglycemia.

[0120] It should be noted that this type of correction strategy also incorporates a two-phase mechanism design, that is, to examine the acute changes in the first half hour after eating and the more stable chronic evolution after one hour, so that the intervention can respond to sudden metabolic disturbances in a timely manner and also take into account the needs of long-term health goals.

[0121] Step S805: The modified cooperative gain factor and antagonistic inhibition factor are transformed into constraint rules for the inverse solution algorithm.

[0122] Specifically, for the application scenarios of synergistic gain factors, the system will set a relationship between carbohydrate demand and the current liver glycogen deficit, ensuring that the recommended carbohydrate intake is always within the optimal range that is most conducive to supporting the opening of the current protein synthesis window. As for the mechanism of action of antagonistic inhibitory factors, it is reflected in setting a minimum protein requirement limit to ensure that even in the face of adverse environmental challenges such as high-fat diets, the negative effects caused by them can be offset by pre-compensating for additional protein supply.

[0123] Understandably, all these rules will eventually be encoded into a strict set of inequalities and injected as boundary conditions into advanced optimization engines such as sequential quadratic programming, so that the output scheme conforms to the actual metabolic operation logic of the human body.

[0124] In the above implementation method, based on a comprehensive understanding of the user's current metabolic state, the system deeply integrates the user's actual absorption capacity of various nutrients and the biochemical interaction patterns between them, and further supplements this with a closed-loop blood glucose response monitoring method to continuously fine-tune various weight coefficients, thereby achieving truly precise health management.

[0125] This application also discloses a dynamic metabolic rate prediction system based on multimodal biological data.

[0126] A dynamic metabolic rate prediction system based on multimodal biological data, specifically comprising: The data acquisition module is used to acquire the user's body composition data, physiological signal data, nutritional intake image data, and blood glucose time series data, and generate a multimodal biological data stream; The time-domain feature extraction module is used to perform time-domain feature extraction on physiological signal data and generate physiological signal time-domain feature data. The image recognition and processing module is used to perform image recognition processing on nutrient intake image data to generate nutrient quantitative data. The spatiotemporal alignment and fusion module is used to input body composition data, physiological signal temporal feature data and nutritional quantification data into the spatiotemporal alignment engine, and generate a timestamp-aligned multimodal fusion dataset through a dynamic time warping algorithm; The metabolic rate calculation module is used to calculate the baseline value of basal metabolic rate, exercise metabolic compensation value and food thermic effect compensation value based on a multimodal fusion dataset, and then superimpose them to generate a real-time metabolic rate sequence. The metabolic state construction module is used to construct metabolic state vectors based on real-time metabolic rate sequences and nutrient quantification data. The reverse solving module is used to generate nutritional requirement ratios based on metabolic state vectors and preset activity plan data using a reverse solving algorithm. The dynamic correction output module is used to dynamically correct the nutritional requirements ratio based on continuously monitored blood glucose time-series data and output the final meal replacement plan.

[0127] The dynamic metabolic rate prediction system based on multimodal biological data in this application embodiment can implement any of the above methods, and the specific working process of each module in the system can refer to the corresponding process in the above method embodiments.

[0128] In the several embodiments provided in this application, it should be understood that the provided methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for example, the division of a certain module is merely a logical functional division, and in actual implementation there may be other division methods, such as multiple modules can be combined or integrated into another system, or some features can be ignored or not executed.

[0129] This application also discloses a computer-readable storage medium.

[0130] A computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described above in any of the methods for dynamic metabolic rate prediction based on multimodal biological data.

[0131] The computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device; the program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0132] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A method for predicting dynamic metabolic rate based on multimodal biological data, characterized in that, The dynamic metabolic rate prediction method includes: Acquire user's body composition data, physiological signal data, nutrient intake image data, and blood glucose time series data to generate a multimodal biological data stream; Perform time-domain feature extraction on the physiological signal data to generate physiological signal time-domain feature data; The nutrient intake image data is processed by image recognition to generate nutrient quantification data; The body composition data, physiological signal temporal feature data, and nutritional quantification data are input into the spatiotemporal alignment engine, and a timestamp-aligned multimodal fusion dataset is generated through a dynamic time warping algorithm. Based on the multimodal fusion dataset, the baseline value of basal metabolic rate, the exercise metabolic compensation value, and the food thermic effect compensation value are calculated and superimposed to generate a real-time metabolic rate sequence. Based on the real-time metabolic rate sequence and the nutrient quantification data, a metabolic state vector is constructed; Based on the metabolic state vector and the preset activity plan data, a nutritional requirement ratio is generated through a reverse algorithm. The nutritional requirements are dynamically corrected based on the continuously monitored blood glucose time-series data, and a final meal replacement plan is output.

2. The method for predicting dynamic metabolic rate based on multimodal biological data according to claim 1, characterized in that, The steps of inputting the body composition data, physiological signal temporal feature data, and nutritional quantification data into the spatiotemporal alignment engine, and generating a timestamp-aligned multimodal fusion dataset through a dynamic time warping algorithm include: Obtain body composition data sequences, including timestamps and corresponding lean body mass values; Acquire time-domain feature data sequences of physiological signals, including timestamps and corresponding heart rate and acceleration values; Obtain a sequence of quantitative nutritional data, including timestamps and corresponding protein, carbohydrate, and fat intake. Using the sampling frequency of the physiological signal time-domain feature data as the reference time axis, a standardized time point sequence with equal intervals is generated, and the heart rate value sequence and acceleration value sequence are output. Perform linear interpolation on the body composition data sequence to generate a lean body mass interpolation sequence corresponding to the standardized time point sequence; Perform sliding window mean processing on the nutrient quantification data sequence to generate a nutrient smoothing sequence corresponding to the standardized time point sequence; The heart rate sequence, acceleration sequence, lean body mass interpolation sequence, and nutritional smoothing sequence are input into a dynamic time warping algorithm to calculate the time offset of each sequence relative to the reference time axis. Based on the time offset, the lean body mass interpolation sequence and the nutritional smoothing sequence are subjected to time-domain translation correction; The corrected heart rate value sequence, acceleration value sequence, fat-free body mass interpolation sequence, and nutritional smoothing sequence are merged to generate a timestamp-aligned multimodal fusion dataset.

3. The method for predicting dynamic metabolic rate based on multimodal biological data according to claim 2, characterized in that, Based on the multimodal fusion dataset, the steps of calculating the baseline value of basal metabolic rate, the exercise metabolic compensation value, and the food thermic effect compensation value, and then superimposing them to generate a real-time metabolic rate sequence include: Extract the timestamp-aligned heart rate value sequence, acceleration value sequence, lean body mass interpolation sequence, and nutritional smoothing sequence from the multimodal fusion dataset; Perform sliding window statistical processing on the lean body mass interpolation sequence to generate a lean body mass mean sequence; Calculate the baseline basal metabolic rate sequence based on the fat-free mean body mass sequence; Based on the acceleration value sequence, the motion intensity is graded to generate an intensity level coding sequence; By combining the intensity level coding sequence and the heart rate value sequence, the exercise metabolic compensation value sequence is calculated; The macronutrient classification and integration of the nutrient smoothing sequence are performed to generate a food heat effect compensation value sequence. The baseline basal metabolic rate sequence, the exercise metabolic compensation sequence, and the food thermic effect compensation sequence are superimposed to generate a real-time metabolic rate sequence.

4. The method for predicting dynamic metabolic rate based on multimodal biological data according to claim 1, characterized in that, The steps for generating nutritional requirement ratios using a reverse algorithm based on the metabolic state vector and preset activity plan data include: Obtain metabolic state vector data, including liver glycogen level indicators, lipolysis rate indicators, and protein turnover rate indicators; Receive and parse preset activity plan data to extract the movement type code and duration for future target time periods; Based on the exercise type code and duration, the exercise energy consumption coefficient database is matched to generate predicted energy consumption values ​​for future periods. With maintaining a stable liver glycogen level as the first constraint, calculate the lower limit of carbohydrate requirements to meet the liver glycogen level index; With maintaining nitrogen balance as the second constraint, calculate the lower limit of protein requirement that satisfies the protein turnover index. Using the maintenance of lipolysis steady state as the third constraint, calculate the lower limit of fat demand that satisfies the lipolysis rate index; Based on the predicted energy consumption values, lower limits of carbohydrate demand, lower limits of protein demand, and lower limits of fat demand for the future period, an optimized ratio of protein demand, carbohydrate demand, and fat demand is generated through a linear programming algorithm. Determine whether a user's special physiological marker is detected; if so, adjust the proportion of macronutrients in the optimized formula according to the type of special physiological marker to obtain the nutritional requirement formula; if not, use the optimized formula directly as the nutritional requirement formula.

5. The method for predicting dynamic metabolic rate based on multimodal biological data according to claim 4, characterized in that, The steps for dynamically correcting the nutritional requirements ratio based on the continuously monitored blood glucose time-series data and outputting the final meal replacement plan include: Acquire continuously monitored blood glucose time-series data and perform sliding window segmentation to generate blood glucose subsequences; Calculate the coefficient of variation and fluctuation entropy value for each blood glucose subsequence to generate a blood glucose stability index; When the blood glucose stability index exceeds the preset threshold, extract the carbohydrate ratio and protein ratio parameters from the nutritional requirement ratio. Construct a spatiotemporal correlation matrix between blood glucose fluctuation patterns and macronutrient intake; The weight allocation of the carbohydrate and protein percentage parameters is adjusted according to the spatiotemporal correlation matrix. The nutritional requirements are recalculated based on the adjusted weighting, and a final meal replacement solution with timestamps is output.

6. A method for predicting dynamic metabolic rate based on multimodal biological data according to any one of claims 1 to 5, characterized in that, After delivering the final meal replacement solution, the following is also included: Execute the final meal replacement program and obtain the new biological data stream after the final meal replacement program is executed; The newly added biological data stream is subjected to spatiotemporal alignment processing to generate a multimodal fusion dataset for validation; Real-time metabolic rate verification sequences and nutrient quantification verification data were extracted from the multimodal fusion dataset used for verification. Calculate the error matrix between the real-time metabolic rate verification sequence and the real-time metabolic rate sequence; The calculation coefficients of the baseline value of basal metabolic rate, exercise metabolic compensation value, and food thermic effect compensation value are updated according to the error matrix. When the Euclidean norm of the error matrix exceeds a preset threshold after a preset number of consecutive iterations, the dimension weight allocation of the metabolic state vector is reconstructed. The updated calculation coefficients and dimensional weights are fed back to the real-time metabolic rate sequence calculation and metabolic state vector construction steps.

7. The method for predicting dynamic metabolic rate based on multimodal biological data according to claim 6, characterized in that, The construction of the error matrix includes: Divide the metabolic time windows into three categories: pre-meal, post-meal, and post-exercise. Within each metabolic scenario time window, the root mean square error between the real-time metabolic rate verification sequence and the real-time metabolic rate sequence is calculated. The root mean square error of the time window for each type of metabolic scenario is combined into a three-dimensional error vector; Add timestamps and scene identifiers to generate a structured error matrix.

8. The method for predicting dynamic metabolic rate based on multimodal biological data according to claim 1, characterized in that, After the step of constructing the metabolic state vector and before the step of executing the inverse solution algorithm, the following steps are also included: Core metabolic indicators, including liver glycogen level, lipolysis rate, and protein turnover, are extracted from the metabolic state vector. Obtain macronutrient absorption efficiency parameters from the user's historical data repository; Based on the core metabolic indicators and macronutrient absorption efficiency parameters, the synergistic gain factor of carbohydrates and proteins, as well as the antagonistic inhibitory factor of fats and proteins, are calculated through a dynamic coupling function. Based on the slope of postprandial blood glucose change in real-time monitored blood glucose time-series data, the synergistic gain factor and antagonistic inhibitory factor are dynamically corrected for time decay. The modified synergistic gain factor and antagonistic inhibition factor are transformed into constraint rules for the inverse solution algorithm.

9. A dynamic metabolic rate prediction system based on multimodal biological data, characterized in that, The prediction system includes: The data acquisition module is used to acquire the user's body composition data, physiological signal data, nutritional intake image data, and blood glucose time series data, and generate a multimodal biological data stream; The time-domain feature extraction module is used to perform time-domain feature extraction on the physiological signal data to generate physiological signal time-domain feature data; The image recognition processing module is used to perform image recognition processing on the nutrient intake image data to generate nutrient quantification data; The spatiotemporal alignment and fusion module is used to input the body composition data, physiological signal time-domain feature data and nutritional quantification data into the spatiotemporal alignment engine, and generate a timestamp-aligned multimodal fusion dataset through a dynamic time warping algorithm; The metabolic rate calculation module is used to calculate the baseline value of basal metabolic rate, the exercise metabolic compensation value, and the food thermic effect compensation value based on the multimodal fusion dataset, and then superimpose them to generate a real-time metabolic rate sequence. A metabolic state construction module is used to construct a metabolic state vector based on the real-time metabolic rate sequence and the nutrient quantification data; The reverse solving module is used to generate the nutritional requirement ratio based on the metabolic state vector and the preset activity plan data through a reverse solving algorithm; The dynamic correction output module is used to dynamically correct the nutritional requirement ratio based on the continuously monitored blood glucose time series data and output the final meal replacement plan.

10. A computer-readable storage medium, characterized in that: The computer program is stored that can be loaded by a processor and executed as described in any one of claims 1 to 8.

Images