Deep learning personalized nutrition recommendation method and system based on user portrait

Abstract

The invention discloses a deep learning personalized nutrition recommendation method and system based on a user portrait, and the method comprises the steps: collecting user physiological indexes, diet behaviors and environment parameters through a wearable device, intelligent tableware and an environment sensor, and obtaining collection data; inputting the collected data into a dynamic causal time sequence model, calculating the time-varying causal intensity of a diet event and a physiological index by adopting a convergent cross-mapping algorithm, and dynamically adjusting the length of a sliding window according to the maximum physiological response delay, a diabetes risk factor and a glycosylated hemoglobin deviation value; and extracting a user portrait vector containing metabolic phase shift through a bidirectional gating network. According to the method, a dynamic nutrition recommendation system of metabolism perception is constructed, and the core problems of dynamic metabolism response modeling and abnormal state real-time regulation and control are solved through a multi-modal data fusion and pathology-driven adaptive learning mechanism.

Description

Technical Field

[0001] This application relates to the technical field of nutritional recommendation, and particularly to a deep learning-based personalized nutritional recommendation method and system based on user portraits. Background Art

[0002] Existing personalized nutritional recommendation technologies generally rely on static physiological parameters and discrete dietary records, and it is difficult to dynamically capture the complex time-varying relationship between dietary behavior and metabolic response. Especially in the management of metabolic diseases such as diabetes, the lag of physiological indicators (such as postprandial blood glucose fluctuations and abnormal accumulation of ketone bodies) and the heterogeneity of multi-modal data (exercise, diet, environment) lead to significant limitations in traditional models:

[0003] On the one hand, a fixed time window is difficult to adapt to the individual-differentiated physiological response delay, resulting in missed detection of key metabolic events; on the other hand, the recommendation strategy lacks the ability to real-time perceive and adaptively adjust metabolic abnormal states (such as ketosis), which is likely to cause a mismatch between nutritional intervention and physiological needs. In addition, existing methods ignore the pathological constraints of medical ontology in cross-modal feature fusion, resulting in insufficient scientificity and safety of nutritional recommendations. Summary of the Invention

[0004] To solve the above problems, an embodiment of the present invention provides a deep learning-based personalized nutritional recommendation method based on user portraits, and the method includes:

[0005] Collect user physiological indicators, dietary behavior, and environmental parameters through wearable devices, smart tableware, and environmental sensors to obtain collected data;

[0006] Input the collected data into a dynamic causal time series model, use the convergent cross-mapping algorithm to calculate the time-varying causal intensity between dietary events and physiological indicators, dynamically adjust the sliding window length according to the maximum physiological response delay, diabetes risk factors, and glycated hemoglobin deviation value, and extract user portrait vectors containing metabolic phase offsets through a bidirectional gated network;

[0007] Input the user portrait vectors into a pathology-aware hybrid neural network model. The hybrid neural network model includes an encoder and a decoder. The encoder uses a medical ontology-constrained multi-head attention mechanism and maps different attention heads to independent physiological dimensions of glucose metabolism, lipid metabolism, and electrolyte balance respectively. The decoder generates a nutritional requirement vector for the next 24 hours through GRU and constrains the ratio range of ω3 fatty acids and ω6 fatty acids;

[0008] Trigger Bayesian online learning through a wearable sweat sensor to detect the ketone body concentration. When the ketone body concentration is abnormal, freeze the encoder parameters and preferentially update the decoder weights related to metabolism.

[0009] Furthermore, the wearable device is configured with a bioimpedance sensor array to construct a dynamic blood glucose map by periodically collecting skin conductivity and subcutaneous tissue dielectric properties and combining with a motion state compensation algorithm;

[0010] In the motion state compensation algorithm, the motion feature, i.e., the motion intensity index MSI, includes:

[0011]

[0012] In the formula, a i is the instantaneous acceleration value at the i-th sampling point, is the average acceleration within the window, ω rms is the root mean square value of the angular velocity, λ is the motion coupling coefficient, and N is the number of sampling points within the sliding window;

[0013] When MSI is less than the motion intensity threshold, only basic impedance measurement is enabled. When MSI is greater than or equal to the motion intensity threshold, the motion state compensation algorithm is activated. The motion state compensation algorithm includes:

[0014] θ corr = θ raw - η·ΔT·ln(MSI + 1);

[0015] In the formula, θ raw is the original phase angle, θ corr is the corrected phase angle; η is the compensation coefficient; ΔT is the difference between the current skin temperature and the baseline temperature in the user's resting state.

[0016] Furthermore, the piezoelectric array of the smart tableware acquires chewing mechanics waveforms by high-frequency sampling, analyzes the chewing pattern through a time-frequency domain feature extraction algorithm to generate chewing mechanics parameters, and the TOF camera reconstructs the three-dimensional point cloud of the food volume using multi-spectral imaging technology.

[0017] Furthermore, the method for dynamically adjusting the sliding window length and generating the user portrait vector includes a window scaling mechanism with multi-parameter coupling, phase-sensitive causal feature extraction, and dynamic portrait verification:

[0018] Window scaling mechanism with multi-parameter coupling:

[0019] Based on the maximum physiological response delay τ max (t), diabetes risk factor R diabetes and glycated hemoglobin deviation value ΔHbA1c(t), the real-time window length L(t) is calculated through a non-linear fusion function. The calculation formula of the real-time window length L(t) includes:

[0020]

[0021] In the formula, L baseis the initial reference length of the sliding window; θ is the normalization threshold; α and β are metabolic inertia adaptation coefficients;

[0022] Phase-sensitive causal feature extraction: In the adjusted sliding window, a bidirectional gated network is used to synchronously process the time-varying causal intensity sequence and the metabolic phase marker to generate a user portrait vector;

[0023] Dynamic portrait verification: The user portrait vector is input into the adversarial generation network to align the distribution with the clinical diagnosis label. If the KL divergence exceeds the divergence threshold, window length recalibration is triggered.

[0024] Furthermore, in the multi-head attention mechanism constrained by the medical ontology, the glucose metabolism dimension is coupled with the insulin secretion dynamic model, the lipid metabolism dimension is bound to the lipoprotein metabolism gradient feature, and the electrolyte balance dimension is associated with the urinary excretion rate time series;

[0025] Among them, the insulin secretion dynamic equation in the insulin secretion dynamic model includes:

[0026]

[0027] In the formula, I(t) is the insulin secretion rate at time t, G(t - τ) is the blood glucose concentration at time t - τ, A is the maximum rate parameter of insulin secretion, B is the blood glucose concentration sensitivity parameter, γ is the insulin degradation rate constant, τ is the delay time from blood glucose stimulation to insulin secretion, is the rate of change of the secretion rate, reflecting the immediate response to blood glucose stimulation.

[0028] Furthermore, the method for constraining the ratio range of ω3 fatty acids and ω6 fatty acids includes:

[0029] The ratio of ω3 fatty acids to ω6 fatty acids needs to satisfy 1:4 to 1:1, that is, the fatty acid ratio constraint term FattyRatio(x), and the mathematical expression includes:

[0030]

[0031] Furthermore, Bayesian online learning uses stochastic gradient variational Bayes to achieve online approximation of the posterior distribution of parameters. When the ketone body concentration is abnormal, the decoder weights are updated directionally based on the Fisher information matrix of metabolic-related parameters. The variational inference framework formula in Bayesian online learning includes:

[0032]

[0033] In the formula, q φ (w|D t ) is the variational posterior distribution; μ t is the mean vector of the weight distribution; Σ t-1 is the covariance matrix of the weight distribution; c is the metabolic sensitivity coefficient; b is the regularization coefficient; F(w t-1 ) is the Fisher information matrix; I is the identity matrix.

[0034] Furthermore, the directional update uses a double-filtering strategy to select the decoder weights. The double-filtering strategy includes a first layer and a second layer:

[0035] The first layer screens a subset of weights that directly participate in the lipid metabolism pathway based on the structural positions of the decoder weights in the metabolic dimension attention heads;

[0036] The second layer retains the top M% of the decoder weights with the highest sensitivity to the current abnormal state of ketone bodies through the relative ranking of Fisher information values, where M is a preset value.

[0037] A deep learning personalized nutrition recommendation system based on a user profile, the system includes:

[0038] A collection module, which collects user physiological indicators, dietary behaviors, and environmental parameters through wearable devices, smart utensils, and environmental sensors to obtain collection data;

[0039] A user profile generation module, which inputs the collection data into a dynamic causal time series model, calculates the time-varying causal intensity between dietary events and physiological indicators using the convergent cross-mapping algorithm, dynamically adjusts the sliding window length according to the maximum physiological response delay, diabetes risk factors, and glycated hemoglobin deviation values, and extracts a user profile vector containing metabolic phase shifts through a bidirectional gated network;

[0040] A nutrition recommendation module, which inputs the user profile vector into a pathology-aware hybrid neural network model. The hybrid neural network model includes an encoder and a decoder. The encoder uses a multi-head attention mechanism constrained by a medical ontology and maps different attention heads to independent physiological dimensions of glucose metabolism, lipid metabolism, and electrolyte balance respectively. The decoder generates a nutrition demand vector for the next 24 hours through GRU and constrains the ratio range of ω3 fatty acids and ω6 fatty acids;

[0041] An anomaly detection module, which triggers Bayesian online learning by detecting the ketone body concentration through a wearable sweat sensor. When the ketone body concentration is abnormal, it freezes the encoder parameters and preferentially updates the decoder weights related to metabolism.

[0042] The technical effects and advantages of the deep learning personalized nutrition recommendation method and system based on a user profile provided by the present invention:

[0043] The present invention constructs a dynamic nutrition recommendation system for metabolic perception, which solves the core problems of dynamic metabolic response modeling and real-time regulation of abnormal states through multimodal data fusion and a pathology-driven adaptive learning mechanism. Based on the window stretching mechanism of causal strength and metabolic risk factors, the present invention achieves precise adaptation of individual physiological response delays, enhancing the sensitivity of metabolic event detection; through a multi-dimensional attention mechanism embedded with medical ontologies, it ensures that nutrition recommendations conform to the medical logic of glucose and lipid metabolism and electrolyte balance, reducing intervention risks; through a Bayesian directed learning strategy triggered by ketone bodies, while maintaining the stability of the model, it enables rapid response to metabolic abnormal states and targeted parameter correction. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a flowchart of the deep learning personalized nutrition recommendation method based on user portraits in Embodiment 1;

[0045] Figure 2 It is a flowchart of the dynamic adjustment of the sliding window length and the method for generating user portrait vectors in Embodiment 1;

[0046] Figure 3 It is a schematic connection diagram of the deep learning personalized nutrition recommendation system based on user portraits in Embodiment 2. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0047] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0048] Embodiment 1:

[0049] Please refer to Figure 1 As shown, the deep learning personalized nutrition recommendation method based on user portraits in this embodiment includes:

[0050] Collect user physiological indicators, dietary behaviors, and environmental parameters through wearable devices, smart tableware, and environmental sensors to obtain collected data;

[0051] Input the collected data into a dynamic causal time series model, calculate the time-varying causal strength between dietary events and physiological indicators using the convergent cross-mapping algorithm, dynamically adjust the sliding window length according to the maximum physiological response delay, diabetes risk factors, and glycated hemoglobin deviation value, and extract user portrait vectors containing metabolic phase offsets through a bidirectional gated network;

[0052] Input the user profile vector into a pathology-aware hybrid neural network model. The hybrid neural network model includes an encoder and a decoder. The encoder adopts a medical ontology-constrained multi-head attention mechanism and maps different attention heads to independent physiological dimensions of glucose metabolism, lipid metabolism, and electrolyte balance respectively. The decoder generates a nutritional requirement vector for the next 24 hours through GRU and constrains the ratio range of ω3 fatty acids to ω6 fatty acids.

[0053] Trigger Bayesian online learning by detecting the ketone body concentration through a wearable sweat sensor. When the ketone body concentration is abnormal, freeze the encoder parameters and preferentially update the weights of the decoder related to metabolism.

[0054] Physiological indicators include continuous glucose monitoring, heart rate variability spectrum, and sweat metabolite concentration. Dietary behaviors include chewing mechanics parameters identified by a piezoelectric array and three-dimensional point cloud of food volume obtained by a TOF camera. Environmental parameters include oxygen saturation correction coefficient and respiratory stress index. Physiological indicators include continuous glucose monitoring, heart rate variability spectrum, and sweat metabolite concentration. Dietary behaviors include chewing mechanics parameters identified by a piezoelectric array and three-dimensional point cloud of food volume obtained by a TOF camera. Environmental parameters include oxygen saturation correction coefficient and respiratory stress index.

[0055] The wearable device is configured with a bioimpedance sensor array. By periodically collecting skin conductivity and subcutaneous tissue dielectric properties and combining with a motion state compensation algorithm, a continuous glucose monitoring is constructed.

[0056] Apply a multi-frequency alternating current signal (10 kHz - 1 MHz) through an electrode array, detect the amplitude and phase angle at 12 frequency points, which are the bioimpedance characteristics. Use a six-axis sensor to collect three-dimensional acceleration and angular velocity in real time, and use a PT1000 sensor to measure the epidermal temperature, which is the temperature characteristic.

[0057] In the motion state compensation algorithm, the motion feature, i.e., the motion intensity index MSI, includes:

[0058]

[0059] where a i is the instantaneous acceleration value at the i-th sampling point, is the average acceleration within the window, ω rms is the root mean square value of the angular velocity, λ is the motion coupling coefficient (default 0.2, adjusted by user calibration), and N is the number of sampling points within the sliding window (default N = 50, corresponding to 0.5 seconds).

[0060] When MSI is less than the motion intensity threshold (default 0.1), only basic impedance measurement is enabled. When MSI is greater than or equal to the motion intensity threshold, the motion state compensation algorithm is activated. The motion state compensation algorithm includes:

[0061] θ corr = θ raw -η·ΔT·ln(MSI + 1);

[0062] In the formula, θ raw is the original phase angle, that is, the uncorrected phase angle directly obtained from bioimpedance measurement, reflecting the electrical properties of biological tissues (such as cell membrane capacitance, extracellular fluid resistance, etc.), and θ corr is the corrected phase angle; η is the compensation coefficient, used to adjust the coupling effect of temperature and motion on the phase angle (dimensionless), determined by user calibration or factory calibration (default value 0.03); ΔT is the difference between the current skin temperature and the baseline temperature under the user's resting state, used to quantify the impact of temperature drift on impedance.

[0063] When the user exercises vigorously (the MSI value increases), the correction term on the right side of the formula (η·ΔT·ln(MSI + 1)) will increase, thereby dynamically reducing the measured value of the original phase angle (θ raw ), suppressing the interference of motion artifacts on the prediction of metabolic parameters.

[0064] When the exercise intensity index is calculated and the motion state compensation algorithm also compensates the bioimpedance characteristics, the impedance characteristics, motion characteristics, and temperature characteristics are input into the pre - constructed metabolic parameter mapping model, and the blood glucose concentration, extracellular fluid sodium ion concentration index, and subcutaneous fat decomposition rate are output. Then, a dynamic blood glucose map is constructed from the blood glucose concentration, extracellular fluid sodium ion concentration index, and subcutaneous fat decomposition rate; through the multi - modal sensing fusion and motion - temperature dynamic compensation mechanism, high - precision and anti - interference continuous blood glucose prediction is achieved.

[0065] The piezoelectric array of the intelligent tableware acquires the chewing mechanics waveform by high - frequency sampling, analyzes the chewing pattern through the time - frequency domain feature extraction algorithm, generates chewing mechanics parameters, and the TOF camera reconstructs the three - dimensional point cloud of the food volume using multi - spectral imaging technology.

[0066] The handle and contact surface of the intelligent tableware are embedded with a piezoelectric sensor array, arranged in a multi - row and multi - column matrix. Through the electromechanical coupling characteristics of the piezoelectric material, the mechanical signals generated during the user's chewing are captured in real - time. The sensor array supports three - dimensional mechanical detection (normal force, tangential force), and completely records the dynamic changes during the chewing process through high - frequency sampling, including the mechanical characteristics of the bite strength, chewing frequency, and swallowing action.

[0067] The time - frequency domain features include time - domain features and frequency - domain features. The time - domain features include the bite pulse width (reflecting the duration of a single chewing) and the peak force change rate (used to distinguish the chewing stage and the swallowing action), and the frequency - domain features include the main frequency energy distribution (identifying the chewing pattern, such as grinding or cutting type) and the degree of harmonic distortion (related to the food hardness, hard foods cause an increase in high - frequency harmonics).

[0068] Chewing mechanics parameters are quantitative indicators generated through multi-dimensional signal analysis. They include both mechanical characteristics (force, energy, and frequency) at the physical level and also derived evaluation parameters for behavior patterns and health status. Their core value lies in transforming complex chewing dynamics behaviors into computable and intervenable digital features.

[0069] The TOF camera calculates the distance between an object and the camera directly by emitting modulated light pulses and measuring the time difference of the light from emission to reflection back to the sensor; multi-spectral imaging captures the reflection characteristics of an object through light of different wavelengths (such as visible light, near-infrared, and short-wave infrared, etc.); an initial three-dimensional point cloud is generated from the TOF depth map, multi-spectral information is mapped to point cloud attributes (such as RGB color and near-infrared reflection intensity), missing points in occluded areas are repaired using multi-spectral texture information (such as the PatchMatch algorithm), and spectral features (such as sudden changes in edge reflectivity) are combined to enhance contour accuracy, thus completing the reconstruction of the three-dimensional point cloud of the food volume.

[0070] The fusion technology of TOF and multi-spectral significantly improves the three-dimensional reconstruction accuracy in complex scenes (such as non-rigid foods and multi-material mixtures) through the dual information of geometric depth + spectral attributes.

[0071] As Figure 2 shown, the dynamic adjustment of the sliding window length and the user portrait vector generation method include a window stretching and shrinking mechanism with multi-parameter coupling, phase-sensitive causal feature extraction, and dynamic portrait verification:

[0072] Window stretching and shrinking mechanism with multi-parameter coupling:

[0073] Based on the maximum physiological response delay τ max (t), diabetes risk factor R diabetes and glycated hemoglobin deviation value ΔHbA1c(t), the real-time window length L(t) is calculated through a non-linear fusion function. The calculation formula for the real-time window length L(t) includes:

[0074]

[0075] In the formula, L base is the initial reference length of the sliding window, usually set based on the physiological cycle or clinical experience; θ is the normalization threshold, used to adjust the influence of the glycated hemoglobin deviation value on the window length; α and β are metabolic inertia adaptation coefficients, used to adjust the influence of the diabetes risk factor and the glycated hemoglobin deviation value on the window length; tanh is the hyperbolic tangent function.

[0076] Phase-sensitive Causal Feature Extraction: Within the adjusted sliding window, a bidirectional gated network is used to synchronously process the time-varying causal intensity sequence and the metabolic phase markers (postprandial state and circadian rhythm), generating a user profile (i.e., user profile vector) that includes the metabolic lag gradient and the intervention sensitivity vector.

[0077] Dynamic Profile Verification: The user profile vector is input into the adversarial generation network to align the distribution with the clinical diagnosis label (such as insulin resistance classification). If the KL divergence exceeds the divergence threshold, window length recalibration is triggered.

[0078] By adjusting the dynamic window, the capture delay of metabolic events is shortened. Combining phase-sensitive feature extraction improves the prediction accuracy of postprandial hyperglycemia. Through adversarial verification, the distribution deviation between the user profile and clinical diagnosis is controlled within the divergence threshold, achieving clinical-level accuracy in diabetes risk warning sensitivity while keeping the daily computational cost as small as possible.

[0079] Through the collaborative action of a multi-head attention encoder constrained by medical ontology and a dynamic constraint decoder, the user physiological data in the user profile vector is mapped to independent dimensions of glucose metabolism, lipid metabolism, and electrolyte balance, generating a 24-hour nutritional requirement vector that conforms to individual pathological characteristics.

[0080] In the multi-head attention mechanism constrained by medical ontology, the glucose metabolism dimension is coupled with the insulin secretion dynamic model, the lipid metabolism dimension is bound to the lipoprotein metabolism gradient characteristics, and the electrolyte balance dimension is associated with the urinary excretion rate time series.

[0081] The insulin secretion dynamic equation in the insulin secretion dynamic model includes:

[0082]

[0083] In the formula, I(t) is the insulin secretion rate at time t, G(t - τ) is the blood glucose concentration at time t - τ, A is the maximum rate parameter of insulin secretion, B is the blood glucose concentration sensitivity parameter, γ is the insulin degradation rate constant, τ is the delay time from blood glucose stimulation to insulin secretion, is the rate of change of the secretion rate, reflecting the immediate response to blood glucose stimulation.

[0084] The lipoprotein metabolism gradient characteristics include a time gradient and a spatial gradient;

[0085] The time gradient is that lipoprotein metabolism has circadian rhythm and postprandial fluctuations. By calculating the 3-hour change rate of non-HDL-C, the postprandial lipid metabolism load (such as the triglyceride peak) can be obtained;

[0086] The spatial gradient is the lipoprotein concentration difference based on carotid artery ultrasound and venous blood samples, which can reflect the risk of lipid deposition on the blood vessel wall.

[0087] The electrolyte (Na + , K + and Ca 2+ ) homeostasis depends on the dynamic balance between glomerular filtration rate (eGFR) and tubular reabsorption. The urinary excretion rate time series of the urinary system can reflect circadian rhythm (such as reduced nocturnal sodium excretion) and acute compensation (such as enhanced H + -K + exchange during acidosis).

[0088] To prevent reverse causality interference, a causal mask matrix M ij is designed based on urinary physiology as follows:

[0089]

[0090] In the formula, i is the index of the current time step (target time step), and j is the index of the historical time step (the time step of interest).

[0091] Specifically, j ≤ i is the causal constraint, which strictly ensures that future information is not leaked. The core goal of the causal mask matrix is to simulate the delay effect of tubular reabsorption and prevent the model from using future information during prediction. Dynamically adjusting the length of the historical time window accessible to the target time step can make the model more in line with the physiological mechanism.

[0092] By decomposing the medical ontology into three major metabolic dimensions of sugar, lipid, and electrolyte, and using dedicated attention heads that conform to the physiological mechanism, a paradigm upgrade from data-driven to knowledge-embedded has been achieved.

[0093] The method for constraining the ratio range of ω3 fatty acid to ω6 fatty acid includes:

[0094] The ratio of ω3 fatty acid to ω6 fatty acid needs to satisfy 1:4 to 1:1, so as to inhibit pro-inflammatory factors (such as IL-6) and optimize metabolic health. That is, the fatty acid ratio constraint term FattyRatio(x), and its mathematical expression includes:

[0095]

[0096] Based on the nutritional requirement vector and the constrained fatty acid ratio, multi-objective optimization search can be performed in the molecular cuisine database. The 3D printing recipe structure parameters are generated with the gastric emptying half-life and the degree of molecular crosslinking as the constraint conditions, or device-free quantification can be carried out through the food portion comparison table, and then ingredient screening and combination are performed through the language large model.

[0097] Bayesian online learning adopts a variational inference framework. When the ketone body concentration is abnormal, the decoder weights are updated directionally based on the Fisher information matrix of the metabolic-related parameters.

[0098] Online approximation of the posterior distribution of parameters is achieved using stochastic gradient variational Bayes. The variational inference framework formulas include:

[0099]

[0100] In the formula, q φ (w|D t ) is the variational posterior distribution, representing the probability distribution of the model weights w given the current dataset D t ; μ t is the mean vector of the weight distribution; Σ t -1 is the covariance matrix of the weight distribution, reflecting the uncertainty of the parameters; c is the metabolic sensitivity coefficient, controlling the influence weight of Fisher information on the covariance; b is the regularization coefficient, preventing the covariance matrix from being singular; F(w t-1 ) is the Fisher information matrix, measuring the sensitivity of the model parameters to the prediction of ketone body concentration; represents the expectation calculation for the current dataset D t ; I is the identity matrix.

[0101] The directional update uses a double filtering strategy to select the decoder weights:

[0102] The first layer filters a subset of weights that directly participate in the lipid metabolism pathway based on the structural position of the decoder weights in the metabolic dimension attention heads;

[0103] The second layer retains the top M% of the decoder weights with the highest sensitivity to the current abnormal state of ketone bodies through the relative ranking of Fisher information values, where M is a preset value.

[0104] Through the metabolically sensitive Bayesian online learning framework, the model can directionally enhance the update intensity of lipid metabolism-related parameters when ketone bodies are abnormal. At the same time, it uses the sparsification screening of the Fisher information matrix to reduce ineffective parameter updates, reducing computational energy consumption while maintaining prediction accuracy and significantly improving the real-time response ability to abnormal metabolic states.

[0105] According to the respiratory stress index in the environmental parameters, the weighted combination of particulate matter concentration and oxidizing gas concentration is fused, and the antioxidant nutrient ratio is dynamically adjusted when the stress index exceeds the safety threshold.

[0106] Example 2:

[0107] As Figure 3 shown, based on the same inventive concept as the deep learning personalized nutrition recommendation method based on user portraits in the foregoing embodiment, this application provides a deep learning personalized nutrition recommendation system based on user portraits. The system in the embodiments of this application and the method embodiments are based on the same inventive concept. Among them, the system includes:

[0108] A collection module, which collects users' physiological indicators, dietary behaviors and environmental parameters through wearable devices, smart tableware and environmental sensors to obtain collection data;

[0109] A user portrait generation module, which inputs the collection data into a dynamic causal time series model, calculates the time-varying causal intensity between dietary events and physiological indicators using the convergent cross mapping algorithm, dynamically adjusts the sliding window length according to the maximum physiological response delay, diabetes risk factors and glycated hemoglobin deviation value, and extracts user portrait vectors containing metabolic phase shifts through a bidirectional gated network;

[0110] A nutrition recommendation module, which inputs the user portrait vectors into a pathologically aware hybrid neural network model. The hybrid neural network model includes an encoder and a decoder. The encoder uses a medical ontology-constrained multi-head attention mechanism and maps different attention heads to independent physiological dimensions of glucose metabolism, lipid metabolism and electrolyte balance respectively. The decoder generates a nutrition demand vector for the next 24 hours through a GRU and constrains the ratio range of ω3 fatty acids to ω6 fatty acids;

[0111] An anomaly detection module, which triggers Bayesian online learning by detecting the ketone body concentration through a wearable sweat sensor. When the ketone body concentration is abnormal, it freezes the encoder parameters and preferentially updates the weights of the decoder related to metabolism.

[0112] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention also intends to include these changes and modifications.

[0113] The above are only the preferred specific embodiments of the embodiments of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed in the present application, according to the technical solution and its concept of the present application, makes equivalent substitutions or changes, and should be covered by the protection scope of the present application.

Claims

1. A deep learning personalized nutrition recommendation method based on user portraits, characterized by: Methods include: Collect user physiological indicators, eating behaviors, and environmental parameters through wearable devices, smart tableware, and environmental sensors to obtain collected data; The collected data is input into a dynamic causal time series model. A convergent cross-mapping algorithm is used to calculate the time-varying causal strength between dietary events and physiological indicators. The sliding window length is dynamically adjusted based on the maximum physiological response delay, diabetes risk factors, and glycated hemoglobin deviation value. A bidirectional gating network is used to extract user profile vectors containing metabolic phase offsets. The user portrait vector is input into a pathology-aware hybrid neural network model. The hybrid neural network model includes an encoder and a decoder. The encoder uses a multi-head attention mechanism constrained by medical ontology and maps different attention heads to independent physiological dimensions of glucose metabolism, lipid metabolism, and electrolyte balance. The decoder uses GRU to generate a nutritional requirement vector for the next 24 hours and constrains the ratio range of ω3 fatty acids to ω6 fatty acids. The ketone body concentration is detected by a wearable sweat sensor to trigger Bayesian online learning. When the ketone body concentration is abnormal, the encoder parameters are frozen and the metabolism-related decoder weights are updated first.

2. The deep learning personalized nutrition recommendation method based on user portrait according to claim 1 is characterized in that: The wearable device is equipped with a bioimpedance sensor array, which periodically collects skin conductivity and subcutaneous tissue dielectric properties, and combines it with a motion state compensation algorithm to construct a dynamic blood glucose map. In the motion state compensation algorithm, the motion characteristics, namely the motion intensity index (MSI), include: Where a i is the instantaneous value of acceleration at the i-th sampling point, is the mean acceleration within the window, ω rms is the root mean square value of the angular velocity, λ is the motion coupling coefficient, and N is the number of sampling points in the sliding window; When the MSI is less than the motion intensity threshold, only basic impedance measurement is enabled. When the MSI is greater than or equal to the motion intensity threshold, the motion state compensation algorithm is activated. The motion state compensation algorithm includes: i corr =θ raw -η·ΔT·ln(MSI+1); Where θ raw is the original phase angle, θ corr is the corrected phase angle; η is the compensation coefficient; ΔT is the difference between the current skin temperature and the user's baseline temperature in the resting state.

3. The deep learning personalized nutrition recommendation method based on user portrait according to claim 1 is characterized in that: The piezoelectric array of the smart tableware uses high-frequency sampling to obtain chewing mechanical waveforms, and analyzes the chewing pattern through time-frequency domain feature extraction algorithm to generate chewing mechanical parameters. The TOF camera uses multispectral imaging technology to reconstruct the three-dimensional point cloud of the food volume.

4. The deep learning personalized nutrition recommendation method based on user portrait according to claim 1 is characterized in that: The method for dynamically adjusting the sliding window length and generating a user profile vector includes a multi-parameter coupled window scaling mechanism, phase-sensitive causal feature extraction, and dynamic profile verification: Multi-parameter coupled window scaling mechanism: Based on the maximum physiological response delay τ max (t), diabetes risk factor R diabetes and the glycated hemoglobin deviation value ΔHbA1c(t), and the real-time window length L(t) is calculated by a nonlinear fusion function. The calculation formula of the real-time window length L(t) includes: Where, L base is the initial reference length of the sliding window; θ is the normalization threshold; α and β are metabolic inertia adaptation coefficients; Phase-sensitive causal feature extraction: A bidirectional gating network is used to synchronously process the time-varying causal strength sequence and metabolic phase markers within an adjusted sliding window to generate a user profile vector. Dynamic portrait verification: The user portrait vector is input into the adversarial generative network and aligned with the clinical diagnosis label. If the KL divergence exceeds the divergence threshold, the window length recalibration is triggered.

5. The deep learning personalized nutrition recommendation method based on user portrait according to claim 1 is characterized in that: In the multi-head attention mechanism constrained by medical ontology, the glucose metabolism dimension is coupled to the dynamic model of insulin secretion, the lipid metabolism dimension is bound to the lipoprotein metabolism gradient characteristics, and the electrolyte balance dimension is associated with the urinary system excretion rate time series; The dynamic equations of insulin secretion in the dynamic model of insulin secretion include: Where I(t) is the insulin secretion rate at time t, G(t-τ) is the blood glucose concentration at time t-τ, A is the maximum rate parameter of insulin secretion, B is the blood glucose concentration sensitivity parameter, γ is the insulin degradation rate constant, and τ is the delay time from blood glucose stimulation to insulin secretion. is the rate of change of secretion rate.

6. The deep learning personalized nutrition recommendation method based on user portrait according to claim 1 is characterized in that Methods for constraining the range of the ratio of ω3 fatty acids to ω6 fatty acids include: The ratio of ω3 fatty acids to ω6 fatty acids must satisfy the range of 1:4 to 1:1, which is the fatty acid ratio constraint term FattyRatio(x). The mathematical expression includes:

7. The method for personalized nutrition recommendation based on deep learning of user portraits according to claim 1, characterized in that: Bayesian online learning uses stochastic gradient variational Bayesian to implement online approximation of the parameter posterior distribution. When ketone body concentration is abnormal, the decoder weights are updated in a targeted manner based on the Fisher information matrix of metabolic-related parameters. The variational inference framework formula in Bayesian online learning includes: Where q φ (w|D t ) is the variational posterior distribution; μ t is the mean vector of weight distribution; Σ t -1 is the covariance matrix of weight distribution; c is the metabolic sensitivity coefficient; b is the regularization coefficient; F(w t-1 ) is the Fisher information matrix; I is the identity matrix.

8. The deep learning personalized nutrition recommendation method based on user portrait according to claim 7 is characterized in that: Directed update uses a dual filtering strategy to select decoder weights. The dual filtering strategy includes the first and second layers: The first layer selects a subset of weights directly involved in the lipid metabolism pathway based on the structural position of the decoder weights in the attention head of the metabolic dimension; The second layer retains the top M% decoder weights that are most sensitive to the current abnormal ketone body state through relative sorting of Fisher information values, where M is a preset value.

9. A deep learning personalized nutrition recommendation system based on user portraits, characterized by: The system includes: A collection module, which collects user physiological indicators, eating behaviors, and environmental parameters through wearable devices, smart tableware, and environmental sensors to obtain collected data; a user profile generation module, which inputs the collected data into a dynamic causal time series model, uses a convergent cross-mapping algorithm to calculate the time-varying causal strength between dietary events and physiological indicators, dynamically adjusts the sliding window length based on the maximum physiological response delay, diabetes risk factors, and glycated hemoglobin deviation value, and extracts a user profile vector containing metabolic phase offset through a bidirectional gating network; A nutrition recommendation module, which inputs the user portrait vector into a pathology-aware hybrid neural network model. The hybrid neural network model includes an encoder and a decoder. The encoder uses a multi-head attention mechanism constrained by medical ontology and maps different attention heads to independent physiological dimensions of glucose metabolism, lipid metabolism, and electrolyte balance. The decoder uses GRU to generate a nutritional requirement vector for the next 24 hours and constrains the ratio range of ω3 fatty acids to ω6 fatty acids. An anomaly detection module, which detects ketone body concentration through a wearable sweat sensor and triggers Bayesian online learning. When the ketone body concentration is abnormal, the encoder parameters are frozen and the metabolism-related decoder weights are preferentially updated.

Images