A multimodal child vital sign assessment system and meal matching method

Abstract

The application discloses a multi-modal child physical sign evaluation system and a dietary matching method, and belongs to the technical field of intelligent detection and data processing. The system comprises: a multi-modal data acquisition module for synchronously acquiring bone age images, multi-frequency body composition impedance and eye feature data; an AI fusion operation engine, a multi-modal fusion model built in a Transformer encoder, the model comprising three independent 128-dimensional modal encoding branches, a 4-head attention cross-modal fusion layer and a classification output layer, a linear interpolation modal compensation algorithm being triggered when some modal data is missing, 1000 groups of simulated missing experiments verifying that the mean of the classification label deviation is 3.2% and the standard deviation is 1.1%; a dietary information matching module for searching for a corresponding dietary scheme through a table lookup algorithm according to a 5-level classification label; and a longitudinal tracking module for dynamically adjusting dietary information when periodic retesting is performed. The application solves the problems of data fragmentation and strong subjectivity of evaluation of existing physical sign detection equipment.

Description

Technical Field

[0001] This invention belongs to the field of intelligent health detection and data processing technology, specifically relating to a multimodal children's vital sign assessment system and dietary matching method. Background Technology

[0002] Currently, equipment for detecting growth-related physical signs in children is generally deployed independently, with bone age imaging devices, body composition analyzers, and eye image acquisition devices belonging to different manufacturers. The data interfaces are not standardized, making joint analysis of physical sign data impossible. Existing equipment can only output single-dimensional detection values, lacking multimodal data fusion, quantification, and grading logic, and the assessment results heavily rely on subjective human judgment. A search revealed several existing technologies for monitoring children's health. For example, a big data-based preschool child dietary management system collects data on the shape of a child's head, head circumference, chest, and legs using a body shape detection module. It then uses the relationship between body shape and micronutrients to determine micronutrient deficiencies and provides dietary recommendations after filtering through a food database. Intelligent evaluation methods for infant and toddler diets involve inputting food intake information and physical development data via a mobile app. These methods utilize a pre-set health evaluation model to obtain evaluation results and provide feeding suggestions. However, the aforementioned existing technologies all have the following shortcomings: First, the acquisition and fusion of multimodal data lack a unified hardware architecture, and different detection devices are still in a separate and independent state; Second, existing assessment models cannot stably output grading results when data for a certain modality is missing, and lack an effective modality compensation mechanism; Third, health intervention programs are often disconnected from detection devices, and lack a closed-loop system that automatically matches standardized dietary information based on multidimensional objective vital signs data. Summary of the Invention

[0003] I. Technical Issues The technical problem that this invention aims to solve is that existing vital sign detection equipment suffers from fragmented data, strong subjective assessment, and a lack of standardized dietary information matching mechanism that links with objective vital sign data. II. Technical Solution To address the above problems, this invention provides a multimodal pediatric vital sign assessment system, comprising: The multimodal data acquisition module is used to simultaneously acquire bone age imaging data, multi-frequency body composition impedance data, and ocular feature image data of the subject. The bone age imaging data acquisition unit adopts a low-dose X-ray imaging architecture to reduce the radiation dose per exposure. The body composition detection unit adopts a multi-frequency multi-electrode contact impedance acquisition architecture to simultaneously measure multi-segment impedance data. The ocular feature acquisition unit adopts a multispectral imaging architecture to automatically extract quantitative ocular feature parameters. The AI ​​fusion computing engine incorporates a multimodal fusion model built on a Transformer encoder. This model comprises the following structure: (a) three independent modal coding branches, corresponding to bone age imaging data, multi-frequency body component impedance data, and eye feature image data, respectively. Each branch contains two fully connected layers with a 128-dimensional output; (b) a cross-modal attention fusion layer employing a multi-head attention mechanism with four attention heads, which weights and fuses the three 128-dimensional feature vectors using a learnable weight matrix to output a 128-dimensional fused feature vector; and (c) a classification output layer containing two fully connected layers. The model integrates feature vectors and maps them to 5-level growth sign grading labels. It is trained on 21,000 sets of desensitized children's growth sign samples, using a cross-entropy loss function, the Adam optimizer, a learning rate of 0.001, 50 training epochs, and a batch size of 32. When a modality is missing, a built-in linear interpolation modality compensation algorithm is triggered. This algorithm interpolates and fills in the missing modality based on the mean and variance of the corresponding modality in the complete dataset. Verified through 1000 simulated missing data experiments, when bone age data is missing, the mean deviation of the grading labels from the complete dataset is 3.2%, the standard deviation is 1.1%, and the maximum deviation does not exceed 5.8%. The dietary information matching module has a built-in multi-dimensional vital signs-dietary formula mapping database. This database contains the correspondence between 5-level graded labels and 5-level dietary formula schemes: Level 1 labels correspond to basic balanced dietary schemes, Level 2 labels correspond to protein-supplemented dietary schemes, Level 3 labels correspond to energy-controlled dietary schemes, Level 4 labels correspond to comprehensive conditioning dietary schemes, and Level 5 labels correspond to intensive intervention dietary schemes. The matching module retrieves the corresponding dietary schemes based on the growth vital signs graded labels output by the AI ​​fusion computing engine through a lookup table matching algorithm, and outputs standardized dietary formula information including ingredient composition, single consumption amount, and consumption period. The longitudinal tracking module stores the subject's historical vital signs data and matching dietary formula information. During subsequent periodic retests, it automatically compares and updates the vital signs grading labels and dynamically adjusts the output dietary formula information. This invention also provides a multimodal method for assessing children's physical characteristics and matching their diet based on the above system, comprising the following steps: First, simultaneously collecting bone age, body composition, and eye feature data of the examinee through a multimodal data acquisition module; Second, inputting the collected multidimensional data into an AI fusion computing engine for normalization and feature fusion, and outputting growth characteristic grading labels; Third, using a dietary information matching module, retrieving a pre-set database based on the growth characteristic grading labels, and outputting corresponding standardized dietary formula information and consumption parameters; Fourth, storing the examinee's historical physical characteristic data and matched dietary formula information through a longitudinal tracking module, automatically comparing and updating the physical characteristic grading labels during subsequent periodic retests, and dynamically adjusting the output dietary formula information. III. Beneficial Effects The beneficial effects of this invention are as follows: it integrates three types of separate detection devices into a unified data bus, solving the problem of data fragmentation; the multimodal fusion model is trained with real datasets, avoiding bias from subjective human judgment; when data of a certain modality is missing, the stability of the graded label output is maintained through the modal compensation algorithm; it realizes the automated linkage between detection data and dietary formula; and it achieves dynamic adjustment through the longitudinal tracking module, forming a complete data closed loop. Detailed Implementation

[0004] Example 1: Multimodal Data Acquisition and Fusion The system in this embodiment includes a bone age detection unit, an eight-electrode multi-frequency body component detection unit, and a one-second eye feature acquisition unit. The bone age detection unit adopts a micro-dose X-ray imaging architecture, with tube voltage adjustable from 40 to 70 kV, single exposure radiation dose ≤ 0.001 mSv, imaging resolution ≥ 20 lp / mm, and outputs structured bone age data conforming to the DICOM 3.0 standard. The body composition detection unit adopts 5kHz / 50kHz / 250kHz three-frequency synchronous detection, with an excitation current ≤90μA, and outputs standardized body composition indicators such as body fat percentage and skeletal muscle mass. The eye feature acquisition unit is equipped with an 8MP medical 4K optical lens, which automatically extracts quantitative parameters such as pupil diameter and iris texture. The AI ​​fusion computing engine normalizes the three modalities of data to the 0-1 range before inputting them into the Transformer encoder. The model is trained on 21,000 sets of desensitized children's physical characteristics samples, aged 3-12 years, with a male-to-female ratio of 1:1.1 and a body fat percentage ranging from 8% to 35%. The model's performance on the test set is: accuracy 89.3%, recall 87.1%, and F1 score 88.2% (based on 3,000 independent test samples). An attention mechanism dynamically allocates the weights of each modality's features. The error control mechanism of the modal compensation algorithm is as follows: When a missing modality is detected, the system first identifies the type of the missing modality (bone age / body composition / eye features); it retrieves the mean and standard deviation of the modality for the current age group and gender from a pre-set complete data set statistics table; it uses the mean as the interpolation filler value for the missing modality and calculates the complete feature vector accordingly; the interpolated feature vector is input into the Transformer model, and the grading label is output. Validated by 1000 sets of simulated missing data experiments, when bone age data is missing, the mean deviation of the grading label from the complete data set is 3.2%, and the standard deviation is 1.1%; among these, 68.4% of the samples have a deviation less than 3%, 24.1% have a deviation between 3% and 5%, 7.5% have a deviation greater than 5%, and the maximum deviation is 5.8%. Example 2: Dietary Information Matching and Longitudinal Tracking The dietary information matching module has a built-in multi-dimensional vital sign-dietary formula mapping database. When the AI ​​engine outputs growth vital sign grading labels, the matching module automatically retrieves the corresponding standardized dietary formula information from the database. Based on the vital sign grading labels, the matching module outputs the corresponding dietary formula information and generates dietary plan parameters including single consumption amount and consumption period. The longitudinal tracking module supports local and cloud data storage. After local structured processing of the raw image data, only numerical indicators are retained. The system performs retests at preset intervals, automatically compares historical data, dynamically updates vital sign grading labels, and synchronously adjusts the output dietary formula information, forming a complete data loop. Example 3: Comparison and verification with existing technologies This embodiment compares and verifies the system of the present invention with existing technologies. One hundred children aged 6-12 years were selected and divided into an experimental group (n=50) and a control group (n=50). The experimental group used the system of the present invention for vital sign assessment and dietary matching, while the control group used the traditional single BMI index assessment method. Results showed that the multimodal data collection completeness rate was 98% in the experimental group and 85% in the control group; the accuracy rate of vital sign assessment was 89.3% in the experimental group and 72.1% in the control group; the dietary formula matching satisfaction rate was 92% in the experimental group and 68% in the control group. The results indicate that the system of the present invention is significantly superior to existing technologies in terms of data collection completeness, assessment accuracy, and dietary matching satisfaction.

Claims

1. A multimodal pediatric vital sign assessment system, characterized in that, include: The multimodal data acquisition module is used to simultaneously acquire bone age imaging data, multi-frequency body component impedance data, and ocular feature image data of the examinee. The bone age imaging data acquisition unit adopts a low-dose X-ray imaging architecture to reduce the radiation dose per exposure; the body composition detection unit adopts a multi-frequency multi-electrode contact impedance acquisition architecture to simultaneously measure multi-segment impedance data; and the eye feature acquisition unit adopts a multispectral imaging architecture to automatically extract quantitative eye feature parameters. The AI ​​fusion computing engine incorporates a multimodal fusion model built on a Transformer encoder. This model comprises the following structure: (a) three independent modal coding branches, corresponding to bone age imaging data, multi-frequency body component impedance data, and eye feature image data, respectively. Each branch contains two fully connected layers with a 128-dimensional output; (b) a cross-modal attention fusion layer employing a multi-head attention mechanism with four attention heads, which weights and fuses the three 128-dimensional feature vectors using a learnable weight matrix to output a 128-dimensional fused feature vector; and (c) a classification output layer containing two fully connected layers. The model integrates feature vectors and maps them to 5-level growth sign grading labels. It is trained on 21,000 sets of desensitized children's growth sign samples, using a cross-entropy loss function, the Adam optimizer, a learning rate of 0.001, 50 training epochs, and a batch size of 32. When a modality is missing, a built-in linear interpolation modality compensation algorithm is triggered. This algorithm interpolates and fills in the missing modality based on the mean and variance of the corresponding modality in the complete dataset. Verified through 1000 simulated missing data experiments, when bone age data is missing, the mean deviation of the grading labels from the complete dataset is 3.2%, the standard deviation is 1.1%, and the maximum deviation does not exceed 5.8%. The dietary information matching module has a built-in multi-dimensional vital signs-dietary formula mapping database. This database contains the correspondence between 5-level graded labels and 5-level dietary formula schemes: Level 1 labels correspond to basic balanced dietary schemes, Level 2 labels correspond to protein-supplemented dietary schemes, Level 3 labels correspond to energy-controlled dietary schemes, Level 4 labels correspond to comprehensive conditioning dietary schemes, and Level 5 labels correspond to intensive intervention dietary schemes. The matching module retrieves the corresponding dietary schemes based on the growth vital signs graded labels output by the AI ​​fusion computing engine through a lookup table matching algorithm, and outputs standardized dietary formula information including ingredient composition, single consumption amount, and consumption period. The longitudinal tracking module stores the subject's historical vital signs data and matching dietary formula information. During subsequent periodic retests, it automatically compares and updates the vital signs grading labels and dynamically adjusts the output dietary formula information.

2. The multimodal pediatric vital sign assessment system according to claim 1, characterized in that, The multimodal data acquisition module includes: The bone age detection unit adopts a micro-dose X-ray imaging architecture, with an adjustable tube voltage of 40~70kV, a single exposure radiation dose ≤0.001mSv, and an imaging resolution ≥20lp / mm. It is used to acquire wrist images and output bone age structured data that conforms to the DICOM 3.0 standard. The body composition detection unit adopts a 5kHz / 50kHz / 250kHz three-frequency multi-electrode segmented contact impedance acquisition architecture with an excitation current ≤90μA. It is used to simultaneously acquire multi-segment impedance data and output body fat percentage and skeletal muscle mass as standardized body composition indicators. The eye feature acquisition unit adopts a multispectral imaging architecture and is equipped with an 8MP medical 4K optical lens to acquire eye images and automatically extract quantitative eye feature parameters such as pupil diameter and iris texture.

3. The multimodal pediatric vital sign assessment system according to claim 1, characterized in that, The AI ​​fusion computing engine's model performance on the test set was: accuracy 89.3%, recall 87.1%, F1 score 88.2%, and the test set consisted of 3000 independent samples.

4. A multimodal method for assessing children's vital signs and matching their diet based on the system described in any one of claims 1 to 3, characterized in that, It includes the following steps: The first step is to simultaneously collect the subject's bone age, body composition, and eye characteristics data using a multimodal data acquisition module; The second step is to input the collected multi-dimensional data into the AI ​​fusion computing engine for normalization and feature fusion, and output growth sign grading labels. The third step involves using the dietary information matching module to retrieve the preset database based on the growth sign grading labels and outputting the corresponding standardized dietary formula information and consumption parameters. The fourth step involves storing the subject's historical vital signs data and matching dietary formula information through a longitudinal tracking module. During subsequent periodic retests, the vital signs classification labels are automatically compared and updated, and the output dietary formula information is dynamically adjusted.