Measurement data auxiliary analysis method for health monitor based on AI large model

Abstract

The invention relates to the technical field of artificial intelligence, and discloses a measurement data auxiliary analysis method for a health monitor based on an AI large model, and the method comprises the steps: obtaining the physiological parameter measurement data of a user through the health monitor; combining the measurement data and the basic information and historical health data of the user into an analysis context; a pre-trained AI large model is utilized to perform deep interpretation on the analysis context, and the AI large model is trained based on mass medical knowledge and can identify potential modes, relevance and risk signals in the data; and finally, generating a comprehensive analysis report including health condition evaluation, trend interpretation, abnormity early warning and personalized health suggestions. The limitation that a traditional health monitor can only provide isolated data is overcome, intelligent and scene-based health insight comparable to professional consultants is provided for users through powerful reasoning and cognitive ability of the AI large model, and the value and practicability of family health monitoring are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method for auxiliary analysis of measurement data in health monitoring instruments based on AI large models. Background Technology

[0002] While current health monitoring devices can collect basic physiological data such as heart rate and blood pressure, they generally suffer from insufficient data fusion capabilities. Existing technologies mostly process single-type data such as numerical values, images, and text separately, lacking an effective multimodal information integration mechanism. This makes it impossible to explore the potential correlations between data from different dimensions, resulting in analysis results that only present isolated data and fail to form a comprehensive user health profile. This single-modal analysis mode ignores the synergistic indicative significance of physiological indicators, vital signs, images, environmental descriptions, and other information, failing to provide in-depth data support for health assessment and limiting the intelligence level of the monitoring device.

[0003] Meanwhile, traditional health monitoring technologies mostly focus on alarms for obvious abnormalities that have already occurred, lacking the ability to dynamically quantify and assess the risk of sub-health transformation. Existing solutions often rely on fixed thresholds for judgment, making it difficult to capture early risk signals that are weak in intensity but have a long duration. They also cannot analyze the evolution trend of risks by combining the user's historical health trajectory, resulting in a lag in early warning of sub-health status, making it difficult for users to take early intervention measures. In addition, existing report generation mechanisms lack personalized adaptation capabilities and cannot transform complex data into actionable health guidance, making it difficult to meet users' needs for precise and scenario-based health management. Therefore, how to improve the personalized adaptation capabilities of the report generation mechanism to meet users' needs for precise and scenario-based health management has become an urgent problem to be solved. Summary of the Invention

[0004] This invention provides a method for auxiliary analysis of measurement data in health monitoring instruments based on AI large models, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a measurement data-assisted analysis method for health monitoring instruments based on AI large-scale models, comprising: S1, through the non-contact and contact sensors of the health monitor, synchronously acquires the user's multimodal physiological measurement data at a specific time point to construct a physiological parameter measurement dataset for the user's current cycle; S2, interactive fusion is performed to parse numerical feature vectors, image feature vectors and text feature vectors from the physiological parameter measurement dataset to obtain an analysis context vector sequence; S3, extract the historical time-series data sequence corresponding to the current test item from the historical health record, input the analysis context vector sequence and the historical time-series data sequence into a pre-trained dedicated AI large model, and output a set of deep interpretation signals, wherein the deep interpretation signals include: current physiological state assessment label, abnormal pattern description, and key potential correlation and weak risk signal clusters identified from the data; S4. Based on the key potential correlations and the weak risk signal clusters, as well as the changing trends of the corresponding signals in the historical time series data sequence, analyze the user's sub-health conversion risk to obtain the sub-health risk index and the corresponding risk level label. S5. Based on the current physiological state assessment labels, abnormal pattern descriptions, sub-health risk index, and risk level labels, a structured, personalized, comprehensive health report is automatically generated according to a preset report template.

[0006] In a preferred embodiment, the multimodal physiological measurement data of the user at a specific time point is simultaneously acquired through the non-contact and contact sensors of the health monitor to construct a physiological parameter measurement dataset for the user's current cycle, including: Based on the integrated contact sensor on the health monitor, a set of physiological parameter values ​​of the user during the monitoring period is obtained, wherein the set of physiological parameter values ​​includes heart rate, blood pressure, blood oxygen saturation, and body surface temperature; Based on the non-contact sensor integrated on the health monitor, user vital sign image data and environmental perception data synchronized with the physiological parameter values ​​are acquired during the monitoring period; The physiological parameter values, user vital sign image data, and environmental perception data are integrated with the timestamps to obtain a time series sequence of multimodal measurement data. The aligned multimodal measurement data are format-standardized and structurally reorganized to construct a user's current cycle physiological parameter measurement dataset containing numerical, image, and descriptive fields.

[0007] In a preferred embodiment, the interactive fusion parses numerical feature vectors, image feature vectors, and text feature vectors from the physiological parameter measurement dataset to obtain an analysis context vector sequence, including: Numerical features of numerical fields in the physiological parameter measurement dataset are extracted to obtain a numerical feature vector; Hierarchical features of image fields in the physiological parameter measurement dataset are extracted to obtain image feature vectors; Natural language understanding and encoding are performed on the descriptive fields in the physiological parameter measurement dataset to obtain text feature vectors; The numerical feature vector, the image feature vector, and the text feature vector are subjected to feature dimension alignment mapping processing. Based on the interactive attention fusion algorithm, the dimension-aligned numerical feature vector, the image feature vector, and the text feature vector are subjected to cross-modal information weighted fusion to obtain the analysis context vector sequence.

[0008] In a preferred embodiment, the step of extracting numerical features from the numerical fields of the physiological parameter measurement dataset to obtain a numerical feature vector includes: Outlier cleaning and dimensional standardization are performed on the numerical fields in the physiological parameter measurement dataset, which includes heart rate, blood pressure, blood oxygen saturation, and body surface temperature, to obtain a standardized numerical sequence. The standardized numerical sequence is subjected to time-domain statistical features and morphological features to obtain a numerical feature vector.

[0009] In a preferred embodiment, the step of extracting hierarchical features from the image-type fields in the physiological parameter measurement dataset to obtain an image feature vector includes: The image and video data contained in the image field are preprocessed to obtain standardized image data; Hierarchical visual feature extraction is performed on the standardized image data to obtain image feature maps that represent the multi-scale semantic information of the images; The image feature map is subjected to spatial dimension compression and vectorization transformation to obtain the image feature vector.

[0010] In a preferred embodiment, the step of performing natural language understanding and encoding on the descriptive fields in the physiological parameter measurement dataset to obtain a text feature vector includes: The original text data contained in the descriptive fields of the physiological parameter measurement dataset is subjected to text normalization processing to obtain a standardized word sequence; Based on a pre-built health domain feature dictionary, feature matching and statistical counting are performed on the standardized word sequence to obtain a word-level sparse feature vector. The word-level sparse feature vector is then weighted and its dimensions are compressed to obtain the text feature vector.

[0011] In a preferred embodiment, the step of extracting historical time-series data sequences corresponding to the current detection item from historical health records, and inputting the analysis context vector sequence and the historical time-series data sequences together into a pre-trained dedicated AI large model, outputs a set of deep interpretation signals, including: Based on the current detection items, the user's historical health records are normalized to generate a standard historical time-series data sequence that is isomorphic to the current physiological parameter measurement dataset. The analysis context vector sequence, representing the multimodal state at the current moment, is fused with the standard historical time series data sequence based on the spatiotemporal fusion algorithm to construct a spatiotemporally enhanced model input tensor; The spatiotemporal augmentation model input tensor is fed into a pre-trained dedicated AI large model, which performs inference through its internal multi-head attention mechanism and decoding layer to output a deep interpretation signal group containing current physiological state assessment labels, abnormal pattern descriptions, and key potential correlations and weak risk signal clusters.

[0012] In a preferred embodiment, the step of analyzing the user's sub-health conversion risk based on the key potential correlations and the weak risk signal clusters, as well as the changing trends of the corresponding signals in the historical time-series data sequence, to obtain a sub-health risk index and corresponding risk level labels includes: Trend quantification is performed on the key potential associations and each signal element in the weak risk signal cluster to obtain a set of quantitative risk indicators containing signal strength values, associated disease weight values, and time series gradient values. The quantitative risk index values ​​are used to conduct a risk quantification assessment to obtain a sub-health risk index that represents the likelihood of a user transitioning to a specific disease or sub-health state. Based on the numerical range of the sub-health risk index and the compositional properties of the weak risk signal cluster, multi-level threshold judgment and semantic mapping are performed to generate corresponding risk level labels.

[0013] In a preferred embodiment, the trend quantification of the key potential associations and each signal element in the weak risk signal cluster yields a set of quantified risk indicators including signal strength values, associated disease weight values, and time-series gradient values, including: The signal strength value is determined based on the deviation of the physiological parameter measurement value corresponding to each signal in the weak risk signal cluster from the individual baseline. Based on a pre-built medical knowledge base, the associated disease weight values ​​are mapped according to the signal strength value and the type of signal element. Based on the historical time series data sequence, the occurrence frequency, duration and numerical change rate of the abnormal pattern corresponding to each signal in the weak risk signal cluster are analyzed, and the time series gradient value is calculated. By combining the signal strength value, the associated disease weight value, and the time-series gradient value, a set of quantitative risk indicators is obtained.

[0014] In a preferred embodiment, based on the current physiological state assessment labels, abnormal pattern descriptions, sub-health risk index, and risk level labels, a structured, personalized, comprehensive health report is automatically generated according to a preset report template, including: The current physiological state assessment labels, abnormal pattern descriptions, sub-health risk indices, and risk level labels are aligned and structured serialized to generate standardized report input data packets. Based on the preset report template, the standardized report input data package is subjected to template matching, content filling and logical decision-making to generate a set of original report content containing key interpretations, risk assessments and personalized suggestions; The original report content set is subjected to format rendering, visualization element insertion, and metadata appending operations to generate and output the structured personalized comprehensive health report.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention proposes a multimodal context encoding mechanism for the first time, effectively solving the core problem of multi-source data fusion in traditional health monitoring technology. This mechanism unifies heterogeneous feature vectors such as numerical values, images, and text into a common feature dimension space through feature dimension alignment mapping. Combined with an interactive attention fusion algorithm, it dynamically calculates the correlation weights between different modalities and achieves weighted fusion of cross-modal information, forming a comprehensive and deeply correlated analysis context vector sequence. This innovation not only breaks the isolation of single-modal data, but also accurately captures the potential correlations between multi-dimensional data, such as the synergistic indicative significance of blood pressure values ​​with facial flushing images and descriptions of high environmental temperature. It significantly improves the comprehensiveness and accuracy of health status assessment, upgrading the analysis results from isolated data presentation to a three-dimensional user health profile.

[0016] 2. This invention focuses on the dynamic quantitative assessment of the "risk of sub-health transformation," overcoming the limitations of traditional monitoring technologies that can only provide warnings of obvious abnormalities. By extracting key potential correlations and weak risk signal clusters from the output of a large AI model, and combining signal strength, associated disease weights, and time-series gradients for multi-dimensional quantification, and then using a risk quantification assessment algorithm to achieve nonlinear weighted fusion and standardized transformation, a dynamic risk index that accurately reflects the probability of sub-health transformation is obtained. At the same time, based on the risk index range and the properties of the signal clusters, targeted risk level labels are generated. This not only enables the early capture of weak risk signals, but also accurately depicts the risk evolution trend through the combined analysis of historical time-series data and the current state. It provides users with full-chain health management support from risk warning to personalized intervention suggestions, greatly improving the foresight and practicality of family health monitoring. Attached Figure Description

[0017] Figure 1This is a flowchart illustrating a method for assisting in the analysis of measurement data in a health monitoring instrument based on an AI large model, according to an embodiment of the present invention. The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] This application provides a method for measurement data-assisted analysis of health monitoring instruments based on AI large-scale models. The execution entity of this method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the method for measurement data-assisted analysis of health monitoring instruments based on AI large-scale models can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0020] Reference Figure 1 The diagram shown is a flowchart illustrating a measurement data-assisted analysis method for a health monitoring device based on an AI large-scale model, according to an embodiment of the present invention. In this embodiment, the measurement data-assisted analysis method for a health monitoring device based on an AI large-scale model includes: S1, through the non-contact and contact sensors of the health monitor, synchronously acquires the user's multimodal physiological measurement data at a specific time point to construct a physiological parameter measurement dataset for the user's current cycle; In this embodiment of the invention, the simultaneous acquisition of multimodal physiological measurement data of the user at a specific time point through non-contact and contact sensors of the health monitoring device to construct a physiological parameter measurement dataset for the user's current cycle includes: Based on the integrated contact sensor on the health monitor, a set of physiological parameter values ​​of the user during the monitoring period is obtained, wherein the set of physiological parameter values ​​includes heart rate, blood pressure, blood oxygen saturation, and body surface temperature; Based on the non-contact sensor integrated on the health monitor, user vital sign image data and environmental perception data synchronized with the physiological parameter values ​​are acquired during the monitoring period; The physiological parameter values, user vital sign image data, and environmental perception data are integrated with the timestamps to obtain a time series sequence of multimodal measurement data. The aligned multimodal measurement data are format-standardized and structurally reorganized to construct a user's current cycle physiological parameter measurement dataset containing numerical, image, and descriptive fields.

[0021] It should be noted that the process of acquiring physiological parameter data involves using sensors that directly contact the user's skin, such as blood pressure cuffs, electrocardiogram electrodes, and blood oxygen saturation probes, to collect directly quantifiable indicators like heart rate, blood pressure, blood oxygen saturation, and body surface temperature. This data provides real-time quantitative feedback on the user's core physiological functions.

[0022] It should be noted that the set of physiological parameter values ​​refers to a set of direct measurement results of user physiological indicators recorded in digital form. It represents the short-term state of a user's specific physiological system at the time of measurement and serves as the basic quantitative basis for subsequent health status assessment.

[0023] It should be noted that the operation of acquiring vital sign image data and environmental perception data is carried out through sensors such as cameras, infrared thermal imagers, and microphones that do not directly contact the user's body. These sensors simultaneously collect facial images of the user for analyzing skin color and mental state, videos of the user's posture for analyzing posture and micro-movements, infrared thermal images for analyzing body surface temperature distribution, environmental audio for analyzing cough and breathing sounds, as well as environmental light, temperature, and humidity. This data provides supplementary information about the user's appearance, behavior, and environment.

[0024] It should be noted that the vital signs image data refers to image and video data containing the user's visual characteristics, reflecting the user's external appearance, mental state, and possible non-contact physiological indicators, such as estimating heart rate through facial video. The environmental perception data refers to data describing the surrounding environment at the time of measurement. It is used to provide context for physiological measurement data and help determine whether data fluctuations are affected by the environment, such as the effect of ambient temperature on body surface measurements.

[0025] It should be noted that timestamp alignment refers to the process of uniformly calibrating all raw data streams collected by contact and non-contact sensors within the monitoring period to a common and accurate time reference point based on their respective timestamp information and using clock synchronization technology. This operation ensures that data from different sensors strictly correspond to the same measurement time in the time dimension.

[0026] It should be noted that the format standardization and structure reorganization operations include: unifying numerical data to preset units of measurement and data precision, compressing or converting image and video data into standard encoding formats, such as JPEG and H.264, and attaching metadata, and converting audio data into standard waveform files; Finally, according to the preset data model, the processed data are classified into numerical fields, image fields, and descriptive fields, thereby generating a structured and queryable data entity.

[0027] Furthermore, the numerical fields include the collected heart rate, blood pressure, blood oxygen saturation, and body surface temperature; Image-based fields include user facial images, user body posture videos, and infrared thermal images; Descriptive fields contain environment-aware data.

[0028] S2, interactive fusion is performed to parse numerical feature vectors, image feature vectors and text feature vectors from the physiological parameter measurement dataset to obtain an analysis context vector sequence; In this embodiment of the invention, the interactive fusion parses numerical feature vectors, image feature vectors, and text feature vectors from the physiological parameter measurement dataset to obtain an analysis context vector sequence, including: Numerical features of numerical fields in the physiological parameter measurement dataset are extracted to obtain a numerical feature vector; Hierarchical features of image fields in the physiological parameter measurement dataset are extracted to obtain image feature vectors; Natural language understanding and encoding are performed on the descriptive fields in the physiological parameter measurement dataset to obtain text feature vectors; The numerical feature vector, the image feature vector, and the text feature vector are subjected to feature dimension alignment mapping processing. Based on the interactive attention fusion algorithm, the dimension-aligned numerical feature vector, the image feature vector, and the text feature vector are subjected to cross-modal information weighted fusion to obtain the analysis context vector sequence.

[0029] It should be noted that feature dimension alignment mapping refers to using a linear projection matrix to map numerical feature vectors, image feature vectors, and text feature vectors from different feature extraction paths, which have different original dimensions, to a unified, pre-defined common feature dimension space. This operation ensures that features from different modalities have the same vector length before subsequent fusion, allowing them to interact and be compared in the same mathematical space.

[0030] Furthermore, the specific method for mapping using a linear projection matrix is ​​as follows: When the original dimensions of the three feature vectors (numerical, image, and text) are inconsistent, the smallest dimension among the three is first uniformly selected as the common target dimension. For feature vectors whose original dimensions are greater than the common target dimension, a truncation method is used, directly taking the first few elements of the common target dimension. For feature vectors whose original dimensions are less than the common target dimension, a mean-filling method is used, calculating the average value of all elements of the vector, and using this average value to fill the common target dimension.

[0031] Furthermore, a specific implementation process of the interactive attention fusion algorithm is described as follows: First, the query vector, key vector, and value vector are calculated; then, the attention weights are calculated; and finally, the weighted sum is calculated to obtain the fusion vector.

[0032] Furthermore, the mathematical expression for the interactive attention fusion algorithm is as follows:

[0033] In the formula, Q m For the query vector, K n V is the key vector. n F is a value vector. m and F n These represent dimension-aligned feature vectors of any two different modalities. , and These are learnable linear transformation matrices corresponding to the query, key, and value, respectively.

[0034] Through formula Calculate attention weight A m,n In the formula, Q m For the query vector, K n Let d be the key vector. k It is the dimension of the key vector.

[0035] Finally, the formula is used. Multiply the attention weights by the value vector to calculate the fusion vector C of mode m enhanced by the information of mode n. m This calculation is repeated for all modal pairs, and each modality is concatenated or fused again with the vectors enhanced by other modalities to form the final analysis context vector sequence.

[0036] In this embodiment of the invention, the step of extracting numerical features from the numerical fields of the physiological parameter measurement dataset to obtain a numerical feature vector includes: Outlier cleaning and dimensional standardization are performed on the numerical fields in the physiological parameter measurement dataset, which includes heart rate, blood pressure, blood oxygen saturation, and body surface temperature, to obtain a standardized numerical sequence. The standardized numerical sequence is subjected to time-domain statistical features and morphological features to obtain a numerical feature vector.

[0037] It should be noted that performing outlier cleaning and dimensional standardization on the original numerical sequence is a preprocessing operation to ensure data quality and comparability. Outlier cleaning refers to identifying and removing or correcting obvious erroneous data points caused by transient interference in measurement, within the physiologically feasible range. Dimensional standardization refers to converting the values ​​of different physiological parameters, such as blood pressure (in mmHg) and heart rate (in bpm), to a unified, unitless numerical scale by subtracting the mean and dividing by the standard deviation. This eliminates differences in dimensions and orders of magnitude between different parameters and improves numerical stability.

[0038] It should be noted that calculating the time-domain statistical and morphological features of a standardized numerical sequence is a process of extracting quantitative descriptive indicators from a one-dimensional sequence. The calculation of time-domain statistical features includes: calculating the mean, variance, maximum and minimum values ​​of the sequence to reflect the average level, volatility and extreme values; the extraction of morphological features involves more refined waveform analysis by identifying peaks and troughs in the sequence.

[0039] Furthermore, the peak-finding algorithm involved in morphological feature extraction is essentially about comparing the numerical relationship between each data point and its neighboring points; a simplified criterion can be expressed as: for a data point x, if x satisfies i >x i-1 And x i >x i+1 If it meets the condition, it can be identified as a candidate peak; otherwise, if x is satisfied... <x i-1 And x i <x i+1 If so, it can be identified as a candidate trough.

[0040] In this embodiment of the invention, the step of extracting hierarchical features of image-type fields from the physiological parameter measurement dataset to obtain image feature vectors includes: The image and video data contained in the image field are preprocessed to obtain standardized image data; Hierarchical visual feature extraction is performed on the standardized image data to obtain image feature maps that represent the multi-scale semantic information of the images; The image feature map is subjected to spatial dimension compression and vectorization transformation to obtain the image feature vector.

[0041] It should be noted that the preprocessing process for the image and video data contained in the image field is as follows: First, the pixel resolution of the input image and video frame is uniformly adjusted to a preset fixed size. Second, the color values ​​of all pixels are converted from the original integer representation to floating-point numbers and normalized to the [0,1] range by linear scaling. At the same time, it is ensured that all data are in RGB three-channel format. For video data, this processing is performed frame by frame. The physical essence of this operation is to eliminate inconsistencies in size, numerical range, and format of the original visual data, generating a regular and uniform digital matrix to provide standard input for subsequent feature extraction algorithms.

[0042] It should be noted that the hierarchical visual feature extraction process is as follows: Standardized image data is input into a deep network consisting of multiple convolutional and pooling layers stacked alternately. Each convolutional layer performs convolution operations on the input through a set of learnable filters to extract local features, and then introduces nonlinearity through a nonlinear activation function. The pooling layer downsamples the feature map, retaining the main features while reducing dimensionality and expanding the receptive field. As the network depth increases, the extracted features are gradually abstracted from the geometric details at the bottom layer to the semantic concepts at the higher layer. The output of this process is a three-dimensional image feature map, whose width and height dimensions correspond to spatial location, while the depth dimension encodes the feature responses at different levels of abstraction.

[0043] It should be noted that the specific operation process of spatial dimension compression and vectorization transformation is as follows: Global average pooling is performed on the above three-dimensional image feature map; specifically, along the spatial dimension of the feature map, the arithmetic mean of all activation values ​​in each feature channel is calculated, which is mathematically expressed as: for a feature map with C channels, the pooling value of the c-th channel... Where W and H are the width and height of the feature map, respectively, and x c (i,j) is the activation value at position (i,j), where i and j are the indices of width and height, respectively. Then, the average values ​​of these C channels are concatenated in order to form a one-dimensional feature vector of length C, which is the final image feature vector.

[0044] In this embodiment of the invention, the step of performing natural language understanding and encoding on the descriptive fields in the physiological parameter measurement dataset to obtain text feature vectors includes: The original text data contained in the descriptive fields of the physiological parameter measurement dataset is subjected to text normalization processing to obtain a standardized word sequence; Based on a pre-built health domain feature dictionary, feature matching and statistical counting are performed on the standardized word sequence to obtain a word-level sparse feature vector. The word-level sparse feature vector is then weighted and its dimensions are compressed to obtain the text feature vector.

[0045] It should be noted that text normalization of raw text data is a fundamental operation that transforms unstructured natural language descriptions into a standardized form that can be used for feature extraction. This process includes, but is not limited to: segmenting continuous text into independent lexical units, removing commonly used function words that do not contribute to actual semantics, restoring words to their basic word forms, and mapping synonyms or near-synonyms to preset standard terms. The purpose of this operation is to eliminate noise, ambiguity, and expression differences in the text, forming a clear lexical sequence composed of standardized terms.

[0046] It should be noted that feature matching and statistical counting based on a pre-built health domain feature dictionary is a rule-based and knowledge-based feature extraction method. The health domain feature dictionary is a predefined structured list that contains key entity categories related to health monitoring, such as symptoms, signs, environmental factors, behaviors, and body parts. This operation traverses the standardized word sequence, performing exact or fuzzy matching between each word and an entry in the feature dictionary. For successfully matched words, the count is accumulated along the corresponding feature category dimension to generate a high-dimensional sparse vector, where each dimension corresponds to a category in the feature dictionary, and its value represents the frequency or presence of words in that category in the text.

[0047] Furthermore, the mathematical representation of feature matching and statistical counting can be described as follows: Let the feature dictionary have a total of The standardized vocabulary sequence includes [number] categories. The word, for the first Each feature category has its corresponding vector element. The calculation is as follows: In the formula, It is an indicator function, when the word term m Category k The value is 1 if it is true, and 0 otherwise.

[0048] It should be noted that the lexical-level sparse feature vector is a high-dimensional numerical vector with most elements being 0. It records the distribution of text content across different health-related semantic dimensions in a structured and quantifiable manner.

[0049] It should be noted that the weight adjustment and dimensionality compression are intended to solve the problem that the original sparse vector has too high a dimension and that the importance of different features is not distinguished. First, different static weights are assigned to the feature according to its category attribute. Then, dimensionality reduction is performed to compress the high-dimensional sparse vector into a low-dimensional, dense real number vector while retaining the most important feature information.

[0050] Furthermore, weight adjustment is achieved by pre-setting an importance weight for each category in the feature dictionary. Then the weighted eigenvectors The weight of each category is a set of configurable parameters that are comprehensively set based on the clinical prior importance of the medical knowledge base and the statistical correlation strength of historical labeled data. For example, the "chest pain" category, which is strongly associated with cardiovascular emergencies, has high prior importance according to clinical guidelines and strong statistical correlation with adverse events in historical data, so it may be assigned a high weight, such as 1.8; while the "fatigue" category, which is a common but non-specific symptom, may have a weight set to the baseline value of 1.0; and the "noise interference" category, which is a general environmental factor, has a low direct contribution to health assessment, so its weight may be set to 0.3.

[0051] Furthermore, dimensionality compression stacks the weighted sparse feature vectors of multiple samples into a matrix, and uses mathematical methods to find two low-dimensional non-negative matrices such that their product can approximate the original matrix. Thus, the high-dimensional features of each sample are represented as a linear combination of these two low-dimensional matrices, resulting in a dense vector after dimensionality reduction.

[0052] S3, extract the historical time-series data sequence corresponding to the current test item from the historical health record, input the analysis context vector sequence and the historical time-series data sequence into a pre-trained dedicated AI large model, and output a set of deep interpretation signals, wherein the deep interpretation signals include: current physiological state assessment label, abnormal pattern description, and key potential correlation and weak risk signal clusters identified from the data; In this embodiment of the invention, the step of extracting the historical time-series data sequence corresponding to the current detection item from historical health records, and inputting the analysis context vector sequence and the historical time-series data sequence together into a pre-trained dedicated AI large model, outputs a set of deep interpretation signals, including: Based on the current detection items, the user's historical health records are normalized to generate a standard historical time-series data sequence that is isomorphic to the current physiological parameter measurement dataset. The analysis context vector sequence, representing the multimodal state at the current moment, is fused with the standard historical time series data sequence based on the spatiotemporal fusion algorithm to construct a spatiotemporally enhanced model input tensor; The spatiotemporal augmentation model input tensor is fed into a pre-trained dedicated AI large model, which performs inference through its internal multi-head attention mechanism and decoding layer to output a deep interpretation signal group containing current physiological state assessment labels, abnormal pattern descriptions, and key potential correlations and weak risk signal clusters.

[0053] It should be noted that performing time-series retrieval and normalization on a user's historical health records is a comprehensive data processing operation. This operation first finds all past measurement records of the same type in the database based on the type of the current test item; then it sorts them by timestamp and performs data cleaning to remove obviously physiologically unreasonable outliers; finally, it uses a time-series interpolation algorithm to fill in the missing time points and resamples all the data to a fixed time interval, thereby generating a uniform, continuous and complete sequence of time points.

[0054] It should be noted that the standard historical time series data sequence is an array that is uniformly sampled in the time dimension and quality-controlled in the numerical dimension. It quantitatively depicts the continuous change trajectory and trend of a user's specific physiological indicators over a period of time, providing a crucial longitudinal comparison baseline for assessing the relative position of the current measurement value in the historical context.

[0055] It should be noted that the spatiotemporal fusion algorithm is the key to realizing the "combined analysis of vertical trend and horizontal snapshot" of this invention. This algorithm aligns and concatenates the time series representing the historical vertical evolution with the context vector representing the current horizontal multimodal snapshot in the feature space and time dimension. The specific implementation is as follows: First, the historical time series is mapped to the same feature dimension as the current context vector through a linear projection layer; then, in the time dimension, the context vector representing the current moment is used as a new time step and concatenated at the end of the historical sequence, thereby forming a two-dimensional tensor that integrates the time context and the current fine features.

[0056] It should be noted that the spatiotemporal augmentation model input tensor is a structured data object. It is no longer isolated single measurement data, but rather models the current measurement within the context of the user's personal historical health trajectory. This enables the AI ​​model to identify "weak risk signals" that appear normal in a single measurement but show signs of deviation in the long-term trend.

[0057] It should be noted that the pre-trained dedicated AI model is a Transformer-type model pre-trained on a large-scale medical knowledge graph and clinical time-series data. Its internal multi-head attention mechanism can adaptively focus on the complex interaction relationships between different time steps and different features in the input tensor. The decoding layer maps this deep spatiotemporal correlation pattern learned by the model into a structured output with clear health semantics.

[0058] The specific process of reasoning through the multi-head attention mechanism and decoding layer inside the pre-trained dedicated AI large model is as follows: First, the pre-trained dedicated AI large model uses the multi-head attention mechanism to enable it to examine the spatiotemporal tensor of the input from multiple different "views" in parallel, dynamically calculate and integrate the complex dependencies between any element in the sequence, such as outliers at a certain time in history and the current multimodal features, thereby constructing a deep, context-dependent internal representation. Subsequently, based on this context-rich internal representation, the decoding layer performs targeted signal synthesis and mapping, gradually generating and outputting a set of deeply interpreted signals, such as status labels, descriptive text, and risk signal clusters. This process from "deep correlation analysis" to "precise signal generation" is the core mechanism by which this dedicated AI model realizes actionable insights from complex data.

[0059] It should be noted that the deep interpretation signal group is a multi-level cognitive output that goes beyond simple binary judgment. The current physiological state assessment label is an instant classification judgment based on comprehensive information. The abnormal pattern description is a qualitative explanation of the abnormality in natural language. The key potential correlation and weak risk signal cluster are a set of early risk indicators that are weak in intensity but have time persistence, which are hidden statistical correlation patterns between multi-dimensional data mined by the model.

[0060] S4. Based on the key potential correlations and the weak risk signal clusters, as well as the changing trends of the corresponding signals in the historical time series data sequence, analyze the user's sub-health conversion risk to obtain the sub-health risk index and the corresponding risk level label. In this embodiment of the invention, the step of analyzing the user's sub-health conversion risk based on the key potential correlations and the weak risk signal clusters, as well as the changing trends of the corresponding signals in the historical time-series data sequence, to obtain a sub-health risk index and corresponding risk level labels, includes: Trend quantification is performed on the key potential associations and each signal element in the weak risk signal cluster to obtain a set of quantitative risk indicators containing signal strength values, associated disease weight values, and time series gradient values. The quantitative risk index values ​​are used to conduct a risk quantification assessment to obtain a sub-health risk index that represents the likelihood of a user transitioning to a specific disease or sub-health state. Based on the numerical range of the sub-health risk index and the compositional properties of the weak risk signal cluster, multi-level threshold judgment and semantic mapping are performed to generate corresponding risk level labels.

[0061] It should be noted that the mathematical expression of the risk quantification assessment algorithm is as follows:

[0062] In the formula, DRI is the sub-health risk index, σ is the Sismgid activation function, N is the total number of identified risk signal elements, i is the index of the risk signal element, and S... i Let W be the intensity value of the i-th signal. i G represents the weight value of the disease associated with the i-th signal. i Let be the temporal gradient value of the i-th signal, tanh be the hyperbolic tangent function, α, β, γ be the harmonic coefficients of intensity, weight, and gradient, respectively, and w be the time gradient value. i Let be the comprehensive weight of the i-th signal, where is the Sismgid activation function used to map the weighted sum to the (0,1) interval, achieving standardization and probabilistic interpretation of the results. The comprehensive weight of the signal can be dynamically adjusted according to the signal's novelty, specificity, or matching degree with the user's personal health record, satisfying the following conditions: .

[0063] It should be noted that performing multi-level threshold judgment and semantic mapping means that according to the preset risk index segmentation interval, that is, [0,0.3) is low risk, [0.3,0.7) is the attention period, and [0.7,1.0] is the early stage of high risk, combined with the specific properties of weak risk signal clusters, the numerical DRI is mapped into text labels with clear health guidance significance.

[0064] Furthermore, semantic mapping not only relies on a fixed threshold, but also makes fine adjustments based on the properties of the signal cluster. For example, even if the DRI is 0.65, if the signal cluster is mainly composed of blood glucose-related indicators, the label may be "prediabetes risk - attention period"; if it is mainly composed of blood pressure and vascular elasticity indicators, the label may be "prehypertension risk - attention period".

[0065] It should be noted that the risk level label is a qualitative and intuitively understandable classification description. It is a semantic encapsulation and clinical interpretation of the sub-health risk index, such as "low risk", "period of attention for abnormal glucose metabolism", and "pre-high cardiovascular risk", providing clear action guidance for users and health managers.

[0066] In this embodiment of the invention, the step of trend quantification of the key potential associations and each signal element in the weak risk signal cluster to obtain a set of quantitative risk indicators including signal strength values, associated disease weight values, and time-series gradient values ​​includes: The signal strength value is determined based on the deviation of the physiological parameter measurement value corresponding to each signal in the weak risk signal cluster from the individual baseline. Based on a pre-built medical knowledge base, the associated disease weight values ​​are mapped according to the signal strength value and the type of signal element. Based on the historical time series data sequence, the occurrence frequency, duration and numerical change rate of the abnormal pattern corresponding to each signal in the weak risk signal cluster are analyzed, and the time series gradient value is calculated. By combining the signal strength value, the associated disease weight value, and the time-series gradient value, a set of quantitative risk indicators is obtained.

[0067] It should be noted that the core calculation logic for determining the signal strength value is to calculate and standardize the relative deviation between the current physiological parameter measurement and the selected baseline value. Specifically, when the baseline is the individual's baseline value and the statistical mean of the current physiological parameter's historical 30-day data, the standardization is achieved using the formula... Where S is the standardized score, and V is the standardized score. c The signal strength value is the absolute value of the standardized score, where μ is the statistical mean of the historical data of the current physiological parameter, σ is the standard deviation of the historical data of the current physiological parameter, and the signal strength value is ultimately the absolute value of the standardized score.

[0068] It should be noted that the mapping of the associated disease weight value based on the signal strength value and the type of signal element is implemented by querying a pre-built mapping table. This mapping table is part of the medical knowledge base, and its structure defines the mapping relationship from signal type and signal strength range to weight coefficient. First, the row of the mapping table is determined according to the type of signal element. Then, the column of the mapping table is determined according to the preset strength range into which the calculated signal strength value falls. Finally, the preset weight value stored at the intersection of the row and column is read.

[0069] It should be noted that the core calculation logic for obtaining the time series gradient value is: to perform a weighted linear combination of multiple time series feature indicators. First, three basic indicators are derived from the historical time series data sequence of the past 30 days: the frequency index F: is the statistical value of the number of times the event occurs per unit time. Duration index D: The average duration of historical events; Rate of change index R: the slope of the physiological parameter value over time obtained through linear regression analysis; Subsequently, these basic indicators are standardized by subtracting the mean from the value, and then dividing this difference by the variance to obtain... Through formula The temporal gradient value G is calculated, where ω1, ω2, and ω3 are pre-set weight coefficients, each taking a value of one-third.

[0070] S5. Based on the current physiological state assessment labels, abnormal pattern descriptions, sub-health risk index, and risk level labels, a structured, personalized, comprehensive health report is automatically generated according to a preset report template.

[0071] In this embodiment of the invention, based on the current physiological state assessment label, abnormal pattern description, sub-health risk index, and risk level label, a structured, personalized, comprehensive health report is automatically generated according to a preset report template, including: The current physiological state assessment labels, abnormal pattern descriptions, sub-health risk indices, and risk level labels are aligned and structured serialized to generate standardized report input data packets. Based on the preset report template, the standardized report input data package is subjected to template matching, content filling and logical decision-making to generate a set of original report content containing key interpretations, risk assessments and personalized suggestions; The original report content set is subjected to format rendering, visualization element insertion, and metadata appending operations to generate and output the structured personalized comprehensive health report.

[0072] It should be noted that performing data alignment and structured serialization operations refers to integrating and formatting heterogeneous input data from different analysis stages, such as classification labels, natural language descriptions, continuous numerical indices, and rank labels. This operation first establishes a unified timestamp and user identifier index for all input data; then, it segments the text descriptions and extracts key information, normalizes or segments the numerical indices, and organizes all data items according to preset key-value pairs or JSON-LD and other structured data patterns to form a data package with consistent internal logic that is easy to process programmatically.

[0073] It should be noted that the preset report template is a framework file that defines the report chapter structure, content slots, logical judgment rules, and writing style. The report generation workflow is as follows: First, based on the main health concern areas mentioned in the risk level label and abnormal pattern description, the most matching basic template is dynamically selected or adjusted from the template library; then, each content slot in the template is traversed, and the corresponding data in the standardized report input data package is filled in; finally, the logical decision rules embedded in the template are executed. For example, when the sub-health risk index exceeds a certain threshold and the signal cluster contains cardiovascular indicators, a set of targeted "diet control and aerobic exercise suggestions" is selected and combined from the suggestion library and filled into the "personalized health guidance" chapter.

[0074] Furthermore, a key logical decision-making process in the report generation algorithm can be described as: personalized recommendation generation rules based on risk level and signal cluster properties. These rules can be expressed as a set of production rules, such as: IF the risk level label is “attention period” AND the abnormal pattern description includes “non-dipping nocturnal blood pressure” THEN the report strengthens the recommendations of “recommending 24-hour ambulatory blood pressure monitoring” and “paying attention to sodium intake”.

[0075] It should be noted that performing format rendering, visualization element insertion, and metadata attachment operations is the final step in transforming machine-readable raw content into a user-friendly document. Format rendering refers to applying the font, layout, and color scheme defined by the template to generate a PDF, HTML, or document in a specific format. Visual element insertion refers to automatically generating line charts based on numerical data and color-coding risk matrices according to risk levels. Metadata attachment refers to embedding information such as the unique ID, timestamp, data source version, and model version of this report into the report to ensure the report's traceability.

[0076] It should be noted that the structured, personalized, comprehensive health report is a final electronic document delivered to the user. It is the final integration and intuitive presentation of all the aforementioned complex data collection, multimodal fusion, AI-powered in-depth interpretation, and risk quantification assessment.

[0077] In the several embodiments provided by this invention, it should be understood that the disclosed method can be implemented in other ways.

[0078] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0079] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, and technology that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0080] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for analyzing measurement data-assisted data in a health monitoring instrument based on an AI large-scale model, characterized in that: The method includes: S1, through the non-contact and contact sensors of the health monitor, synchronously acquires the user's multimodal physiological measurement data at a specific time point to construct a physiological parameter measurement dataset for the user's current cycle; S2, interactive fusion is performed to parse numerical feature vectors, image feature vectors and text feature vectors from the physiological parameter measurement dataset to obtain an analysis context vector sequence; S3, extract the historical time-series data sequence corresponding to the current test item from the historical health record, input the analysis context vector sequence and the historical time-series data sequence into a pre-trained dedicated AI large model, and output a set of deep interpretation signals, wherein the deep interpretation signals include: current physiological state assessment label, abnormal pattern description, and key potential correlation and weak risk signal clusters identified from the data; S4. Based on the key potential correlations and the weak risk signal clusters, as well as the changing trends of the corresponding signals in the historical time series data sequence, analyze the user's sub-health conversion risk to obtain the sub-health risk index and the corresponding risk level label. S5. Based on the current physiological state assessment labels, abnormal pattern descriptions, sub-health risk index, and risk level labels, a structured, personalized, comprehensive health report is automatically generated according to a preset report template.

2. The measurement data-assisted analysis method for a health monitoring instrument based on an AI large model as described in claim 1, characterized in that, The health monitor uses non-contact and contact sensors to simultaneously acquire multimodal physiological measurement data of the user at specific time points to construct a physiological parameter measurement dataset for the user's current cycle, including: Based on the integrated contact sensor on the health monitor, a set of physiological parameter values ​​of the user during the monitoring period is obtained, wherein the set of physiological parameter values ​​includes heart rate, blood pressure, blood oxygen saturation, and body surface temperature; Based on the non-contact sensor integrated on the health monitor, user vital sign image data and environmental perception data synchronized with the physiological parameter values ​​are acquired during the monitoring period; The physiological parameter values, user vital sign image data, and environmental perception data are integrated with the timestamps to obtain a time series sequence of multimodal measurement data. The aligned multimodal measurement data are format-standardized and structurally reorganized to construct a user's current cycle physiological parameter measurement dataset containing numerical, image, and descriptive fields.

3. The measurement data-assisted analysis method for a health monitoring instrument based on an AI large model as described in claim 1, characterized in that, The interactive fusion extracts numerical feature vectors, image feature vectors, and text feature vectors from the physiological parameter measurement dataset to obtain an analysis context vector sequence, including: Numerical features of numerical fields in the physiological parameter measurement dataset are extracted to obtain a numerical feature vector; Hierarchical features of image fields in the physiological parameter measurement dataset are extracted to obtain image feature vectors; Natural language understanding and encoding are performed on the descriptive fields in the physiological parameter measurement dataset to obtain text feature vectors; The numerical feature vector, the image feature vector, and the text feature vector are subjected to feature dimension alignment mapping processing. Based on the interactive attention fusion algorithm, the dimension-aligned numerical feature vector, the image feature vector, and the text feature vector are subjected to cross-modal information weighted fusion to obtain the analysis context vector sequence.

4. The measurement data-assisted analysis method for a health monitoring instrument based on an AI large model as described in claim 3, characterized in that, The step of extracting numerical features from the numerical fields in the physiological parameter measurement dataset to obtain a numerical feature vector includes: Outlier cleaning and dimensional standardization are performed on the numerical fields in the physiological parameter measurement dataset, which includes heart rate, blood pressure, blood oxygen saturation, and body surface temperature, to obtain a standardized numerical sequence. The standardized numerical sequence is subjected to time-domain statistical features and morphological features to obtain a numerical feature vector.

5. The measurement data-assisted analysis method for a health monitoring instrument based on an AI large model as described in claim 3, characterized in that, The step of extracting hierarchical features from the image-type fields in the physiological parameter measurement dataset to obtain image feature vectors includes: The image and video data contained in the image field are preprocessed to obtain standardized image data; Hierarchical visual feature extraction is performed on the standardized image data to obtain image feature maps that represent the multi-scale semantic information of the images; The image feature map is subjected to spatial dimension compression and vectorization transformation to obtain the image feature vector.

6. The measurement data-assisted analysis method for a health monitoring instrument based on an AI large model as described in claim 3, characterized in that, The process of performing natural language understanding and encoding on the descriptive fields in the physiological parameter measurement dataset to obtain text feature vectors includes: The original text data contained in the descriptive fields of the physiological parameter measurement dataset is subjected to text normalization processing to obtain a standardized word sequence; Based on a pre-built health domain feature dictionary, feature matching and statistical counting are performed on the standardized word sequence to obtain a word-level sparse feature vector. The word-level sparse feature vector is then weighted and its dimensions are compressed to obtain the text feature vector.

7. The method for measurement data-assisted analysis in a health monitoring instrument based on an AI large model as described in claim 1, characterized in that, The process involves extracting historical time-series data sequences corresponding to the current testing item from historical health records, inputting the analysis context vector sequence and the historical time-series data sequences into a pre-trained dedicated AI model, and outputting a set of deep interpretation signals, including: Based on the current detection items, the user's historical health records are normalized to generate a standard historical time-series data sequence that is isomorphic to the current physiological parameter measurement dataset. The analysis context vector sequence, representing the multimodal state at the current moment, is fused with the standard historical time series data sequence based on the spatiotemporal fusion algorithm to construct a spatiotemporally enhanced model input tensor; The spatiotemporal augmentation model input tensor is fed into a pre-trained dedicated AI large model, which performs inference through its internal multi-head attention mechanism and decoding layer to output a deep interpretation signal group containing current physiological state assessment labels, abnormal pattern descriptions, and key potential correlations and weak risk signal clusters.

8. The method for measurement data-assisted analysis of a health monitoring instrument based on an AI large model as described in claim 7, characterized in that, The method analyzes the user's sub-health conversion risk based on the key potential correlations and weak risk signal clusters, as well as the changing trends of corresponding signals in the historical time-series data sequence, to obtain a sub-health risk index and corresponding risk level labels, including: Trend quantification is performed on the key potential associations and each signal element in the weak risk signal cluster to obtain a set of quantitative risk indicators containing signal strength values, associated disease weight values, and time series gradient values. The quantitative risk index values ​​are used to conduct a risk quantification assessment to obtain a sub-health risk index that represents the likelihood of a user transitioning to a specific disease or sub-health state. Based on the numerical range of the sub-health risk index and the compositional properties of the weak risk signal cluster, multi-level threshold judgment and semantic mapping are performed to generate corresponding risk level labels.

9. The measurement data-assisted analysis method for a health monitoring instrument based on an AI large model as described in claim 8, characterized in that, The process of trend quantification of each signal element in the key potential associations and the weak risk signal cluster yields a set of quantitative risk indicators containing signal strength values, associated disease weight values, and time-series gradient values, including: The signal strength value is determined based on the deviation of the physiological parameter measurement value corresponding to each signal in the weak risk signal cluster from the individual baseline. Based on a pre-built medical knowledge base, the associated disease weight values ​​are mapped according to the signal strength value and the type of signal element. Based on the historical time series data sequence, the occurrence frequency, duration and numerical change rate of the abnormal pattern corresponding to each signal in the weak risk signal cluster are analyzed, and the time series gradient value is calculated. By combining the signal strength value, the associated disease weight value, and the time-series gradient value, a set of quantitative risk indicators is obtained.

10. The method for measurement data-assisted analysis in a health monitoring instrument based on an AI large model as described in claim 1, characterized in that, Based on the current physiological state assessment labels, abnormal pattern descriptions, sub-health risk index, and risk level labels, a structured, personalized, comprehensive health report is automatically generated according to a preset report template, including: The current physiological state assessment labels, abnormal pattern descriptions, sub-health risk indices, and risk level labels are aligned and structured serialized to generate standardized report input data packets. Based on the preset report template, the standardized report input data package is subjected to template matching, content filling and logical decision-making to generate a set of original report content containing key interpretations, risk assessments and personalized suggestions; The original report content set is subjected to format rendering, visualization element insertion, and metadata appending operations to generate and output the structured personalized comprehensive health report.

Images