CN-121999986-A - Post-inspection health promotion management method and system based on multi-mode data fusion

CN121999986ACN 121999986 ACN121999986 ACN 121999986ACN-121999986-A

Abstract

The invention discloses a post-inspection health promotion management method and system based on multi-mode data fusion, and relates to the technical field of medical health information. The health task is interesting, scenic and fragmented, and the abstract health index is converted into an apparent daily task, so that health promotion becomes interesting and easy to persist by combining real-time feedback and forward excitation. The transition from primary physical examination to long-term health management is achieved, forming a continuously optimized closed loop of assessment-planning-execution-monitoring-feedback-adjustment. By the automatic evaluation and recommendation system, scientificity and efficiency of health scheme making are greatly improved, and accuracy and individuation degree of scheme recommendation are ensured. By learning the optimal intervention strategy from a large amount of user data, the recommendation algorithm is continuously improved, so that the health promotion scheme is more and more accurate.

Inventors

JIN TAO
LIAO YOULIANG
YUAN JING
CHEN JIA

Assignees

四川程康科技有限公司
中国人民解放军东部战区总医院

Dates

Publication Date: 20260508
Application Date: 20260306

Claims (10)

1. The method for promoting and managing the health after inspection based on multi-mode data fusion is characterized by comprising the following steps of: s1, acquiring patient information, wherein the patient information comprises physical examination data and medical history acquired by a physical examination system of a butt joint hospital, and acquired movement, diet, sleep and psychological multi-mode data; The multi-mode data comprises acquisition of activity, exercise type, exercise intensity, joint activity and heart rate through a wearable device or an exercise sensor, identification of food images, types and components input through user terminal images or manually, and daily nutrient intake obtained by analysis in combination with a nutrition database; s2, processing and comprehensively evaluating the acquired multi-mode data to generate health risk scores and user health portraits; the processing and comprehensive evaluation comprises the following steps: The method comprises the steps of obtaining physical examination data, adopting Z-score standardization based on age and gender to extract time domain, frequency domain and nonlinear characteristics from behavior time sequence data, constructing a self-adaptive weighted risk index model based on the standardized and characteristic extracted data, and calculating risk indexes of all health dimensions; The method comprises the steps of performing feature selection and labeling on multi-mode data subjected to standardization and feature extraction by combining health risk scores according to preset health dimension classification and risk threshold values, and finally forming a structured user health portrait representing the current comprehensive health state of a user; S3, based on the user health portrait and health risk score, combining an expert knowledge base constructed by a medical ontology, and generating a recommended health intervention scheme through content filtering, collaborative filtering and Bayesian network reasoning; S4, a doctor receives the recommended health intervention scheme, inputs adjustment information according to a clinical evaluation result to optimize the recommended health intervention scheme, generates a target health scheme and sends the target health scheme to a patient; S5, monitoring the execution data of the patient on the target health scheme, and collecting the execution condition in real time through a sensor and a user log; And S6, judging whether to trigger an early warning rule or deviate from a scheme target according to the execution condition, if so, adjusting the scheme, and repeating the steps S5-S6 until the health condition of the patient is improved or reaches the health condition.
2. The post-inspection health promotion management method based on multi-modal data fusion according to claim 1, wherein the physical examination data is normalized by adopting a Z-score normalization formula, time domain, frequency domain and nonlinear characteristics are extracted from behavior time sequence data, an adaptive weighted risk index model is constructed, and a final health risk score is obtained through defuzzification calculation.
3. The method for post-inspection health promotion management based on multi-modal data fusion according to claim 2, wherein the constructing the adaptive weighted risk index model comprises: Risk index for each dimension The calculation is as follows: ; Wherein, the Represent the first Total risk index, weight of individual dimensions Representation of the first by dynamic calculation of entropy weight In the third dimension The importance of the individual indicators to the risk, Is an activation function, and the index value after Z-score normalization Smoothly map to the (0, 1) interval when At negative values, the output is near 0, indicating low risk, when When the output is positive, the output is close to 1, the risk is high, the output value is in the (0, 1) interval, and the output value can be used for explaining the relative risk probability of the index prompt, and the specific process comprises the following steps: will be the original index value Mapping between (0, 1), representing a risk probability or a normalized risk value, Is the first In the third dimension The original or intermediate calculated value of the individual indicators, Is the first Information entropy calculated by the item index based on historical user population data distribution and used for representing variation degree of index data, the first The weights of the item indexes are as follows: ; Wherein, the Is the first The entropy value or error value of the individual indicators, Representing the slave to a variable =1 Add to Sum of m, the first The first dimension is one of four health dimensions of exercise, diet, sleep and psychology The individual indexes are specific measured values under corresponding dimensions; is the numerical value of the index after standardized treatment; the function is used for mapping the normalized index value into a single risk probability between 0 and 1, and the closer the value is to 1, the higher the health risk of the index prompt is; The weight obtained by dynamic calculation according to the data variation degree of the index in the user group through the entropy weight method is used for representing the contribution degree of the index to the comprehensive health risk of the current user.
4. The method for post-inspection health promotion management based on multimodal data fusion of claim 1, wherein the generating the health risk score and the user health representation comprises: Vector the risk index of each dimension Inputting into a Takagi-Sugeno type fuzzy inference system, obtaining final output risk scores through defuzzification calculation, wherein the elements in the risk index vectors of each dimension are according to the following steps Calculating to obtain; The fuzzy inference system comprises M fuzzy rules, each rule comprising an activation strength And a local output value The final output risk score is obtained by defuzzification calculations: ; Wherein, the For the health risk score of the final output, Is the first The activation strength of the bar fuzzy rule based on the input vector Is used for the calculation of the membership function of the (c), Is the first The local output value of the fuzzy rule is determined by a linear function preset in a fuzzy rule base; the user health representation is expressed as: ; Wherein, the Risk indexes respectively representing four dimensions of exercise, diet, sleep and psychology are expressed by the formula Calculating to obtain; is a label vector determined based on a risk threshold.
5. The post-inspection health promotion management method based on multi-modal data fusion according to claim 1, wherein the content filtering comprises the steps of calculating matching degree of user health labels and scheme descriptions through a cosine similarity formula, judging similar user groups through an Euclidean distance formula through collaborative filtering, calculating scheme suitability through a posterior probability formula through a Bayesian network, and finally screening an optimal health scheme through a utility function.
6. The method for post-inspection health promotion management based on multimodal data fusion according to claim 5, wherein the final screening of the optimal health plan by the utility function comprises: (1) After the user health portrait P is received, content-based filtering is carried out, TF-IDF is utilized to vectorize user health labels and scheme description, cosine similarity is calculated to carry out preliminary screening, the content matching degree between user health portrait vectors and scheme description vectors is represented, and the higher the score is, the more relevant the scheme is to the user health labels: ; Wherein, the Portrait vector for user health And the first Personal health promotion plan description vector The similarity score between the two values is [ -1,1], the score is between [0,1], Representing the dot product of the two vectors; (2) Based on the user health portrait by adopting collaborative filtering idea Searching a historical user group with similar portraits, judging the similarity through Euclidean distance, and representing the similarity distance between the health portraits of two users, wherein the smaller the distance is, the more similar the health condition and risk characteristics of the two users are, and the formula is as follows: ; Wherein, the Health portrait vectors respectively representing two users, wherein each vector consists of risk indexes and labels of a plurality of health dimensions; 、 respectively represent vectors And The value of the i-th feature in (a); The Euclidean distance between the health portraits of the two users is represented, and the smaller the distance is, the more similar the health conditions and risk characteristics of the two users are represented; (3) User health based portrayal The extracted health evidence E processes uncertainty reasoning through a Bayesian network, the posterior probability of adopting a scheme module M under the given user evidence E is calculated, the adaptive probability of adopting the scheme module M under the condition that the current health evidence E of a user is known is represented, and the uncertainty matching degree between a scheme and the health state of the user is quantified by the following formula: ; (4) Final integrated content similarity Group feedback obtained by collaborative filtering and Bayesian posterior probability Generating a sports prescription result with the highest utility score for the target T: Set the candidate health intervention scheme set as Wherein each scheme pi is associated with a scheme description vector Historical user population feedback scoring Scheme module Posterior probability of (2) ; Computing a composite utility value for each scheme pi The calculation formula is as follows: ; Wherein, the Portrait vector for user health Scheme description vector The degree of cosine similarity between the two, Performing an effect average score on the history of the scheme for similar user groups obtained through collaborative filtering; to be in the health evidence of a given user Lower scheme module Bayesian posterior probability of (2); , , Is a preset weight coefficient, meets the following conditions + + =1; By traversing all candidate schemes Comparing the utility values, and selecting the scheme with the maximum utility value as the optimal recommended scheme: ; obtaining a health plan set In the current user health state And health goals Under the scheme with highest utility value , wherein, A specific prescription for an exercise is shown, Representing a set of all possible sports prescriptions, The object is represented by a set of objects, Is shown in given And Under the condition of (a) prescription Is used to determine the utility score of (c) for a given application, Representation is such that The one that takes the maximum value 。
7. The method for post-test health promotion management based on multimodal data fusion of claim 1, wherein monitoring the patient's performance data of the target health regimen comprises: monitoring exercise posture, diet data, sleep and psychological state data in real time; And after data acquisition is executed, the scheme is dynamically adjusted through a closed loop feedback mechanism, and a model is generated through a multi-task deep learning loss function and cGAN by a self-adaptive AI optimization engine so as to continuously optimize the scheme recommendation strategy through a Q-learning algorithm of the DQN.
8. The post-inspection health promotion management method based on multi-modal data fusion according to claim 7, wherein the Q-learning algorithm continuous optimization scheme recommendation strategy for generating a model and DQN by the adaptive AI optimization engine through a multi-task deep learning loss function, cGAN comprises: Adopting a multi-task deep learning model, sharing features at the bottom layer, arranging special risk prediction branches at the upper layer, and carrying out accurate parallel prediction of multiple health risks through joint optimization weighting loss functions; Introducing a condition generation countermeasure network cGAN, and synthesizing a personalized scheme by a generator under the condition of the health state of a user; and pre-evaluating and screening the candidate schemes by adopting a depth Q network DQN, learning an optimal recommendation strategy by a Q-learning algorithm, and dynamically adjusting the health scheme by combining the real-time monitored user execution data.
9. The method for post-inspection health promotion management based on multi-modal data fusion according to claim 8, wherein the pre-evaluation and screening of candidate solutions by adopting a deep Q network DQN, learning an optimal recommendation strategy by a Q-learning algorithm, comprises: The regimen with the highest long-term expected utility is preferentially selected: ; Wherein the state is Action of Rewarding Time sequence differential learning for improving user health status, scheme adjustment options and compliance/index respectively Is shown in the state Down selection action A kind of electronic device Values, representing long-term expected utility, The learning rate is represented, the updating step length is controlled, the value is between 0 and 1, Indicating immediate rewards, executing actions The immediate return obtained after that is, Representing a discount factor, having a value between 0 and 1, for measuring the importance of future rewards, Indicating in the next state All possible actions are as follows Maximum of (3) Values representing the best long-term utility in the future.
10. A post-inspection health promotion management system based on multi-modal data fusion, comprising: The acquisition module is used for acquiring patient information, including physical examination data and medical history acquired by a physical examination system of a butt joint hospital, and acquired movement, diet, sleep and psychological multi-mode data; the processing module is used for processing and comprehensively evaluating the acquired multi-mode data to generate a health risk score and a user health portrait; the scheme generation module is used for generating a recommended health intervention scheme through content filtering, collaborative filtering and Bayesian network reasoning based on the user health portrait and health risk score and in combination with an expert knowledge base constructed by a medical ontology; The optimizing module is used for receiving the recommended health intervention scheme by a doctor, inputting adjusting information according to a clinical evaluation result to optimize the recommended health intervention scheme, generating a target health scheme and sending the target health scheme to a patient; The execution module is used for monitoring the execution data of the patient on the target health scheme and collecting the execution condition in real time through the sensor and the user log; and the feedback adjustment module is used for judging whether to trigger the early warning rule or deviate from the scheme target according to the execution condition, and adjusting the scheme if the early warning rule or deviate from the scheme target is triggered until the health condition of the patient is improved or reaches the health condition.

Description

Post-inspection health promotion management method and system based on multi-mode data fusion Technical Field The invention relates to the technical field of medical health information, in particular to a post-inspection health promotion management method and system based on multi-mode data fusion. Background Currently, existing health management systems can be mainly divided into the following three categories: The physical examination center information system mainly realizes the generation of physical examination reports and the preliminary interpretation of abnormal indexes, and the functions of the physical examination center information system are limited to the evaluation and presentation of health conditions. However, such systems terminate the service upon generating the report, lacking a technical means to translate the assessment conclusion into a personalized health action plan that is tangible, executable and has continuous tracking capabilities. Independent health management applications (e.g., keep, mint health, etc.) such applications focus on the management and recording of a single health dimension of exercise, diet, etc. The core disadvantage is that the advice provided by the method is mutually disjointed with the clinical physical examination data of the user, so that the advice lacks medical authority for specific health conditions (such as chronic diseases and abnormal indexes) of the user. Meanwhile, the application is technically difficult to realize linkage analysis and collaborative intervention of multidimensional data such as movement, diet, psychology, sleep and the like. The hospital follow-up system is generally in an interactive mode of active initiation by doctors and passive response by patients, and is specifically represented by telephone review reminding, medication reminding and the like. The mode is essentially a one-way, low frequency management flow, rather than a two-way, high frequency, immersive acceleration process. The technical limitation is that a professional platform for supporting multi-role cooperative guidance of a nutritionist, a convalescence engineer, a psychologist and the like is lacked in a mechanism for collecting continuous data of a daily life style of a patient. Based on the analysis of the prior art, the method can be concluded that the method mainly has the following technical defects that (1) data are split, a unified health portrait cannot be formed, and firm data islands exist among various systems. Clinical physical examination data, daily behavior data (exercise, sleep), nutritional data and psychological data are independent of each other and have different formats, and an effective fusion and intercommunication mechanism is lacked, so that a uniform and dynamically updated personal health portrait cannot be constructed based on the whole data. (2) The health intervention scheme is single and lacks authority basis, and advice provided by the existing health management application is generally based on a general model and is seriously disjointed with the clinical health condition (such as abnormal physical examination items and medical history) of the individual user. The solution generation process lacks personalized decision logic which is dominant by clinicians and strictly depends on medical evidence, so that the pertinence and the credibility of the solution are insufficient. (3) The management flow is closed and cyclized and cross-specialized collaborative deletion, the prior art fails to construct a complete closed-loop management system of health assessment, scheme planning, user execution, effect monitoring, data feedback and scheme adjustment. Meanwhile, each health professional role (such as clinicians, nutritionists and psychologists) lacks efficient collaborative platform support in technology, presents states of each battle, and cannot form team service resultant force with clinicians as a core. (4) The user compliance guarantee mechanism is weak, the traditional health management mode depends on boring education content and unidirectional reminding notification, and the technology lacks of an effective user participation incentive strategy and an immersive interactive guiding mechanism, so that long-term adherence and behavior compliance of the user are generally poor, and the overall effect of health intervention is finally influenced. (5) AI decision support capability is lacking-existing health management systems generally lack automated assessment and recommendation capabilities based on artificial intelligence. The system cannot intelligently analyze massive health data, cannot automatically identify health risk modes, cannot automatically generate personalized intervention scheme suggestions based on a medical knowledge base, depends on manual judgment, is low in efficiency and is easy to miss key risk factors. Therefore, how to provide a method and a system for post-inspection health promotion management based on mu