Intelligent guiding method for chronic disease staged behavior change

Abstract

The invention relates to a chronic disease staged behavior change intelligent guiding method, and belongs to the technical field of chronic disease health management. The method comprises the steps of capturing behavior, psychological, environmental and physiological data to construct a four-dimensional data set, filling missing values by adopting similar patient behavior mode interpolation, filtering abnormal values by adopting a decision tree branch pruning rule, and generating a patient health portrait vector; constructing a training set based on longitudinal behavior data, screening key dimensions through feature importance, and optimizing a behavior change dynamic decision tree in combination with an incremental dynamic iterative training mechanism; inputting the four-dimensional data set into a decision tree, generating a personalized obstacle analysis report and predicting a future stage retrograde risk; and calling a corresponding strategy library, and dynamically adjusting the guide strategy by taking the task completion rate and the index improvement amplitude as feedback signals. Accurate judgment of the chronic disease behavior change stage is achieved, the problems of single data dimension and strategy homogenization in traditional intervention are solved, and the patient behavior change compliance and the long-term management effect are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of chronic disease health management, and particularly relates to a chronic disease stage behavior change intelligent guidance method. BACKGROUND

[0002] The incidence of chronic diseases (such as diabetes, hypertension, coronary heart disease, etc.) continues to rise, and patients need to adjust their health behaviors (such as regular exercise, sugar control diet, and timely medication) for a long time to control the progression of the disease and reduce the risk of complications, so scientific behavior intervention has become the core link of chronic disease management. Traditional solutions rely on medical staff to conduct manual assessment during outpatient follow-up, and only use single physiological index detection results (such as blood glucose value, blood pressure value) or static paper questionnaires (such as behavior change stage scale) to judge the behavior state of patients, which cannot capture real-time patient daily behavior data (such as exercise frequency, diet structure), psychological fluctuations (such as changes in self-management confidence), and environmental influences (such as family support and living area exercise facilities), resulting in one-sided stage determination and difficulty in matching the real behavior change needs of patients. Although some digital tools attempt to introduce wearable devices to collect data, technical limitations are still significant: the stage determination model is mostly fixed and static, such as using a preset threshold to divide the stage, which cannot dynamically optimize the determination rules with the addition of new health data of the patient (such as changes in long-term medication adherence, exercise habit adjustments due to seasonal factors); intervention strategies are generally homogeneous, for example, the same health guidance is pushed to patients in the preparation stage and the action stage, without considering the differences in patient education level (such as understanding ability of medical terminology) and life scene (such as whether to balance work and family care); at the same time, there is a lack of effective feedback loop, most tools can only record task completion, and do not associate task completion rate, physiological index improvement amplitude and strategy adjustment depth, when patients encounter execution difficulties (such as being unable to complete exercise tasks due to busy work), the system cannot switch to an adaptive solution in time, leading to a gradual decline in patient adherence, and ultimately making it difficult to achieve long-term behavior change and support the clinical goal of fine and personalized management of chronic diseases. SUMMARY

[0003] To solve the above problems in the prior art, the application provides a chronic disease stage behavior change intelligent guidance method, The object of the application can be achieved by the following technical solutions: S1: capture behavior data, collect psychological data, record environmental data, and synchronize physiological indicators to build a four-dimensional data set; for missing values, use interpolation filling based on the behavior patterns of similar patients, and for abnormal values, use decision tree branch pruning rules for identification and filtering, and finally generate a patient health portrait vector; S2: Based on the longitudinal behavior data of chronic disease patients, a training set is constructed, the behavior change stage transition is taken as the target variable, the four-dimensional data set is taken as the split feature, and the importance screening mechanism is used to lock the key dimensions that contribute to the stage determination; through the incremental dynamic iterative training mechanism, the branch threshold is dynamically adjusted, and after the new patient data is included, the tree structure optimization is triggered; S3: The four-dimensional data set is input into the behavior change dynamic decision tree, and the behavior change stage result is output through three-level node reasoning, while the key features on the decision path are automatically extracted, and a personalized obstacle analysis report with hierarchical visualization design is generated; based on the historical data correlation analysis of the decision tree leaf nodes, the future stage regression risk is predicted; S4: According to the behavior change stage result output by the decision tree, the corresponding strategy library is called; the patient task completion rate and the index improvement range are taken as feedback signals to dynamically adjust the guidance strategy weight, and the adaptive scheme is switched for patients who have not completed the task.

[0004] Specifically, the four-dimensional data set is used to integrate the multi-dimensional health information of the patient, and the specific process is: through the mobile terminal sensor to track the motion path, the diet type and frequency behavior data; combined with dynamic questionnaire to collect self-efficacy, emotional ups and downs psychological data; with the help of positioning function and environmental sensing components to record temperature and humidity, place attribute environment data; through wearable devices to synchronize heart rate, blood glucose level physiological indicators; four kinds of data are associated according to time stamp, active and passive behavior is distinguished, physiological indicator fluctuation is marked, scene label is added, forming continuous and mutually associated four-dimensional data set.

[0005] Specifically, the missing value filling is used to perfect the four-dimensional data set, and the specific process is: screening the same patient group similar to the target patient in disease, complications, behavior mode and age, extracting the typical data mode of the same patient group in the corresponding missing period, and interpolating and supplementing according to the behavior trend of the target patient before and after the period. After filling, the data is compared and verified with the complete data of the patient in other periods.

[0006] Specifically, the abnormal value identification and filtering is used to purify the four-dimensional data set, and the specific process is: based on the historical data statistics of the same patient to determine the regular fluctuation range of each data, the values exceeding the range are marked as abnormal, combined with the actual scene of the patient to determine the misrecord, and the abnormal data is filtered out, the abnormal time and possible reasons are recorded and archived separately.

[0007] Specifically, the patient health portrait vector generation is used to extract core health features, and the specific process is: extracting key features from the cleaned four-dimensional data, including exercise weekly duration, blood glucose fluctuation amplitude, emotional stability index, and environment adaptation degree of frequently visited places, eliminating data redundancy through feature fusion algorithm, classifying according to physiological, behavioral, psychological and environmental dimensions, and combining the quantified features to form a vector containing core dimensions, and dynamically adjusting the vector value with new data.

[0008] Specifically, the training set construction is used to provide training data for the decision tree model, and the specific process is: collecting longitudinal behavior data of different chronic disease patients in the period matched with the key period of disease management, covering behavior execution, physiological index change, psychological state transition and environmental feature, cleaning the data to remove repeated and contradictory records, dividing training and verification subsets, and labeling the behavior change stage transition of each sample by medical staff combined with clinical judgment.

[0009] Specifically, the importance screening mechanism is used to determine the key split features of the decision tree, and the specific process is: calculating the information gain rate of each split feature, analyzing the relevance of the feature and the stage transition, sorting by gain rate, selecting key dimensions, setting redundancy feature judgment standard, and retaining features through small test set.

[0010] Specifically, the incremental dynamic iterative training mechanism is used to optimize the structure of the decision tree, and the specific process is: setting the new data accumulation trigger condition, starting the training process when the new data meets the condition, merging the new data with the historical data, recalculating the key dimension feature weight, adjusting the branch node judgment condition and threshold, retaining the branch structure with accurate judgment in the historical model, and comparing the results of the new and old models on the same data set.

[0011] Specifically, the personalized obstacle analysis report generation is used to present the stage judgment results and influencing factors, and the specific process is: extracting the core basis of stage judgment from the four-dimensional data set and the decision tree inference results, including physiological index compliance, behavior execution records, psychological state fluctuation characteristics, and environmental support for behavior, comparing the core basis with the standards to be achieved in the behavior change stage, selecting the unqualified feature items as the obstacle factors affecting the stage transition, collecting the performance of these obstacle factors in the patient's historical data, integrating the stage judgment results, obstacle factor list and corresponding historical data into the report content, and intuitively displaying the influence degree of each obstacle factor through charts to reflect the specific problems that hinder the patient from entering the next stage.

[0012] Specifically, the three-level node reasoning is used to accurately determine the patient behavior change stage, and the specific process is: according to the clinical basic control target set in the authoritative guide of chronic disease management, the physiological indicators of the patient are evaluated to determine whether the basic control standard is reached, and on the basis of the basic control standard, the behavior execution rate of the patient is calculated, that is, the ratio of the actual number of times of completing the related health management task to the planned number of times, and the preliminary division result is combined to further narrow the determination range, and different subgroups with different behavior specification degrees are distinguished from the group with physiological indicators meeting the standard. Then, the psychological characteristics of the patient in the change intention are combined for comprehensive analysis, and finally the specific behavior change stage of the patient is locked.

[0013] As a preferred technical solution of the present application, S3 first inputs the four-dimensional data set associated with the time stamp after being integrated and cleaned by S1 into the behavior change dynamic decision tree optimized by S2, and starts the data reasoning process; first, the three-level node reasoning determines the patient behavior change stage, the first level node evaluates the physiological indicators in the four-dimensional data according to the clinical basic control target set in the authoritative guide of chronic disease management, and distinguishes the groups with physiological indicators meeting the standard and not meeting the standard, and completes the preliminary range screening; the second level node extracts the behavior execution records in the four-dimensional data on the basis of the first level determination result, calculates the ratio of the actual number of times of completing the health management tasks such as exercise, medication and diet to the planned number of times, and further divides the behavior specification and behavior improvement subgroups from the group with physiological indicators meeting the standard, and narrows the stage determination range; the third level node continues to call the psychological data in the four-dimensional data, analyzes the stability of the patient's change intention (such as whether the tendency of actively adjusting behavior in the continuous period is continuous), and analyzes the persistence of the health management confidence (such as the belief strength when facing execution difficulties), and combines the determination results of the previous two levels to lock the specific behavior change stage of the patient, and outputs the stage determination result.

[0014] Specifically, while outputting the stage determination result, the system automatically extracts the key features on the decision path in the decision tree reasoning process, compares the key features with the standard one by one in combination with the standard requirements corresponding to the chronic disease behavior change stage, and screens out the feature items not meeting the standard as the obstacle factors affecting the stage transition of the patient; then, the performances of these obstacle factors in the historical four-dimensional data of the patient are collected, the stage determination result, the obstacle factor list and the corresponding historical data are integrated into a personalized obstacle analysis report, and the report is presented by visual design, and the influence degree of each obstacle factor is directly displayed by charts, so that medical staff and patients can clearly master the specific problems hindering the stage transition.

[0015] Specifically, the future stage regression risk prediction is used for early identification of potential risks, and the specific process is: extracting historical patient data similar to the current patient stage and characteristics in the decision tree leaf node, analyzing the trigger factors of stage regression of these patients in the subsequent period, correlating the recent data of the current patient, judging the possibility of risk occurrence through historical correlation rules, and ranking the risk factors according to the influence degree to include in the report.

[0016] Specifically, the guidance strategy library calling is used for matching the staged intervention scheme, and the specific process is: in the pre-intention stage, the content complexity of the disease and behavior related animation popularization is adjusted according to the education level of the patient; in the preparation stage, the target task is decomposed into daily executable small steps, synchronized to the intelligent terminal and reminded according to the work and rest period; in the action stage, the correct medication and exercise posture are demonstrated through technology to generate a virtual scene, and the patient can shoot his own operation and demonstration for comparison and correction; in the maintenance stage, the type of health service exchanged by behavior points is set according to the patient's preference.

[0017] The beneficial effects of the present application are: (1) By setting the physiological-behavioral-psychological-environmental four-dimensional data set construction mechanism, the missing value filling method based on the behavior pattern of similar patients, the abnormal value identification filtering process of the decision tree branch pruning rule and the patient health portrait vector generation module of the core feature fusion, on the one hand, the multi-dimensional real-time data of the patient can be integrated, and the one-sidedness of the traditional intervention relying only on single physiological or static questionnaire data can be avoided; on the other hand, the missing values are interpolated and filled by the behavior pattern of similar patients, and the abnormal values are filtered in combination with the actual scene, which effectively guarantees the data integrity and accuracy, and then the core health features such as exercise week average duration and blood glucose fluctuation amplitude are extracted by feature fusion, forming a health portrait dynamically updated with new data, providing comprehensive and accurate data support for subsequent behavior change stage determination, solving the problem of insufficient judgment basis caused by single data dimension and static evaluation in traditional manual evaluation; (2) By setting the dynamic decision tree model integrating the behavior change theory, the future stage regression risk prediction module based on historical data correlation analysis and the guidance strategy library and feedback adjustment mechanism adapted to the stage, firstly, the three-level node reasoning verifies from the physiological index standard situation, the behavior execution rate and the psychological feature stability layer by layer, and combines the incremental dynamic iterative training, so that the behavior change stage determination is more suitable for the real-time state of the patient, solving the problem of low determination accuracy and inability to adapt to dynamic data changes of the traditional static model; secondly, the risk prediction module can identify potential regression risks in advance and sort them by analyzing the regression trigger factors of similar historical patients and correlating the recent data of the current patient; thirdly, the guidance strategy library provides personalized schemes according to the stage, and dynamically adjusts the strategy weight according to the feedback signal of task completion rate and index improvement amplitude, solving the problem of low patient compliance caused by the homogenization of traditional intervention strategies and the lack of closed-loop adjustment, effectively improving the sustainability of patient behavior change and the intervention effect. BRIEF DESCRIPTION OF DRAWINGS

[0018] For the convenience of those skilled in the art to understand, the present application is further described below in conjunction with the drawings.

[0019] Fig. 1 The framework diagram of a chronic disease stage behavior change intelligent guidance method of the present application; Fig. 2 The data flow diagram of a chronic disease stage behavior change intelligent guidance method of the present application. DETAILED DESCRIPTION

[0020] In order to further illustrate the technical means and effects adopted by the present application to achieve the predetermined application purposes, the specific embodiments, structures, features and effects according to the present application are described in detail below in conjunction with the drawings and preferred embodiments.

[0021] Please refer to Figs. 1-2 A chronic disease stage behavior change intelligent guidance method S1: Capture behavior data, collect psychological data, record environmental data, and synchronize physiological indicators to build a four-dimensional data set; for missing values, use interpolation filling based on the behavior pattern of similar patients, and for abnormal values, use decision tree branch pruning rules for identification filtering, and finally generate a patient health portrait vector; S2: Based on the longitudinal behavior data of chronic disease patients, a training set is constructed, the behavior change stage transition is taken as the target variable, the four-dimensional data set is taken as the splitting feature, and the importance screening mechanism is used to lock the key dimensions that contribute to the stage determination; through the incremental dynamic iterative training mechanism, the branch threshold is dynamically adjusted, and after the new patient data is included, the tree structure optimization is triggered; S3: Input the four-dimensional data set into the behavior change dynamic decision tree, output the behavior change stage result through three-level node reasoning, automatically extract the key features on the decision path, and generate a hierarchical visual design personalized obstacle analysis report; based on the historical data association analysis of the decision tree leaf node, predict the future stage regression risk; S4: According to the behavior change stage result output by the decision tree, call the corresponding strategy library; take the patient task completion rate and index improvement amplitude as the feedback signal, dynamically adjust the guidance strategy weight, and switch the adaptive scheme for patients who have not completed the task.

[0022] Specifically, the four-dimensional data set is constructed for integrating multi-dimensional health information of the patient, and the specific process is as follows: the movement path, the food type and the frequency behavior data are tracked through the mobile terminal sensor; the self-efficacy and the emotional fluctuation psychological data are collected by combining the dynamic questionnaire; the temperature, the humidity, the place attribute and the environment data are recorded by means of the positioning function and the environment sensing component; the heart rate and the blood glucose level physiological indicators are synchronized through the wearable device; the four types of data are associated according to the time stamp, the active and passive behaviors are distinguished, the physiological indicator fluctuation is marked, the scene label is added, and the continuous and mutually associated four-dimensional data set is formed.

[0023] Specifically, the missing value filling is used for perfecting the four-dimensional data set, and the specific process is as follows: a similar patient group similar to the target patient in terms of disease type, complication condition, behavior mode and age stage is screened, a typical data mode of the similar patient group in the corresponding missing period is extracted, interpolation is supplemented by referring to the behavior trend of the target patient before and after the period, and the data is compared and verified with the complete data of the patient in other periods after filling.

[0024] In the embodiment, the related calculation formula of the similar patient similarity weighted filling is as follows: , Wherein, is the filling value of the index j of the current patient i, n is the number of similar patients screened, is the similarity weight of the current patient i and the similar patient k, is the effective value of the index j of the similar patient k, which is weighted calculated by the disease matching degree (0.4), the behavior mode similarity (0.3) and the age stage matching degree (0.3).

[0025] In the embodiment, for the scene that the “daily exercise duration” (index j) of the middle-aged type 2 diabetes patient (current patient i) is missing for 1 day, the formula is applied for filling: Screening of similar patients k: from the hospital chronic disease management database (containing 5000+ type 2 diabetes patient data), the similar patients k completely matched with the current patient i in terms of disease type a ik =1, the behavior mode similarity b ik =0.8, such as taking steps as the main exercise mode, and the age stage being 45-55 years old c ik =1, a total of n=5 people; the effective value x kj of the “daily exercise duration” of the similar patient k: the exercise duration in the corresponding missing period is extracted from the historical records of the wearable devices of the five patients, and the similarity weight w is obtained by the related calculation formula of the similar patient similarity weighted filling.

[0026] Specifically, the outlier identification filter is used to purify the four-dimensional data set, and the specific process is: based on the statistical data of the same patient history, the fluctuation range of each data is determined, the values exceeding the range are marked as abnormal, combined with the actual scene of the patient to determine the misrecord, the confirmed abnormal data is filtered out, and the abnormal time and possible reasons are recorded separately.

[0027] Specifically, the patient health portrait vector generation is used to extract core health features, and the specific process is: extracting key features from the cleaned four-dimensional data, including exercise weekly duration, blood glucose fluctuation amplitude, emotional stability index, and environment adaptation degree of frequently visited places, eliminating data redundancy through feature fusion algorithm, classifying according to physiological, behavioral, psychological and environmental dimensions, and combining the quantified features to form a vector containing core dimensions, and the vector value is dynamically adjusted with new data.

[0028] Specifically, the training set construction is used to provide training data for the decision tree model, and the specific process is: collecting longitudinal behavior data of different chronic disease patients in the period matching the key period of disease management, covering behavior execution, physiological index change, psychological state transition and environmental feature situation in each stage, cleaning data to remove repeated and contradictory records, dividing training and verification subsets, and labeling the behavior change stage transition of each sample by medical staff combined with clinical judgment.

[0029] Specifically, the importance screening mechanism is used to determine the key split features of the decision tree, and the specific process is: calculating the information gain rate of each split feature, analyzing the relevance of the feature and the stage transition, sorting by gain rate, selecting key dimensions, setting redundant feature judgment standard, and testing reserved features through small range test set.

[0030] Specifically, the incremental dynamic iterative training mechanism is used to optimize the structure of the decision tree, and the specific process is: setting the new data accumulation trigger condition, starting the training process when the new data meets the condition, merging the new data with the historical data, recalculating the key dimension feature weight, adjusting the branch node judgment condition and threshold, retaining the branch structure with accurate judgment in the historical model, and comparing the results of the new and old models on the same data set.

[0031] Specifically, the personalized obstacle analysis report generation is used for presenting the stage determination result and influencing factors, and the specific process is as follows: extracting the core basis of stage determination from the four-dimensional data set and the decision tree inference result, including the physiological index reaching standard situation, the specific record of behavior execution, the fluctuation characteristics of psychological state, and the support degree of environment on behavior, comparing the core basis with the standard to be reached in the behavior change stage, screening out the unqualified characteristic items as the obstacle factors influencing the stage transition, collecting the performance of these obstacle factors in the historical data of the patient, integrating the stage determination result, the obstacle factor list, and the corresponding historical data support into the report content, and intuitively displaying the influence degree of each obstacle factor through a chart to reflect the specific problems hindering the patient from entering the next stage.

[0032] Specifically, the three-level node reasoning is used for accurately determining the behavior change stage of the patient, and the specific process is as follows: evaluating the physiological index of the patient according to the clinical basic control target set in the authoritative guideline for chronic disease management, judging whether the basic control standard is reached, calculating the behavior execution rate of the patient, that is, the ratio of the actual number of times of completing the related health management task to the planned number of times, and further narrowing the determination range in combination with the preliminary division result to distinguish the subgroups with different behavior standardization degrees from the group with the physiological index reaching the standard. Then, the psychological characteristics of the patient in the change willingness are combined for comprehensive analysis, and finally, the specific behavior change stage of the patient is locked.

[0033] As a preferred technical solution of the present application, the S3 first inputs the four-dimensional data set associated with the time stamp after being integrated and cleaned by the S1 into the behavior change dynamic decision tree optimized by the S2, and starts the data inference process; first, the behavior change stage of the patient is determined by three-level node reasoning, the first level node evaluates the physiological index in the four-dimensional data according to the clinical basic control target set in the authoritative guideline for chronic disease management, and distinguishes the groups with the physiological index reaching the standard and not reaching the standard, to complete the preliminary range screening; the second level node extracts the behavior execution record in the four-dimensional data on the basis of the first level determination result, calculates the ratio of the actual number of times of completing the health management tasks such as exercise, medication, and diet to the planned number of times, and further divides the subgroups with behavior standardization and behavior improvement from the group with the physiological index reaching the standard in combination with the ratio, to narrow the stage determination range; the third level node continues to call the psychological data in the four-dimensional data, analyzes the stability of the change willingness of the patient (such as whether the tendency of actively adjusting the behavior is continuous in the continuous period), the persistence of the health management confidence (such as the belief strength when facing the execution difficulty), combines the determination results of the previous two levels, locks the specific behavior change stage of the patient, and outputs the stage determination result.

[0034] In this embodiment, for the scenario of limited computing power of terminal devices of primary medical institutions (such as outpatient tablet computers and community health management terminals), a lightweight decision tree inference engine based on pruning optimization and quantization compression is deployed on the terminal for the S3 three-level node inference link; the specific process is as follows: first, the S2 training optimized dynamic decision tree model is pruned to remove branches with very low contribution to the stage judgment, thereby reducing the model size and computing power consumption; during inference, the quantized decision tree model is first loaded, four-dimensional data input by S1 is received, and the three-level node logic is inferred step by step: the first level node only calls the model branch related to physiological data to quickly judge whether the standard is met; the second level node calls the behavior data branch based on the first level result to calculate the execution rate and narrow the range; the third level node calls the core branch of the psychological and environmental data to finally output the stage result; during the inference process, the engine records the computing time of each node in real time.

[0035] Specifically, while outputting the stage judgment result, the system automatically extracts the key features on the decision path in the decision tree inference process, compares the key features with the standard one by one according to the standard requirements corresponding to the stage of behavior change of chronic diseases, and screens out the features that do not meet the standard as obstacle factors affecting the stage transition of the patient; then, the performances of these obstacle factors in the historical four-dimensional data of the patient are collected, the stage judgment result, the obstacle factor list, and the corresponding historical data are integrated into a personalized obstacle analysis report, and the report is presented by visual design, so that the influence degree of each obstacle factor is directly displayed by charts, and the medical staff and the patient can clearly understand the specific problems that hinder the stage transition.

[0036] In this embodiment, a middle-aged patient with type 2 diabetes (with a long course of disease and in the behavior change preparation stage) is taken as an example to perform S3: first, the four-dimensional data set (containing daily fasting blood glucose, postprandial blood glucose, and other core monitoring indicators of physiological indicators; containing daily exercise, staple food intake record, and other key behavior information affecting blood glucose of behavior data; containing periodic self-efficacy evaluation results, reflecting the willingness strength of the patient to actively adjust behavior of psychological data; containing family diet environment adaptability, community exercise facility accessibility, and associating with the patient's daily behavior execution scene of environmental data) associated by time stamp after S1 cleaning and integration is input into the behavior change dynamic decision tree trained and optimized by S2; Start three-level node inference: the first level node refers to the basic control standard of physiological indicators in the guidelines for the prevention and treatment of type 2 diabetes, evaluates that the patient's recent physiological indicators are in the standard range, and preliminarily divides them into the standard group; the second level node combines the actual execution of the exercise and diet tasks in the behavior data, finds that the patient's exercise task completion needs to be improved, and further divides them into the behavior improvement sub-group from the standard group; the third level node calls the psychological data, observes that the patient's recent self-efficacy evaluation results show a downward trend, and finally locks the patient in the transition period from the preparation stage to the action stage based on the results of the previous two levels.

[0037] At the same time of outputting the stage results, the system automatically extracts the key features on the decision path - insufficient exercise performance and fluctuation of self-efficacy, compares these features with the behavior and psychological standards that should be achieved during the transition period, and screens out the above two as obstacle factors affecting the stage transition; then it collects the occurrence rules of insufficient exercise performance and the correlation records of fluctuation of self-efficacy and behavior adjustment in the patient's historical data, integrates the stage determination results, the obstacle factor list and the historical data support, generates a visual report, and intuitively presents the influence degree of each obstacle factor on the stage transition.

[0038] Finally, based on the historical patient data stored in the decision tree leaf nodes, similar samples are screened out from the current patient stage and the core features (physiological standard, exercise to be improved, fluctuation of self-efficacy), the main trigger factors (mostly related to unfavorable exercise conditions in the environment) of subsequent stage regression of such samples are analyzed; the current patient's recent environmental data is associated to find similar unfavorable conditions, based on which the risk of future stage regression is predicted, and the unfavorable condition is listed as the primary risk factor and included in the report; here, when selecting four-dimensional data, the physiological index focuses on blood glucose because blood glucose control is the core of type 2 diabetes management, the behavior data selects exercise and diet because they directly affect blood glucose fluctuations, the psychological data selects self-efficacy because the strength of will determines the degree of behavior adherence, and the environmental data associates with the diet and exercise scene because home and community environment is the main place for patients' daily behavior, which can truly reflect the external conditions of behavior execution.

[0039] In this embodiment, the edge server of the basic community health service center is taken as the core, and the patient end wearable device (Huawei bracelet, etc.), intelligent medicine box (Xiaomi intelligent medicine box), household medical grade sensor (temperature and humidity sensor) and basic outpatient tablet computer are matched to build a distributed data collection and calculation hardware system; the software environment is based on Linux / Ubuntu operating system, MySQL / PostgreSQL database is used to store four-dimensional data and historical samples of chronic disease patients, Python is used for data cleaning and formula calculation, Scikit-learn framework is used to build a decision tree model, TensorFlowLite is used to deploy a lightweight inference engine, a federal learning framework is used for multi-institution data privacy protection, AR technology supports strategy pre-demonstration, and the overall environment and technology stack adapt to the needs of basic computing power, multi-source data synchronization and privacy protection, etc., providing stable technical support for each link (missing value filling, feature screening, risk prediction, strategy pushing) of the method.

[0040] Specifically, the future stage regression risk prediction is used to identify potential risks in advance, and the specific process is as follows: the historical patient data similar to the current patient stage and characteristics in the decision tree leaf node are extracted, the triggering factors of stage regression of these patients in the subsequent period are analyzed, the recent data of the current patient are associated, the possibility of risk occurrence is judged through the historical association rule, and the risk factors are sorted according to the influence degree and included in the report.

[0041] In this embodiment, the risk probability P(R) of the current patient to occur stage regression in the future period is calculated, and the formula is as follows: , Wherein, M is the number of groups of historical patient samples similar to the core characteristics of the current patient in the decision tree leaf node, is the feature matching score of the mth group of historical samples and the current patient, is the total sum of feature matching scores of all groups of historical samples, is the proportion of stage regression (i.e. the ratio of the number of regression to the total number) in the mth group of historical samples.

[0042] In this embodiment, for the future 1 month stage regression risk prediction scene of the stable period of chronic obstructive pulmonary disease (current patient i), the risk probability P(R) is calculated by the formula: Historical sample screening: from the respiratory department chronic disease management system (containing 2000+ chronic obstructive pulmonary disease patient data), the historical samples similar to the core characteristics (oxygen saturation 92%-95%, daily pulmonary function exercise 10-15min, home humidity 40%-50%) of the current patient in the decision tree leaf node are screened, and there are M=3 groups; Feature matching score : The score is calculated by the system according to the matching degree of physiology (0.4), behavior (0.3) and environment (0.3) dimensions; Regression proportion : The regression proportions of the three groups of samples are obtained from the follow-up records (once every two weeks, for three months) of the historical samples; finally, the risk probability formula is calculated.

[0043] Specifically, the guided strategy library calling is used to match the stage intervention scheme, and the specific process is as follows: in the pre-intention stage, the content complexity of the disease and behavior related animation popular science is adjusted according to the education level of the patient; in the preparation stage, the target task is divided into daily executable small steps, which is synchronized to the intelligent terminal and reminded according to the work and rest period; in the action stage, the correct medication and exercise posture are demonstrated through technology to support the patient to shoot his own operation and demonstrate comparison correction; in the maintenance stage, the health service type of behavior points exchange is set according to the patient's preference.

[0044] In the embodiment, after the result of the S3 output behavior change stage is output, an augmented reality (AR) real-time guidance technology is used to preview the guidance strategy of the corresponding stage, helping the patient to intuitively understand the strategy content and improve the subsequent execution willingness. The specific process is as follows: after the result of the stage is output, the core strategy of the stage in the strategy library is automatically associated; the AR module is started, the patient scans the home environment through the camera of the mobile terminal, and the AR system superimposes a virtual guidance scene in the screen based on the environmental recognition result, such as displaying a 3D virtual mannequin demonstration of “jumping jacks” in the living room area, marking the action standard (such as “knees not exceeding the tip of the foot”), and displaying the step animation of “bar stretching” in the balcony area; at the same time, the AR system combines the behavior data of S1 (such as the action deviation of the patient during the previous exercise), and highlights the problems that need to be paid attention to in the demonstration (such as “the arm lifting is insufficient this time, please lift to the shoulder level”); the patient can adjust the viewing angle in real time to view the virtual demonstration from different angles, and can also click the screen to pause and view the text explanation of a certain step; after the preview demonstration is completed, the system automatically saves the AR guidance content as a video for the patient to review later.

[0045] The above is only a preferred embodiment of the present application, and does not limit the present application in any form. Although the present application has been disclosed as above, it is not intended to limit the present application. Any person skilled in the art can make some changes or modifications to the above disclosed technical content without departing from the scope of the technical solution of the present application, and any brief introduction, modification, equivalent change and modification of the above embodiments according to the technical essence of the present application are still within the scope of the technical solution of the present application.

Claims

1. A chronic disease stage-based behavior change intelligent guidance method, characterized in that, The application comprises the following steps: S1: capture behavior data, collect psychological data, record environmental data, synchronize physiological indicators, and construct a four-dimensional data set; For missing values, interpolation filling based on the behavior pattern of patients with similar diseases is adopted, and abnormal values are identified and filtered through decision tree branch pruning rules, and finally a patient health portrait vector is generated; S2: Based on the longitudinal behavior data of chronic disease patients, a training set is constructed, the behavior change stage transition is taken as the target variable, the four-dimensional data set is taken as the splitting feature, the importance screening mechanism is adopted to lock the key dimensions that contribute to the stage determination, and the branch threshold is dynamically adjusted through incremental dynamic iterative training mechanism, and the tree structure is optimized after new patient data is added; S3: The four-dimensional data set is input into the behavior change dynamic decision tree, and the behavior change stage result is output through three-level node reasoning, and the key features on the decision path are automatically extracted, and a hierarchical visual design personalized obstacle analysis report is generated; Based on the historical data correlation analysis of the decision tree leaf node, the future stage regression risk is predicted; S4: According to the behavior change stage result output by the decision tree, the corresponding strategy library is called; the patient task completion rate and the index improvement range are taken as feedback signals, the guiding strategy weight is dynamically adjusted, and the adaptive scheme is switched for patients who have not completed the task.

2. The method according to claim 1, characterized in that, The four-dimensional data set is used to integrate the multi-dimensional health information of patients, and the specific process is as follows: the movement path, diet type and frequency behavior data are tracked through the mobile terminal sensor; the psychological data is collected by combining dynamic questionnaire; the temperature and humidity, place attribute environment data are recorded by means of positioning function and environment sensing component; the heart rate, blood glucose level physiological indicators are synchronized through wearable device; the four types of data are associated according to time stamp, the active and passive behaviors are distinguished, the physiological indicator fluctuation is marked, the scene label is added, and the continuous and mutually associated four-dimensional data set is formed.

3. The method according to claim 1, characterized in that, The missing value filling is used to perfect the four-dimensional data set, and the specific process is as follows: similar patient groups with similar diseases, complications, behavior patterns and age groups of the target patient are screened, typical data patterns of the similar patient groups in the corresponding missing period are extracted, and interpolation is supplemented by referring to the behavior trend of the target patient before and after the period, and the data after filling is compared and verified with the complete data of the patient in other periods.

4. The method according to claim 1, characterized by, The abnormal value identification and filtering is used to purify the four-dimensional data set, and the specific process is as follows: based on the historical data statistics of similar patients, the normal fluctuation range of each data is determined, the values beyond the range are marked as abnormal, the misrecordings are determined according to the actual scene, the abnormal data are filtered and removed, and the abnormal occurrence time and possible reasons are separately archived.

5. The method of claim 1, characterized in that, The patient health portrait vector generation is used to extract core health features, and the specific process is as follows: the key features are extracted from the cleaned four-dimensional data, the data redundancy is eliminated through feature fusion algorithm, the features are classified according to physiological, behavioral, psychological and environmental dimensions, the features are quantified and combined to form a vector containing core dimensions, and the vector value is dynamically adjusted with new data.

6. The method of claim 1, characterized in that, The training set is constructed to provide training data for the decision tree model, and the specific process is as follows: collecting longitudinal behavior data of different chronic disease patients in the period matched with the key period of disease management, covering the behavior execution situation, physiological index change, psychological state transition, and environmental feature situation in each stage, cleaning the data to remove repeated and contradictory records, dividing the training and verification subsets, and marking the behavior change stage transition of each sample by medical staff combined with clinical judgment.

7. The method of claim 1, characterized in that, The importance screening mechanism is used to determine the key split features of the decision tree, and the specific process is as follows: calculating the information gain rate of each split feature, analyzing the relevance of the feature and the stage transition, sorting by the gain rate, selecting the key dimension, setting the redundant feature judgment standard, and testing the reserved features through a small range of test sets.

8. The method of claim 1, characterized in that, The incremental dynamic iteration training mechanism is used to optimize the structure of the decision tree, and the specific process is as follows: setting the new data accumulation trigger condition, starting the training process when the new data meet the condition, merging the new data with the historical data, recalculating the key dimension feature weight, adjusting the branch node judgment condition and threshold, retaining the branch structure with accurate judgment in the historical model, and comparing the results of the new and old models on the same data set.

9. The method of claim 1, characterized in that, The personalized obstacle analysis report generation is used to present the stage judgment results and influencing factors, and the specific process is as follows: extracting the core basis of stage judgment from the four-dimensional data set and the decision tree reasoning results, including the physiological index standard, the specific record of behavior execution, the fluctuation characteristics of psychological state, and the support degree of environment to behavior, comparing the core basis with the standard to be reached in the behavior change stage, screening out the unqualified feature items as the obstacle factors affecting the stage transition, collecting the performance of these obstacle factors in the historical data of the patient, integrating the stage judgment results, obstacle factor list, and corresponding historical data into the report content, and intuitively displaying the influence degree of each obstacle factor through charts to reflect the specific problems that hinder the patient from entering the next stage.

10. The method of claim 1, characterized in that, The three-level node reasoning is used to accurately determine the behavior change stage of the patient, and the specific process is as follows: evaluating the physiological index of the patient according to the clinical basic control target set in the authoritative guideline of chronic disease management, judging whether the basic control standard is reached, calculating the behavior execution rate of the patient, i.e. the ratio of the actual number of times of completing related health management tasks to the planned number of times, combining the preliminary division result to further narrow down the judgment range, and distinguishing the subgroups with different behavior standardization degrees from the group with physiological index meeting the standard. Then, the psychological characteristics of the patient in the change intention are comprehensively analyzed, and the specific behavior change stage of the patient is finally locked.

11. The method of claim 1, characterized in that, The future stage regression risk prediction is used to identify potential risks in advance, and the specific process is as follows: extracting the historical patient data similar to the current patient stage and features from the decision tree leaf node, analyzing the trigger factors of stage regression of these patients in the subsequent period, associating the recent data of the current patient, judging the possibility of risk occurrence through historical association rules, and ranking the risk factors according to the influence degree and including them in the report.

12. The method of claim 1, characterized in that, The guiding strategy library calls for matching the phased intervention scheme, and the specific process is as follows: in the pre-intention stage, the content complexity of the disease and behavior related animation popular science is adjusted according to the education level of the patient; in the preparation stage, the target task is divided into small steps executable every day, synchronized to the intelligent terminal and reminded according to the working hours; In the action stage, a virtual scene is generated by technology to demonstrate correct medication and exercise posture, and the patient can take photos of his own operation and demonstration for comparison and correction; in the maintenance stage, the type of health services exchanged by behavior points is set according to the patient's preference.

Images