Intelligent Screening Method for Chronic Diseases Based on AI Model Data Fusion

Abstract

The present invention belongs to the technical field of AI-driven multimodal data fusion. Specifically, a chronic disease intelligent screening method based on AI model data fusion is provided, including: dynamically associating and analyzing health data, behavioral characteristics, and environmental data through the use of a multimodal data fusion hierarchical architecture, improving the comprehensiveness of screening; dynamically optimizing the alarm strategy to achieve the technical effect of adaptively adjusting the alarm mode by using dynamic scene perception and alarm decision tree to analyze the accompanying status, time nodes, and fluctuations in health indicators in the actual scene of patients in real time, which helps to increase the real-time dynamic monitoring ability of patients, reduce the burden of artificial accompaniment, and improve the pertinence of emergency response of patients; quantitatively analyzing the model deviation characteristics and prediction uncertainty through the use of federated learning and Monte Carlo confidence intervals to achieve the technical effect of dynamically correcting model parameters, which helps to reduce the model misjudgment rate.

Description

Technical Field

[0001] The present invention belongs to the technical field of AI-driven multimodal data fusion, and relates to a chronic disease intelligent screening method based on AI model data fusion. Background Art

[0002] Chronic diseases have become one of the major challenges in global public health. It is crucial to screen the real-time health status of patients with chronic diseases such as hypertension, chronic obstructive pulmonary disease, and cerebral infarction. Traditional chronic disease screening methods mainly rely on regular physical examinations and patient self-reporting, which have problems such as fragmented data and insufficient real-time performance, and are difficult to meet the needs of precise and personalized health management. With the development of wearable devices, the Internet of Things, and artificial intelligence (AI) technology, chronic disease intelligent screening methods based on multimodal data fusion have gradually become a research hotspot.

[0003] In the prior art, some solutions collect physiological indicators (such as heart rate, blood pressure) through wearable devices and combine machine learning models for risk prediction; other solutions use environmental sensors or patient behavior data for auxiliary analysis. For example, the patent with the Chinese patent publication number CN106845113A discloses a chronic disease remote management method and its management system based on blood pressure monitoring. By wirelessly uploading the basic chronic disease data and blood pressure data of patients to the system platform, the blood pressure of patients is screened to identify hypertensive patients among them; and chronic disease patients are classified and risk assessment analysis is carried out according to the risk level of hypertension for chronic diseases, and a chronic disease risk assessment report is given. The present invention uses the chronic disease system platform and the "1 + 4" mode of "doctors, experts, family members, and doctor assistants" and remote detection of blood pressure to achieve comprehensive management of hypertensive patients, thereby improving the treatment compliance and compliance rate of hypertensive patients, and further reducing the incidence and mortality of cardiovascular events.

[0004] The above solutions also have the following limitations: (1) The prior art has deficiencies in multi-source data fusion and comprehensive screening, resulting in limited recognition ability of the screening model for multiple types of chronic diseases. (2) The prior art lacks or is insufficient in dynamic scenario adaptation and precise alarm. It does not adjust the alarm intensity according to the real-time scenario of the patient (such as night mode, lack of escort), and makes the patient highly dependent on artificial escort. (3) The prior art lacks in model adaptive optimization and result credibility. It lacks correction of data deviation and misjudgment rate of prediction results, resulting in a decrease in the credibility of the patient's health status assessment results. Summary of the Invention

[0005] To overcome the above-mentioned defects of the prior art and to achieve the above object, the present invention proposes the following technical solutions: A chronic disease intelligent screening method based on AI model data fusion, comprising the following steps: S1. Real-time monitor the health data of patients through wearable monitoring devices and evaluate the health indicators of patients.

[0006] S2. When the health indicators of the patient are lower than the normal values, collect the interaction data of the space where the patient is located, and set the alarm mode of the device accordingly. The alarm mode includes strong alarm and weak alarm.

[0007] S3. When the health indicators of the patient are not lower than the normal values, monitor the daily activity data of the patient, and construct a multi-modal data fusion model based on the health data and the daily activity data. The multi-modal data fusion model includes a bottom-layer fusion unit, a middle-layer fusion unit, and a top-layer prediction unit.

[0008] S4. Retrieve the historical screening records, obtain the screening time slice sequence and its corresponding daily activity data, and adaptively train the multi-modal data fusion model accordingly.

[0009] S5. Derive the chronic disease health status reduction indicators of the patient from the adaptively trained multi-modal data fusion model.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) By using the multi-modal data fusion hierarchical architecture to dynamically analyze the correlation between health data, behavioral characteristics and environmental data, the present invention integrates multi-dimensional parameters, improves the comprehensiveness of chronic disease health status screening, and thus significantly improves the recognition accuracy of the corresponding health status of chronic disease types.

[0011] (2) By using the dynamic scene perception and alarm decision tree to analyze the accompanying status, time nodes and health indicator fluctuations in the actual scene of the patient in real time, and dynamically optimize the alarm strategy to achieve adaptive adjustment of the alarm mode, the present invention helps to increase the real-time dynamic monitoring ability of the patient, reduce the artificial accompanying burden, and improve the pertinence of the patient's emergency response.

[0012] (3) By using federated learning and Monte Carlo confidence intervals to quantitatively analyze the model deviation characteristics and prediction uncertainties, the present invention achieves the technical effect of dynamically correcting the model parameters, helps to reduce the model misjudgment rate, and at the same time enhances the credibility and practical applicability of the patient health status evaluation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0014] Figure 1 This is a schematic diagram of the implementation steps of the method of the present invention. Specific implementation manner

[0015] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0016] Please refer to Figure 1 As shown, the chronic disease intelligent screening method based on AI model data fusion proposed by the present invention includes the following steps: S1. Real-time monitor the health data of patients through wearable monitoring devices and evaluate the health indicators of patients.

[0017] In a preferred implementation manner, the content of S1 includes: collecting the health data of patients within a period of time, and the patient health data covers physiological indicators and various characteristic parameters of sleep monitoring.

[0018] The physiological index data covers heart rate-related characteristics and respiration-related characteristics. The heart rate-related characteristics at least include parameters such as heart rate and blood pressure, and the respiration-related characteristics at least include parameters such as respiration rate and expiratory flow rate; the sleep monitoring data at least includes parameters such as movement amplitude, movement change frequency, number of sleep-wake times, and sleep duration.

[0019] By analyzing the deviation change speed and deviation amplitude between the corresponding collected values of each characteristic parameter of the health data and the health pre-limit value, the change trend of the health data is judged, and the change trend includes a downward trend and a stable trend.

[0020] Specifically, establish a time window sequence corresponding to the time period, and obtain the instantaneous deviation between the corresponding collected values of each characteristic parameter of the health data and the health pre-limit value in each time window , calculate the deviation change speed between the corresponding collected values of each characteristic parameter of the health data and the health pre-limit value , where represents the instantaneous deviation between the corresponding collected value of the th characteristic parameter of the health data and the health pre-limit value in the th time window, represents the preset instantaneous deviation reference value, represents the number of each characteristic parameter of the health data, , represents the number of each time window, .

[0021] Meanwhile, calculate the deviation amplitude between the corresponding collected values of each characteristic parameter of the health data and the health pre-limit value , where and respectively represent the minimum and maximum values among the instantaneous deviations of the corresponding collected values of each characteristic parameter of the health data from the health pre-limit value in each time window.

[0022] If , then determine that the change trend of the health data is a downward trend, otherwise it is a stable trend, and accordingly determine the change trend of the health data, where is the corresponding baseline value of the preset deviation change speed, is the corresponding baseline value of the preset deviation amplitude,<s is the existential operator, is the logical AND operator.

[0023] The said and are obtained through training with historical data. Example: Select the data of the first 72 hours when the final diagnosis is the deterioration of chronic diseases in the historical training data, and determine the optimal threshold by the ROC curve.

[0024] When the health data shows a downward trend, activate the behavior monitoring unit of the device, collect the current behavior of the patient, and identify whether there is a diseased behavior. The diseased behavior includes excessive forced breathing, long-term continuous cough with abnormal sputum, etc.; if there is a diseased behavior, then set the health index of the patient below the normal value. Example: The normal value is 0.7, then when there is a diseased behavior, set the health index of the patient to 0.1.

[0025] For example, activate the millimeter-wave radar sensor corresponding to the behavior monitoring unit of the wearable monitoring device and the accelerometer corresponding to the inertial measurement unit to collect the following behavior characteristics: the movement amplitude of the respiratory muscle group, the standard deviation of the trunk acceleration, and the cough action frequency; construct a diseased behavior recognition model through a convolutional neural network: the input layer receives the said behavior characteristics, after extracting the temporal characteristics through the LSTM network, the output layer calculates the diseased probability P through the Sigmoid function; when any of the following conditions is met, it is determined that there is a diseased behavior: P>0.85 and lasts for more than 5 minutes, the movement amplitude of the respiratory muscle group exceeds 200% of the baseline value and is accompanied by the standard deviation of the trunk acceleration greater than 3m / s², the cough action frequency is greater than 15 times per minute and the proportion of the vocal cord energy spectrum in the 100 - 500Hz frequency band is greater than 40%.

[0026] S2. When the health index of the patient is lower than the normal value, collect the interaction data of the space where the patient is located, and accordingly set the alarm mode of the device. The alarm mode includes strong alarm and weak alarm. The interaction data includes whether there is a caregiver and their caregiving status, and the time node.

[0027] In a preferred embodiment, the content of S2 includes: activating the infrared sensing unit of the device, scanning the space where the patient is located, and determining whether there is a caregiver and their caregiving status, where the caregiving status includes a sleeping state and an awake state.

[0028] Specifically, if there are other people in the space where the patient is located, it is determined that there is a caregiver; further capturing the action sequence of the caregiver, if there is at least a continuous number of action changes at several time stamps in the action sequence, it is determined that the corresponding caregiving status of the caregiver is the awake state, otherwise it is the sleeping state.

[0029] The action sequence is marked by action codes. Example: The action codes for "standing", "walking", and "bending down" are marked as J01, J02, and J03 respectively.

[0030] Collect the current time node of the space where the patient is located, match the current time node with the time range defined for the night mode. If the current time node is within the corresponding time range of the night mode, it is determined that the time node belongs to the night mode.

[0031] The corresponding time range of the night mode can be adjusted and set by the caregiver according to their own sleep time, such as set to 22:00 - 6:00.

[0032] In the case of at least one of the following: there is no caregiver, the caregiving status of the caregiver is the sleeping state, or the time node belongs to the night mode, set the alarm mode to strong alarm, otherwise it is weak alarm.

[0033] The present invention uses dynamic scene perception and alarm decision tree to perform real-time analysis on the caregiving status, time node (such as night mode), and fluctuations in health indicators in the actual scenario of the patient, dynamically optimizing the alarm strategy to achieve adaptive adjustment of the alarm mode (strong / weak alarm), which helps to increase the real-time dynamic monitoring ability of the patient, reduce the burden of manual caregiving, and improve the pertinence of the patient's emergency response.

[0034] S3. When the patient's health indicators are not lower than the normal values, monitor the patient's daily activity data, including dynamic characteristics of the activity scene and activity behavior characteristics, and construct a multi-modal data fusion model based on the health data and daily activity data. The multi-modal data fusion model includes a bottom-layer fusion unit, a middle-layer fusion unit, and a top-layer prediction unit. Among them, the bottom-layer fusion unit is used to map the correlation between the patient's health data and activity behavior, the middle-layer fusion unit is used to dynamically allocate the health impact weights of the deviation degrees of different parameters included in the daily activity data, and the top-layer prediction unit is used to output health status reduction indicators for various types of chronic diseases such as hypertension, chronic obstructive pulmonary disease, and limb movement disorders due to cerebral infarction.

[0035] In a preferred embodiment, when the patient's health indicators are not lower than the normal values, the daily activity data of the patient is monitored, and the content includes: collecting the position space data and air environment data of the corresponding daily activity scenarios of the patient through the scene positioning unit of the device, and generating dynamic characteristics of the activity scenarios.

[0036] For example, the scene positioning unit includes an optical three-dimensional scanning sensor and an air quality sensor. The position space data includes, but is not limited to, the position space volume and its residence duration, which can be collected by the optical three-dimensional scanning sensor, and the air environment data is the clean air index collected by the air quality sensor.

[0037] Collect the patient's daily medication records and action records through the behavior marking unit of the device, analyze the medication compliance parameters and quantified parameters of living habits of the patient, and generate activity behavior characteristics.

[0038] For example, the behavior marking unit includes RFID (Radio Frequency Identification) technology and an accelerometer. The daily medication record at least includes the medication time, medication type and dosage of the patient, and can scan the medicine box label through RFID technology when the patient takes medicine. The action record includes the patient's daily walking actions or exercise actions, and the accelerometer can measure the acceleration changes of an object in three axes to determine the patient's behavior actions. The identification methods of the above daily medication records and action records are all prior arts and will not be elaborated here.

[0039] On the one hand, match the actual medication timestamp recorded by RFID with the doctor's order planned medication timestamp, and set an allowable error time window ; read the drug ID through the NFC tag and compare it with the prescription database in the electronic medical record to verify the consistency of the drug types.

[0040] Analyze the patient's medication compliance parameters through the following steps (1)-(3): (1) Calculate the time compliance index , where represents the serial number of the daily medication times, is the number of times of medication to be taken on the day, represents the th actual medication timestamp, represents the th doctor's order planned medication timestamp, , when the condition in the parentheses is satisfied, the value of is 1, otherwise it is 0. For example, if the deviation between a certain actual medication time and the planned time is within the allowable range, that is, , then Output 1 (indicating "taking medicine on time"), otherwise output 0 (indicating "not taking medicine on time").

[0041] (2) Calculate the dosage accuracy index , where represents the number of the types of medicine taken on the day, , is the number of the types of medicine taken on the day, represents the actual dosage of the -th type of medicine taken by the patient (obtained by scanning the RFID / NFC medicine box or recording by the sensor), represents the prescribed dosage of the -th type of medicine prescribed by the doctor (from the electronic medical record system), is the maximum prescribed dosage for a single type of medicine, which is pre-imported into the wearable monitoring device by hospital staff or equipment R & D management personnel.

[0042] (3) Calculate the medication compliance parameter of the patient: , where , respectively represent the corresponding preset weights of the time compliance index and the dosage accuracy index, which are set by empirical fitting, such as , .

[0043] On the other hand, the triaxial accelerations of the patient's exercise actions in multiple time windows are collected by the accelerometer, the exercise intensity index of each time window is calculated based on the Euclidean distance formula, and the duration of the continuous time window sequence in which the exercise intensity index exceeds the preset exercise intensity index threshold is obtained, which is recorded as the effective exercise duration; furthermore, the exercise intensity index and the effective exercise duration are constructed into a quantified parameter of living habits.

[0044] Integrate the dynamic features of the activity scene and the activity behavior features into the patient's daily activity data.

[0045] In a further preferred implementation manner, the multi-modal data fusion model constructed based on the health data and the daily activity data includes: aligning the time stamps of the health data and the daily activity data, and establishing a dynamic association matrix between the health data and the daily activity data by using the cross-attention mechanism to generate a bottom-layer fusion unit including the behavior-health dynamic association matrix.

[0046] Specifically, the behavior-health dynamic association matrix is composed of the corresponding association elements of the parameters included in the daily activity data and the feature parameters included in the health data, reflecting the dynamic association relationship between the daily activity data and the health data, where the association element refers to the degree of association between the two parameters and is generated by training the historical data based on the cross-attention mechanism.

[0047] Example: The behavior-health dynamic association matrix is , where represents the corresponding association element between the first parameter (such as the position space volume) in the daily activity data and the first characteristic parameter (such as the breathing frequency) included in the health data, represents the corresponding association element between the first parameter in the daily activity data and the i-th characteristic parameter included in the health data, represents the corresponding association element between the x-th parameter in the daily activity data and the first characteristic parameter included in the health data, represents the corresponding association element between the x-th parameter in the daily activity data and the i-th characteristic parameter included in the health data, where x represents the corresponding number of each parameter included in the daily activity data, .

[0048] Each parameter included in the daily activity data consists of the position data and air environment data of the patient's corresponding daily activity scenarios, the patient's medication compliance parameter, and the quantified parameter of living habits.

[0049] In the middle-layer fusion unit, the outpatient medical record of the patient is imported to determine the matching type of the patient's chronic disease, and the corresponding gating index of the health status reduction index of this type of chronic disease and the corresponding health impact weight of the unit deviation of each parameter included in the daily activity data are imported.

[0050] Specifically, each type of chronic disease has different sensitivities to daily activity parameters. For example: Hypertension may have a higher weight for exercise volume and diet and living habits; Chronic obstructive pulmonary disease may have a higher weight for air environment parameters. These weights are obtained through domain knowledge or data-driven learning.

[0051] In the top-layer prediction unit, the corresponding output data of the bottom-layer fusion unit and the middle-layer fusion unit are received, and the health status reduction index corresponding to the matching type of the patient's chronic disease is output; among them, the health status reduction indexes corresponding to different types of chronic diseases are independently calculated using the Sigmoid activation function.

[0052] Specifically, through hierarchical fusion and dynamic weight allocation, the prediction ability of the model for different chronic diseases is optimized, and the accuracy of the screening results is improved.

[0053] In a further preferred implementation manner, the health status reduction index corresponding to the chronic disease includes: screening out the association elements greater than 0 from the dynamic association matrix of health data and daily activity data, counting the parameters to which they belong in the daily activity data, and marking them as each associated parameter.

[0054] Specifically, if the associated element in the dynamic association matrix is greater than 0, it indicates that the associated element is related to the health data. Therefore, it can be used as an association parameter and further as a reference parameter for analyzing the reduction index of the corresponding health status of chronic diseases in the follow-up.

[0055] Define the ratio of the deviation between the collected value and the standard value of the characteristic parameter exceeding the preset deviation range as the deviation degree, and obtain the deviation degree of each associated parameter accordingly.

[0056] Accumulate the deviation degree of each associated parameter and the corresponding health impact weight of the unit deviation degree of the corresponding parameter included in the daily activity data, and then import the accumulated value into the Sigmoid activation function to obtain the reduction index of the corresponding health status of chronic diseases.

[0057] For example, the volume of the space where the patient is located is , and its standard value is , then is the deviation degree of the volume of the space where the patient is located, and is the corresponding preset deviation range of the space volume. Since staying in a narrow space for a long time has an adverse effect on patients with chronic obstructive pulmonary disease, the greater the deviation between the volume of the space where the patient with chronic obstructive pulmonary disease is located and the standard value, the greater the deviation degree of the space volume parameter is reflected.

[0058] The present invention performs dynamic association analysis on health data, behavior characteristics, and environmental data by using a multi-modal data fusion hierarchical architecture (bottom layer, middle layer, top layer), integrates multi-dimensional parameters (such as respiratory rate, medication compliance, space volume, etc.), improves the comprehensiveness of chronic disease health status screening, and then significantly improves the recognition accuracy of the corresponding health status of chronic disease types.

[0059] S4. Retrieve historical screening records, obtain the screening time slice sequence and its corresponding daily activity data, and adaptively train the multi-modal data fusion model accordingly.

[0060] In a preferred embodiment, the obtaining of the screening time slice sequence and its corresponding daily activity data includes: performing time slice segmentation on a certain monitoring period to obtain the health data of the patient in each time slice.

[0061] Screen out the time slices with a certain fluctuation range in the health data, combine them into a screening time slice sequence, and then obtain the daily activity data of the patient corresponding to the screening time slice sequence from the historical screening records.

[0062] For example, if the deviation degree of a certain parameter in the health data exceeds the preset first-level threshold of the deviation degree in a continuous number of time slices, then import the continuous number of time slices into the screening time slice sequence.

[0063] In a further preferred embodiment, the adaptive training of the multi-modal data fusion model includes: obtaining the deviation degrees of the parameters included in the daily activity data of the patient in each time slice of the screening time slice sequence, and starting the federated learning update mechanism to correct the errors of the deviation degrees of the parameters. This helps to correct the deviation of the fusion model caused by the differences in different parameters corresponding to different patients under the same chronic disease type, enabling the global model to adapt to different groups, and thus improving the accuracy of the multi-modal data fusion model.

[0064] Deploy a Monte Carlo layer in the top-level prediction unit to output the 95% confidence interval of the prediction of the health status reduction index corresponding to the chronic disease in real time. When the width of the confidence interval is greater than the preset width baseline value, manual review is required.

[0065] The Monte Carlo Dropout is an approximation method in Bayesian deep learning. By keeping the Dropout layer activated during the test phase (inference) and propagating the input data forward multiple times, the distribution of the prediction results is obtained. Retaining the prediction results of multiple samplings of Dropout during testing is equivalent to sampling from the approximate posterior distribution, thereby estimating the uncertainty of the prediction.

[0066] For example, for the same input of daily activity data, perform R forward propagations (each time Dropout randomly discards different neurons) to obtain the set of health status reduction index outputs ; calculate the mean and standard deviation of this set. The mean reflects the predicted value of the health status reduction index, and the standard deviation reflects the uncertainty of the health status reduction index. Based on this, the 95% confidence interval is obtained. Assuming that the predicted value follows a normal distribution, 1.96 is the 97.5 quantile of the standard normal distribution.

[0067] Obtain the width of the 95% confidence interval for the prediction of the health status reduction index through the general calculation formula of the confidence width. The 95% confidence interval provides the fluctuation range of the model prediction value, which helps to improve the reliability of the prediction of the health status reduction index. Example: If the mean is , it means that there is a 95% probability that the true value of the health status reduction index falls within this interval.

[0068] If the set width baseline value is set to 0.2 (clinical studies have shown that the misjudgment rate increases significantly when the width is greater than 0.2), then when the width of the confidence interval is greater than 0.2, the misjudgment rate of the model increases. At this time, feedback to the device terminal is required to request manual review.

[0069] S5. Derive the chronic disease health status reduction index of the patient from the multi-modal data fusion model after adaptive training.

[0070] In a preferred embodiment, the S5 is specifically as follows: Derive the chronic disease matching type of the patient and its health status reduction index from the top-level prediction unit of the multi-modal data fusion model, compare it with the corresponding gating index of the health status reduction index of this type of chronic disease, and when its health status reduction index exceeds its corresponding gating index, extract the deviation degrees of the parameters included in the patient's daily activity data in each time slice of the screening time slice sequence, and jointly export them to the device display end. This helps to intuitively display the chronic disease risk of the patient and facilitates medical staff or patients to take intervention measures in a timely manner.

[0071] The present invention quantifies and analyzes the model deviation characteristics and prediction uncertainty by using federated learning and Monte Carlo confidence intervals, achieving the technical effect of dynamically correcting model parameters, which helps to reduce the model misjudgment rate and at the same time enhances the credibility and practical applicability of the assessment of the patient's health status.

[0072] It should be noted that for the formulas mentioned above, through the principle of dimensional consistency and mathematical standardization means (such as normalization processing, dimensionless parameter conversion, or unit system unification), physical quantities with different attributes can be translated into unitless standard values or superimposable parameters of the same dimension, so as to eliminate the interference of different dimensions on the operation logic, and make the formulas have mathematical operation rationality and objective law adaptability while retaining the original data distribution characteristics. The above are only exemplary embodiments of the present invention and should not be used to limit the scope of the present invention.

Claims

1. A chronic disease intelligent screening method based on AI model data fusion, characterized in that Including the following steps: S1. Real-time monitor the patient's health data through a wearable monitoring device and evaluate the patient's health indicators; S2. When the patient's health indicators are lower than the normal values, collect the interaction data of the space where the patient is located, and set the alarm mode of the device accordingly. The alarm mode includes strong alarm and weak alarm; S3. When the patient's health indicators are not lower than the normal values, monitor the patient's daily activity data, and construct a multi-modal data fusion model based on the health data and daily activity data. The multi-modal data fusion model includes a bottom-layer fusion unit, a middle-layer fusion unit, and a top-layer prediction unit; The construction of the multi-modal data fusion model based on the health data and daily activity data includes: aligning the time stamps of the health data and the daily activity data, and using a cross-attention mechanism to establish a dynamic association matrix between the health data and the daily activity data, and generating a bottom-layer fusion unit containing a behavior-health dynamic association matrix; Import the patient's outpatient medical record list into the middle-layer fusion unit, determine the chronic disease matching type of the patient, and import the corresponding gating indicators of the health status reduction indicators of this type of chronic disease and the health impact weights corresponding to the unit deviation degrees of the parameters included in the daily activity data; The top-layer prediction unit receives the corresponding output data of the bottom-layer fusion unit and the middle-layer fusion unit, and outputs the health status reduction indicators corresponding to the chronic disease matching type of the patient; among them, the health status reduction indicators corresponding to different types of chronic diseases are independently calculated using the Sigmoid activation function; The health status reduction indicators corresponding to the chronic disease include: Screen out the association elements greater than 0 from the dynamic association matrix of the health data and the daily activity data, count the parameters to which they belong in the daily activity data, and mark them as each associated parameter; Define the ratio of the deviation between the collected value and the standard value of the characteristic parameter exceeding the preset deviation range as the deviation degree, and thus obtain the deviation degree of each associated parameter; Accumulate the deviation degrees of each associated parameter and the health impact weights corresponding to the unit deviation degrees of the corresponding parameters included in the daily activity data, and then import the accumulated value into the Sigmoid activation function to obtain the health status reduction indicators corresponding to the chronic disease; S4. Retrieve the historical screening records, obtain the screening time slice sequence and the corresponding daily activity data, and adaptively train the multi-modal data fusion model accordingly; S5. Export the health status reduction indicators of the patient's chronic disease from the adaptively trained multi-modal data fusion model.

2. The intelligent chronic disease screening method based on AI model data fusion according to claim 1, wherein, The content of S1 includes: Collect the health data of the patient within a period of time. The patient's health data covers physiological indicators and each characteristic parameter corresponding to sleep monitoring; By analyzing the deviation change speed and deviation amplitude between the collected value of each characteristic parameter to which the health data belongs and the health pre-limit value, judge the change trend of the health data. The change trend includes a downward trend and a stable trend; When the health data shows a downward trend, activate the behavior monitoring unit of the device, collect the patient's current behavior, and identify whether there is a diseased behavior; if there is a diseased behavior, set the patient's health indicators below the normal value.

3. The chronic disease intelligent screening method based on AI model data fusion according to claim 1, wherein, The content of S2 includes: Activate the infrared sensing unit of the device to scan the space where the patient is located, and determine whether there is a caregiver and their caregiving status. The caregiving status includes a sleeping state and an awake state; Collect the current time node of the space where the patient is located, and match the current time node with the time range defined for the night mode. If the current time node is within the corresponding time range of the night mode, it is determined that the time node belongs to the night mode; Set the alarm mode to strong alarm in at least one of the following cases: when there is no caregiver, or the caregiver's caregiving status is in the sleeping state, or the time node belongs to the night mode; otherwise, it is a weak alarm.

4. The intelligent chronic disease screening method based on AI model data fusion according to claim 1, wherein, When the patient's health indicators are not lower than the normal values, monitor the patient's daily activity data, including: Collect the position space data and air environment data of the corresponding daily activity scene of the patient through the scene positioning unit of the device to generate dynamic features of the activity scene; Collect the patient's daily medication records and action records through the behavior marking unit of the device, analyze the medication compliance parameters and quantified parameters of living habits of the patient to generate activity behavior features; Integrate the dynamic features of the activity scene and the activity behavior features into the patient's daily activity data.

5. The chronic disease intelligent screening method based on AI model data fusion according to claim 1, wherein, The obtaining of the screening time slice sequence and its corresponding daily activity data includes: Perform time slice segmentation on a certain monitoring period to obtain the patient's health data in each time slice; Screen out the time slices with a certain fluctuation range in the health data, combine them into a screening time slice sequence, and then obtain the corresponding patient's daily activity data from the historical screening records for the screening time slice sequence.

6. The chronic disease intelligent screening method based on AI model data fusion according to claim 5, characterized in that, The adaptive training of the multi-modal data fusion model includes: obtaining the deviation degrees of the parameters included in the patient's daily activity data in each time slice corresponding to the screening time slice sequence, and starting the federated learning update mechanism to correct the errors of the deviation degrees of the parameters; Deploy a Monte Carlo layer in the top-level prediction unit to continuously output the width of the 95% confidence interval for the prediction of the health status reduction index of the chronic disease. When the width of the confidence interval is greater than the preset width baseline value, manual review is required.

7. The chronic disease intelligent screening method based on AI model data fusion according to claim 6, characterized in that The S5 is specifically: derive the chronic disease matching type of the patient and its health status reduction index from the top-level prediction unit of the multi-modal data fusion model, compare it with the corresponding gating index of the health status reduction index of this type of chronic disease. When its health status reduction index exceeds its corresponding gating index, extract the deviation degrees of the parameters included in the patient's daily activity data in each time slice corresponding to the screening time slice sequence, and jointly export them to the device display terminal.

Images