HFpEF subtype recognition and MAE prediction method and system based on machine learning

Abstract

The invention belongs to the technical field of medical information, and particularly discloses an HFpEF subtype recognition and MAE prediction method and system based on machine learning, and the method comprises the following steps: collecting the electronic health record and echocardiography data of an HFpEF patient from a plurality of medical centers, and carrying out the preprocessing; carrying out clustering analysis on the data of the HFpEF patients by using a K-prototypes algorithm, and identifying different HFpEF subtypes; and based on the identified subtype, constructing an MAE prediction model by using a machine learning algorithm, carrying out interpretability analysis by using SHAP, and outputting an MAE prediction result. By adopting the technical scheme, different HFpEF subtypes are identified by using more parameters, and the MAE prediction result is obtained.

Description

Technical Field

[0001] The present invention belongs to the field of medical information technology, and relates to a method and system for identifying HFpEF subtypes and predicting MAE based on machine learning. Background Art

[0002] Heart failure (HF) is a syndrome caused by heart dysfunction, which can lead to congestion. Approximately 64.3 million people worldwide are affected by it, including about 8.9 million in China. Due to its high hospitalization rate and high mortality rate, it has brought a heavy burden to the public health system.

[0003] Left ventricular ejection fraction (LVEF) can distinguish the clinical manifestations of HF and guide treatment decisions, classifying patients into heart failure with reduced ejection fraction (HFrEF) (LVEF < 40%), heart failure with mildly reduced ejection fraction (HFmrEF) (40 ≤ LVEF < 50%), and heart failure with preserved ejection fraction (HFpEF) (LVEF ≥ 50%).

[0004] HFpEF accounts for approximately half of HF cases, and its incidence is on the rise. However, the results of clinical trials have been less than satisfactory, indicating that HFpEF has greater subtype heterogeneity compared to HFrEF. Understanding this heterogeneity helps in risk stratification and clinical management of HFpEF patients.

[0005] Although previous studies have classified HFpEF into multiple subtypes, the sample sizes used for unsupervised clustering are small, relying on clinical experience and limited biomarkers, and these subtypes have not been widely validated, nor have their prognostic situations been deeply analyzed. Summary of the Invention

[0006] The purpose of the present invention is to provide a method and system for identifying HFpEF subtypes and predicting MAE based on machine learning, accurately identifying HFpEF subtypes and predicting the occurrence probability of major adverse events (MAE).

[0007] To achieve the above purpose, the basic solution of the present invention is: A method for identifying HFpEF subtypes and predicting MAE based on machine learning, comprising the following steps:

[0008] Collect electronic health records and echocardiogram data of HFpEF patients from multiple medical centers and perform preprocessing;

[0009] Use the K-prototypes algorithm to perform clustering analysis on the data of HFpEF patients to identify different HFpEF subtypes;

[0010] Based on the identified subtypes, a MAE prediction model is constructed using machine learning algorithms, and SHAP is used for interpretability analysis to output the MAE prediction results.

[0011] The working principle and beneficial effects of this basic solution are as follows: This technical solution starts from data collection, processes the original echocardiogram text records and structured clinical data in the electronic health record (EHR), converts the text into structured data and cleans it.

[0012] Apply k-prototype clustering (generate subtypes) to the data of HFpEF patients to identify different HFpEF subtypes. The operation is simple and conducive to data processing. Construct a MAE prediction model to predict the occurrence probability of major adverse events (MAE) and realize the analysis of the prognosis situation.

[0013] Furthermore, the optimal number of clusters is determined by the elbow method, silhouette coefficient and gap statistic, and the clustering results are internally and externally validated. Specifically:

[0014] Divide the data of HFpEF patients into a derivation cohort, validation cohort 1 and validation cohort 2. Apply k-prototype clustering to the derivation cohort to generate subtypes, and conduct external validation on validation cohorts 1 and 2 to obtain different survival curve analyses to verify the prognostic differences between subtypes;

[0015] Apply the clustering algorithm based on partitioning around medoids to the derivation cohort to identify and ensure the robustness of the results.

[0016] Comprehensively use the elbow method, silhouette coefficient and gap statistic for evaluation. Consider the number of clusters from 1 to 8, and carefully evaluate the performance of each clustering setting. Use the silhouette score, and a higher score indicates that the clustering is clearer and tighter. The gap statistic measures the difference between the clustering and random data, and the higher the value, the better the clustering structure. The elbow method finds the optimal clustering by observing the slowdown of the reduction rate of clustering variance, which indicates that the clustering is optimal.

[0017] Furthermore, the MAE prediction model is a random forest or a support vector machine.

[0018] Select the model according to needs. The model has good performance and is conducive to use.

[0019] Furthermore, based on the MAE prediction results, judge the risk level of the patient:

[0020] Low risk level: MAE prediction result ≤ 0.5;

[0021] Medium risk level: MAE prediction result is between 0.5 and 0.7;

[0022] High risk level: MAE prediction result > 0.7.

[0023] Obtain the risk level of the patient, which is beneficial to prognosis.

[0024] Furthermore, based on the patient's risk level and HFpEF subtype results, auxiliary training is carried out:

[0025] Based on different risk levels, preset corresponding physical preset values, collect the patient's physical data, and compare it with the physical preset values corresponding to the risk levels to determine whether the patient's current body is abnormal;

[0026] If the patient's physical data is abnormal, alarm data is sent to the patient terminal and the medical staff terminal. On the contrary, if the patient's physical data is normal, an auxiliary training plan is carried out, specifically:

[0027] According to the patient's risk level, exercise items are initially screened;

[0028] Based on the patient's HFpEF subtype, determine the patient's exercise tolerance and limiting factors, secondarily screen the exercise items, and according to the intensity, duration, and frequency parameters of the secondarily screened exercise items, assign weights to each parameter to calculate the training scores of each exercise item, sort the training scores of each exercise item, and select the exercise item with the largest training score as the final auxiliary training item.

[0029] Based on the patient's risk level and HFpEF subtype results, auxiliary training is carried out to screen the optimal auxiliary training items so that the patient can train to achieve better training effects.

[0030] Furthermore, when the patient performs the final auxiliary training item, the patient's body movement images are collected in real time, the patient's body joint angles and movement trajectories are extracted, and similarity calculations are performed with the preset images to determine whether the patient's real-time training meets the standards. If it does not meet the standards, a prompt signal is output, and at the end of one training, it is calculated whether the overall training of this training meets the standards;

[0031] Taking one week as a cycle period, record the number of training times of the patient within one week. If the number of training times meets the preset number and the overall training pass rate of the patient in one week is above 80%, then add 1 to the preset number and add 10 minutes to the duration of the exercise item; otherwise, an alarm signal is output to the patient terminal and the medical staff terminal;

[0032] Similarity calculation method:

[0033] For two joint connection lines and respectively represent the vectors from joint A to joint B and from joint B to joint C, and the joint angle θ is:

[0034]

[0035] Among them, is the dot product of two vectors, and are the magnitudes of two vectors;

[0036] Let the joint angle vectors of the patient and the preset image be θ a and θ r respectively, and the similarity between them is calculated by cosine similarity:

[0037]

[0038] where, θ a ·θ r is the dot product of two vectors, |θ a | and |θ r | are the magnitudes of two vectors;

[0039] Let the running trajectory be obtained by the joint position transformation between consecutive frames, represented by the two-dimensional or three-dimensional coordinates of each joint. If the patient's body makes a movement in the image sequence, the joint trajectory is the sequence of joint positions at each moment;

[0040] If the trajectories and of the patient and the preset image are composed of a series of points, the Euclidean distance is used to measure the difference D between them:

[0041]

[0042] where, k is the frame index in the trajectory, represents the Euclidean distance of the vector;

[0043] The final similarity S combines the angle similarity and the trajectory-based similarity through weighted average:

[0044]

[0045] where, w1 and w2 are weights, satisfying w1 + w2 = 1.

[0046] Judge whether the patient's training is up to standard, so as to make timely improvements, optimize the training, and gradually increase the training duration at the same time, and the training effect will be better.

[0047] The present invention also provides an HFpEF subtype identification and MAE prediction system based on the method described in the present invention, including a data acquisition module, a subtype identification module, and an MAE prediction module;

[0048] The data acquisition module is used to collect the electronic health records and echocardiogram data of HFpEF patients from multiple medical centers and transmit them to the subtype identification module;

[0049] The subtype recognition module is used to perform clustering analysis on the data of HFpEF patients using the K-prototypes algorithm to identify different HFpEF subtypes;

[0050] The input end of the MAE prediction module is connected to the output end of the subtype recognition module, and the MAE prediction module is used to output the MAE prediction result.

[0051] This system utilizes larger and more diverse real-world data, combines unsupervised machine learning clustering methods to find more robust HFpEF subtypes, and uses machine learning methods to predict the probability of major adverse events (MAE) in each subtype.

[0052] Furthermore, it also includes an auxiliary training device, which includes a support base, a human-computer interaction module, various sports equipment, a mobile sensor, an acceleration sensor, a blood pressure sensor, a pulse sensor, a temperature sensor, and a wearable;

[0053] The human-computer interaction module is connected to the support base through a bracket, the human-computer interaction module is connected to the subtype recognition module and the MAE prediction module, and an ID tag recognition module is provided in the human-computer interaction module;

[0054] The treadmill among the various sports equipment is arranged on the support base. A vertical threaded rod is connected to the side wall of the support base. A plurality of placement boxes are provided on the threaded rod. The placement boxes are provided with threaded holes and are threadedly connected to the threaded rod through the threaded holes. The sports equipment can be respectively placed in the placement boxes, and an electric valve is provided on one side of the placement box;

[0055] The mobile sensor, acceleration sensor, blood pressure sensor, pulse sensor, and temperature sensor are all arranged on the wearable. The wearable is worn on the patient. An ID tag is provided on the wearable. The human-computer interaction module recognizes the patient's ID tag, receives the corresponding HFpEF subtype recognition and MAE prediction information, and the auxiliary training item control signal output end of the human-computer interaction module is connected to the control end of the electric valve;

[0056] The output ends of the mobile sensor, acceleration sensor, blood pressure sensor, pulse sensor, and temperature sensor are all connected to the human-computer interaction module, the patient terminal, and the medical staff terminal.

[0057] The human-computer interaction module is set up to facilitate data interaction and can also display corresponding data. When the patient is training, it can also provide corresponding training guidance videos. The treadmill is arranged on the support base. When the patient needs to walk slowly or quickly for training, it can be carried out on the treadmill.

[0058] A threaded rod is provided on the support base, and a placement box is provided on the threaded rod, which is conducive to storing a variety of sports equipment and can be used at any time. Moreover, the placement box is threadedly connected to the threaded rod, and the placement height of the placement box can be arbitrarily adjusted by rotating the placement box, which is convenient for taking and placing. It can also be used as a seat for patients to rest temporarily or an armrest for patients during exercise.

[0059] Using a variety of sensors to collect patients' body data in real time is conducive to monitoring the patients' physical conditions and improving the safety during training.

[0060] Furthermore, the auxiliary training device further includes an inner rod, a guardrail, a camera and an inflation and deflation mechanism;

[0061] At least two threaded rods are provided on both sides of the support base. The top of the threaded rod is open and hollow inside. The bottom of the inner rod extends into the threaded rod from the top of the threaded rod, and the outer wall of the inner rod is in sealed sliding connection with the inner wall of the threaded rod. The inner bottom of the threaded rod is communicated with the inflation and deflation mechanism through a trachea;

[0062] The guardrail is annular, horizontally placed and installed on the top of the inner rod. There is an annular track on the side of the guardrail facing the patient. The camera is installed on the mobile trolley through a universal cloud platform, and the mobile trolley moves along the annular track.

[0063] In this way, when the inflation and deflation mechanism inflates or deflates the threaded rod, the inner rod can be controlled to move up or down along the inner wall of the threaded rod to adjust the height of the guardrail, and it can be flexibly adjusted for patients of different heights. At the same time, the height of the camera on the guardrail changes, and image information of patients at different heights is collected, which is conducive to comprehensive analysis. The camera can collect the whole body images of the patient during training by moving along the annular track on the mobile trolley, and can also move to the required position to collect image information at a specific angle.

[0064] Furthermore, a plurality of air bags are provided inside the guardrail, and the air bags are communicated with the inflation and deflation mechanism. After being inflated, the air bags extend towards the side where the patient is located.

[0065] After being inflated, the air bags are closer to the patient. Without affecting the patient's movement, the air bag material is elastic and soft, which can protect the patient. The air bags can surround the patient to play roles such as preventing falls and collisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 is a flow diagram of the method for HFpEF subtype recognition and MAE prediction based on machine learning of the present invention;

[0067] Figure 2 is a front view of the auxiliary training device of the HFpEF subtype recognition and MAE prediction system of the present invention;

[0068] Figure 3It is a top view of the protective fence of the auxiliary training equipment of the HFpEF subtype identification and MAE prediction system of the present invention.

[0069] The reference numerals in the drawings of the specification include: support base 1, human-computer interaction module 2, threaded rod 3, placement box 4, inner rod 5, guardrail 6, camera 7, treadmill 8, airbag 9, bracket 10. DETAILED DESCRIPTION

[0070] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0071] In the description of the present invention, it is necessary to understand that the terms "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.

[0072] In the description of the present invention, unless otherwise specified and limited, it should be noted that the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a mechanical connection or an electrical connection, or it can be the internal connection between two components. It can be a direct connection or an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to the specific circumstances.

[0073] The present invention discloses a method for HFpEF subtype identification and MAE prediction based on machine learning, which uses larger and more diverse real-world data, combined with unsupervised machine learning clustering methods, to find more robust HFpEF subtypes and develop potential clinical pathways to enhance impact (using machine learning methods to predict the probability of major adverse events (MAE) in each subtype). Figure 1 As shown in Figure 2, the HFpEF subtype identification and MAE prediction method based on machine learning includes the following steps:

[0074] Electronic health records and echocardiographic data of HFpEF patients were collected from multiple medical centers and preprocessed: features with more than 30% missing data were removed, and the missing values ​​of numerical features were interpolated using the random forest (RF) algorithm, and the missing values ​​of categorical features were processed using iterative interpolation. In addition, features with correlation coefficients > 0.8 were filtered out, and the features with the greatest clinical significance were retained.

[0075] The K-prototypes algorithm is used to perform cluster analysis on the data of HFpEF patients to identify different HFpEF subtypes, such as male low-risk type (younger age, most with a smoking history; relatively low incidence of other comorbidities except hyperlipidemia); male atherosclerosis type (characteristics - including older individuals, almost all with a smoking history, and a relatively high incidence of atherosclerotic factors); female diabetes type (composed of older individuals, with the highest prevalence of diabetes (99.7%)); female atrial fibrillation type (including the oldest individuals, with a relatively high incidence of atrial fibrillation); female low-risk type (composed of younger individuals, with a relatively low incidence of other comorbidities except hyperlipidemia);

[0076] Based on the identified subtypes, a MAE prediction model is constructed using a machine learning algorithm (preferably the MAE prediction model is a random forest or a support vector machine), and SHAP (SHapley Additive exPlanations, SHAP is a game-theory-based machine learning model interpretation tool used to quantify the contribution of features to model predictions) is used for interpretability analysis, and the MAE prediction results are output (the prognosis results cover in-hospital mortality and major adverse events (MAE), where the latter is defined as in-hospital all-cause death, liver failure, renal failure, respiratory failure, or myocardial infarction).

[0077] In a preferred embodiment of the present invention, the optimal number of clusters is determined by the elbow method, silhouette coefficient, and gap statistic, and internal and external validation of the clustering results is performed, specifically:

[0078] The data of HFpEF patients is divided into a derivation cohort, validation cohort 1, and validation cohort 2. The k-prototype clustering is applied to the derivation cohort to generate subtypes, and external validation is performed on validation cohorts 1 and 2 to obtain different survival curve analyses to verify the prognostic differences between subtypes;

[0079] The Partitioning Around Medoids clustering algorithm (internal validation) is applied to the derivation cohort to identify and ensure the robustness of the results.

[0080] The elbow method, silhouette coefficient, and gap statistic are comprehensively used for evaluation. Consider the number of clusters from 1 to 8, and carefully evaluate the performance of each clustering setting. Using the silhouette score, a higher score indicates clearer and tighter clustering. The gap statistic measures the difference between the clustering and random data, and the higher the value, the better the clustering structure. The elbow method finds the optimal number of clusters by observing the slowdown in the rate of decrease in clustering variance, indicating that the clustering is optimal.

[0081] In a preferred embodiment of the present invention, based on the MAE prediction results, the risk level of the patient is determined:

[0082] Low risk level: The MAE prediction result ≤ 0.5, indicating a low probability of MAE occurring in patients with the corresponding subtype category, belonging to the low-risk population;

[0083] Medium risk level: The MAE prediction result is between 0.5 and 0.7;

[0084] High risk level: The MAE prediction result > 0.7.

[0085] Obtaining the risk level of the patient is beneficial for prognosis.

[0086] In a preferred embodiment of the present invention, based on the risk level of the patient and the HFpEF subtype results, auxiliary training is carried out:

[0087] Based on different risk levels, corresponding physical preset values are preset, the patient's physical data is collected, and compared with the physical preset values corresponding to the risk levels to determine whether the patient's current body is abnormal;

[0088] If the patient's physical data is abnormal, alarm data is sent to the patient terminal and the medical staff terminal. Conversely, if the patient's physical data is normal, an auxiliary training plan is carried out, specifically:

[0089] According to the risk level of the patient, exercise items are initially screened;

[0090] Based on the HFpEF subtype of the patient, the exercise tolerance and limiting factors of the patient are determined, the exercise items are secondarily screened, and according to the intensity, duration, and frequency parameters of the secondarily screened exercise items, weights are assigned to each parameter to calculate the training scores of each exercise item. The training scores of each exercise item are sorted and the exercise item with the largest training score is selected as the final auxiliary training item.

[0091] Based on the risk level of the patient and the HFpEF subtype results, auxiliary training is carried out to screen the optimal auxiliary training item so that the patient can achieve better training effects during training.

[0092] In a preferred embodiment of the present invention, when the patient performs the final auxiliary training item, the patient's body movement images are collected in real time, the body joint angles and movement trajectories of the patient are extracted, and similarity calculations are performed with the preset images to determine whether the patient's real-time training meets the standards. If it does not meet the standards, a prompt signal is output, and at the end of one training, it is calculated whether the overall training of this training meets the standards;

[0093] Taking one week as a cycle period, record the number of training sessions of the patient within one week. If the number of training sessions meets the preset number and the overall training compliance rate of the patient in a week is above 80%, then add 1 to the preset number and add 10 minutes to the duration of the exercise program; otherwise, output an alarm signal to the patient terminal and the medical staff terminal.

[0094] Similarity calculation method:

[0095] For two joint connection lines and respectively represent the vectors from joint A to joint B and from joint B to joint C, and the joint angle θ is:

[0096]

[0097] Among them, is the dot product of the two vectors, and are the magnitudes of the two vectors;

[0098] Let the joint angle vectors of the patient and the preset image be θ a and θ r respectively, and the similarity between the two is calculated by cosine similarity:

[0099]

[0100] Among them, θ a ·θ r is the dot product of the two vectors, |θ a | and |θ r | are the magnitudes of the two vectors;

[0101] Suppose the movement trajectory is obtained through the transformation of joint positions between consecutive frames and is represented by the two-dimensional or three-dimensional coordinates of each joint. If the patient's body makes a movement in the image sequence, the joint trajectory is the sequence of joint positions at each moment;

[0102] If the trajectories and of the patient and the preset image are composed of a series of points, then the Euclidean distance is used to measure the difference D between the two:

[0103]

[0104] Among them, k is the frame index in the trajectory, represents the Euclidean distance of the vector;

[0105] The final similarity S combines the angle similarity and the trajectory-based similarity through weighted average:

[0106]

[0107] Among them, w1 and w2 are weights, satisfying w1 + w2 = 1.

[0108] Judge whether the patient's training is up to standard, so as to optimize the training in a timely manner, and at the same time gradually increase the training duration, and the training effect is better.

[0109] The present invention also provides an HFpEF subtype recognition and MAE prediction system based on the method of the present invention, including a data acquisition module, a subtype recognition module, and an MAE prediction module;

[0110] The data acquisition module is used to collect the electronic health records and echocardiogram data of HFpEF patients from multiple medical centers and transmit them to the subtype recognition module;

[0111] The subtype recognition module is used to perform cluster analysis on the data of HFpEF patients using the K-prototypes algorithm to identify different HFpEF subtypes;

[0112] The input end of the MAE prediction module is connected to the output end of the subtype recognition module, and the MAE prediction module is used to output the MAE prediction result.

[0113] This system utilizes larger and more diverse real-world data, combines unsupervised machine learning clustering methods to find more robust HFpEF subtypes, and uses machine learning methods to predict the probability of major adverse events (MAEs) in each subtype.

[0114] In a preferred embodiment of the present invention, as Figure 2 shown, the HFpEF subtype recognition and MAE prediction system further includes an auxiliary training device, which includes a support base 1, a human-computer interaction module 2, a variety of sports equipment, a mobile sensor, an acceleration sensor, a blood pressure sensor, a pulse sensor, a temperature sensor, and a wearable.

[0115] The human-computer interaction module 2 is connected to the support base 1 through a bracket 10. The bracket 10 can be a metal bracket 10, and structures such as a protective housing can be provided on the outside. The bottom of the bracket 10 is welded or riveted to the support base 1, and the human-computer interaction module 2 is welded or riveted to the top or the body of the bracket 10. Preferably, the human-computer interaction module 2 can be an intelligent computer, a touch screen, etc.

[0116] The human-computer interaction module 2 is electrically connected to the subtype recognition module and the MAE prediction module. The human-computer interaction module 2 is provided with an ID tag recognition module (such as an RFID tag reader, etc.) for ID tag recognition.

[0117] A treadmill 8 in various sports equipment (such as barbells, dumbbells, resistance bands, fitness balls, etc.) is arranged (such as directly placed, or riveted, etc.) on a support base 1. A vertical threaded rod 3 is fixedly connected (such as welded, riveted, etc.) to the side wall of the support base 1. A plurality of placement boxes 4 are provided on the threaded rod 3. The placement boxes 4 are provided with threaded holes, and the placement boxes 4 are threadedly connected to the threaded rod 3 through the threaded holes. Sports equipment can be respectively placed in the placement boxes 4. An electric valve is provided on one side of the placement box 4. A control button can be installed on the outer wall of the placement box 4. The output end of the control button is electrically connected to the control end of the electric valve, so as to manually press the control button to control the opening and closing of the electric valve.

[0118] A motion sensor, an acceleration sensor, a blood pressure sensor, a pulse sensor, and a temperature sensor are all arranged (integrated and installed) on a wearable. The wearable (such as a watch, a smart belt, a smart vest, etc.) is worn on the patient. An ID tag (such as an RFID tag, etc.) is provided on the wearable. The human-computer interaction module 2 identifies the patient's ID tag and receives the corresponding HFpEF subtype identification and MAE prediction information. The auxiliary training item control signal output end of the human-computer interaction module 2 is electrically connected to the control end of the electric valve.

[0119] The output ends of the motion sensor, the acceleration sensor, the blood pressure sensor, the pulse sensor, and the temperature sensor are all electrically connected to the human-computer interaction module 2, the patient terminal, and the medical staff terminal.

[0120] The human-computer interaction module 2 is provided to facilitate data interaction and can also display corresponding data. When the patient is training, corresponding training guidance videos can also be provided. The treadmill 8 is arranged on the support base 1. When the patient needs to walk slowly or quickly for training, he can do it on the treadmill 8.

[0121] The threaded rod 3 is arranged on the support base 1, and the placement box 4 is arranged on the threaded rod 3, which is beneficial for storing various sports equipment and can be used at any time. Moreover, the placement box 4 is threadedly connected to the threaded rod 3, and the placement height of the placement box 4 can be arbitrarily adjusted by rotating the placement box 4, which is convenient for taking and placing. It can also be used as a seat for the patient to rest temporarily or an armrest for the patient during exercise.

[0122] Using a variety of sensors to collect the patient's body data in real time is beneficial for monitoring the patient's physical condition and the safety during training is higher.

[0123] In a preferred solution of the present invention, as Figure 3 shown, the auxiliary training device further includes an inner rod 5, a guardrail 6, a camera 7, and an inflation and deflation mechanism.

[0124] On both sides of the support base 1, there are at least two threaded rods 3 (it can also be four or more). The top of the threaded rod 3 is open and hollow inside. The bottom of the inner rod 5 extends into the threaded rod 3 from the top of the threaded rod 3, and the outer wall of the inner rod 5 is in sealed sliding connection with the inner wall of the threaded rod 3. The inner bottom of the threaded rod 3 is connected to the air charging and discharging mechanism through an air pipe. The air charging and discharging mechanism can adopt a two-way fan, etc., and can be placed outside the support base 1 or inside the support base 1.

[0125] The guardrail 6 is annular. The guardrail 6 is horizontally placed and installed on the top of the inner rod 5. On the side of the guardrail 6 facing the patient, there is an annular track. The camera 7 is installed on the mobile trolley through a universal cloud platform, and the mobile trolley moves along the annular track.

[0126] In this way, when the air charging and discharging mechanism inflates or deflates the threaded rod 3, the inner rod 5 can be controlled to move up or down along the inner wall of the threaded rod 3, adjusting the height of the guardrail 6, and flexibly adjusting for patients of different heights to use. At the same time, the height of the camera 7 on the guardrail 6 changes, collecting image information of patients at different heights, which is conducive to comprehensive analysis. The camera 7 can collect the whole body image of the patient during training by moving along the annular track with the mobile trolley, or can also move to the required position to collect image information at a specific angle.

[0127] In a preferred solution of the present invention, a plurality of air bags 9 are arranged inside the guardrail 6. The air bags 9 are connected to the air charging and discharging mechanism, and after being inflated, the air bags 9 extend towards the side where the patient is located. One side of the air bag 9 can be bonded to the top or bottom of the guardrail 6, without affecting the movement of the mobile trolley along the annular track.

[0128] After the air bag 9 is inflated, it is closer to the patient. Without affecting the movement of the patient, the material of the air bag 9 is elastic and soft, which can protect the patient. The air bag 9 can surround the patient on all sides, playing roles such as preventing falls and preventing bumps.

[0129] In the description of this specification, the description with reference to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples.

[0130] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and purposes of the present invention. The scope of the present invention is defined by the claims and their equivalents.

Claims

1. A method for identifying HFpEF subtypes and predicting MAE based on machine learning, characterized in that It includes the following steps: Collect the electronic health records and echocardiogram data of HFpEF patients from multiple medical centers and perform preprocessing; Use the K-prototypes algorithm to perform cluster analysis on the data of HFpEF patients to identify different HFpEF subtypes; Based on the identified subtypes, use a machine learning algorithm to construct an MAE prediction model and perform interpretability analysis using SHAP, and output the MAE prediction results.

2. The method for HFpEF subtype identification and MAE prediction based on machine learning according to claim 1, wherein Determine the optimal number of clusters through the elbow method, silhouette coefficient, and gap statistic, and perform internal and external validation on the clustering results. Specifically: Divide the data of HFpEF patients into a derivation cohort, validation cohort 1, and validation cohort 2. Apply k-prototype clustering to the derivation cohort to generate subtypes, and perform external validation on validation cohorts 1 and 2 to obtain different survival curve analyses to verify the prognostic differences between subtypes; Apply the clustering algorithm based on partitioning around medoids to the derivation cohort to identify and ensure the robustness of the results.

3. The method for HFpEF subtype recognition and MAE prediction based on machine learning according to claim 1, characterized in that, The MAE prediction model is a random forest or a support vector machine.

4. The method for HFpEF subtype identification and MAE prediction based on machine learning according to claim 1, wherein Based on the MAE prediction results, judge the risk level of the patient: Low risk level: MAE prediction result ≤ 0.5; Medium risk level: MAE prediction result is between 0.5 and 0.7; High risk level: MAE prediction result > 0.

7.

5. The method for HFpEF subtype identification and MAE prediction based on machine learning according to claim 4, wherein Based on the risk level of the patient and the HFpEF subtype results, perform auxiliary training: Based on different risk levels, preset corresponding physical preset values, collect the patient's physical data, and compare it with the physical preset values corresponding to the risk levels to judge whether the patient is currently physically abnormal; If the patient's physical data is abnormal, send alarm data to the patient terminal and the medical staff terminal. Conversely, if the patient's physical data is normal, perform an auxiliary training plan. Specifically: According to the risk level of the patient, initially screen sports items; Based on the HFpEF subtype of the patient, determine the patient's exercise tolerance and limiting factors, secondarily screen sports items, and assign weights to the intensity, duration, and frequency parameters of the secondarily screened sports items to calculate the training scores of each sports item. Sort the training scores of each sports item and select the sports item with the highest training score as the final auxiliary training item.

6. The method for HFpEF subtype identification and MAE prediction based on machine learning according to claim 4, characterized in that, When the patient performs the final auxiliary training item, collect the patient's body movement images in real time, extract the patient's body joint angles and movement trajectories, calculate the similarity with the preset images, judge whether the patient's real-time training meets the standard. If it does not meet the standard, output a prompt signal. After one training session, calculate whether the overall training of this training session meets the standard; Taking one week as a cycle, record the number of training sessions of the patient in one week. If the number of training sessions meets the preset number and the overall training pass rate of the patient in one week is above 80%, then add 1 to the preset number and add 10 minutes to the duration of the sports item; otherwise, output an alarm signal to the patient terminal and the medical staff terminal; Similarity calculation method: For the two-section joint connection line and respectively represent the vectors from joint A to joint B and from joint B to joint C. The joint angle θ is: wherein, is the dot product of two vectors, and are the magnitudes of two vectors; Let the joint angle vectors of the patient and the preset image be θ a and θ r respectively, and the similarity between them is calculated by cosine similarity: where, θ a ·θ r is the dot product of two vectors, |θ a | and |θ r | are the magnitudes of the two vectors; Suppose the running trajectory is obtained through the transformation of joint positions between consecutive frames, represented by the two-dimensional or three-dimensional coordinates of each joint. If the patient's body makes a movement in the image sequence, the joint trajectory is the sequence of joint positions at each moment; If the trajectories of the patient and the preset image and are composed of a series of points, the Euclidean distance is used to measure the difference D between them: where k is the frame index in the trajectory, represents the Euclidean distance of the vectors; The final similarity S combines the angle similarity and the trajectory-based similarity through weighted average: Among them, w1 and w2 are weights, and w1 + w2 = 1.

7. An HFpEF subtype identification and MAE prediction system based on the method according to any one of claims 1-6, characterized in that, It includes a data collection module, a subtype identification module, and an MAE prediction module; The data acquisition module is used to collect the electronic health records and echocardiogram data of HFpEF patients from multiple medical centers and transmit them to the subtype identification module; The subtype identification module is used to perform clustering analysis on the data of HFpEF patients using the K-prototypes algorithm to identify different HFpEF subtypes; The input end of the MAE prediction module is connected to the output end of the subtype identification module, and the MAE prediction module is used to output the MAE prediction result.

8. The HFpEF subtype recognition and MAE prediction system according to claim 7, wherein It further includes an auxiliary training device, which includes a support base, a human-computer interaction module, various sports equipment, a mobile sensor, an acceleration sensor, a blood pressure sensor, a pulse sensor, a temperature sensor, and a wearable; The human-computer interaction module is connected to the support base through a bracket, the human-computer interaction module is connected to the subtype identification module and the MAE prediction module, and an ID tag recognition module is provided in the human-computer interaction module; The treadmill in the various sports equipment is arranged on the support base. A vertical threaded rod is connected to the side wall of the support base. A plurality of placement boxes are provided on the threaded rod. The placement box is provided with a threaded hole, and the placement box is threadedly connected to the threaded rod through the threaded hole. The sports equipment can be respectively placed in the placement box, and an electric valve is provided on one side of the placement box; The mobile sensor, acceleration sensor, blood pressure sensor, pulse sensor, and temperature sensor are all arranged on the wearable. The wearable is worn on the patient. An ID tag is provided on the wearable. The human-computer interaction module recognizes the patient's ID tag, receives the corresponding HFpEF subtype identification and MAE prediction information, and the auxiliary training item control signal output end of the human-computer interaction module is connected to the control end of the electric valve; The output ends of the mobile sensor, acceleration sensor, blood pressure sensor, pulse sensor, and temperature sensor are all connected to the human-computer interaction module, the patient terminal, and the medical staff terminal.

9. The HFpEF subtype recognition and MAE prediction system according to claim 8, characterized in that, The auxiliary training device further includes an inner rod, a guardrail, a camera, and an inflation and deflation mechanism; At least two threaded rods are provided on both sides of the support base. The top of the threaded rod is open and hollow inside. The bottom of the inner rod extends into the threaded rod from the top of the threaded rod, and the outer wall of the inner rod is hermetically slidably connected to the inner wall of the threaded rod. The inner bottom of the threaded rod is communicated with the inflation and deflation mechanism through an air pipe; The guardrail is annular, the guardrail is horizontally placed and installed on the top of the inner rod. An annular track is provided on the side of the guardrail facing the patient. The camera is installed on a mobile trolley through a universal gimbal, and the mobile trolley moves along the annular track.

10. The HFpEF subtype identification and MAE prediction system according to claim 9, characterized in that, A plurality of air bags are arranged inside the guardrail. The air bags are communicated with the inflation and deflation mechanism, and extend toward the patient side after being inflated.

Images