Heart failure patient home care intelligent monitoring method combined with deep learning

Abstract

The invention relates to the field of deep learning, and discloses a heart failure patient home care intelligent monitoring method combined with deep learning, which comprises the following steps: continuously acquiring the heart rate, blood pressure, breathing and body movement states of a patient in a home environment; performing conjoint analysis of time and activity states on the data streams after priority sorting, and forming a dynamic health feature set in combination with historical health records of patients and environment variable information; comprehensively evaluating the dynamic health feature set by using a deep learning model, and identifying a potential abnormal mode of short-time change through the deep learning model; marking and grouping the physiological parameters and related behavior scenes corresponding to the candidate abnormal events, locally evaluating the current health state of the patient based on grouping information, and identifying a potential high-risk condition; and according to the local evaluation result and the overall health trend of the patient, generating real-time early warning and personalized nursing suggestions. The method has the advantage of improving the identification capability of short-time abnormity.

Description

Technical Field

[0001] This invention relates to the field of deep learning, specifically to an intelligent monitoring method for home care of heart failure patients that incorporates deep learning. Background Technology

[0002] With the increasing number of heart failure patients each year, home care has become an important auxiliary means in addition to clinical intervention. In existing technologies, most home monitoring systems collect basic parameters such as heart rate, blood pressure, and weight through wearable devices and issue alarms based on preset thresholds. However, monitoring blind spots still exist in specific scenarios. For example, patients often experience transient shortness of breath or fluctuations in blood oxygen levels when taking medication and engaging in light walking or getting up. Since existing systems primarily rely on timed static sampling, they cannot continuously capture such transient abnormalities, easily leading to the neglect of early deterioration signals and increasing the risk of sudden events. To address this issue, deep learning demonstrates advantages in temporal pattern recognition and multimodal data fusion, enabling dynamic modeling of body movement data, respiratory rhythm, and heart rate changes, thereby compensating for the shortcomings of traditional monitoring methods in specific activity scenarios. Therefore, it is essential to design an intelligent home care monitoring method for heart failure patients that incorporates deep learning to improve the ability to identify short-term abnormalities. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides an intelligent monitoring method for home care of heart failure patients that incorporates deep learning. This method has the advantage of improving the ability to identify short-term abnormalities and solves the problems mentioned in the background technology.

[0004] To achieve the aforementioned goal of improving the ability to identify short-term abnormalities, this invention provides the following technical solution: a smart monitoring method for home care of heart failure patients combining deep learning, comprising the following steps: The heart rate, blood pressure, respiration and body movement of patients in their home environment were continuously acquired, and the data segments were prioritized according to the sensitivity of symptoms. The data streams after priority ranking are subjected to joint analysis of time and activity status, and combined with the patient's historical health records and environmental variable information to form a dynamic set of health characteristics; A deep learning model is used to comprehensively evaluate a dynamic set of health features, identify potential abnormal patterns that change in the short term, and generate a list of candidate abnormal events. The physiological parameters and related behavioral scenarios corresponding to candidate abnormal events are labeled and grouped, and the current health status of patients is locally assessed based on the grouping information to identify potential high-risk situations. Based on the local assessment results and the patient's overall health trend, the signal acquisition frequency, model processing priority, and home care intervention strategies are adjusted to generate real-time early warnings and personalized care suggestions.

[0005] Preferably, the process of prioritizing data segments according to symptom sensitivity is as follows: Heart rate, blood pressure, respiratory rate and body movement amplitude are collected synchronously by wearable devices and environmental monitoring terminals to form a multi-source data stream; A sensitivity matching algorithm based on a clinical case knowledge base was used to calculate the similarity index between different data fragments and typical abnormal patterns of heart failure. By combining the patient's individualized risk assessment form, corresponding sensitivity weight factors are generated, and various physiological data fragments are weighted and sorted, outputting data fragments arranged by priority.

[0006] Preferably, the process of performing joint analysis of time and activity status on the prioritized data stream is as follows: Synchronize high-priority data segments with patients' daily activity records over time; An event-driven time segmentation mechanism is adopted to fuse multi-source data fragments within the same activity segment; External environmental variables are introduced during the fusion process and participate in the calculation as correction factors, outputting joint analysis results with dual attributes of time dimension and activity scenario dimension.

[0007] Preferably, the process of forming a dynamic set of health characteristics is as follows: Heart rate variability, blood pressure fluctuation amplitude, respiratory cycle stability, and body movement coordination were extracted from the joint analysis results. Model the correlation between indicators and construct a feature vector group that reflects multidimensional physiological state; A data integrity correction mechanism is used to interpolate and repair missing segments and remove abnormal noise points to obtain a dynamically updated set of health characteristics that covers multiple physiological and environmental factors.

[0008] Preferably, the process of using a deep learning model to comprehensively evaluate a dynamic set of health features is as follows: The set of health features is input into a deep learning model composed of convolutional units and recurrent units; Priority labels are introduced during the model encoding stage to dynamically adjust the weights of different features in training and inference. The intermediate hidden layer states of the model output are monitored to extract short-term sensitive change features and output a set of candidate risk distribution probabilities.

[0009] Preferably, the process of generating a list of candidate exception events is as follows: The risk distribution probability output by the comprehensive assessment is compared with the patient's individual health threshold database; If the probability of risk in a consecutive period exceeds a set threshold, the consecutive period will be marked as a potential anomaly. An event merging algorithm is used to integrate adjacent time periods to generate a list of candidate abnormal events that includes event timestamps, ranges of physiological parameter changes, and scene descriptions.

[0010] Preferably, the process of labeling and grouping the physiological parameters and related behavioral scenarios corresponding to candidate abnormal events is as follows: The heart rate, blood pressure, respiration, and body movement indicators involved in the candidate abnormal event list are labeled. Group abnormal events under different tags according to similar behavioral scenarios; During the grouping process, the duration, frequency, and severity of abnormal events are recorded, and a set of candidate abnormal events with grouping labels and contextual information is output.

[0011] Preferably, the process of performing a local assessment of the patient's current health status based on grouping information is as follows: For each abnormal event group, the fluctuation range of physiological parameters and the degree of difference from the normal reference value are calculated; By combining the patient's historical health records, we can analyze whether the grouped events are part of a continuous deterioration trend; The local assessment model outputs a risk level reflecting the urgency, generating a health risk profile for the patient in the current time period.

[0012] Preferably, the process for generating real-time alerts and personalized care recommendations is as follows: The risk level obtained from the local assessment is compared with the patient's overall health trend to determine whether an early warning needs to be triggered. If the risk level exceeds the set threshold, a warning message will be pushed to the patient's terminal and the medical care platform in real time, and personalized care suggestions will be generated in conjunction with the patient's daily care plan.

[0013] Compared with existing technologies, this invention provides an intelligent monitoring method for home care of heart failure patients that combines deep learning, which has the following beneficial effects: This invention continuously collects multi-source physiological and behavioral data from heart failure patients in their home environment and prioritizes these data based on symptom sensitivity, enabling the system to focus on potentially high-risk indicators and thus improve the ability to identify short-term abnormalities. Through joint analysis of time and activity status, combined with the patient's historical health records and environmental variables, a dynamic health feature set can be generated, comprehensively reflecting the subtle fluctuations and dynamic trends in the patient's physiological state in daily life. A deep learning model is used to comprehensively evaluate this feature set, which can not only identify short-term, sudden abnormal patterns but also quantify the probability and risk level of abnormalities, providing a scientific basis for generating candidate abnormal events. By labeling and grouping candidate abnormal events according to physiological parameters and behavioral scenarios, and conducting local health status assessments based on the grouping information, potential high-risk situations in patients can be identified, distinguishing between occasional fluctuations and continuous deterioration trends, providing precise references for personalized intervention. Based on the local assessment results and overall health trends, the signal acquisition frequency and model processing priority are dynamically adjusted, generating real-time warnings and personalized nursing suggestions, achieving intelligent, refined, and proactive management of home care. It can significantly improve the speed and accuracy of response to short-term anomalies, reduce the risks caused by delayed intervention, reduce the pressure on medical resources, improve the safety of patients' home life and the efficiency of health management, and has good value for promotion and application. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the method of the present invention; Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] Example 1: Please refer to Figure 1 As shown in the embodiment of the present invention, a smart monitoring method for home care of heart failure patients combining deep learning includes the following steps: S1: Continuously acquire the patient's heart rate, blood pressure, respiration and body movement status in the home environment, and prioritize the data segments according to the sensitivity of symptoms.

[0017] The process of prioritizing data segments according to symptom sensitivity in S1 is as follows: Wearable devices and environmental monitoring terminals simultaneously collect heart rate, blood pressure, respiratory rate, and body movement amplitude to form a multi-source data stream. During home care, patients wear smart bracelets equipped with ECG signals, pulse waves, and motion sensing modules. At the same time, an environmental monitoring terminal with temperature, humidity, and air quality detection functions is deployed indoors. The wearable devices can record the patient's heart rate change curve, blood pressure fluctuations, respiratory rate, and body movement amplitude in real time at a sampling frequency of seconds, and upload them to the data processing platform wirelessly. The environmental monitoring terminal simultaneously records parameters such as indoor temperature, humidity, and air pollution index. The multi-source data streams are timestamped to form multi-source continuous data that can be analyzed. A sensitivity matching algorithm based on a clinical case knowledge base is used to calculate the similarity index between different data segments and typical abnormal patterns of heart failure. A case knowledge base constructed from a large amount of clinical follow-up data of heart failure patients is pre-stored. The knowledge base contains physiological data patterns of heart failure patients when they experience typical symptoms such as dyspnea, nocturnal apnea, and a sharp increase in heart rate. After the currently collected multi-source data stream is divided into fixed time segments, the sensitivity matching algorithm is called to compare the degree of matching between each data segment and the typical abnormal patterns in the knowledge base. For example, when the heart rate fluctuates beyond a certain reference range in a short period of time, accompanied by irregular changes in respiratory rate, the similarity between this segment and the acute exacerbation heart failure pattern will be marked as high. Through this matching process, a similarity index corresponding to each data segment can be generated. Based on the patient's individualized risk assessment form, corresponding sensitivity weighting factors are generated, and various physiological data segments are weighted and ranked, outputting data segments arranged by priority. An individualized risk assessment form is established upon patient enrollment, assessing factors including age, medical history, baseline blood pressure, left ventricular ejection fraction, renal function, and previous hospitalization for acute heart failure. These factors are converted into a baseline risk score using a pre-defined quantitative model. The risk score is then interactively calculated with the aforementioned similarity indicators to generate individualized sensitivity weighting factors. For example, elderly patients with multiple hospitalizations for acute heart failure will receive higher weights for their blood pressure abnormality segments during sensitivity weighting, enhancing the ability to prioritize the identification of this type of risk. Based on the generated sensitivity weighting factors, all physiological data segments are ranked. The ranking process considers not only the abnormal magnitude of a single parameter but also the interaction between multiple parameters. For instance, segments with only increased heart rate without accompanying increased respiratory rate receive a lower ranking, while segments with both heart rate and respiratory rate significantly deviating from the normal range are assigned higher priority. Through this weighted sorting, the system can prioritize the segments that truly reflect acute deterioration of heart failure. After sorting, a sequence of data segments with priority labels will be generated and stored in the monitoring platform's cache. Each segment in the data segment sequence contains a timestamp, the corresponding original data content, a similarity index with the clinical knowledge base, and a sensitivity weight value.

[0018] S2: Perform joint analysis of time and activity status on the prioritized data stream, and combine patient historical health records and environmental variable information to form a dynamic health feature set.

[0019] The process of performing joint analysis of time and activity status on the priority-sorted data stream in S2 is as follows: High-priority data segments are synchronized with patients' daily activity records. While collecting patients' physiological signals, daily activities such as resting, eating, walking, going up and down stairs, or taking medication are recorded through smart bracelets, bedside monitors, or mobile applications. The activity records have precise timestamp information. High-priority physiological data segments are filtered and matched with activity records of the corresponding time period by timestamp alignment. This ensures that each abnormal data segment has a clear behavioral background, thereby avoiding misjudging short-term fluctuations caused by normal activities as pathological abnormalities. The system employs an event-driven time segmentation mechanism to fuse multi-source data fragments within the same activity segment. Instead of relying solely on fixed time intervals to divide data, the system uses activity events as the basis for segmentation, such as the start and end of eating, the duration of walking, or 30 minutes after taking medication. Within each activity segment, the system fuses multi-source data fragments such as heart rate, blood pressure, respiratory rate, and body movement amplitude to form a comprehensive data unit with activity context. This event-driven segmentation mechanism can better reflect the relationship between physiological changes and specific behaviors. External environmental variables are introduced during the fusion process and used as correction factors in the calculation, resulting in a joint analysis with dual attributes of time and activity scenario. External environmental variables, including parameters such as indoor temperature, humidity, air quality, and noise level, are also introduced during data fusion and used as correction factors in the calculation. For example, in a high-temperature environment, mild physical movement by a patient can lead to increased heart rate fluctuations. Such situations are corrected by environmental factors to avoid being misjudged as abnormal. The output joint analysis results not only contain continuous characteristics in the time dimension but also annotations in the activity scenario dimension, thus generating a data structure with dual attributes.

[0020] The process of forming a dynamic health feature set in S2 is as follows: The combined analysis results extract heart rate variability, blood pressure fluctuation amplitude, respiratory cycle stability, and body movement coordination. After obtaining the combined analysis results of time and activity status, short-term and long-term statistical processing is performed on the heart rate signal to extract heart rate variability indicators reflecting autonomic nervous function. The difference and fluctuation range between adjacent measurement points of blood pressure data are calculated to form the blood pressure fluctuation amplitude parameter. For respiratory data, respiratory cycle stability indicators are extracted by detecting whether there is irregular prolongation or shortening of the respiratory cycle. For body movement data, the continuity of acceleration and posture changes is analyzed to obtain body movement coordination indicators. In this way, the extracted indicators can comprehensively reflect the patient's physiological dynamic characteristics. The correlation between indicators is modeled to construct a feature vector group that reflects multidimensional physiological states. By calculating the correlation between different indicators, the multidimensional coupling state is reflected. For example, there is a certain synchronicity between heart rate variability and respiratory cycle stability, and there is a causal relationship between blood pressure fluctuation and body movement coordination. Using the correlation between these indicators, multidimensional feature vectors are generated, and vectors from different time periods are combined to form a feature vector group to comprehensively describe the patient's physiological state in a specific scenario. In this way, the constructed feature set not only includes a single indicator, but also reflects the interaction between physiological signals, thereby improving the ability to characterize potential health risks. A data integrity correction mechanism is employed to interpolate and repair missing segments and remove abnormal noise points, resulting in a dynamically updated set of health features that encompasses multiple physiological and environmental factors. During data acquisition, network interruptions, poor sensor contact, or environmental interference may lead to some data loss or anomalies. Time series interpolation methods are used to reasonably repair short-term missing segments to maintain data continuity. Abnormal noise points that significantly deviate from the physiologically reasonable range are removed using threshold detection and statistical discrimination methods. After correction and cleaning, multi-source features such as heart rate, blood pressure, respiration, body movement, and environmental variables are integrated to form a dynamically updated set of health features that can be expanded and corrected in real time as the patient's daily condition changes.

[0021] S3: Utilize deep learning models to comprehensively evaluate dynamic health feature sets, identify potential abnormal patterns of short-term changes through deep learning models, and generate a list of candidate abnormal events.

[0022] The process of comprehensively evaluating the dynamic health feature set using a deep learning model in S3 is as follows: The health feature set is input into a deep learning model composed of convolutional and recurrent units. The health feature set, after dynamic updating and integrity correction, is fed into the deep learning model containing convolutional neural network units and recurrent neural network units. The convolutional units are mainly used to capture local change patterns in the feature set, such as the fluctuation features of heart rate or blood pressure in a very short time. The recurrent units are used to identify long-term dependencies in the time series, ensuring that the model can remember the continuous evolution features of patients in different activity scenarios and physiological states. Through the combination of convolutional and recurrent units, the model can simultaneously take into account short-term sensitivity and long-term trends, thereby achieving a multi-level representation of the patient's health features. Priority labels are introduced during the model encoding stage to dynamically adjust the weights of different features in training and inference. When encoding the input feature set, the priority labels generated in the previous steps are also introduced as control signals and embedded into the weight adjustment mechanism of the deep learning model. In this way, features with higher priority (such as sudden increases in heart rate, rapid fluctuations in blood pressure, etc.) will receive higher computational weights during training and inference, thus having a greater impact on the model results; while features with lower priority (such as daily minor physical movements or stable environmental data) are weakened to reduce noise interference. This dynamic weight adjustment mechanism ensures that the model can highlight key features when processing complex data, thereby improving its sensitivity to potential health risks. The intermediate hidden layer states of the model output are monitored to extract short-term sensitive change features and output a set of candidate risk distribution probabilities. In the inference process of the deep learning model, it does not only rely on the final classification or regression output, but also continuously monitors the state of the intermediate hidden layer. The hidden layer state can often more directly reflect the small fluctuations of the input features in the time series, such as a sudden decrease in heart rate variability, a brief disorder of the respiratory cycle, or abnormal cessation of body movement. By dynamically extracting these intermediate states, short-term sensitive changes can be detected at an early stage, and these changes can be mapped to a set of candidate risk distribution probabilities. The candidate risk distribution probabilities not only identify the possible types of health abnormalities, but also quantify the probability of their occurrence.

[0023] The process of generating the candidate abnormal event list in S3 is as follows: The risk distribution probability output by the comprehensive assessment is compared with the patient's individual health threshold database. The candidate risk distribution probability output by the deep learning model is compared one by one with the individualized health threshold database established for each patient at the time of enrollment. The threshold database includes the patient's normal heart rate range, upper and lower limits of blood pressure, respiratory cycle fluctuation range, and baseline values ​​of body movement level. Through comparison, the risk probability of each time segment is mapped to the corresponding physiological parameter threshold. If the probability is in the high-risk range, the time segment is initially marked as potentially abnormal, ensuring that the determination of candidate events takes into account both the model prediction results and the patient's individual physiological characteristics, thereby improving the accuracy of abnormality identification. If the risk probability of a continuous period exceeds a set threshold, the continuous period is marked as a potential abnormality. Continuity analysis is performed on high-risk probability segments, that is, the intervals in which the risk probability exceeds the preset threshold at multiple consecutive sampling time points are identified. If a certain continuous period meets the condition, the entire interval is marked as a potential abnormal event in order to distinguish between occasional short-term fluctuations and real health abnormalities. The marking not only records the start and end times of the event, but also retains information on the changes in various physiological parameters within the continuous period. The event merging algorithm integrates adjacent time periods to generate a list of candidate abnormal events containing event timestamps, ranges of physiological parameter changes, and scene descriptions. Adjacent or overlapping potential abnormal time periods are merged. The event merging algorithm integrates events with similar time and nature into a single candidate abnormal event. Each candidate abnormal event contains the start and end timestamps of the event, the minimum and maximum values ​​of physiological parameters within the interval, and the fluctuation range. At the same time, it combines the patient's activity scene records (such as rest, eating, or mild exercise) and environmental conditions (such as room temperature and noise level) to form a complete event description.

[0024] S4: Mark and group the physiological parameters and related behavioral scenarios corresponding to candidate abnormal events, and conduct a local assessment of the patient's current health status based on the grouping information to identify potential high-risk situations.

[0025] The process of marking and grouping the physiological parameters and related behavioral scenarios corresponding to candidate abnormal events in S4 is as follows: The heart rate, blood pressure, respiration, and body movement indicators involved in the candidate abnormal event list are labeled. Each event in the candidate abnormal event list is subjected to feature analysis, and physiological parameters such as heart rate, blood pressure, respiratory rate, and body movement amplitude are labeled according to the abnormality type. For example, rapid increase in heart rate, excessive short-term fluctuation in blood pressure, irregular respiratory cycle, or abnormal pause in body movement are all marked as corresponding physiological abnormality labels. Each label not only describes the abnormality type but also adds an abnormality level, such as mild, moderate, or severe. Through labeling, the physiological characteristics of the event can be clearly classified. Abnormal events under different labels are grouped according to similar behavioral scenarios. Using behavioral scenario information recorded by candidate abnormal events, such as rest, eating, light exercise, and post-medication status, abnormal events with similar behavioral backgrounds are grouped into the same group. For events in the same group, they are ensured to have a high degree of consistency in activity type, time period, or environmental conditions in order to analyze the relationship between abnormal events and specific behavioral scenarios. Through this grouping method, short-term abnormal patterns related to specific daily behaviors can be identified. During the grouping process, the duration, frequency, and severity of abnormal events are recorded, and a candidate abnormal event set with grouping labels and contextual information is output. Within each group, all events are statistically analyzed, and the duration, frequency of occurrence within the observation period, and corresponding severity level of each event are recorded. The behavioral scenario, environmental state, and time information corresponding to each event are retained, forming a candidate abnormal event set with rich context. The output event set not only shows the physiological characteristics of the events but also provides behavioral and environmental background.

[0026] The process of performing a local assessment of the patient's current health status based on grouping information in S4 is as follows: For each abnormal event group, the fluctuation range of physiological parameters and their difference from normal reference values ​​are calculated; statistical analysis is performed on heart rate, blood pressure, respiratory rate and body movement amplitude within each abnormal event group to calculate their maximum, minimum and average fluctuation range during the event. These fluctuation values ​​are compared with the patient's individualized normal reference values ​​to quantify the degree to which each physiological parameter deviates from the normal range. Through this difference calculation, the potential impact of each event group on the patient's physiological state can be identified. By combining the patient's historical health records, we can analyze whether the grouped events are part of a continuous deterioration trend. We can also compare the current abnormal event grouping with the patient's previous health records longitudinally, including the long-term trends of heart rate, blood pressure, respiration and body movement. By observing the consistency between the abnormal event grouping and the historical trend, we can determine whether the grouping reflects the continuous deterioration or periodic fluctuation of the patient's health status. If the event grouping shows a deterioration pattern that is consistent with the historical trend, it can indicate that the patient is in a high-risk stage. The local assessment model outputs a risk level reflecting the urgency, generating a health risk profile for the patient in the current time period. The fluctuation range, deviation, and matching results with historical trends of each event group are input into the local assessment model. The model generates a quantitative risk level, such as low, medium, high, or urgent, through comprehensive calculation. The assessment results of all event groups are summarized and combined with time period and behavioral scenario information to form a health risk profile for the patient in the current time period. This profile not only shows the abnormalities of various physiological indicators but also reflects the distribution of risk in different activity scenarios.

[0027] S5: Based on the local assessment results and the patient's overall health trend, adjust the signal acquisition frequency, model processing priority, and home care intervention strategies to generate real-time early warnings and personalized care suggestions.

[0028] The process of generating real-time early warnings and personalized care suggestions in S5 is as follows: The risk level obtained from the local assessment is compared with the patient's overall health trend to determine whether an early warning needs to be triggered. The risk level of each event group output by the local assessment model is compared longitudinally with the patient's long-term health trend, including the trends of changes in heart rate, blood pressure, respiration, and body movement. If the abnormal amplitude shown by the local risk level is significantly higher than the patient's historical baseline level, or deviates significantly from the overall health trend, the state will be marked as potentially having an urgent health risk, thereby determining whether a real-time early warning needs to be triggered. This comparison mechanism ensures that the early warning not only relies on instantaneous abnormalities, but also takes into account the patient's long-term health background, improving the accuracy of the warning. If the risk level exceeds the set threshold, a real-time warning message is pushed to the patient's terminal and the medical care platform. Combined with the patient's daily care plan, personalized care suggestions are generated. When the local risk level exceeds the preset threshold, the system will immediately generate real-time warning information, including abnormal physiological parameters, risk level, time of event occurrence and related behavioral scenarios. This information is pushed to the patient through a mobile application, wearable device or home care terminal, and simultaneously uploaded to the medical management platform or the medical staff's terminal. Based on the patient's daily care plan and individualized health record, further analysis and intervention measures are taken, such as adjusting rest arrangements, reminding medication, guiding light exercise or arranging medical staff follow-up visits. This generates targeted personalized care suggestions, realizing closed-loop management from real-time monitoring to intervention. It can proactively alert risks and provide actionable care plans in the early stages.

[0029] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0030] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent monitoring of home care for heart failure patients combining deep learning, characterized in that, Includes the following steps: The heart rate, blood pressure, respiration and body movement of patients in their home environment were continuously acquired, and the data segments were prioritized according to the sensitivity of symptoms. The data streams after priority ranking are subjected to joint analysis of time and activity status, and combined with the patient's historical health records and environmental variable information to form a dynamic set of health characteristics; A deep learning model is used to comprehensively evaluate a dynamic set of health features, identify potential abnormal patterns that change in the short term, and generate a list of candidate abnormal events. The physiological parameters and related behavioral scenarios corresponding to candidate abnormal events are labeled and grouped, and the current health status of patients is locally assessed based on the grouping information to identify potential high-risk situations. Based on the local assessment results and the patient's overall health trend, the signal acquisition frequency, model processing priority, and home care intervention strategies are adjusted to generate real-time early warnings and personalized care suggestions.

2. The intelligent monitoring method for home care of heart failure patients combining deep learning according to claim 1, characterized in that, The process of prioritizing data segments according to their symptom sensitivity is as follows: Heart rate, blood pressure, respiratory rate and body movement amplitude are collected synchronously by wearable devices and environmental monitoring terminals to form a multi-source data stream; A sensitivity matching algorithm based on a clinical case knowledge base was used to calculate the similarity index between different data fragments and typical abnormal patterns of heart failure. By combining the patient's individualized risk assessment form, corresponding sensitivity weight factors are generated, and various physiological data fragments are weighted and sorted, outputting data fragments arranged by priority.

3. The intelligent monitoring method for home care of heart failure patients combining deep learning according to claim 2, characterized in that, The process of performing joint analysis of time and activity status on the prioritized data stream is as follows: Synchronize high-priority data segments with patients' daily activity records over time; An event-driven time segmentation mechanism is adopted to fuse multi-source data fragments within the same activity segment; External environmental variables are introduced during the fusion process and participate in the calculation as correction factors, outputting joint analysis results with dual attributes of time dimension and activity scenario dimension.

4. The intelligent monitoring method for home care of heart failure patients combining deep learning according to claim 3, characterized in that, The process of forming a dynamic set of health characteristics is as follows: Heart rate variability, blood pressure fluctuation amplitude, respiratory cycle stability, and body movement coordination were extracted from the joint analysis results. Model the correlation between indicators and construct a feature vector group that reflects multidimensional physiological state; A data integrity correction mechanism is used to interpolate and repair missing segments and remove abnormal noise points to obtain a dynamically updated set of health characteristics that covers multiple physiological and environmental factors.

5. The intelligent monitoring method for home care of heart failure patients combining deep learning according to claim 4, characterized in that, The process of using deep learning models to comprehensively evaluate a dynamic set of health features is as follows: The set of health features is input into a deep learning model composed of convolutional units and recurrent units; Priority labels are introduced during the model encoding stage to dynamically adjust the weights of different features in training and inference. The intermediate hidden layer states of the model output are monitored to extract short-term sensitive change features and output a set of candidate risk distribution probabilities.

6. The intelligent monitoring method for home care of heart failure patients combining deep learning according to claim 5, characterized in that, The process of generating a list of candidate exception events is as follows: The risk distribution probability output by the comprehensive assessment is compared with the patient's individual health threshold database; If the probability of risk in a consecutive period exceeds a set threshold, the consecutive period will be marked as a potential anomaly. An event merging algorithm is used to integrate adjacent time periods to generate a list of candidate abnormal events that includes event timestamps, ranges of physiological parameter changes, and scene descriptions.

7. The intelligent monitoring method for home care of heart failure patients combining deep learning according to claim 6, characterized in that, The process of labeling and grouping the physiological parameters and related behavioral scenarios corresponding to candidate abnormal events is as follows: The heart rate, blood pressure, respiration, and body movement indicators involved in the candidate abnormal event list are labeled. Group abnormal events under different tags according to similar behavioral scenarios; During the grouping process, the duration, frequency, and severity of abnormal events are recorded, and a set of candidate abnormal events with grouping labels and contextual information is output.

8. The intelligent monitoring method for home care of heart failure patients combining deep learning according to claim 7, characterized in that, The process of conducting a local assessment of a patient's current health status based on grouping information is as follows: For each abnormal event group, the fluctuation range of physiological parameters and the degree of difference from the normal reference value are calculated; By combining the patient's historical health records, we can analyze whether the grouped events are part of a continuous deterioration trend; The local assessment model outputs a risk level reflecting the urgency, generating a health risk profile for the patient in the current time period.

9. The intelligent monitoring method for home care of heart failure patients combining deep learning according to claim 8, characterized in that, The process of generating real-time alerts and personalized care recommendations is as follows: The risk level obtained from the local assessment is compared with the patient's overall health trend to determine whether an early warning needs to be triggered. If the risk level exceeds the set threshold, a warning message will be pushed to the patient's terminal and the medical care platform in real time, and personalized care suggestions will be generated in conjunction with the patient's daily care plan.

Images