Heart rate monitoring method and system based on deep learning

Abstract

The invention relates to the technical field of health monitoring, in particular to a heart rate monitoring method and system based on deep learning. The method comprises the steps of obtaining multi-modal physiological data collected by sensor equipment worn by a user, generating a multi-modal physiological data sequence, and preprocessing the multi-modal physiological data sequence to obtain a preprocessed data sequence; extracting time sequence features of the data to form a training data set, training the training data set through LSTM and GRU deep learning models, and learning a heart rate change mode in a resting and motion state transition period to obtain a pre-trained heart rate fluctuation mode recognition model; inputting a heart rate signal acquired in real time into the pre-training model, and generating a detection result in combination with the motion state data; and on the basis of a detection result, abnormal fluctuation is evaluated in combination with a rule engine, and personalized health feedback is generated. The accuracy of heart rate fluctuation recognition is improved, personalized health management and real-time feedback are supported, and the method is suitable for health monitoring requirements of complex multiple scenes.

Description

Technical Field

[0001] The present invention relates to the field of health monitoring technology, and in particular to a heart rate monitoring method and system based on deep learning. Background Art

[0002] With the widespread application of smart health monitoring devices, the importance of heart rate monitoring technology in health management and medical fields has become increasingly prominent. Heart rate monitoring devices can reflect the health status of the human body in real time by collecting the user's physiological data. However, existing heart rate monitoring technology mainly relies on fixed threshold rules to determine whether the heart rate is abnormal. This method has significant deficiencies in dealing with short-term heart rate fluctuations during the transition period between resting and exercise states. Specifically, when the user transitions from a resting state to an exercise state, the heart rate usually rises steadily in a short period of time. This normal physiological fluctuation is easily misjudged by traditional methods as arrhythmia or other health problems, resulting in a high false alarm rate. In addition, it is difficult for existing technologies to comprehensively analyze the correlation between heart rate signals, exercise status data, and environmental data, and lacks the ability to accurately evaluate individual health status, which limits the application effect of heart rate monitoring technology in multi-scenario health management. Summary of the invention

[0003] The present invention provides a heart rate monitoring method and system based on deep learning to solve the problem of how to accurately analyze heart rate fluctuations during the transition period between resting and exercising states through a deep learning model based on heart rate signals, motion status data and environmental data, avoid the false alarm of a stable rising heart rate in a short period of time as arrhythmia, and evaluate the health status of the user's heart rate in real time to provide personalized health feedback.

[0004] In order to solve the above technical problems, the present invention provides a heart rate monitoring method based on deep learning, comprising:

[0005] Acquire multimodal physiological data collected by a sensor device worn by a user, generate a multimodal physiological data sequence, and preprocess the multimodal physiological data sequence to obtain a preprocessed multimodal physiological data sequence;

[0006] Extracting the temporal features of the preprocessed multimodal physiological data sequence, and forming a training data set according to the extracted temporal features; training the training data set by using LSTM and GRU deep learning models to learn the heart rate change pattern during the transition period between resting and exercise states, and obtaining a pre-trained heart rate fluctuation pattern recognition model;

[0007] Inputting the real-time collected heart rate signal into the pre-trained heart rate fluctuation pattern recognition model to perform heart rate fluctuation pattern recognition, combining the exercise state data to determine whether the current heart rate fluctuation is normal fluctuation or abnormal fluctuation, and generating corresponding detection results;

[0008] Based on the detection results, the abnormal fluctuations are evaluated and confirmed in combination with the rule engine to generate personalized health feedback, and the user is informed of the current heart rate status and whether corresponding health management measures need to be taken in real time through the feedback mechanism.

[0009] Furthermore, the step of obtaining multimodal physiological data collected by the sensor device worn by the user specifically includes:

[0010] Acquire heart rate signals, motion state data and environmental data collected by sensor devices worn by the user; integrate the heart rate signals, motion state data and environmental data into a multimodal physiological data sequence.

[0011] Furthermore, the step of preprocessing the multimodal physiological data sequence specifically includes:

[0012] Performing denoising processing on the multimodal physiological data sequence;

[0013] The multimodal physiological data sequence is interpolated, supplemented and normalized to obtain a preprocessed multimodal physiological data sequence.

[0014] Furthermore, the step of extracting the time series features of the multimodal physiological data and forming a training data set according to the extracted time series features specifically includes:

[0015] Dividing the preprocessed multimodal physiological data sequence into time windows, and extracting the time series features of the heart rate signal from the time windows;

[0016] The time series features of the motion state data and the environmental data are extracted from the time window, and the time series features of the motion data and the environmental data are combined with the time series features of the heart rate signal to form a multidimensional feature vector and a training data set is formed based on the multidimensional feature vector.

[0017] Furthermore, the step of training the training data set by using the LSTM and GRU deep learning models specifically includes:

[0018] Input the training data set into the LSTM deep learning model to perform time series modeling on the heart rate signal;

[0019] Inputting the training data set into the GRU deep learning model to capture the dynamic change pattern of the heart rate signal;

[0020] The outputs of the LSTM and GRU deep learning models are combined to generate the pre-trained heart rate fluctuation pattern recognition model.

[0021] Furthermore, the step of inputting the real-time collected heart rate signal into the pre-trained heart rate fluctuation pattern recognition model specifically includes:

[0022] Acquire the real-time collected heart rate signal and motion status data and process them into feature vector format;

[0023] The feature vector is input into the pre-trained heart rate fluctuation pattern recognition model to generate a recognition result of the current heart rate fluctuation pattern.

[0024] Furthermore, the step of judging whether the current heart rate fluctuation is normal fluctuation or abnormal fluctuation in combination with the exercise status data specifically includes:

[0025] According to the heart rate fluctuation pattern recognition result, combined with the exercise state data, it is determined whether the current state is resting or exercising;

[0026] According to the heart rate fluctuation amplitude and the preset threshold range, it is further judged whether the heart rate fluctuation is a normal physiological fluctuation or an abnormal fluctuation.

[0027] Furthermore, the steps of evaluating and confirming abnormal fluctuations in combination with the rule engine specifically include:

[0028] Analyzing whether abnormal fluctuations meet the conditions of real health risks based on the heart rate fluctuation pattern recognition results and exercise status data;

[0029] If the abnormal fluctuation does not meet the real health risk conditions, it is judged as a false alarm and the normal result is output; if it meets the conditions, the abnormal health risk is output.

[0030] Furthermore, the step of generating personalized health feedback and informing the user in real time through the feedback mechanism specifically includes:

[0031] Generate feedback information of current heart rate status based on abnormal fluctuation assessment results;

[0032] The feedback information is sent to the user device via wireless communication in real time, and health management suggestions are displayed.

[0033] Furthermore, a heart rate monitoring system based on deep learning is characterized by comprising:

[0034] The data acquisition and preprocessing module acquires multimodal physiological data acquired by the sensor device worn by the user; generates a multimodal physiological data sequence, and preprocesses the multimodal physiological data sequence to obtain a preprocessed multimodal physiological data sequence;

[0035] A deep learning feature extraction and training module extracts the temporal features of the preprocessed multimodal physiological data sequence, and forms a training data set based on the extracted temporal features; trains the training data set through LSTM and GRU deep learning models, learns the heart rate change pattern during the transition period between resting and exercise states, and generates a pre-trained heart rate fluctuation pattern recognition model;

[0036] The real-time recognition and health assessment module inputs the real-time collected heart rate signal into the pre-trained heart rate fluctuation pattern recognition model, recognizes the current heart rate fluctuation pattern, and determines the health risk in combination with the exercise status data;

[0037] Personalized feedback and user interaction module, which generates personalized health feedback information and notifies users of their current heart rate status and health management suggestions in real time through their devices;

[0038] Compared with the prior art, the main beneficial effects of the present invention are as follows:

[0039] (1) Improving the accuracy of heart rate fluctuation analysis: The present invention trains heart rate signals through LSTM and GRU deep learning models, which can accurately capture the heart rate fluctuation patterns during the transition period between resting and exercise states. Compared with the traditional method based on fixed thresholds, it can effectively reduce the false alarm of arrhythmia caused by a stable increase in heart rate in a short period of time, thereby improving the reliability of monitoring results.

[0040] (2) Real-time and multimodal fusion: The present invention utilizes heart rate signals, motion status data and environmental data collected in real time to fuse multimodal information, eliminating the limitations of a single signal monitoring mode in complex scenarios and achieving dynamic evaluation of the user's current heart rate status. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A flowchart of a heart rate monitoring method based on deep learning provided in an embodiment of the present application;

[0042] Figure 2 A structural block diagram of a heart rate monitoring system based on deep learning provided in an embodiment of the present application. DETAILED DESCRIPTION

[0043] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by technicians in the technical field of the present application; the terms used in the specification of the application herein are only for the purpose of describing specific embodiments and are not intended to limit the present application; the terms "including" and "having" and any variations thereof in the specification and claims of the present application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of the present application or the above-mentioned drawings are used to distinguish different objects, not to describe a specific order.

[0044] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0045] Example 1: Reference Figure 1 , an embodiment of the present invention provides a heart rate monitoring method based on deep learning, the method may at least include steps S100-S400:

[0046] S100: Acquire multimodal physiological data collected by a sensor device worn by a user, generate a multimodal physiological data sequence, and preprocess the multimodal physiological data sequence to obtain a preprocessed multimodal physiological data sequence.

[0047] S200, extracting the temporal features of the preprocessed multimodal physiological data, and forming a training data set based on the extracted temporal features; training the training data set through LSTM and GRU deep learning models to learn the pattern of heart rate changes during the transition period between resting and exercise states, and obtaining a pre-trained heart rate fluctuation pattern recognition model.

[0048] S300, input the real-time collected heart rate signal into the pre-trained heart rate fluctuation pattern recognition model to perform heart rate fluctuation pattern recognition, combine the motion state data, determine whether the current heart rate fluctuation is normal fluctuation or abnormal fluctuation, and generate corresponding detection results.

[0049] S400. Based on the detection results, the abnormal fluctuations are evaluated and confirmed in combination with the rule engine, and personalized health feedback is generated. The user is informed of the current heart rate status and whether corresponding health management measures need to be taken in real time through the feedback mechanism.

[0050] Furthermore, step S100 at least includes steps S110-S130:

[0051] S110: Acquire multimodal physiological data collected by a sensor device worn by the user, including heart rate signals, motion data, and environmental data.

[0052] First, multimodal physiological data is collected through sensor devices worn by users (smart watches, chest straps, etc.). Specifically, the sensor devices include heart rate signal acquisition sensors (such as ECG or PPG sensors), motion state sensors (accelerometers or gyroscopes), and environmental sensors (temperature and humidity sensors). The above data will be used as system input.

[0053] Furthermore, the heart rate signal can be obtained through a PPG (photoplethysmography) or ECG (electrocardiogram) sensor. Specifically, the ECG signal can be expressed as x ECG (t), where t is the timestamp, x ECG (t) is a continuous signal of the electrocardiogram. The motion data is obtained by an accelerometer sensor and can be obtained by three-axis acceleration x acc (t),y acc (t),z acc (t) is represented by x, y, and z, where x, y, and z are the components of acceleration in the three axes. The environmental data includes environmental parameters such as temperature and humidity, which can be represented by x env (t), where x env (t) is the environmental data at time t.

[0054] The above three types of physiological data (heart rate signal, motion data, and environmental data) are integrated in time series. Assume that the collected heart rate signal, motion data, and environmental data are x ECG (t),x acc (t), and x env (t), then at time point t, the obtained multimodal physiological data sequence can be expressed as:

[0055] X(t) = [x ECG (t),x acc (t),x env (t)];

[0056] Wherein, X(t) is a multimodal physiological data sequence; the multimodal physiological data sequence will be used as an input for subsequent preprocessing.

[0057] S120: Preprocessing the multimodal physiological data sequence to eliminate interference of noise on subsequent analysis.

[0058] For the multimodal physiological data sequence X(t), denoising is performed. Specifically, the heart rate signal is usually affected by motion artifacts and electrical interference, so filtering technology, low-pass filtering or high-pass filtering, is required. Assume that the heart rate signal x ECG (t) is low-pass filtered, and the filter transfer function is H(f), then the heart rate signal after filtering is It can be expressed as:

[0059]

[0060] Likewise, other sensor data (acceleration and environmental data) can also be denoised using appropriate filtering methods.

[0061] Furthermore, each type of data is standardized to eliminate the dimension differences between different sensors. The standardization process is completed by normalizing each data sequence by mean or standard deviation. Specifically, for each data sequence x data (t), the normalized data sequence It can be expressed as:

[0062]

[0063] Among them, μ data and σ data are the mean and standard deviation of the data series respectively. The standardized data will eliminate the dimensionality impact of different sensor data and ensure that the data can be effectively processed by the deep learning model.

[0064] Furthermore, an interpolation algorithm (linear interpolation or spline interpolation) is used to supplement the missing parts (there are missing values ​​in the multimodal physiological data sequence, especially when the motion data or environmental data is missing). env (t), the interpolation process can be expressed as:

[0065]

[0066] After interpolation and completion, the obtained multimodal physiological data sequence is complete and continuous.

[0067] S130: Format the preprocessed multimodal physiological data sequence in a time series and integrate it into a data format suitable for deep learning model processing.

[0068] First, the preprocessed multimodal physiological data sequence is divided into time windows W of a certain size, so that each window contains a certain number of sampling points (10 seconds or 30 seconds of data). Assuming that each window contains N time points, the data in each window can be expressed as:

[0069] X W =[X(t1),X(t2),…,X(t N )];

[0070] Among them, X W is a sequence of multimodal physiological data within each window. The window will become the input for subsequent deep learning model training.

[0071] Furthermore, from the multimodal physiological data sequence X in each time window W Extract time series features. Specifically, statistical features (mean, standard deviation, maximum value, minimum value, etc.) and frequency domain features (power spectrum density) can be extracted. For example, the feature extraction of heart rate signal can be expressed as:

[0072]

[0073] Among them, mean(x ECG (t)) represents the feature extraction of the heart rate signal. After feature extraction for each data window, the obtained feature sequence will be used as the input of the deep learning model.

[0074] Furthermore, the extracted feature sequence is formatted into a shape suitable for deep learning model input. For example, after feature extraction, the features of heart rate data, motion data, and environmental data can be combined into a matrix X formatted , each row of the matrix represents the data of a time window, and the columns represent different features:

[0075] X formatted =[mean(x ECG ),std(x acc ),…];

[0076] The preprocessed matrix X formatted It will be input into the subsequent deep learning model to identify the heart rate fluctuation pattern.

[0077] Step S200 at least includes steps S210-S230:

[0078] S210: extracting time series features of the preprocessed multimodal physiological data sequence based on the preprocessed multimodal physiological data sequence.

[0079] Specifically, for the heart rate signal x ECG (t), the extracted temporal features include:

[0080] ① Mean: reflects the overall level of the heart rate signal, which can be expressed by the formula:

[0081]

[0082] Among them, μ ECG is the mean of the heart rate signal, t i is the timestamp, and N is the number of sample points in the window.

[0083] ②Standard deviation: measures the fluctuation amplitude of the heart rate signal. The formula is:

[0084]

[0085] Among them, σ ECG is the standard deviation of the heart rate signal.

[0086] ③ Maximum and minimum values: reflect the extreme changes of the heart rate signal:

[0087] max(xECG )=max(x ECG (t1),x ECG (t2),…,x ECG (t N ));

[0088] min(x ECG )=min(x ECG (t1),x ECG (t2),…,x ECG (t N ));

[0089] The above time series features will provide basic data for the subsequent training process.

[0090] ④ Comprehensive multimodal features: In addition to the heart rate signal x ECG (t), motion data x acc (t) and environmental data x env (t) also participates in feature extraction. Specifically, by performing the same statistical processing on the motion data, the features related to the motion (mean value, standard deviation of acceleration, etc.) are extracted, and the feature extraction of the environmental data is performed. The feature extraction formula of each data is similar to the feature extraction of the heart rate signal.

[0091] ⑤ Combined feature vector: Finally, the time series features extracted from different types of data are integrated into a multidimensional feature vector F W :

[0092] F W =[μ ECG ,σ ECG ,max(x ECG ),min(x ECG ),μ acc ,σ acc ,μ env ,σ env ];

[0093] The multidimensional feature vector contains the temporal features of heart rate, motion and environmental data in each time window, and will serve as input for deep learning model training.

[0094] S220: Based on the extracted time series features, the heart rate signal is trained through LSTM and GRU deep learning models to learn the heart rate fluctuation pattern during the transition period between resting state and exercise state.

[0095] First, the multidimensional feature vector F W As the input of the deep learning model. For each time window W, input the multidimensional feature vector F W, including all features extracted from heart rate signals, motion data, and environmental data. Assume that the training set composed of features extracted from K time windows is Each multidimensional feature vector F W The training data set will be composed together with the corresponding labels (such as resting state or movement state).

[0096] Furthermore, based on the above training data set, LSTM and GRU deep learning models are trained respectively. Specifically, LSTM (Long Short-Term Memory Network) and GRU (Gated Recurrent Unit) networks are both a type of recurrent neural network (RNN), which are suitable for processing time series data and performing pattern recognition by capturing long-term dependencies in sequence data.

[0097] LSTM deep learning model training: Update its state through the following recursive formula:

[0098] i t =σ(W i ·F W +b i );

[0099] f t =σ(W f ·F W +b f );

[0100] o t =σ(W o ·F W +b o );

[0101] c t =f t ·c t-1 +i t tanh(W c ·F W +b c )

[0102] h t =o t ·tanh(c t )

[0103] Among them, i t ,f t ,o t They are input gate, forget gate and output gate, c t and h t are the cell state and hidden state respectively, W and b are the learned weights and biases, and F W is the multidimensional feature vector of the input.

[0104] GRU deep learning model training:

[0105] r t =σ(W r ·F W +b r );

[0106] z t =σ(W z ·F W +b z );

[0107]

[0108] Among them, r t and z t They are reset gate and update gate, is a candidate hidden state, h t is the updated hidden state.

[0109] Through the training of LSTM and GRU deep learning models, it is possible to learn the fluctuation patterns of heart rate signals during the transition period between resting state and exercise state, providing an effective training model for subsequent pattern recognition.

[0110] Furthermore, using the extracted multidimensional feature vector F W and the corresponding label (resting state or motion state) as input for supervised learning. Depending on the specific task type (classification or regression), the loss function uses the cross entropy loss function or the mean square error (MSE). The deep learning model is trained using the gradient descent optimization algorithm (Adam optimizer), and the deep learning model parameters are iteratively updated until the loss function converges.

[0111] S230: Obtaining a pre-trained heart rate fluctuation pattern recognition model that has been trained.

[0112] First, after training, the performance of LSTM and GRU deep learning models is evaluated through cross-validation and other methods. The evaluation indicators are accuracy, precision, recall, etc., and appropriate indicators are selected according to the actual task. Specifically, for each training sample multi-dimensional vector F W , predicted heart rate fluctuation pattern It will be compared with the true label yw, and the accuracy of the model will be evaluated through the loss function.

[0113] Furthermore, after sufficient training and evaluation, the best performing model (LSTM or GRU deep learning model) is selected and saved as a pre-trained heart rate fluctuation pattern recognition model. This model can be directly applied in subsequent real-time heart rate fluctuation pattern recognition.

[0114] Furthermore, the obtained pre-trained heart rate fluctuation pattern recognition model MHR It will be used in the pattern recognition phase of real-time data. Specifically, the model M HR It can be expressed as:

[0115] M HR ={LSTM Model or GRU Model};

[0116] The pre-trained heart rate fluctuation pattern recognition model will be used for subsequent real-time data processing to identify the transition between the resting state and the exercise state.

[0117] Step S300 at least includes steps S310-S330:

[0118] S310: Input the real-time collected heart rate signal and exercise state data into the pre-trained heart rate fluctuation pattern recognition model, perform heart rate fluctuation pattern recognition through the pre-trained heart rate fluctuation pattern recognition model, and identify whether the current fluctuation belongs to the resting state or the exercise state. Specifically:

[0119] First, the user's heart rate signal x is obtained in real time from the sensor device worn ECG (t) and motion state data x acc (t), namely the user's heart rate signal and acceleration (or motion sensor output). The data is transmitted to the system through the wireless connection between the sensor and the mobile terminal to form a time series data sequence.

[0120] Furthermore, the real-time collected heart rate signal and motion state data are preprocessed into an input data sequence F in the same format as the training data. W The input data sequence F W It consists of the following steps:

[0121] F W =[μ ECG 1,σ ECG 1,max(x ECG ),min(x ECG ),μ acc1 ,σ acc1 ];

[0122] Among them, μ ECG 1 is the mean of the heart rate signal, σ ECG1 is the standard deviation of the heart rate signal, μ acc1 and σ acc1 are the mean and standard deviation of the motion signal respectively. The real-time data is converted into the above multidimensional feature vector F W For model input.

[0123] Furthermore, the processed multidimensional feature vector F WInput to the previously trained pre-trained heart rate fluctuation pattern recognition model M HR , which is an LSTM or GRU deep learning model. The model has learned to recognize the fluctuation pattern of heart rate signals under different exercise states.

[0124] The pre-trained heart rate fluctuation pattern recognition model processes the input multi-dimensional feature vector through structures such as convolutional layers or recurrent layers, and outputs a prediction result. This result indicates whether the heart rate fluctuation belongs to the "resting" state or the "exercise" state. Specifically, the pre-trained heart rate fluctuation pattern recognition model calculates the probability value and determines the heart rate fluctuation pattern based on the preset threshold:

[0125]

[0126] in, Represents the recognized heart rate fluctuation pattern (resting state or motion state). If the probability of resting state is greater than the probability of motion state, it is considered to be in resting state, otherwise it is in motion state.

[0127] Furthermore, based on the output of the pre-trained heart rate fluctuation pattern recognition model, a label of the current fluctuation pattern is generated. This label determines whether an assessment based on different health risks is required next.

[0128] S320: Based on the current exercise status data, determine whether the current heart rate fluctuation is a normal physiological fluctuation or an abnormal fluctuation, and evaluate the heart rate health status. Specifically:

[0129] First, in step S310, the current heart rate fluctuation pattern (resting state or exercise state) has been identified. According to the identified heart rate fluctuation pattern, combined with the current exercise state data (such as exercise intensity, acceleration and other indicators), it is evaluated whether the heart rate fluctuation is a normal physiological fluctuation. Specifically, if the current state is at rest and the heart rate fluctuation amplitude is large, it may be an abnormal fluctuation; if the state is in exercise, the heart rate fluctuation amplitude is large, which is normal.

[0130] Further, determine the current heart rate fluctuation x ECG (t) Whether it conforms to the normal physiological fluctuation range. If the fluctuation exceeds the normal physiological range (based on the threshold of the previously trained model), it is considered an abnormal fluctuation. The formula can be used to determine whether the fluctuation exceeds the threshold:

[0131] if|x ECG (t)-μ ECG1 ∣>δ thresholdthen ,Then abnormal;

[0132] Among them, δ thresholdthenIt is the threshold of the normal heart rate fluctuation range, usually derived from historical data statistics.

[0133] Furthermore, the heart rate health status is further evaluated by combining the exercise status data and the heart rate fluctuation range. If the heart rate fluctuation exceeds the normal fluctuation range during exercise, it is evaluated as "normal physiological fluctuation"; if there is a sharp fluctuation in the resting state, it is evaluated as "abnormal fluctuation". This step determines the heart rate health status through the rule engine or logical judgment based on deep learning.

[0134] Furthermore, the health assessment results are output: the health assessment results are generated based on the above judgments

[0135]

[0136] in, An assessment result representing health status, which can be "normal" or "abnormal".

[0137] S330: Generate corresponding detection results according to the identified heart rate fluctuation pattern, and provide real-time feedback to the user to determine whether further health management measures or interventions are needed. Specifically:

[0138] First, based on the health status assessment result in step S320 Combined with the aforementioned heart rate fluctuation mode (resting or exercising), the corresponding test results are generated. If the current heart rate fluctuation is within the normal range, the test result "normal heart rate" is generated; if it exceeds the normal fluctuation range, the test result "abnormal heart rate fluctuation" is generated. Formula:

[0139]

[0140] Furthermore, according to the detection result, health feedback is provided to the user in real time. Specifically, if the heart rate is within the abnormal range, the user is reminded that there may be heart health risks and is advised to further monitor or seek medical treatment; if the heart rate is within the normal range, the user is fed back that the heart rate is stable and continues to maintain the current lifestyle.

[0141] Furthermore, personalized health management suggestions are generated based on the evaluation results. If the test result is abnormal heart rate fluctuation, the system may recommend health interventions (rest, relaxation training, medical treatment, etc.) to the user. Specific health management suggestions can be generated by the rule engine and pushed to the user's terminal device in real time.

[0142] Furthermore, according to the detection result, timely feedback is provided to the user. The system feedback includes but is not limited to real-time display of heart rate status, health risk assessment, exercise status and other information.

[0143] Step S400 at least includes steps S410-S430:

[0144] S410: Based on the results of the heart rate fluctuation pattern recognition, the rule engine is combined to further evaluate the abnormal fluctuations to determine whether it is a real health risk and eliminate false positives. Specifically:

[0145] First, S320 has completed the health status assessment of heart rate fluctuations, and outputted the determination result of whether the heart rate fluctuations are normal or abnormal. If it is judged as "abnormal", enter this step for further evaluation.

[0146] Furthermore, in order to improve the accuracy of heart rate health status assessment, the rule engine is used to further analyze abnormal fluctuations. Specifically, the rule engine evaluates abnormal fluctuations based on the following key parameters: current heart rate value x ECG (t), heart rate fluctuation amplitude Δx ECG , motion state data x acc (t).

[0147] The rule engine combines the parameters to determine whether the heart rate fluctuations truly reflect health risks. For example, if the heart rate fluctuates too much in a resting state, it may be a false alarm; while in a state of exercise, large fluctuations are usually normal fluctuations.

[0148] Combined with the preset threshold, the rule engine determines whether the current abnormal fluctuation meets the conditions of real health risk. ECG Greater than the set threshold δ HR , and accompanied by an abnormal acceleration x acc (t) When it is displayed as a resting state, it is assessed as a health risk. Example:

[0149] if|x ECG (t)-μ ECG1 ∣>δ HR and x acc (t)<δ acc1 then health risk;

[0150] Among them, δ HR is the threshold of heart rate fluctuation, δ acc is the threshold of motion state.

[0151] Furthermore, the rule engine outputs the final health risk determination result based on the analysis results. If the fluctuation is confirmed to be abnormal and there is a health risk, "Abnormal Health Risk" is output; if the assessment is a false alarm, "Normal" is output.

[0152]

[0153] S420: Generate personalized health feedback based on the evaluation result to guide the user whether to take immediate action or further monitor the heart rate.

[0154] First, based on the health risk assessment results obtained in S410 Generate personalized health feedback. If the assessment is "abnormal health risk", it will recommend that the user take immediate intervention measures (rest, seek medical attention, etc.); if it is "normal", it will recommend that the user continue to maintain the current activity and monitor the heart rate status regularly.

[0155] Furthermore, health feedback should include the following elements:

[0156] Health status description: According to "Current heart rate is normal" or "There is a health risk."

[0157] Recommended actions: If an abnormality is detected, the prompt is "Please rest and monitor your heart rate" or "It is recommended to seek medical attention"; if normal, the prompt is "Continue to maintain current activities and check regularly."

[0158] Personalized health advice: Generate personalized health management advice based on the user's historical data (past heart rate fluctuation trends, exercise intensity, etc.). If the user has a recent history of abnormal fluctuations, stricter monitoring may be recommended.

[0159] Generation of healthy feedback: When generating feedback, the urgency of the recommendation can be defined by the following formula:

[0160]

[0161] At this time, depending on the type of feedback, the system will generate an instant feedback message for the user and notify the user in real time through the user's device (mobile phone, watch, etc.).

[0162] S430: Feedback the health status to the user in real time through the feedback mechanism. Specifically:

[0163] First, based on the health feedback results generated in step S420, the system will push a message to the user's device through a real-time feedback mechanism to inform the user of the current heart rate health status. The system transmits the feedback information to the user's terminal device (smart watch, mobile phone, etc.) through a wireless communication interface (Bluetooth, Wi-Fi, etc.).

[0164] Specifically, the feedback message may include the following:

[0165] Current heart rate status: According to the test results Display the user's heart rate status in real time ("normal heart rate" or "abnormal").

[0166] Health management suggestions: Based on the feedback on health status, real-time suggestions on whether to take action or monitor heart rate are pushed.

[0167] Monitoring Guide: If continued monitoring is required, the system will provide a quick entry to the heart rate monitoring function and display the heart rate data in real time on the user's device.

[0168] Instant reminder: When the user's heart rate fluctuation exceeds the preset health threshold, the system will push a high-priority health alert to ensure that the user can be aware of health problems in time. For example, when the heart rate fluctuation Δx ECG When the threshold is exceeded, an "abnormal fluctuation, please take a rest" reminder will be pushed immediately.

[0169] Furthermore, to enhance the user experience, users can interact through terminal devices, such as confirming that they have received feedback or learning more about their health status. Feedback mechanisms include push messages, vibration prompts, voice reminders, and other methods.

[0170] Furthermore, the system will continue to monitor the heart rate in real time, and re-evaluate based on the heart rate data collected subsequently. If there are new changes in health status, the system will promptly update the feedback and push it to the user terminal again.

[0171] The present invention trains heart rate signals through LSTM and GRU deep learning models, which can accurately capture the heart rate fluctuation pattern during the transition period between resting and exercise states. Compared with the traditional method based on fixed threshold, it can effectively reduce the false alarm of arrhythmia caused by a stable increase in heart rate in a short period of time, and improve the reliability of monitoring results.

[0172] Compared with the prior art, the main beneficial effects of the present invention are as follows:

[0173] (1) Improving the accuracy of heart rate fluctuation analysis: The present invention trains heart rate signals through LSTM and GRU deep learning models, which can accurately capture the heart rate fluctuation patterns during the transition period between resting and exercise states. Compared with the traditional method based on fixed thresholds, it can effectively reduce the false alarm of arrhythmia caused by a stable increase in heart rate in a short period of time, thereby improving the reliability of monitoring results.

[0174] (2) Real-time and multimodal fusion: The present invention utilizes heart rate signals, motion status data and environmental data collected in real time to fuse multimodal information, eliminating the limitations of a single signal monitoring mode in complex scenarios and achieving dynamic evaluation of the user's current heart rate status.

[0175] Embodiment 2: Figure 2 It is shown that a heart rate monitoring system based on deep learning is provided according to an embodiment of the present invention, and the system includes:

[0176] The data acquisition and preprocessing module 10 acquires multimodal physiological data from the sensor devices (smart watches, chest straps, etc.) worn by the user, including heart rate signals, motion data, and environmental data. The collected data is denoised, interpolated, supplemented, and normalized to generate a preprocessed multimodal physiological data sequence to ensure the integrity and consistency of the data.

[0177] Specifically, the heart rate signal x is collected by the sensor ECG (t), motion state data x acc (t), environmental data x env (t), integrated into a multimodal physiological data sequence X(t) = [x ECG (t),x acc (t),x env (t)]. Filtering technology (such as low-pass filtering) is used to denoise the data, and the standardization formula is as follows:

[0178]

[0179] Among them, μ ECG and σ ECG are the mean and standard deviation respectively.

[0180] The preprocessed multimodal physiological data sequence is divided into time windows W to form a multidimensional feature vector F W For subsequent models.

[0181] The deep learning feature extraction and training module 20 extracts time series features from the preprocessed multimodal physiological data sequence. The deep learning model (LSTM and GRU) is used to train the time series feature data to learn the heart rate change pattern during the transition period between resting and exercise states. Specifically:

[0182] Based on multimodal physiological data sequence F W , extract the time series features of the heart rate signal (such as mean, standard deviation) and the time series features of the motion and environmental data, and combine them into a training data set.

[0183] The training data set is input into the LSTM and GRU deep learning models to build a recurrent neural network to capture temporal dependencies. The update formula of LSTM is as follows:

[0184] c t =f t ·c t-1 +i t tanh(W c ·FW +b c );

[0185] Among them, f t ,i t ,c t They are forget gate, input gate and cell state respectively.

[0186] After the training is completed, the pre-trained heart rate fluctuation pattern recognition model M is generated. HR , used for real-time heart rate fluctuation recognition.

[0187] The real-time recognition and health assessment module 30 inputs the real-time collected heart rate signal into the pre-trained heart rate fluctuation pattern recognition model M HR In the process of identifying the heart rate fluctuation pattern, combined with the exercise status data, it is determined whether the heart rate fluctuation is normal or abnormal, and a health assessment is performed. Specifically:

[0188] Get heart rate signals and exercise data in real time And input the pre-trained heart rate fluctuation pattern recognition model M HR Perform pattern recognition.

[0189] According to the pattern recognition result y W To determine the current heart rate state (resting or exercising), the formula is as follows:

[0190]

[0191] Combined with current motion data x acc (t), assess whether there is a health risk. If the fluctuation exceeds the threshold range, it is marked as "abnormal":

[0192] if|x ECG (t)-μ ECG ∣>δ thresholdthen ,Then abnormal;

[0193] The personalized feedback and user interaction module 40 generates a health assessment report, provides real-time feedback on the user's health status, provides personalized health advice, and pushes health notifications through mobile devices. Specifically:

[0194] Generates feedback based on the assessment results, including current heart rate status (normal or abnormal) and recommended actions.

[0195] Health reports are pushed to user devices in real time, displaying current heart rate status and recommended actions via Bluetooth or Wi-Fi.

[0196] The system provides an interactive interface where users can query historical heart rate data and obtain detailed health management suggestions.

[0197] The present invention improves the accuracy of heart rate monitoring, especially the fluctuation judgment during the transition between resting and exercise states, through the time series feature extraction and pattern recognition capabilities of the deep learning model. It generates personalized health feedback by combining the user's multimodal physiological data and historical data to improve the user experience. The system can identify heart rate fluctuation patterns in real time, and dynamically evaluate abnormal fluctuations through the rule engine to provide highly reliable health status assessment. Through the interaction and feedback mechanism with the user terminal, the user's heart rate status is notified in a timely manner and intervention suggestions are provided, which helps the user to conduct scientific health management.

[0198] Obviously, the embodiments described above are only some embodiments of the present application, rather than all embodiments. The preferred embodiments of the present application are given in the accompanying drawings, but they do not limit the patent scope of the present application. The present application can be implemented in many different forms. On the contrary, the purpose of providing the embodiments is to make the understanding of the disclosure of the present application more thorough and comprehensive. Although the present application is described in detail with reference to the aforementioned embodiments, for those skilled in the art, it is still possible to modify the technical solutions recorded in the aforementioned specific implementation methods, or to replace some of the technical features therein with equivalents. Any equivalent structure made using the contents of the specification and drawings of this application, directly or indirectly used in other related technical fields, is similarly within the scope of patent protection of this application.

Claims

1. A heart rate monitoring method based on deep learning, characterized in that: The steps include: Acquire multimodal physiological data collected by a sensor device worn by a user, generate a multimodal physiological data sequence, and preprocess the multimodal physiological data sequence to obtain a preprocessed multimodal physiological data sequence; Extracting the temporal features of the preprocessed multimodal physiological data sequence, and forming a training data set according to the extracted temporal features; training the training data set by using LSTM and GRU deep learning models to learn the heart rate change pattern during the transition period between resting and exercise states, and obtaining a pre-trained heart rate fluctuation pattern recognition model; Inputting the real-time collected heart rate signal into the pre-trained heart rate fluctuation pattern recognition model to perform heart rate fluctuation pattern recognition, combining the exercise state data to determine whether the current heart rate fluctuation is normal fluctuation or abnormal fluctuation, and generating corresponding detection results; Based on the detection results, the abnormal fluctuations are evaluated and confirmed in combination with the rule engine to generate personalized health feedback, and the user is informed of the current heart rate status and whether corresponding health management measures need to be taken in real time through the feedback mechanism.

2. The heart rate monitoring method based on deep learning according to claim 1, characterized in that: The step of obtaining multimodal physiological data collected by a sensor device worn by a user specifically includes: Acquire heart rate signals, motion state data and environmental data collected by sensor devices worn by the user; integrate the heart rate signals, motion state data and environmental data into a multimodal physiological data sequence.

3. The heart rate monitoring method based on deep learning according to claim 2, characterized in that: The step of preprocessing the multimodal physiological data sequence specifically includes: Performing denoising processing on the multimodal physiological data sequence; The multimodal physiological data sequence is interpolated, supplemented and normalized to obtain a preprocessed multimodal physiological data sequence.

4. The heart rate monitoring method based on deep learning according to claim 3, characterized in that: The step of extracting the time series features of the multimodal physiological data and forming a training data set according to the extracted time series features specifically includes: Dividing the preprocessed multimodal physiological data sequence into time windows, and extracting the time series features of the heart rate signal from the time windows; The time series features of the motion state data and the environmental data are extracted from the time window, and the time series features of the motion data and the environmental data are combined with the time series features of the heart rate signal to form a multidimensional feature vector and a training data set is formed based on the multidimensional feature vector.

5. The heart rate monitoring method based on deep learning according to claim 4, characterized in that: The step of training the training data set by using the LSTM and GRU deep learning models specifically includes: Input the training data set into the LSTM deep learning model to perform time series modeling on the heart rate signal; Inputting the training data set into the GRU deep learning model to capture the dynamic change pattern of the heart rate signal; The outputs of the LSTM and GRU deep learning models are combined to generate the pre-trained heart rate fluctuation pattern recognition model.

6. The heart rate monitoring method based on deep learning according to claim 5, characterized in that: The step of inputting the real-time collected heart rate signal into the pre-trained heart rate fluctuation pattern recognition model specifically includes: Acquire the real-time collected heart rate signal and motion status data and process them into feature vector format; The feature vector is input into the pre-trained heart rate fluctuation pattern recognition model to generate a recognition result of the current heart rate fluctuation pattern.

7. The heart rate monitoring method based on deep learning according to claim 6, characterized in that: The step of judging whether the current heart rate fluctuation is normal fluctuation or abnormal fluctuation in combination with the exercise status data specifically includes: According to the heart rate fluctuation pattern recognition result, combined with the exercise state data, it is determined whether the current state is resting or exercising; According to the heart rate fluctuation amplitude and the preset threshold range, it is further judged whether the heart rate fluctuation is a normal physiological fluctuation or an abnormal fluctuation.

8. The heart rate monitoring method based on deep learning according to claim 7, characterized in that: The steps for evaluating and confirming abnormal fluctuations in combination with the rule engine include: Analyzing whether abnormal fluctuations meet the conditions of real health risks based on the heart rate fluctuation pattern recognition results and exercise status data; If the abnormal fluctuation does not meet the real health risk conditions, it is judged as a false alarm and the normal result is output; if it meets the conditions, the abnormal health risk is output.

9. The heart rate monitoring method based on deep learning according to claim 8, characterized in that: The steps of generating personalized health feedback and informing the user in real time through the feedback mechanism specifically include: Generate feedback information of current heart rate status based on abnormal fluctuation assessment results; The feedback information is sent to the user device via wireless communication in real time, and health management suggestions are displayed.

10. A heart rate monitoring system based on deep learning, characterized in that: include: The data acquisition and preprocessing module acquires multimodal physiological data collected by the sensor devices worn by the user; and generating a multimodal physiological data sequence, and preprocessing the multimodal physiological data sequence to obtain a preprocessed multimodal physiological data sequence; A deep learning feature extraction and training module extracts the temporal features of the preprocessed multimodal physiological data sequence and forms a training data set based on the extracted temporal features; trains the training data set through LSTM and GRU deep learning models to learn the heart rate change pattern during the transition period between resting and exercise states, and generates a pre-trained heart rate fluctuation pattern recognition model; The real-time recognition and health assessment module inputs the real-time collected heart rate signal into the pre-trained heart rate fluctuation pattern recognition model, recognizes the current heart rate fluctuation pattern, and determines the health risk in combination with the exercise status data; Personalized feedback and user interaction module, which generates personalized health feedback information and notifies users of their current heart rate status and health management suggestions in real time through their devices; To implement a heart rate monitoring method based on deep learning as described in any one of claims 1-9.

Images