Blood oxygen saturation prediction method and device and intelligent wearable equipment
By collecting user health data through smart wearable devices and utilizing technologies such as deep learning neural networks and multi-head self-attention modules, blood oxygen saturation can be predicted and alerted in real time and accurately. This solves the problems of invasiveness and equipment specialization of traditional detection methods, and enables rapid and accurate blood oxygen saturation detection and alerting in non-professional environments.
Patent Information
- Application Number
- CN202510882490.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-27
- Publication Date
- 2025-11-07
AI Technical Summary
Traditional methods for measuring blood oxygen saturation are highly invasive, require specialized equipment, and are difficult to implement in a timely manner in emergency situations, especially when resources and the environment are limited, making it difficult to provide accurate blood oxygen saturation readings.
Multiple continuous time-series health signals and static health information are collected by smart wearable devices. A blood oxygen saturation prediction model trained by deep learning neural networks is used for prediction. Combined with technologies such as multi-head self-attention modules and residual connection modules, blood oxygen saturation is predicted in real time and accurately, and an early warning is issued when it is below the threshold.
It enables rapid and accurate prediction of blood oxygen saturation in non-professional environments and provides timely warnings when blood oxygen is low. It is suitable for users who need long-term monitoring, such as patients with respiratory or cardiovascular diseases.
Smart Images
Figure CN120899238A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of computer, in particular to a blood oxygen saturation prediction method and device and an intelligent wearable device. BACKGROUND
[0002] Blood oxygen saturation (SaO2) refers to the percentage of oxygen-bound hemoglobin capacity in total hemoglobin capacity in human blood, that is, the oxygen content in human blood. Many respiratory diseases, such as asthma and pneumonia, can cause the blood oxygen saturation in human blood to decrease. Blood oxygen saturation is an important physiological parameter of human respiratory circulation and plays an important role in health care and disease diagnosis. Hypoxia can cause organ dysfunction or even failure; in severe cases, the blood oxygen saturation of healthy people can decrease sharply due to infection with viruses, and this condition must be immediately treated.
[0003] Traditional blood oxygen saturation detection methods are divided into two types: invasive method and non-invasive method. The invasive method uses physical and chemical methods to detect the oxygen content in the blood sample, which can accurately measure blood oxygen saturation, but can cause pain or infection to the user and cannot provide continuous blood oxygen saturation values. The non-invasive method mainly uses pulse oximeters, but pulse oximeters are usually difficult to obtain as special medical equipment. Even in emergency situations, timely blood oxygen saturation detection is usually difficult to implement due to limited resources and environment. Therefore, how to provide accurate blood oxygen saturation to users is a problem that needs to be solved at present. SUMMARY
[0004] In view of the defects of the prior art, the present application provides a blood oxygen saturation prediction method, device and intelligent wearable device, which can quickly provide accurate blood oxygen saturation prediction results for users.
[0005] The present application provides a blood oxygen saturation prediction method, comprising: determining user health data; the user health data is fused with multiple continuous time series health signals and static health information of the user; the multiple continuous time series health signals are physiological signals of the user collected by an intelligent wearable device; the static health information is health history information of the user; inputting the user health data into a blood oxygen saturation prediction model at a current time to obtain a blood oxygen saturation prediction value output by the blood oxygen saturation prediction model at the current time; wherein the blood oxygen saturation prediction model at the current time is a model obtained by adjusting the blood oxygen saturation prediction model at the previous time based on comparison and analysis of the blood oxygen saturation prediction value at the previous time and a measured blood oxygen saturation value at the previous time according to error conditions; the blood oxygen saturation prediction model is obtained by training a deep learning neural network based on blood oxygen saturation prediction training samples; and the blood oxygen saturation prediction training samples include user health data samples and blood oxygen saturation prediction value samples.
[0006] According to the blood oxygen saturation prediction method provided by the application, the determination of the user health data comprises: acquiring the multiple continuous time sequence health signals and the static health information; the multiple continuous time sequence health signals comprise a plurality of combinations of heart rate signals, breathing signals, pulse signals, skin temperatures and motion data; the static health information comprises a plurality of combinations of user's age, gender, race and medical history; the multiple continuous time sequence health signals and the static health information are respectively subjected to vectorization fusion processing to obtain a multiple continuous time sequence health fusion vector and a static health information fusion vector; and the multiple continuous time sequence health fusion vector and the static health information fusion vector are spliced to obtain the user health data.
[0007] According to the blood oxygen saturation prediction method provided by the application, the multiple continuous time sequence health signals and the static health information are respectively subjected to vectorization fusion processing to obtain a multiple continuous time sequence health fusion vector and a static health information fusion vector, which comprises: pre-processing the multiple continuous time sequence health signals to obtain pre-processed health signals; the pre-processing mode comprises uniform sampling frequency and filtering sampling; the pre-processed health signals are segmented and sorted according to sampling time to obtain segmented health signals; the segmented health signals are subjected to vectorization fusion using a signal vectorization fusion method to obtain the multiple continuous time sequence health fusion vector; the signal vectorization fusion method comprises average pooling, rearrangement mapping and mapping addition; the static health information is classified and converted into an encoded representation; the code is subjected to feature vector mapping and splicing to obtain the static health information fusion vector.
[0008] According to the blood oxygen saturation prediction method provided by the application, the blood oxygen saturation prediction model adopts a LSTM, a GNN, a Bert, a GPT, and a T5 model based on a Transformer; the user health data is input into the blood oxygen saturation prediction model at the current time to obtain a blood oxygen saturation prediction value output by the blood oxygen saturation prediction model at the current time, which comprises: adding the user health data and a position vector to obtain an input signal vector; the input signal vector is input into an encoder of the blood oxygen saturation prediction model to obtain a state vector of a preset length output by the encoder as a predicted current blood oxygen saturation sequence result; wherein the encoder comprises a multi-head self-attention module, a residual connection module, a normalization module, a full connection layer, and an activation layer; the multi-head self-attention module is used to perform parallel processing on the time sequence features in the input signal vector through a multi-head self-attention mechanism, extract feature information in different subspaces, and output multi-head self-attention features; the residual connection module is used to perform residual connection on the multi-head self-attention features and the time sequence features in the input signal vector to output features after residual connection; the normalization module is used to perform normalization processing on the features after residual connection to output normalized features; the full connection layer is used to integrate and convert the normalized features to output blood oxygen saturation related features; and the activation layer is used to introduce a non-linear factor into the blood oxygen saturation related features through a non-linear activation function to output the predicted current blood oxygen saturation sequence result.
[0009] According to the blood oxygen saturation prediction method provided by the application, the user health data is input into the blood oxygen saturation prediction model at the current time to obtain a blood oxygen saturation prediction value output by the blood oxygen saturation prediction model at the current time, which further comprises: in the case that the blood oxygen saturation prediction model has a decoder, inputting the state vector output by the encoder into the decoder to obtain a first blood oxygen saturation future prediction value sequence generated by the decoder in an autoregressive cycle; in the case that the blood oxygen saturation prediction model does not have a decoder, performing average pooling on the state vector output by the encoder and inputting the state vector into a pre-trained linear layer to obtain a second blood oxygen saturation future prediction value sequence.
[0010] According to the blood oxygen saturation prediction method provided by the application, after the user health data is input into the blood oxygen saturation prediction model at the current time to obtain a blood oxygen saturation prediction value output by the blood oxygen saturation prediction model at the current time, it further comprises: in the case that the blood oxygen saturation prediction value is less than a preset low blood oxygen threshold, performing a low blood oxygen warning.
[0011] The application further provides a blood oxygen saturation prediction device, comprising: a data determination module configured to determine user health data; the user health data is fused with multiple continuous time series health signals and static health information of the user; the multiple continuous time series health signals are physiological signals of the user collected by an intelligent wearable device; the static health information is health history information of the user; and a prediction module configured to input the user health data into a blood oxygen saturation prediction model at a current time to obtain a blood oxygen saturation prediction value output by the blood oxygen saturation prediction model at the current time; wherein the blood oxygen saturation prediction model at the current time is a model obtained by adjusting a blood oxygen saturation prediction model at a previous time based on a comparison analysis of a blood oxygen saturation prediction value at the previous time and a measured blood oxygen saturation value at the previous time according to an error condition; the blood oxygen saturation prediction model is obtained by training a deep learning neural network based on blood oxygen saturation prediction training samples; and the blood oxygen saturation prediction training samples comprise user health data samples and blood oxygen saturation prediction value samples.
[0012] The application further provides an intelligent wearable device comprising the blood oxygen saturation prediction device.
[0013] The application further provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the blood oxygen saturation prediction method according to any one of the above when executing the computer program.
[0014] The application further provides a non-transitory computer readable storage medium having a computer program stored thereon, wherein the computer program is executable by a processor to implement the blood oxygen saturation prediction method according to any one of the above.
[0015] The blood oxygen saturation prediction method, device and intelligent wearable device provided by the application first determine user health data; the user health data is fused with multiple continuous time series health signals and static health information of the user; input the user health data into a blood oxygen saturation prediction model at a current time to obtain a blood oxygen saturation prediction value output by the blood oxygen saturation prediction model at the current time; wherein the blood oxygen saturation prediction model at the current time is a model obtained by adjusting a blood oxygen saturation prediction model at a previous time based on a comparison analysis of a blood oxygen saturation prediction value at the previous time and a measured blood oxygen saturation value at the previous time according to an error condition; the blood oxygen saturation prediction model is obtained by training a deep learning neural network based on blood oxygen saturation prediction training samples; and the blood oxygen saturation prediction training samples comprise user health data samples and blood oxygen saturation prediction value samples. The application can quickly obtain accurate blood oxygen saturation prediction results for the user. BRIEF DESCRIPTION OF DRAWINGS
[0016] In order to more clearly illustrate the technical solutions of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings described below are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.
[0017] Figure 1 is a flowchart of a blood oxygen saturation prediction method provided by the present application.
[0018] Figure 2 is a structural schematic diagram of a blood oxygen saturation prediction device provided by the present application.
[0019] Figure 3 is a structural schematic diagram of an electronic device provided by the present application. DETAILED DESCRIPTION
[0020] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions of the present application will be described clearly and completely below in combination with the drawings in the present application. Obviously, the described embodiments are some embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.
[0021] The blood oxygen saturation represents the oxygen concentration in the human blood, and its mathematical calculation formula is: , Among them, represents the blood oxygen saturation, represents the concentration of oxygenated hemoglobin, represents the concentration of reduced hemoglobin.
[0022] The blood oxygen saturation represents the oxygen carrying capacity of the human body. When the oxygen carrying capacity of the human body is insufficient, various diseases may occur, such as insufficient blood supply of heart and brain, chronic hypoxemia, etc.
[0023] At present, the blood oxygen saturation is mainly monitored by pulse oxygen saturation meter and blood gas analysis in the clinic, both of which need special detection instruments. The former is relatively portable, but is affected by skin pigment, bilirubin and other factors, and has a measurement error of 2-3%; the latter is the "gold standard" of blood oxygen saturation monitoring, but needs invasive blood sampling and application of specific blood gas analyzer to analyze data, and cannot be widely applied to home monitoring of the population.
[0024] Please refer to Figure 1 , Figure 1 is a flowchart of a blood oxygen saturation prediction method provided by the present application.
[0025] The application provides a blood oxygen saturation prediction method, comprising: 101: determining user health data; the user health data is fused with multiple continuous time series health signals and static health information of the user; the multiple continuous time series health signals are physiological signals of the user collected by an intelligent wearable device; the static health information is health history information of the user; 102: inputting the user health data into a blood oxygen saturation prediction model at a current moment to obtain a blood oxygen saturation prediction value output by the blood oxygen saturation prediction model at the current moment; wherein the blood oxygen saturation prediction model at the current moment is a model obtained after parameter adjustment of a blood oxygen saturation prediction model at a previous moment based on comparison and analysis of a blood oxygen saturation prediction value at the previous moment and a blood oxygen saturation measured value at the previous moment; the blood oxygen saturation prediction model is obtained by training a deep learning neural network based on blood oxygen saturation prediction training samples; the blood oxygen saturation prediction training samples comprise user health data samples and blood oxygen saturation prediction value samples.
[0026] To solve the technical problems existing in the prior art, the application provides a blood oxygen saturation prediction method, which first determines user health data. The user health data includes two parts: multiple continuous time series health signals and static health information. Among them, the multiple continuous time series health signals are collected by intelligent wearable devices (such as smart bracelets and smart watches), including but not limited to physiological signals such as heart rate, electrocardiogram, respiratory rate, and skin temperature. These signals are continuously collected and recorded at a certain time frequency (such as once per second). The static health information is derived from the health history of the user, such as age, gender, medical history, etc. These information are used as auxiliary input in the prediction process.
[0027] Before inputting the user health data into the prediction model, the data needs to be preprocessed. For the multiple continuous time series health signals, first, data cleaning is performed to remove outliers and noise. For example, the heart rate signal is smoothed by the sliding window average method to remove possible mutation points. For static health information, if it is a continuous variable (such as age), standardization processing is performed; if it is a classification variable (such as gender), one-hot encoding is performed.
[0028] The preprocessed user health data is input into the blood oxygen saturation prediction model at the current moment. The model is constructed based on a deep learning neural network, specifically adopting a long short-term memory network (LSTM) architecture, which can effectively process time series data. The model input includes the multiple continuous time series health signals and static health information at the current moment, and the output is the blood oxygen saturation prediction value at the current moment.
[0029] The blood oxygen saturation prediction model of the current time is a model based on the blood oxygen saturation prediction value and the measured value of the last time. The specific parameter adjustment process is: after each prediction, the prediction value is compared and analyzed with the actually measured blood oxygen saturation value, and the error is calculated. If the error exceeds the preset error threshold (such as the error is greater than 5%), the model parameters are adjusted according to the error condition. For example, in addition to modifying the number of each batch of training data, the learning rate and other conventional hyperparameters, the loss function can be customized, and the new loss function is added to the setting file. Through continuous iteration and optimization, the model gradually improves the accuracy in the continuous prediction process.
[0030] The training samples of the blood oxygen saturation prediction model include user health data samples and blood oxygen saturation prediction value samples. The user health data samples cover various physiological signals and static information, and the blood oxygen saturation prediction value samples are obtained through historical data labeling. For example, from a large amount of historical data of users, signals such as heart rate and breathing rate are extracted, and the corresponding measured values of blood oxygen saturation are obtained as training samples to input the model for training.
[0031] In actual application, the method can predict the blood oxygen saturation of the user in real time. If the prediction value is lower than the normal range (such as lower than 90%), the system automatically issues a low blood oxygen warning to remind the user to take timely measures. This method is especially suitable for users who need to monitor blood oxygen saturation for a long time, such as patients with respiratory system diseases or cardiovascular diseases.
[0032] Through the present application, physiological signals and static health information of users collected by intelligent wearable devices can be used to accurately predict blood oxygen saturation in real time, and warnings can be given according to the prediction results, thereby providing strong support for health monitoring of users.
[0033] As a preferred embodiment, determining the user health data comprises: acquiring multiple continuous time series health signals and static health information; the multiple continuous time series health signals include a plurality of combinations of heart rate signals, breathing signals, pulse signals, skin temperatures and motion data; the static health information includes a plurality of combinations of age, gender, race and medical history of the user; the multiple continuous time series health signals and the static health information are respectively subjected to vectorization fusion processing to obtain multiple continuous time series health fusion vectors and static health information fusion vectors; and the multiple continuous time series health fusion vectors and the static health information fusion vectors are spliced to obtain the user health data.
[0034] As a preferred embodiment, the multiple continuous time sequence health signals and static health information are respectively vectorized and fused to obtain a multiple continuous time sequence health fusion vector and a static health information fusion vector, including: preprocessing the multiple continuous time sequence health signals to obtain preprocessed health signals; the preprocessing mode includes uniform sampling frequency and filtering sampling; segmenting and sorting the preprocessed health signals according to sampling time to obtain segmented health signals; using a signal vectorization fusion method to vectorize and fuse the segmented health signals to obtain a multiple continuous time sequence health fusion vector; the signal vectorization fusion method includes average pooling, rearrangement mapping and mapping addition; classifying the static health information into an encoded representation; performing feature vector mapping on the code to obtain a static health information fusion vector.
[0035] In the embodiment, the user health data includes multiple continuous time sequence health signals and static health information. The multiple continuous time sequence health signals are collected by intelligent wearable devices, and specifically include heart rate signals, breathing signals, pulse signals, skin temperature and motion data, and the like in various combinations. For example, a smart bracelet can monitor the heart rate and pulse of a user in real time, while recording the skin temperature and motion data (such as the number of steps, exercise intensity, and the like). The static health information includes the age, gender, race, medical history, and the like of the user, which can be input by the user or obtained from a medical health record.
[0036] The vectorization fusion process of the multiple continuous time sequence health signals is as follows: An oxygen saturation label is usually attached to the multiple continuous time sequence health signals. For ease of data processing, data points with an oxygen saturation label less than 10% in the multiple continuous time sequence health signals are identified as sampling errors. If a data point has a sampling error, search forward until the first normally sampled data point is found. The frequencies of different health signals are unified to the same frequency, and over-sampling and under-sampling methods are used to unify the sampling frequencies of all health signals to a certain reference frequency. By preprocessing the multiple continuous time sequence health signals, filtering the unqualified signal characteristics of the sampling, the robustness and accuracy of the prediction are further improved.
[0037] The preprocessed health signals are then segmented according to the sampling time, and the multiple health signals in each time segment are taken out and subjected to health signal fusion operation. According to the characteristics of the segmented health signals, the input multiple health information is fused into a multiple continuous time sequence health fusion vector of uniform size using methods such as average pooling, rearrangement mapping, mapping addition, and the like.
[0038] For W multiple continuous time sequence health signals X ={ X 1 , X2 ,..., X W}, W ⩾1, according to the characteristics of multiple continuous time sequence health signals, a suitable fusion mode is selected.
[0039] The average pooling method is mostly used in the case that multiple continuous time sequence health signals are similar, and the formula is: , Among them, X i is the signal value in the sequence. i
[0040] The rearrangement and mapping method rearranges the multiple continuous time sequence health signals in the order of sampling time, and then maps through a linear layer to extract important representation information in multiple signals: W , , Among them, is the multiple continuous time sequence health signals rearranged in the order of sampling time, is health signal fusion.
[0041] The mapping and addition method maps the health signals to the target feature space through a linear layer, and the formula is: .
[0042] The blood oxygen saturation prediction method provided by the embodiment of the application is more flexible and more expandable through the separation of the fusion algorithm and the model architecture. Various health signals are fused into an input signal feature vector, known information is fully utilized, and the prediction accuracy is improved.
[0043] For the continuous case of static health information, discretization operation is performed to obtain static health information categories, and the classification is converted into One-hot encoding representation; On this basis, a trainable mapping layer can be added to map the encoding into a continuous static feature vector, and the complete static health information fusion vector is obtained by splicing.
[0044] Suppose that the input of the static information fusion module is M static information categories of the user: , Correspondingly, the total number of the M kinds of static attributes is: .
[0045] The static information vector obtained by One-hot encoding representation is: .
[0046] On this basis, a mapping layer is added to each static information to obtain a static health information fusion vector: .
[0047] The multiple continuous time sequence health fusion vector and the static health information fusion vector are spliced to obtain user health data that fuses the multiple continuous time sequence health signals and the static health information of the user.
[0048] The blood oxygen saturation prediction method provided in the embodiment of the application fuses static health information into an input signal vector, the steps are parallel to the fusion of multiple continuous time sequence signals, a static vector that can incorporate static information of a user is obtained, and the model is helped to adjust prediction according to the personal condition of the user in model training and testing, and the accuracy is improved.
[0049] As a preferred embodiment, the blood oxygen saturation prediction model adopts an LSTM, a GNN, a Bert, a GPT, or a T5 model based on a Transformer; the user health data is input into the blood oxygen saturation prediction model at the current time to obtain a blood oxygen saturation prediction value output by the blood oxygen saturation prediction model at the current time, including: adding the user health data and a position vector to obtain an input signal vector; inputting the input signal vector into an encoder of the blood oxygen saturation prediction model to obtain a state vector of a preset length output by the encoder as a predicted current blood oxygen saturation sequence result; wherein the encoder includes a multi-head self-attention module, a residual connection module, a normalization module, a full connection layer, and an activation layer; the multi-head self-attention module is used to perform parallel processing on time sequence features in the input signal vector through a multi-head self-attention mechanism, extract feature information in different subspaces, and output multi-head self-attention features; the residual connection module is used to perform residual connection on the multi-head self-attention features and the time sequence features in the input signal vector, and output features after residual connection; the normalization module is used to perform normalization processing on the features after residual connection, and output normalized features; the full connection layer is used to integrate and convert the normalized features, and output features related to blood oxygen saturation; and the activation layer is used to introduce a non-linear factor into the features related to blood oxygen saturation through a non-linear activation function, and output the predicted current blood oxygen saturation sequence result.
[0050] In the embodiment, the user health data is added to a position vector to embody time sequence information, obtain an input signal vector, and input into an encoder of the blood oxygen saturation prediction model.
[0051] By the method of deep learning, a Transformer module is constructed by combining and connecting a multi-head attention module, a residual connection module, a normalization module, a full connection layer, an activation layer, etc., and the Transformer module is stacked into a sequence model network structure composed of an encoder and a decoder.
[0052] According to the input signal vector and the pre-trained blood oxygen saturation prediction model, a current blood oxygen saturation is predicted, and a state vector with a length of l is output by the hidden state of the final layer of the model as the predicted current blood oxygen saturation sequence result.
[0053] The blood oxygen saturation prediction method provided by the embodiment of the application improves the convenience of the process by inputting the signal vector and the pre-trained model to obtain the predicted current blood oxygen saturation value sequence of the set sequence length.
[0054] As a preferred embodiment, the user health data is input into the blood oxygen saturation prediction model at the current time to obtain the blood oxygen saturation prediction value output by the blood oxygen saturation prediction model at the current time, and further comprising: in the case that the blood oxygen saturation prediction model has a decoder, inputting the state vector output by the encoder into the decoder to obtain a first blood oxygen saturation future prediction value sequence generated by the decoder in a self-recurrent cycle; in the case that the blood oxygen saturation prediction model does not have a decoder, performing average pooling on the state vector output by the encoder and inputting the pre-trained linear layer to obtain a second blood oxygen saturation future prediction value sequence.
[0055] In the embodiment, when the blood oxygen saturation prediction model has a decoder, the future blood oxygen saturation is predicted according to the input signal vector and the pre-trained blood oxygen saturation prediction model, and the output representing the future blood oxygen saturation is generated in a self-recurrent cycle, that is, the output value of the previous time is input into the decoder at each time to obtain the output at this time, and the first blood oxygen saturation future prediction value sequence is generated iteratively.
[0056] When the blood oxygen saturation prediction model does not have a decoder, the state vector output by the encoder is average-pooled and input into the pre-trained linear layer to obtain the second blood oxygen saturation future prediction value sequence.
[0057] When the output future blood oxygen saturation is less than 90%, a low blood oxygen warning is performed.
[0058] The blood oxygen saturation prediction method provided by the embodiment of the application improves the simplicity and rapidity of the process by inputting the signal vector and the pre-trained model to obtain the predicted future blood oxygen saturation value sequence of the set sequence length.
[0059] The loss function of the blood oxygen saturation prediction model is defined as the L2 loss of the real blood oxygen saturation sequence and the predicted blood oxygen saturation sequence. To improve the prediction accuracy of low blood oxygen samples, a reweighting method is used to increase the training weight of low blood oxygen samples in the loss function. Specifically, the loss function is: wherein, n is the total number of samples, is the predicted value of the model for the i sample, is the true value of the i sample, X is the input data, Y is the target data.
[0060] As a preferred embodiment, after inputting the user health data into the blood oxygen saturation prediction model at the current time and obtaining the blood oxygen saturation prediction value output by the blood oxygen saturation prediction model at the current time, the method further includes: in the case that the blood oxygen saturation prediction value is less than a preset low blood oxygen threshold, performing low blood oxygen early warning.
[0061] In this embodiment, after obtaining the blood oxygen saturation prediction value at the current time, the system checks whether the prediction value is lower than the preset low blood oxygen threshold. The low blood oxygen threshold is usually set to 90%, because below this value may indicate that the user is in a low blood oxygen state and needs timely intervention.
[0062] A window with a length of is defined. When calculating the low blood oxygen recall rate indicator, for a low blood oxygen point at time , if any time in the window period centered on is correctly detected as low blood oxygen, it is considered that the low blood oxygen data point at time is successfully detected. Similarly, when calculating the low blood oxygen accuracy rate indicator, for a data point with a blood oxygen saturation prediction value less than 90% at time , if any time in the window period centered on has a blood oxygen saturation label value less than 90%, it is considered that the low blood oxygen prediction at time is accurate.
[0063] The early warning process includes generating a low blood oxygen early warning signal, which can be a simple alarm sound, vibration or visual prompt. For example, the smart bracelet can remind the user by vibration and display a red warning icon. Push the low blood oxygen early warning information to the user or relevant medical personnel. For example, send a notification through the mobile phone application, or send the early warning information to the medical monitoring system. Record the low blood oxygen early warning event, including time, predicted value and user's health data. These records can be used for subsequent analysis and research to help improve the model and early warning mechanism. Some basic recommended measures can be provided, such as suggesting the user to rest, take deep breaths or seek medical help. These measures can be provided to the user in the form of voice prompts or text messages.
[0064] Suppose the user is sleeping at night, and the smart bracelet monitors that his blood oxygen saturation predicted value is 88%, which is lower than the preset 90% low blood oxygen threshold. Trigger low blood oxygen early warning immediately: Vibration reminder: the smart bracelet vibrates to remind the user.
[0065] Visual prompt: a red warning icon is displayed on the bracelet screen, and the words "low blood oxygen early warning" are displayed.
[0066] Mobile phone notification: send a notification through the mobile phone application, content: "Your blood oxygen saturation is lower than normal level, please pay attention to rest or seek medical help." Record the event: record this low blood oxygen event, including time, predicted value and user's health data.
[0067] Recommended measures: prompt the user through voice "please try deep breathing, if the symptoms persist, please contact the doctor." It should be noted that the preferred embodiments given in this embodiment can be freely combined without conflict in logic or structure, and the present application does not limit this.
[0068] The blood oxygen saturation prediction device provided by the present application is described below, and the blood oxygen saturation prediction device described below can be referred to each other.
[0069] Please refer to Figure 2 , Figure 2 The structure diagram of the blood oxygen saturation prediction device provided by the present application.
[0070] The application further provides a blood oxygen saturation prediction device, comprising: a data determination module 201 configured to determine user health data; the user health data comprises multiple continuous time-series health signals and static health information of the user; the multiple continuous time-series health signals are physiological signals of the user collected by an intelligent wearable device; the static health information is health history information of the user; a prediction module 202 configured to input the user health data into a blood oxygen saturation prediction model at a current time to obtain a blood oxygen saturation prediction value output by the blood oxygen saturation prediction model at the current time; wherein the blood oxygen saturation prediction model at the current time is a model obtained by adjusting the blood oxygen saturation prediction model at a previous time based on a comparison analysis of a blood oxygen saturation prediction value at the previous time and a measured blood oxygen saturation value at the previous time; the blood oxygen saturation prediction model is obtained by training a deep learning neural network based on blood oxygen saturation prediction training samples; and the blood oxygen saturation prediction training samples comprise user health data samples and blood oxygen saturation prediction value samples.
[0071] As a preferred embodiment, the data determination module 201 is specifically configured to: acquire the multiple continuous time-series health signals and the static health information; the multiple continuous time-series health signals comprise a plurality of combinations of a heart rate signal, a breathing signal, a pulse signal, skin temperature and motion data; the static health information comprises a plurality of combinations of age, gender, race and medical records of the user; perform vectorization fusion processing on the multiple continuous time-series health signals and the static health information respectively to obtain a multiple continuous time-series health fusion vector and a static health information fusion vector; and splice the multiple continuous time-series health fusion vector and the static health information fusion vector to obtain the user health data.
[0072] As a preferred embodiment, the vectorization fusion processing on the multiple continuous time-series health signals and the static health information respectively to obtain the multiple continuous time-series health fusion vector and the static health information fusion vector comprises: pre-processing the multiple continuous time-series health signals to obtain pre-processed health signals; the pre-processing mode comprises uniform sampling frequency and filtering sampling; segmenting and sorting the pre-processed health signals according to sampling time to obtain segmented health signals; performing vectorization fusion on the segmented health signals using a signal vectorization fusion method to obtain the multiple continuous time-series health fusion vector; the signal vectorization fusion method comprises average pooling, rearrangement mapping and mapping addition; classifying the static health information into an encoded representation; performing feature vector mapping on the code to obtain the static health information fusion vector.
[0073] As a preferred embodiment, the blood oxygen saturation prediction model adopts a LSTM, a GNN, a Bert, a GPT or a T5 model based on a Transformer; the prediction module 202 is specifically configured to: add the user health data and the position vector to obtain an input signal vector; input the input signal vector into an encoder of the blood oxygen saturation prediction model to obtain a state vector of a preset length output by the encoder as a predicted current blood oxygen saturation sequence result; wherein the encoder comprises a multi-head self-attention module, a residual connection module, a normalization module, a full connection layer and an activation layer; the multi-head self-attention module is configured to perform parallel processing on the time sequence features in the input signal vector through a multi-head self-attention mechanism, extract feature information in different subspaces, and output multi-head self-attention features; the residual connection module is configured to perform residual connection on the multi-head self-attention features and the time sequence features in the input signal vector to output features after residual connection; the normalization module is configured to perform normalization processing on the features after residual connection to output normalized features; the full connection layer is configured to integrate and convert the normalized features to output blood oxygen saturation related features; and the activation layer is configured to introduce a non-linear factor into the blood oxygen saturation related features through a non-linear activation function to output the predicted current blood oxygen saturation sequence result.
[0074] As a preferred embodiment, the prediction module 202 is specifically further configured to: in the case that the blood oxygen saturation prediction model has a decoder, input the state vector output by the encoder into the decoder to obtain a first blood oxygen saturation future prediction value sequence generated by the decoder in a self-recurrent cycle; and in the case that the blood oxygen saturation prediction model does not have the decoder, perform average pooling on the state vector output by the encoder and input into a pre-trained linear layer to obtain a second blood oxygen saturation future prediction value sequence.
[0075] As a preferred embodiment, the application further comprises a warning module, which is configured to perform low blood oxygen warning in the case that the blood oxygen saturation prediction value is less than a preset low blood oxygen threshold.
[0076] The intelligent wearable device provided by the application is described below, and the intelligent wearable device described below can be correspondingly referred to the blood oxygen saturation prediction method described above.
[0077] The application further provides an intelligent wearable device comprising the blood oxygen saturation prediction device described above.
[0078] Figure 3 An example of an entity structure schematic diagram of an electronic device is shown in FIG. 1. Figure 3As shown, the electronic device can include a processor 301, a communications interface 302, a memory 303, and a communications bus 304, wherein the processor 301, the communications interface 302, and the memory 303 complete mutual communication through the communications bus 304. The processor 301 can invoke a logical instruction in the memory 303 to execute a blood oxygen saturation prediction method, which includes determining user health data; the user health data is fused with multiple continuous time series health signals and static health information of the user; the multiple continuous time series health signals are physiological signals of the user collected through an intelligent wearable device; the static health information is health history information of the user; inputting the user health data into a blood oxygen saturation prediction model at a current time to obtain a blood oxygen saturation prediction value output by the blood oxygen saturation prediction model at the current time; wherein the blood oxygen saturation prediction model at the current time is a model after parameter adjustment of a blood oxygen saturation prediction model at a previous time based on comparison and analysis of a blood oxygen saturation prediction value at the previous time and a blood oxygen saturation measured value at the previous time; the blood oxygen saturation prediction model is obtained by training a deep learning neural network based on blood oxygen saturation prediction training samples; and the blood oxygen saturation prediction training samples include user health data samples and blood oxygen saturation prediction value samples.
[0079] In addition, the logical instruction in the memory 303 described above can be implemented in the form of a software functional unit and sold or used as an independent product, and can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or part of the technical solutions can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the methods described in various embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various program code storage media.
[0080] In another aspect, the present application also provides a computer program product comprising a computer program, which can be stored on a non-transitory computer readable storage medium, and the computer program can be executed by a processor to enable a computer to perform the blood oxygen saturation prediction method provided by the above-mentioned methods, which comprises: determining user health data; the user health data fuses multiple continuous time series health signals and static health information of the user; the multiple continuous time series health signals are physiological signals of the user collected by an intelligent wearable device; the static health information is health history information of the user; inputting the user health data into a blood oxygen saturation prediction model at a current time to obtain a blood oxygen saturation prediction value output by the blood oxygen saturation prediction model at the current time; wherein the blood oxygen saturation prediction model at the current time is a model after parameter adjustment of a blood oxygen saturation prediction model at a previous time based on comparative analysis of the blood oxygen saturation prediction value at the previous time and a measured blood oxygen saturation value at the previous time; the blood oxygen saturation prediction model is obtained by training a deep learning neural network based on blood oxygen saturation prediction training samples; and the blood oxygen saturation prediction training samples comprise user health data samples and blood oxygen saturation prediction value samples.
[0081] In yet another aspect, the present application also provides a non-transitory computer readable storage medium having a computer program stored thereon, which can be executed by a processor to implement the blood oxygen saturation prediction method provided by the above-mentioned methods, which comprises: determining user health data; the user health data fuses multiple continuous time series health signals and static health information of the user; the multiple continuous time series health signals are physiological signals of the user collected by an intelligent wearable device; the static health information is health history information of the user; inputting the user health data into a blood oxygen saturation prediction model at a current time to obtain a blood oxygen saturation prediction value output by the blood oxygen saturation prediction model at the current time; wherein the blood oxygen saturation prediction model at the current time is a model after parameter adjustment of a blood oxygen saturation prediction model at a previous time based on comparative analysis of the blood oxygen saturation prediction value at the previous time and a measured blood oxygen saturation value at the previous time; the blood oxygen saturation prediction model is obtained by training a deep learning neural network based on blood oxygen saturation prediction training samples; and the blood oxygen saturation prediction training samples comprise user health data samples and blood oxygen saturation prediction value samples.
[0082] The device embodiments described above are only schematic, wherein the units shown as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e., can be located in one place, or can be distributed on multiple network units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the present embodiment scheme. Those skilled in the art can understand and implement it without creative labor.
[0083] Those skilled in the art can clearly understand the implementation of the embodiments by the description of the above embodiments, and the embodiments can be implemented by means of software and necessary universal hardware platforms, and of course, can also be implemented by hardware. Based on such understanding, the above technical solutions can be embodied in the form of a software product, and the computer software product can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, etc., and includes a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the method described in each embodiment or some parts of the embodiment.
[0084] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for some technical features thereof; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.
Claims
1. A method of blood oxygen saturation prediction, characterized by, The method comprises the following steps: determining user health data; the user health data is fused with multiple continuous time series health signals of the user and static health information of the user; the multiple continuous time series health signals are physiological signals of the user collected by an intelligent wearable device; the static health information is health history information of the user; inputting the user health data into a blood oxygen saturation prediction model at a current time to obtain a blood oxygen saturation prediction value output by the blood oxygen saturation prediction model at the current time; wherein, the blood oxygen saturation prediction model at the current time is a model obtained by adjusting the blood oxygen saturation prediction model at the previous time based on a comparison and analysis of a blood oxygen saturation prediction value at the previous time and a measured blood oxygen saturation value at the previous time; the blood oxygen saturation prediction model is obtained by training a deep learning neural network based on blood oxygen saturation prediction training samples; the blood oxygen saturation prediction training samples include user health data samples and blood oxygen saturation prediction value samples.
2. The method of claim 1, wherein, The determination of the user health data comprises: obtaining the multiple continuous time series health signals and the static health information; the multiple continuous time series health signals include a combination of multiple types of signals such as heart rate signals, respiration signals, pulse signals, skin temperature and motion data; the static health information includes a combination of multiple types of information such as age, gender, race and medical history of the user; vectorizing and fusing the multiple continuous time series health signals and the static health information respectively to obtain a multiple continuous time series health fusion vector and a static health information fusion vector; splicing the multiple continuous time series health fusion vector and the static health information fusion vector to obtain the user health data.
3. The method of claim 1, wherein, The vectorizing and fusing of the multiple continuous time series health signals and the static health information respectively to obtain the multiple continuous time series health fusion vector and the static health information fusion vector comprises: preprocessing the multiple continuous time series health signals to obtain preprocessed health signals; the preprocessing method includes uniform sampling frequency and filtering sampling; segmenting and sorting the preprocessed health signals according to sampling time to obtain segmented health signals; using a signal vectorization fusion method to vectorize and fuse the segmented health signals to obtain the multiple continuous time series health fusion vector; the signal vectorization fusion method includes average pooling, rearrangement mapping and mapping addition; classifying and converting the static health information into an encoded representation; mapping the encoded representation into a feature vector and splicing to obtain the static health information fusion vector.
4. The method of claim 1, wherein, The blood oxygen saturation prediction model adopts LSTM, GNN, Bert, GPT and T5 models based on Transformer; the inputting of the user health data into the blood oxygen saturation prediction model at the current time to obtain the blood oxygen saturation prediction value output by the blood oxygen saturation prediction model at the current time comprises: adding the user health data and a position vector to obtain an input signal vector; Input the input signal vector into an encoder of the blood oxygen saturation prediction model, obtain a state vector of a preset length output by the encoder as a predicted current blood oxygen saturation sequence result; The encoder comprises a multi-head self-attention module, a residual connection module, a normalization module, a full connection layer and an activation layer. The multi-head self-attention module is configured to perform parallel processing on the time sequence features in the input signal vector through a multi-head self-attention mechanism, extract feature information in different subspaces, and output multi-head self-attention features. The residual connection module is configured to perform residual connection between the multi-head self-attention features and the time sequence features in the input signal vector, and output the residual connected features. The normalization module is configured to perform normalization processing on the residual connected features, and output normalized features. The full connection layer is configured to integrate and convert the normalized features, and output blood oxygen saturation related features. The activation layer is configured to introduce a non-linear factor into the blood oxygen saturation related features through a non-linear activation function, and output the predicted current blood oxygen saturation sequence result.
5. The method of claim 4, wherein, In the case that the blood oxygen saturation prediction model has a decoder, input the state vector output by the encoder into the decoder, and obtain a first blood oxygen saturation future prediction value sequence generated by the decoder in an autoregressive cycle. In the case that the blood oxygen saturation prediction model does not have a decoder, perform average pooling on the state vector output by the encoder, and input it into a pre-trained linear layer to obtain a second blood oxygen saturation future prediction value sequence. After inputting the user health data into the blood oxygen saturation prediction model at the current time, obtaining the blood oxygen saturation prediction value output by the blood oxygen saturation prediction model at the current time, further comprising:
6. The method of claim 1 to 5, wherein, In the case that the blood oxygen saturation prediction value is less than a preset low blood oxygen threshold, a low blood oxygen warning is performed. Comprising:
7. A blood oxygen saturation prediction apparatus characterized by comprising: A data determination module configured to determine user health data; The user health data is fused with multiple continuous time sequence health signals and static health information of a user; the multiple continuous time sequence health signals are physiological signals of the user collected by an intelligent wearable device; The static health information is health history information of the user; A prediction module configured to input the user health data into a blood oxygen saturation prediction model at a current time, and obtain a blood oxygen saturation prediction value output by the blood oxygen saturation prediction model at the current time; The blood oxygen saturation prediction model at the current time is a model obtained by adjusting parameters of a blood oxygen saturation prediction model at a previous time based on a comparison analysis of the blood oxygen saturation prediction value at the previous time and a measured blood oxygen saturation value at the previous time; the blood oxygen saturation prediction model is obtained by training a deep learning neural network based on blood oxygen saturation prediction training samples; the blood oxygen saturation prediction training samples comprise user health data samples and blood oxygen saturation prediction value samples. 8. An intelligent wearable device, characterized by, The blood oxygen saturation prediction device according to claim 7.
9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and running on the processor, characterized in that, The processor realizes the blood oxygen saturation prediction method according to any one of claims 1 to 6 when executing the computer program.
10. A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program realizes the blood oxygen saturation prediction method according to any one of claims 1 to 6 when executed by the processor.