Hemodynamics prediction model based on self-supervised learning and construction method and application thereof
Through self-supervised learning technology, using waveform segmented data and sliding time windows, the problem of scarcity of data in the field of hemodynamics is solved, efficient and accurate CO monitoring of cardiac output, and the generalization of the model is improved. sex and predictive effects.
Patent Information
- Application Number
- CN202410739976.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-07
- Publication Date
- 2024-11-08
AI Technical Summary
In the existing hemodynamic monitoring, AI modeling relies on a large number of labeled data, especially in the field of hemodynamics, with fewer data samples and difficult acquisition, resulting in poor generalization performance of deep learning models and difficult to achieve accurate cardiac output. CO monitoring.
Using a hemodynamic prediction model based on self-supervised learning, the prediction model is trained through waveform segmented data, and fine-tuned with a small amount of label data. Combined with sliding time window and random shading technology, the model performance is improved so that it can be achieved under a small amount of data. The effect of training large amounts of data.
It significantly improves the generalization and prediction effect of the model, is inexpensive and efficient, and can accurately and quickly obtain hemodynamic parameters, especially the prediction of cardiac output CO, to achieve clinically acceptable accuracy.
Smart Images

Figure CN118919081A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of hemodynamics, and in particular relates to a hemodynamic prediction model based on self-supervised learning, and a construction method and application thereof. Background Art
[0002] Hemodynamic monitoring is a quantitative, dynamic, and continuous measurement, analysis, and study of the regularity of blood movement in the body's circulatory system. It is a functional indicator that reflects the oxygen supply and consumption of the heart, blood vessels, blood, and tissues. It can provide a digital basis for clinical anesthesia and clinical treatment, and plays an important role in perioperative, intensive care, and emergency care. It is an important tool for the diagnosis and management of critically ill patients.
[0003] In hemodynamic monitoring, in addition to heart rate HR, blood pressure BP (systolic blood pressure SBP, mean arterial pressure MAP, diastolic blood pressure DBP) and routine vital sign parameter monitoring, the most important parameter is CO (cardiac output), as shown in Table 1. Other parameters such as cardiac index CI, cardiac stroke volume SV, stroke volume index SI, systemic vascular resistance SVR, systemic vascular resistance index SVRI, pulse pressure variation PPV, stroke volume variation SVV, etc. can be derived based on the above parameters through definitions and formulas. Most parameters are related to CO values. Therefore, accurate monitoring of CO parameters determines the accuracy of the entire hemodynamic monitoring.
[0004] Table 1 Calculation methods for different hemodynamic monitoring parameters
[0005]
[0006] CO monitoring can be divided into invasive, minimally invasive and non-invasive monitoring according to whether the monitoring requires invasion of the body; according to the monitoring principle, it can be divided into Fick method, thermodilution method, reagent dilution method, arterial pressure analysis method, electrical impedance method, etc. Among them, the thermodilution method is an invasive method that requires the insertion of a pulmonary artery floating catheter for measurement. It has high accuracy and is the "gold standard" method for monitoring CO. The data measured by the thermodilution method is usually expressed as COTD, but the insertion of a pulmonary artery floating catheter in the body is more traumatic and is not conducive to the patient's prognosis. In clinical practice, only extremely critical patients will be monitored by thermodilution method.
[0007] At present, the most common method for monitoring hemodynamic parameters in clinical practice is based on arterial pressure analysis, such as Edwards Lifesciences' EV1000 (minimally invasive), Vigileo (non-invasive), LiDCO's Rapid (minimally invasive, non-invasive), CNAP, PiCCO, etc. The distinction between minimally invasive and non-invasive lies mainly in whether the method of obtaining arterial pressure waveform is minimally invasive (generally radial artery blood pressure monitoring) or non-invasive. The arterial pressure waveform analysis method mainly finds the relationship between the pressure signal and the flow signal to estimate CO.
[0008] See also Figure 1 As shown in the figure, there are generally two modeling methods for arterial pressure waveform analysis: one is to use fluid mechanics modeling (simulating blood vessels as water pipes) or electrical modeling (voltage-blood pressure, current-blood flow, resistance-vascular pressure), and obtain pressure-flow estimates through professional knowledge, experience and a large amount of experimental data. This is also a common modeling solution used by hemodynamic monitor manufacturers. The other is to use AI modeling, machine learning or deep learning models, input arterial pressure waveform data, and predict hemodynamic parameters such as CO. AI modeling does not require much professional modeling knowledge and experience. Since 2019, more and more researchers have begun to study AI modeling.
[0009] Currently, the arterial pressure waveform analysis solutions using AI modeling can be divided into two types: one is feature engineering + machine learning, and the other is deep learning.
[0010] Although feature engineering + machine learning does not rely on large amounts of data, feature engineering requires human consideration of feature extraction, which has certain limitations and cannot be used to process abnormal waveform data.
[0011] Deep learning can complete abnormal waveform processing by performing a series of tasks such as denoising, feature extraction, and prediction through the end-to-end model, without being affected or limited by human factors. However, deep learning models rely on a large amount of labeled data, namely the gold standard CO-COTD. The collection and accumulation of labeled data requires a lot of resources, time, and energy. At the same time, the cost is high and the cycle is long. Specifically: First, data collection in the medical field needs to rely on clinical trials, which consumes more resources than other industries. Therefore, high-quality labeled data has always been a difficult problem in the field of AI medical care; second, in the field of hemodynamics, only extremely critically ill patients will monitor the gold standard CO monitoring, resulting in fewer data samples; third, in order to avoid causing new trauma to patients, minimally invasive monitoring or even non-invasive monitoring is the current trend of medical monitoring, resulting in a low utilization rate of CO monitoring, and the collection of gold standard data will become increasingly difficult. It is difficult to obtain enough labeled data to meet the needs of deep learning. Only limited labeled data is used for deep learning, and the resulting model has poor generalization performance, is prone to overfitting, and is unstable. Summary of the invention
[0012] In order to improve the deficiencies of the prior art, the present invention provides a hemodynamic prediction model based on self-supervised learning and a construction method and application thereof. The present invention improves the model performance through a small amount of labeled data, so that the model can also achieve the effect of large-scale data training.
[0013] In a first aspect, the present invention further provides a method for constructing a hemodynamic prediction model based on self-supervised learning, comprising the following steps:
[0014] S1. Convert the existing hemodynamic parameters into waveform segmentation data, add the waveform segmentation data to the basic model for training to generate a waveform prediction model;
[0015] S2. Using the waveform prediction model method, determine whether the result predicted by the waveform prediction model in step S1 is less than or equal to a threshold.
[0016] According to an embodiment of the present invention, step S2 is selected from at least one of the following steps:
[0017] S21. Import the existing hemodynamic parameters collected under continuous t1 into the waveform prediction model, and predict the waveform data of (t1+δt1); compare the predicted waveform data of (t1+δt) with the actual waveform data of (t1+δt1), and determine that the error is less than or equal to the threshold, and the current waveform prediction model is a qualified prediction model.
[0018] S22. Randomly cover a part of the existing hemodynamic parameters collected under continuous t1, import the unmasked part into the waveform prediction model, predict the data of the covered part, compare the predicted data with the actual data, and determine that the error is less than or equal to the threshold, and the current waveform prediction model is a qualified prediction model.
[0019] According to the embodiment of the present invention, the determination error is less than or equal to the threshold value using conventional evaluation methods in the art, such as one or more of MAE, MAE, and RMSE.
[0020] According to an implementation scheme of the present invention, step S2 also includes the following steps: comparing the predicted waveform data (t1+δt) with the actual waveform data (t1+δt1), determining that the error is greater than a threshold, the current waveform prediction model is an unqualified prediction model, and proceeding to step S1.
[0021] According to an embodiment of the present invention, step S2 further includes the following steps:
[0022] S3. Import the continuous arterial blood pressure data of the t2 time period into a qualified prediction model for training, predict one or more label data at the (t2+δt2) time point, obtain a regression model, determine that the error of the regression model is less than or equal to a threshold, and the current regression model reaches the fine-tuning target, and obtain a fine-tuning model.
[0023] According to an embodiment of the present invention, the label data is selected from one or more of CO, HR, MBP, SBP, DBP, SV, and COTD, and is preferably selected from one or more of CO, HR, MBP, SBP, DBP, SV, COTD+HR, COTD+MBP, COTD+SBP, COTD+DBP, and COTD+SV.
[0024] According to an implementation scheme of the present invention, determining whether the error of a regression model is less than or equal to a threshold value comprises the following steps: comparing one or more label data at the time point (t2+δt2) with the label data corresponding to the actual time point (t2+δt2), determining whether the error is less than or equal to the threshold value, and the current regression model reaches the fine-tuning target, thereby obtaining a fine-tuning model.
[0025] As an example, step S3 includes the following steps: taking the moment when the label data appears as the last time point, taking several consecutive seconds of arterial blood pressure data as samples, adding the samples to a qualified prediction model for training, obtaining the label data corresponding to the last time point, obtaining a regression model, determining that the error of the regression model is less than or equal to a threshold, and obtaining a fine-tuning model.
[0026] According to an implementation scheme of the present invention, step S3 also includes the following steps: comparing one or more label data at the time point (t2+δt2) with the label data corresponding to the actual time point (t2+δt2), determining that the error is greater than a threshold, the current regression model has not reached the fine-tuning target, and repeating step S3.
[0027] According to an embodiment of the present invention, the arterial blood pressure waveform data is, for example, the gold standard CO-COTD.
[0028] According to the embodiment of the present invention, the determination error is less than or equal to the threshold value using conventional evaluation methods in the art, such as one or more of MAE, MAE, and RMSE.
[0029] According to the implementation scheme of the present invention, the preset value is set according to actual needs, for example, MAE, MAE, RMSE and other methods are used for determination.
[0030] According to an implementation scheme of the present invention, step S3 also includes the following steps: importing gold standard data into a waveform prediction model for training to obtain a fine-tuning model, using the fine-tuning model to predict the CO value, determining that the error between the predicted CO value and the actual CO value is greater than a threshold, and proceeding to step S1.
[0031] According to an embodiment of the present invention, step S3 further includes the following steps:
[0032] S4. Import the hemodynamic parameters into the fine-tuning model for training to obtain the initial prediction value, import the initial prediction value into the learning model for training to obtain the final prediction value.
[0033] According to an embodiment of the present invention, step S4 also includes the following steps: importing hemodynamic parameters into the fine-tuning model for training to obtain initial prediction values, importing the initial prediction values and additional feature information into the learning model for training to obtain final prediction values.
[0034] According to an embodiment of the present invention, the t1 and t2 are independently selected from 1 to 30 s, preferably the t1 and t2 are 1 to 20 s, for example, 10 s.
[0035] According to an embodiment of the present invention, the δt1 and δt2 are independently selected from 1 to 5 s, for example, 1 s.
[0036] According to an embodiment of the present invention, generating waveform segmentation data according to existing hemodynamic parameters includes the following steps: using existing hemodynamic parameters to generate waveform segmentation data through a sliding time window.
[0037] According to an embodiment of the present invention, the initial prediction value includes feature dimension information and / or time information.
[0038] According to an embodiment of the present invention, the qualified prediction model includes an input layer, a feature extraction layer, a coding layer, and an output layer.
[0039] According to an embodiment of the present invention, the additional feature information includes one or more of basic patient information, case information, and medical imaging information, such as features analyzed from electronic case texts, features analyzed from medical images, etc.
[0040] According to an embodiment of the present invention, the learning model is selected from any machine learning model, preferably randomforest, XGBoost, etc.
[0041] According to an embodiment of the present invention, the basic model is selected from one or more of an RNN-based model (a model based on a recurrent neural network), a CNN-based model (a convolutional neural network model), a Transformer-based model (a pre-trained language model), an LLM model (a large language model), a Unified or universal time series model (a large time series model), and the like.
[0042] According to an embodiment of the present invention, the RNN-based model is selected from LSTM (Long Short-Term Memory) and the like.
[0043] According to an embodiment of the present invention, the CNN-based model is selected from Inception Time, TCN (Temporal Convolutional Network) and the like.
[0044] According to an embodiment of the present invention, the Transformer-based model is selected from TST (Time Series Transformer), patchTST (patch time series transformer), Informer (a deep learning model for time series prediction), etc.
[0045] According to an embodiment of the present invention, the LLM model is selected from, for example, a model of the Large Language Models for Time Series class (large language model for processing time series) and the like.
[0046] According to an embodiment of the present invention, the existing hemodynamic parameters are selected from one or more of heart rate HR, systolic blood pressure SBP, mean arterial pressure MBP, diastolic blood pressure DBP, vital sign parameters, and cardiac output CO.
[0047] According to an embodiment of the present invention, the final predicted value is selected from one or more of cardiac index CI, cardiac stroke volume SV, stroke volume index SI, systemic vascular pressure SVR, systemic vascular pressure index SVRI, pulse pressure variability PPV, and stroke volume variability SVV.
[0048] In a second aspect, the present invention provides a hemodynamic prediction model based on self-supervised learning, comprising:
[0049] A pre-training module, wherein the pre-training module is used to generate waveform segmentation data according to existing hemodynamic parameters, add the waveform segmentation data to the basic model for training to generate a waveform prediction model, import the existing hemodynamic parameters collected under continuous t1 into the waveform prediction model, and predict the waveform data of (t1+δt1);
[0050] A judgment module is connected to the pre-training module, and compares the predicted waveform data (t1+δt) with the actual waveform data (t1+δt1) to determine whether the error is less than or equal to a threshold. If so, the current waveform prediction model is determined to be qualified, and the current waveform prediction model is used to predict unknown hemodynamic parameters to obtain initial prediction values.
[0051] According to an embodiment of the present invention, the t1 is 1 to 30 s, preferably the t1 is 1 to 20 s, for example, 10 s.
[0052] According to an embodiment of the present invention, the δt1 is 1 to 5 s, for example, 1 s.
[0053] According to an embodiment of the present invention, the initial prediction value includes feature dimension information and / or time information.
[0054] According to an embodiment of the present invention, the waveform prediction model includes an input layer, a feature extraction layer, a coding layer, and an output layer.
[0055] According to an embodiment of the present invention, the prediction system also includes a fine-tuning module, which is used to modify at least one of the input layer, feature extraction layer, encoding layer or output layer of the qualified waveform prediction model to obtain a modified model, and output an initial prediction value predicted by the modified model.
[0056] According to an implementation scheme of the present invention, the fine-tuning module is used to import the arterial blood pressure data under continuous t2 into the waveform prediction model for training to obtain a regression model, predict one or more label data at the time point (t2+δt2), compare the predicted one or more label data at the time point (t2+δt2) with the label data corresponding to the actual time point (t2+δt2), determine that the error is less than or equal to a threshold, and the current regression model reaches the fine-tuning target, thereby obtaining a fine-tuning model.
[0057] According to an embodiment of the present invention, the t2 is 1 to 30 s, preferably the t2 is 1 to 20 s, for example, 10 s.
[0058] According to an embodiment of the present invention, the δt2 is 1 to 5 s, for example, 1 s. According to an embodiment of the present invention, the prediction system further includes a learning model, the fine-tuning model imports the initial prediction value obtained by prediction into the learning model, and the learning model is used to train the prediction data and obtain the final prediction value.
[0059] According to an embodiment of the present invention, the fine-tuning module is also used to import additional feature information into the learning model.
[0060] According to an embodiment of the present invention, the additional feature information includes one or more of basic patient information, case information, and medical imaging information, such as features analyzed from electronic case texts, features analyzed from medical images, etc.
[0061] According to an embodiment of the present invention, the learning model is selected from any machine learning model, preferably randomforest, XGBoost, CNN, etc.
[0062] According to an embodiment of the present invention, the basic model is selected from one or more of an RNN-based model (a model based on a recurrent neural network), a CNN-based model (a convolutional neural network model), a Transformer-based model (a pre-trained language model), an LLM model (a large language model), a Unified or universal time series model (a large time series model), and the like.
[0063] According to an embodiment of the present invention, the RNN-based model is selected from LSTM (Long Short-Term Memory) and the like.
[0064] According to an embodiment of the present invention, the CNN-based model is selected from Inception Time, TCN (Temporal Convolutional Network) and the like.
[0065] According to an embodiment of the present invention, the Transformer-based model is selected from TST (Time Series Transformer), patchTST (patch time series transformer), Informer (a deep learning model for time series prediction), etc.
[0066] According to an embodiment of the present invention, the LLM model is selected from, for example, a model of the Large Language Models for Time Series class (large language model for processing time series) and the like.
[0067] According to an embodiment of the present invention, the existing hemodynamic parameters are selected from one or more of heart rate HR, systolic blood pressure SBP, mean arterial pressure MBP, diastolic blood pressure DBP, vital sign parameters, and cardiac output CO. According to an embodiment of the present invention, the final predicted value is selected from one or more of cardiac index CI, cardiac stroke volume SV, stroke index SI, systemic vascular resistance SVR, systemic vascular resistance index SVRI, pulse pressure variability PPV, and stroke volume variability SVV.
[0068] The present invention also provides a hemodynamic prediction method based on self-supervised learning, comprising the following steps:
[0069] The known hemodynamic parameters of the patient to be predicted are imported into the prediction system to obtain the final predicted value.
[0070] Beneficial Effects
[0071] 1) The hemodynamic prediction system based on self-supervised learning in the present invention performs deep learning of the self-supervised process through a fine-tuning module. The model is fully trained, the performance difference between the validation set and the training set is not much, the model has good generalization, and the prediction effect is significantly improved, especially in terms of followability. The radial LOA can reach less than 30 degrees, which is within the clinically acceptable range.
[0072] 2) The hemodynamic prediction system based on self-supervised learning in the present invention can accurately and quickly obtain the final prediction value through a small amount of existing hemodynamic parameters, which is not only low-cost but also highly efficient. BRIEF DESCRIPTION OF THE DRAWINGS
[0073] Figure 1 A schematic diagram of the existing modeling method for arterial pressure waveform analysis;
[0074] Figure 2 It is a structural diagram of a method for constructing a hemodynamic prediction model based on self-supervised learning;
[0075] Figure 3 Schematic diagram of sliding time window and continuous measurement;
[0076] Figure 4 A structural diagram of a method for constructing a hemodynamic prediction model based on self-supervised learning;
[0077] Figure 5 A is a schematic diagram of the training data of the pre-training module; Figure 5 B is a schematic diagram of the training data of the fine-tuning module;
[0078] Figure 6 A flowchart of a method for constructing a hemodynamic prediction model based on self-supervised learning;
[0079] Figure 7 Flowchart for predicting data for the hemodynamic prediction model based on self-supervised learning;
[0080] Figure 8 The amount of data used to make predictions using existing forecasting models;
[0081] Fig. 9 The amount of data used for prediction using Example 4;
[0082] Fig.10 The predicted results of HR, MBP, SBP and DBP parameters in Example 4
[0083] Fig.11 A diagram comparing the training process of the hemodynamic prediction model based on self-supervised learning and the prediction model of existing products;
[0084] Fig.12 This is a comparison chart of the training results of the hemodynamic prediction model based on self-supervised learning and the existing product prediction model. DETAILED DESCRIPTION
[0085] The model of the present invention will be further described in detail below in conjunction with specific embodiments. It should be understood that the following embodiments are only exemplary illustrations and explanations of the present invention and should not be construed as limiting the scope of protection of the present invention. All technologies implemented based on the above content of the present invention are included in the scope of protection that the present invention is intended to protect.
[0086] In the description of the present invention, it should be noted that the terms "center", "up", "down", "left", "right", "vertical", "horizontal", "inside", "outside", etc. indicate directions or positional relationships based on the directions or positional relationships shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction. Therefore, they cannot be understood as limitations on the present invention.
[0087] The patient data used in the following embodiments are from VitalDB or MIMIC-IV&II database. VitalDB data is mainly used to verify the consistency of the model, and the generalization of the model needs to be demonstrated by comparing the data of VitalDB and MIMIC-IV&II.
[0088] Example 1
[0089] S101. Collect arterial blood pressure waveform data, cut the waveform data through a 1s sliding time window, and generate initial waveform segmented data. Each segment of initial waveform segmented data corresponds to a continuous 11s waveform (100Hz). If null values or outliers appear, or the frequency of the waveform does not meet 100Hz, fill them with waveform fitting to obtain 1100 continuous and complete data points; use all initial waveform segmented data as initial data, add them to the waveform prediction model for training, and obtain an initial prediction model.
[0090] S102. The waveform data is segmented through a 1s sliding time window: 10s of blood pressure waveform data (100Hz), i.e. 1000 data points, are extracted in segments; if null values or outliers appear, or the waveform does not meet the 100Hz requirement, they are filled with waveform fitting to obtain 1000 continuous and complete data points; the current 10s time window is added to the initial prediction model to predict the next 1s waveform and segmented data, i.e., based on the 1000 continuous and complete data points corresponding to the continuous 10s, the 100 continuous data points corresponding to 10+1s are predicted.
[0091] S103. Evaluate the prediction model based on MAE, determine that for the 100 consecutive data points predicted in step S2, the error MSE of each data point is less than 2 mmHg, and proceed to step S104.
[0092] S104. Using the initial prediction model, 900 known and continuous data points are used to sequentially predict the values of the next 100 data points. MAE is used to evaluate the prediction model. The error MSE of each point is <2 mmHg. It is believed that the prediction model's ability to interpret waveforms can meet the requirements.
[0093] S105. Use label data, i.e., CO, HR, MBP, SBP, and DBP. Take the moment when CO, HR, MBP, SBP, and DBP appear as the time point, and take the arterial blood pressure waveform data of the previous 9 seconds, i.e., 900 blood pressure data points, as a sample; use the multi-parameter regression model of the waveform (the parameters of the waveform prediction model are used to initialize the feature extraction layer, and the parameters of the task layer are randomly initialized), add the samples of the above 900 data points to the initial prediction model, train and predict the values of CO, HR, MBP, SBP, and DBP corresponding to 9+1s (the parameters of the feature extraction layer remain unchanged, and only the parameters of the task layer are trained and updated), and obtain the regression model. Use MAE to evaluate the regression model. If CO<1L / min, HR<2bpm, and MBP / SBP / DBP<2mmHg are satisfied at the same time, it is considered that the regression model has achieved the fine-tuning goal.
[0094] Example 2
[0095] This embodiment adopts the same method as step S1 of embodiment 1 to obtain the initial prediction model.
[0096] S201. Segment the waveform data through a 1s sliding time window: extract 10s of blood pressure waveform data (100Hz), i.e. 1000 data points, in segments: if null values or outliers appear, fill them with the previous and next values.
[0097] S202. Use a 1s sliding time window to randomly mask 10s waveform segments, that is, for 1000 data points, randomly mask a segment of 100 consecutive data points.
[0098] S203. Use the initial waveform prediction model to predict the values of the 100 masked data points through the remaining 900 data points, and use MAE to evaluate the prediction model. If the error MSE of each point is <2 mmHg, it is considered that the prediction model's ability to interpret the waveform can meet the requirements.
[0099] S204. Use labeled data, COTD, HR, MBP, SBP, DBP, take the moment when COTD and HR, or COTD and MBP, or COTD and SBP, or COTD and DBP appear as the last time point, and take the arterial blood pressure waveform data of the previous 9 seconds continuously, that is, 900 blood pressure data points, as a sample; use the multi-parameter regression model of the waveform (the parameters of the waveform prediction model are used to initialize the feature extraction layer, and the parameters of the task layer are randomly initialized), through the above 900 blood pressure data points, train and predict the values of CO, HR, MBP, SBP, and DBP at the last moment (the parameters of the feature extraction layer remain unchanged, only the parameters of the task layer are trained and updated), obtain the regression model, use MAE to evaluate the regression model, if CO<1L / min, HR<2bpm and MBP / SBP / DBP<2mmHg are satisfied at the same time, it is considered that the regression model has achieved the fine-tuning goal.
[0100] Example 3
[0101] This embodiment adopts the same method as step S1 of embodiment 1 to obtain the initial prediction model.
[0102] S301. Segment the waveform data through a 1s sliding time window: extract 10s of blood pressure waveform data in segments. If the blood pressure waveform is not 100Hz, use a waveform fitting solution to scale the waveform up / down to 100Hz, and then perform segmented extraction to ensure that the extracted data points are 1000.
[0103] S302. Use a 1s sliding time window to randomly mask 10s waveform segments, that is, for 1000 data points, randomly mask a segment of 100 consecutive data points.
[0104] S303. Use the initial waveform prediction model to predict the values of the 100 masked continuous data points through the remaining 900 data points, and use MAE to evaluate the prediction model. If the error MSE of each point is <2 mmHg, it is considered that the prediction model's ability to interpret the waveform can meet the requirements.
[0105] S304. Use labeled data, namely COTD and HR, MBP, SBP, DBP, and take the moment when COTD and HR, MBP, SBP, DBP appear as the last time point, and take the arterial blood pressure waveform data of the previous 9 seconds, namely 900 blood pressure data points, as a sample; use the multi-parameter regression model of the waveform (the parameters of the waveform prediction model are used to initialize the feature extraction layer, and the parameters of the task layer are randomly initialized), add the samples of the above 900 data points to the initial prediction model, train and predict the values of CO and HR, MBP, SBP, DBP corresponding to 9+1s (the parameters of the feature extraction layer remain unchanged, and only the parameters of the task layer are trained and updated), and obtain the regression model. Use MAE to evaluate the regression model. If CO<1L / min, HR<2bpm and MBP / SBP / DBP<2mmHg are satisfied at the same time, it is considered that the regression model has achieved the fine-tuning goal.
[0106] Example 4
[0107] See also Figure 8 As shown, 6388 cases in the VitalDB database were used as initial data, 2881 cases without ABP waveforms were deleted, 7 cases with noisy ABP waveforms were deleted, and 3500 cases with ABP waveforms that met the requirements were extracted; the training models used in this comparative example were the FC layers and Random forest models in the self-supervised learning SSHemo model, and the two models were used for training respectively.
[0108] Self-supervised learning: Based on the patient classification mapping table, 3500 medical records were split into a training set (including 2800 medical records), a validation set (including 350 medical records), and a test set (including 350 medical records). The ABP waveforms in the training set, validation set, and test set were sliced, and the slice results of the training set and validation set were used as training samples, and the slice results of the test set were used as test samples. The distribution was added to the training model for training, see Fig.11 As shown, the total number of training times is 111, and the total training time is 8 hours, 7 minutes and 15 seconds; the comparative model is fully trained (it has fully learned the information in the data), and the performance of the validation set and the training set is not much different (the model has good generalization).
[0109] See also Fig.12As shown, in the test results of this embodiment, the mean square errors (MSE) of the two self-supervised training models are 3.87 (L / min)2 and 2.55 (L / min)2, respectively, the root mean square errors (RMSE) are 1.97L / min and 1.60L / min, respectively, and the mean absolute errors (MAE) are 1.57L / min and 1.25L / min, respectively.
[0110] Comparative Example 1
[0111] See also Fig. 9 As shown, 6388 cases in the VitalDB database were used as initial data, 2881 cases without ABP waveforms were deleted, 7 cases with noisy ABP waveforms were deleted, and 3500 cases with ABP waveforms that met the requirements were extracted.
[0112] Based on the patient classification mapping table, 3500 medical records were split into a training set (including 47 medical records), a validation set (including 5 medical records), and a test set (including 5 medical records). The ABP waveforms in the training set, validation set, and test set were sliced, and the slice results of the training set and validation set were used as training samples, and the slice results of the test set were used as test samples. The distribution was added to the supervised learning model for training, see Fig.11 As shown, the total number of training times is 13 times, and the total training time is 20 minutes and 53 seconds; the model training in this embodiment is quickly fitted, the training is insufficient (the information in the data is not fully learned), and the performance of the verification set fluctuates and does not converge (the model generalization is poor).
[0113] See also Fig.12 As shown, in the test results of this comparative example, the mean square error (MSE) is 4.81 (L / min)2, the root mean square error (RMSE) is 2.19 L / min, and the mean absolute error (MAE) is 1.63 L / min.
[0114] The above examples are used to illustrate the specific implementation of the present invention. However, the protection scope of the present invention is not limited to the above exemplary implementation. Any modification, equivalent substitution, improvement, etc. made by those skilled in the art within the spirit and principle of the present invention shall be included in the protection scope of the claims of the present invention.
Claims
1. A method for constructing a hemodynamic prediction model based on self-supervised learning, characterized in that: The steps include: S1. Convert the existing hemodynamic parameters into waveform segmentation data, add the waveform segmentation data to the basic model for training to generate a waveform prediction model; S2. Using the waveform prediction model method, determine whether the result predicted by the waveform prediction model in step S1 is less than or equal to a threshold. Preferably, step S2 is selected from at least one of the following steps: S21. Import the existing hemodynamic parameters collected under continuous t1 into the waveform prediction model, and predict the waveform data of t1+δt1; compare the predicted waveform data of t1+δt with the actual waveform data of t1+δt1, and determine that the error is less than or equal to the threshold, and the current waveform prediction model is a qualified prediction model. S22. Randomly cover a part of the existing hemodynamic parameters collected under continuous t1, import the unmasked part into the waveform prediction model, predict the data of the covered part, compare the predicted data with the actual data, and determine that the error is less than or equal to the threshold, and the current waveform prediction model is a qualified prediction model.
2. The method for constructing a hemodynamic prediction model based on self-supervised learning according to claim 1, characterized in that: After step S2, the following steps are also included: S3. Import the continuous arterial blood pressure data of the t2 time period into a qualified prediction model for training, predict one or more label data at the (t2+δt2) time point, obtain a regression model, determine that the error of the regression model is less than or equal to a threshold, and the current regression model reaches the fine-tuning target, and obtain a fine-tuning model. Preferably, the label data is selected from one or more of CO, HR, MBP, SBP, DBP, SV, and COTD. Preferably, determining whether the error of the regression model is less than or equal to a threshold value includes the following steps: comparing one or more label data at the time point (t2+δt2) with the label data corresponding to the actual time point (t2+δt2), determining that the error is less than or equal to the threshold value, and the current regression model reaches the fine-tuning target, thereby obtaining a fine-tuning model.
3. The method for constructing a hemodynamic prediction model based on self-supervised learning according to claim 1, characterized in that: After step S3, the following steps are also included: S4. Import the hemodynamic parameters into the fine-tuning model for training to obtain the initial prediction value, import the initial prediction value into the learning model for training to obtain the final prediction value. Preferably, step S4 also includes the following steps: importing hemodynamic parameters into the fine-tuning model for training to obtain initial prediction values, importing the initial prediction values and additional feature information into the learning model for training to obtain final prediction values.
4. The method for constructing a hemodynamic prediction model based on self-supervised learning according to any one of claims 1 to 3, characterized in that: Step S2 also includes the following steps: comparing the predicted waveform data of t1+δt with the actual waveform data of t1+δt1, determining that the error is greater than a threshold, and the current waveform prediction model is an unqualified prediction model, and proceeding to step S1.
5. The method for constructing a hemodynamic prediction model based on self-supervised learning according to any one of claims 1 to 3, characterized in that: Generating waveform segmentation data according to existing hemodynamic parameters includes the following steps: using existing hemodynamic parameters to generate waveform segmentation data through a sliding time window. Preferably, the initial prediction value includes feature dimension information and / or time information. Preferably, the additional characteristic information includes one or more of basic patient information, case information, and medical imaging information. Preferably, the existing hemodynamic parameters are selected from one or more of heart rate HR, systolic blood pressure SBP, mean arterial pressure MBP, diastolic blood pressure DBP, vital sign parameters, and cardiac output CO. Preferably, the final predicted value is selected from one or more of cardiac index CI, cardiac stroke volume SV, stroke volume index SI, systemic vascular pressure SVR, systemic vascular pressure index SVRI, pulse pressure variability PPV, and stroke volume variability SVV.
6. A hemodynamic prediction model based on self-supervised learning, characterized in that: The prediction model includes: A pre-training module, wherein the pre-training module is used to generate waveform segmentation data according to existing hemodynamic parameters, add the waveform segmentation data to the basic model for training to generate a waveform prediction model, import the existing hemodynamic parameters collected under continuous t1 into the waveform prediction model, and predict the waveform data of t1+δt1; A determination module is connected to the pre-training module and is used to compare the waveform data of the predicted t1+δt with the waveform data of the actual t1+δt1 to determine whether the error is less than or equal to a threshold. If so, the current waveform prediction model is determined to be qualified, and the current waveform prediction model is used to predict unknown hemodynamic parameters to obtain initial prediction values.
7. The hemodynamic prediction model based on self-supervised learning according to claim 6, characterized in that: The t1 is 1 to 30 s, and the δt1 is 1 to 5 s. Preferably, the waveform prediction model includes an input layer, a feature extraction layer, a coding layer, and an output layer. Preferably, the prediction model also includes a fine-tuning module, which is used to modify at least one of the input layer, feature extraction layer, encoding layer or output layer of the qualified waveform prediction model to obtain a modified model, and output an initial prediction value predicted by the modified model. Preferably, the fine-tuning module is used to import the arterial blood pressure data under continuous t2 into the waveform prediction model for training to obtain a regression model, predict one or more label data at the time point (t2+δt2), compare the predicted one or more label data at the time point (t2+δt2) with the label data corresponding to the actual time point (t2+δt2), determine that the error is less than or equal to a threshold, and the current regression model reaches the fine-tuning target, thereby obtaining a fine-tuning model.
8. The hemodynamic prediction model based on self-supervised learning according to claim 7, characterized in that: The prediction model also includes a learning module. The fine-tuning model imports the initial prediction value obtained by prediction into the learning module. The learning module is used to train the prediction data and obtain the final prediction value.
9. The hemodynamic prediction model based on self-supervised learning according to claim 7, characterized in that: The fine-tuning module is also used to import additional feature information into the learning module, and the additional feature information includes one or more of basic patient information, case information, and medical imaging information. Preferably, the learning module is selected from any machine learning model, preferably random forest, XGBoost, or CNN. Preferably, the existing hemodynamic parameters are selected from one or more of heart rate HR, systolic blood pressure SBP, mean arterial pressure MBP, diastolic blood pressure DBP, vital sign parameters, and cardiac output CO. Preferably, the final predicted value is selected from one or more of cardiac index CI, cardiac stroke volume SV, stroke volume index SI, systemic vascular resistance SVR, systemic vascular resistance index SVRI, pulse pressure variability PPV, and stroke volume variability SVV.
10. A method for predicting hemodynamic parameters based on the prediction model according to any one of claims 6 to 9, comprising the following steps: The known hemodynamic parameters of the patient to be predicted are imported into the prediction model to obtain the final prediction value.
Citation Information
Cited By
Method, system and equipment for calculating hemodynamic parameters
CN117954107A
