Personalized intraoperative hypotension prediction method and system based on continuous learning, electronic equipment and readable storage medium
Through personalized instance multi-task fine-tuning and self-supervised learning methods, the personalized adaptability and real-time problems of the intraoperative hypotension prediction model are solved, and efficient and flexible hypotension prediction is achieved, which is suitable for different patients and clinical scenarios.
Patent Information
- Application Number
- CN202510754305.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-06
- Publication Date
- 2025-09-19
AI Technical Summary
Existing intraoperative hypotension prediction models lack personalized adaptability, poor model generalization, and insufficient real-time and continuous learning capabilities, making it difficult to adapt to the differences between different patients and clinical scenarios.
By fine-tuning the time series model based on personalized instance multi-tasks, combining self-supervised learning and online continuous learning, and using the patient's previous and real-time blood pressure data to update the model parameters, dynamic adaptation to individual physiological characteristics and time series data can be achieved.
It improves the accuracy and robustness of hypotension prediction, can update and optimize the model in real time, adapt to changes in the patient's physiological state, reduce data requirements and computing costs, and enhance the applicability of the model in different clinical scenarios.
Smart Images
Figure CN120673984A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a personalized intraoperative hypotension prediction method, system, electronic device, and readable storage medium based on continuous learning, and belongs to the technical field of medical data processing. Background Art
[0002] Intraoperative hypotension (IOH) refers to a drop in a patient's arterial blood pressure to a clinically unsafe level during surgery. Hypotension often leads to hypoperfusion of vital organs during surgery, increasing the risk of adverse postoperative outcomes such as acute kidney injury, cardiac complications, and a higher postoperative mortality rate. Therefore, early prediction and intervention of intraoperative hypotension are crucial. Its essence lies in monitoring and analyzing the patient's physiological signals to predict the time and probability of hypotension in real time, thereby providing anesthesiologists and surgical teams with timely intervention recommendations to ensure patient safety.
[0003] However, most existing studies tend to train a unified model on a fixed, universal hypotension dataset and apply this model to all patients to predict the occurrence of hypotension. Although such methods can provide basic predictive capabilities for some patients, they ignore individual differences. In particular, the performance of the model is limited when there is a shift in the distribution of the patient's physiological characteristics and intraoperative time series data. Therefore, such methods often fail to provide accurate predictions when faced with personalized data, and the effectiveness of the model is also relatively limited in practical applications, failing to fully capture the unique physiological and time series characteristics of each patient.
[0004] Current research status and limitations:
[0005] Research on intraoperative hypotension has primarily focused on building models based on physiological data collected during surgery (such as blood pressure, heart rate, and respiratory rate) to predict the risk of hypotension. Existing research can be broadly categorized into the following:
[0006] Prediction models based on statistical methods: These methods use traditional statistical techniques, such as regression analysis and time series analysis, to predict hypotensive events. Although these methods perform well in simple scenarios, they often have difficulty capturing complex nonlinear relationships in complex and dynamically changing surgical environments.
[0007] Machine learning-based prediction models include support vector machines (SVMs), random forests (RFs), and gradient boosted trees (GBTs). These models learn patterns from large amounts of patient data to make predictions, resulting in higher accuracy than statistical methods. However, these models typically rely on fixed datasets for training and lack the ability to adapt to individual differences.
[0008] Deep learning-based prediction models, such as recurrent neural networks (RNNs), long short-term memory networks (LSTMs), and convolutional neural networks (CNNs), can automatically extract complex temporal features, significantly improving prediction performance. However, they also suffer from insufficient generalization and poor adaptability to personalized data, making them difficult to effectively transfer and adapt across different patients.
[0009] Although some progress has been made in the study of intraoperative hypotension prediction, current technology has obvious limitations in the following aspects:
[0010] Insufficient personalization capabilities: Most existing models lack the ability to adapt to individual differences and are difficult to adjust according to each patient's unique physiological characteristics and surgical dynamics, resulting in limited prediction accuracy.
[0011] Poor model generalization: Existing models are usually trained on specific datasets and are difficult to adapt to the data distribution of different hospitals, different surgical types, and different patient groups, limiting their widespread application in actual clinical practice.
[0012] Insufficient real-time and continuous learning capabilities: During surgery, the patient's physiological state may change rapidly. Existing models often lack the ability to update in real time and continuously learn, and are unable to adjust prediction results in time to cope with dynamic changes.
[0013] Inconsistent definitions of hypotension: Intraoperative hypotension is not uniformly defined across different clinical scenarios, making it difficult to directly transfer existing models to different tasks. Models often need to be trained from scratch for different hypotension thresholds, otherwise they will struggle to complete new tasks. This not only increases development costs but also reduces the model's universality and clinical value. Summary of the Invention
[0014] In response to the problems existing in the above-mentioned prior art, the present invention provides a personalized intraoperative hypotension prediction method, system, electronic device, and readable storage medium based on continuous learning. The present invention solves the shortcomings of the above-mentioned existing intraoperative hypotension prediction methods in terms of personalized adaptability, model generalization, real-time update capability, and applicability under different clinical definitions.
[0015] The technical solution of the present invention is: a personalized intraoperative hypotension prediction method based on continuous learning, the method comprising:
[0016] Step 1: Fine-tune the time series model based on personalized instance multi-task: Update the pre-trained time series model parameters by using the patient's blood pressure time series data before surgery;
[0017] Step 2, online continuous learning: Before intraoperative hypotension prediction, the time series model after updating the model parameters continues to use the reconstruction self-supervised learning algorithm to further fine-tune the time series model parameters using real-time blood pressure data;
[0018] Step 3. Prediction of intraoperative hypotension: The blood pressure sequence to be predicted is input into the fine-tuned time series model. The fine-tuned time series model outputs the predicted blood pressure sequence. The blood pressure sequence output by the fine-tuned time series model is then checked according to the standard definition of hypotension to obtain a prediction of whether hypotension will occur.
[0019] Furthermore, the Step 1 includes:
[0020] Step 1.1, transfer learning: use the pre-trained time series model Units model for fine-tuning;
[0021] Step 1.2, Model fine-tuning: Use the patient's previous blood pressure time series data to fine-tune the parameters of the pre-trained Units model to ensure that the model adapts to individual physiological characteristics and time series data distribution;
[0022] Step 1.3. Introduce self-supervised learning tasks: During the fine-tuning process, combine it with the mask reconstruction task, that is, randomly mask the input data and let the Units model learn to reconstruct the masked data.
[0023] Furthermore, in Step 1.3, each training sample (x, y) of the self-supervised learning task includes two training losses:
[0024] Supervised learning loss: The Units model predicts the future time series based on the input data x, obtains the predicted value y′, and calculates the mean squared error between the predicted value y′ and the true value y;
[0025] Self-supervised reconstruction loss: Mask the input data x, reconstruct it using the Units model, and calculate the mean square error between the reconstructed data and the original data;
[0026] Taking into account the losses of the time series prediction task and the mask reconstruction task, the loss of fine-tuning the Units model based on personalized instance multi-task is formally expressed as follows:
[0027]
[0028] Among them, θ units Represents the backbone network of the Units model; θ fore ,θ imp Represents the learnable prompt tokens for time series prediction and mask reconstruction tasks respectively; N refers to N training samples; θ fore,unitsRepresents the model parameters used for time series prediction tasks; θ imp,units Represents the model parameters used for mask reconstruction tasks; θ fore,units (x) represents the process of time series prediction by the model; θ imp,units (x') represents the process of mask reconstruction task; MSE is the mean square error, which calculates the square of the error between the true value and the predicted value at each moment and then takes the average. The expression of MSE is: where Y i It represents the true value at the i-th moment, Y i ' is the predicted value at the i-th moment.
[0029] Furthermore, the Step 2 includes:
[0030] The input data x is copied, which is the data of a single observation window, to the number of batch size, and then randomly masking the data of the entire batch size, the backbone network θ units Update the weights on a batch of data to get the updated backbone network θ' units , the learning loss is as follows:
[0031]
[0032] Updated backbone network θ' units +θ fore Make the final low blood pressure prediction to achieve online continuous learning.
[0033] Furthermore, in Step 3, the intraoperative hypotension prediction includes:
[0034] The data of the blood pressure observation window of a predefined length is input into the fine-tuned time series model. The fine-tuned time series model predicts a blood pressure sequence of a specific length. The blood pressure value at each time point in the predicted sequence is checked according to the standard definition of hypotension, and the prediction of whether hypotension will occur is output.
[0035] Furthermore, in Step 3, the fine-tuned time series model processes the blood pressure sequence to be predicted by:
[0036] The time series model uses the Units model. The Units model converts raw time series data into three types of tokens to achieve unified processing of various tasks: sequence token, prompt token, and task token.
[0037] The token types are described as follows:
[0038] Sequence tokens correspond to the original time series data. The original sequence is divided into multiple patches by using non-overlapping windows. Each patch is mapped through a multi-layer perceptron (MLP) to generate an embedded representation of the sequence token. The dimensionality is formally expressed as follows:
[0039] Z x ∈R s×v×d
[0040] Among them, s is the number of segments after being divided into patches, v is the number of time series variables, and d is the dimension of the embedding vector;
[0041] Prompttoken generates several learnable prompt embeddings and concatenates them with sequence tokens to absorb contextual information in the sequence based on task characteristics. Each task corresponds to its own unique prompt embedding. The dimensions are formally expressed as follows:
[0042] Z p ∈R p×v×d
[0043] Among them, p is the number of prompttokens, v is the number of time series variables, and d is the dimension of the embedding vector.
[0044] Task tokens are used to unify different tasks, including Gentokens and CLS tokens. Gentokens is used to unify time series prediction and time series filling. The results generated by Gen are prediction results. CLS tokens are used for tasks related to time series classification. The dimensions are formally expressed as follows:
[0045] Z m ∈R 1×v×d
[0046] For tasks involving prediction, the GEN token Z is copied according to the desired prediction length m f times, get Z m ∈R f ×v×d ; These features are then concatenated with the sample and prompt tags and fed into the Units model for time series prediction; the data dimensions fed into the Units model are formally expressed as follows:
[0047] Z fore =Unit(CA(Z p ,Z x ,Z m ,))∈R (p+s+f)×v×d .
[0048] The present invention also provides a personalized intraoperative hypotension prediction system based on continuous learning, which includes: a module for executing the personalized intraoperative hypotension prediction method based on continuous learning.
[0049] The present invention also 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 personalized intraoperative hypotension prediction method based on continuous learning when executing the program.
[0050] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for predicting personalized intraoperative hypotension based on continuous learning is implemented.
[0051] The present invention also provides a computer program product, comprising a computer program, which, when executed by a processor, implements the personalized intraoperative hypotension prediction method based on continuous learning.
[0052] Unlike existing methods, the present invention does not perform predictions after uniformly training and fixing model parameters on a dataset of all patients. Instead, it performs personalized fine-tuning on the basis of a pre-trained time series model, and uses the time series representation capabilities learned by the model on time series data in different fields for transfer learning. Specifically, personalized fine-tuning uses the patient's early blood pressure data during surgery to dynamically update the parameters of the pre-trained time series model, thereby achieving predictions of the patient's later blood pressure time series. In addition, before the model makes a prediction, the method of the present invention also uses a self-supervised learning algorithm to fine-tune some of the model's parameters on the input observation window data, so as to further improve the model's adaptability and robustness.
[0053] In the implementation of the online continuous learning algorithm, the present invention adopts the Test-time Training strategy, which enables the model parameters to be continuously learned and gradually optimized online during the operation. When using the patient's previous surgical time series data for fine-tuning, the present invention applies a supervised learning method to fit the patient's blood pressure data. In addition, it also combines self-supervised learning technology to enhance the model's learning ability for time series data through random masking of data. Through the self-supervised learning mechanism, the model can not only better understand the distribution of blood pressure time series data, but also adapt to the dynamic changes of intraoperative time series data, thereby improving the accuracy and robustness of the prediction.
[0054] The beneficial effects of the present invention are:
[0055] 1. The present invention's personalized prediction performance, based on continuous learning, is superior. Through a personalized example fine-tuning strategy, the present invention dynamically adjusts the prediction model during surgery, enabling real-time optimization based on each patient's intraoperative blood pressure data. This approach breaks through the traditional unified model and improves the accuracy and personalization of predictions based on individual patient differences.
[0056] 2. Good real-time performance and strong continuous learning capabilities: Using a test-time training strategy, the prediction model can be updated, continuously learned, and optimized in real time during surgery, adapting to rapid changes in the patient's physiological state and ensuring the real-time and accuracy of prediction results. In addition, the introduction of self-supervised learning enhances the prediction model's understanding and adaptability to time series data, further improving the accuracy and robustness of hypotension prediction.
[0057] 3. Efficient data utilization: The method of the present invention only requires a very small amount of data from each patient for personalized fine-tuning, significantly reducing data requirements and computing costs, and achieving efficient data utilization and prediction model optimization;
[0058] 4. Flexible Hypotension Determination: Through the configurable hypotension determination module, the hypotension prediction problem is regarded as a time series prediction task, which outputs a continuous blood pressure sequence and uses user-defined hypotension criteria. This allows users to customize the hypotension determination criteria according to different clinical needs and can automatically adjust the determination logic according to different clinical criteria, enhancing the applicability and universality of the prediction model in different clinical scenarios. BRIEF DESCRIPTION OF THE DRAWINGS
[0059] Figure 1 Schematic diagram of the personalized intraoperative hypotension prediction framework based on continuous learning in the present invention;
[0060] Figure 2 The user can customize the hypotension prediction framework diagram in the present invention;
[0061] Figure 3 The following is a comparison chart of the performance results of comparative experiments of the embodiments of the present invention. DETAILED DESCRIPTION
[0062] Example 1: Figure 1-Figure 3 As shown, a personalized intraoperative hypotension prediction method based on continuous learning, the method comprising:
[0063] Step 1: Fine-tune the time series model based on personalized instance multi-task: Update the pre-trained time series model parameters by using the patient's blood pressure time series data in the early stage of surgery (the first 80%);
[0064] Furthermore, the Step 1 includes:
[0065] Step 1.1, Transfer Learning: Fine-tune the Units model, a time series model pre-trained in fields such as weather, finance, and transportation. The pre-trained model has powerful time series feature extraction and prediction capabilities and can be applied to medical data.
[0066] Step 1.2, Model fine-tuning: Use the patient's previous blood pressure time series data to fine-tune the parameters of the pre-trained Units model to ensure that the model adapts to individual physiological characteristics and time series data distribution;
[0067] Step 1.3. Introducing self-supervised learning tasks: During the fine-tuning process, combined with the masking reconstruction task, that is, randomly masking the input data (25%), allowing the Units model to learn to reconstruct the masked data.
[0068] Furthermore, in Step 1.3, each training sample (x, y) of the self-supervised learning task includes two training losses:
[0069] Supervised learning loss: The Units model predicts the future time series based on the input data x, obtains the predicted value y′, and calculates the mean squared error (MSE) between the predicted value y′ and the true value y;
[0070] Self-supervised reconstruction loss: Mask the input data x, reconstruct it using the Units model, and calculate the mean square error (MSE) between the reconstructed data and the original data;
[0071] Taking into account the losses of the time series prediction task and the mask reconstruction task, the loss of fine-tuning the Units model based on personalized instance multi-task is formally expressed as follows:
[0072]
[0073] Among them, θ units Represents the backbone network of the Units model; θ fore ,θ imp Represents the learnable prompt tokens for time series prediction and mask reconstruction tasks respectively; N refers to N training samples; θ fore,units Represents the model parameters used for time series prediction tasks; θ imp,units Represents the model parameters used for mask reconstruction tasks; θ fore,units (x) represents the process of time series prediction by the model; θ imp,units (x') represents the process of mask reconstruction task; MSE is the mean square error, which calculates the square of the error between the true value and the predicted value at each moment and then takes the average. The expression of MSE is: where Y iIt represents the true value at the i-th moment, Y i ' is the predicted value at the i-th moment.
[0074] Step 2. Online continuous learning: After completing personalized instance multi-task fine-tuning, the model has the ability to predict intraoperative hypotension for specific patients. To achieve higher accuracy, before predicting intraoperative hypotension, the time series model after updating the model parameters continues to use the reconstruction self-supervised learning algorithm and further fine-tunes the time series model parameters using real-time blood pressure data. Since the input data x of the time series model after the model parameters are updated at this time is the data of a single observation window, only mask reconstruction self-supervised learning will be performed.
[0075] Furthermore, the Step 2 includes:
[0076] The input data x is copied, which is the data of a single observation window, to the number of batch size, and then randomly masking the data of the entire batch size, the backbone network θ units Update the weights on a batch of data to get the updated backbone network θ' units , the learning loss is as follows:
[0077]
[0078] Updated backbone network θ' units +θ fore Make the final low blood pressure prediction to achieve online continuous learning.
[0079] Step 3: Predict intraoperative hypotension. Different from the traditional classification method, this paper regards the intraoperative hypotension prediction problem as a time series prediction task, such as Figure 2 As shown in the figure, the process of intraoperative hypotension prediction is as follows: the blood pressure sequence to be predicted is input into the fine-tuned time series model, the fine-tuned time series model outputs the predicted blood pressure sequence, and then the blood pressure sequence is predicted according to the standard definition of hypotension (i.e. Figure 2 Users can define hypotension, which is the standard definition of normal hypotension) and check the blood pressure sequence output by the fine-tuned time series model to predict whether hypotension will occur. This method can flexibly adapt to different hypotension judgment criteria and improve the applicability of the model.
[0080] Furthermore, in Step 3, the intraoperative hypotension prediction includes:
[0081] The data of the blood pressure observation window of a predefined length is input into the fine-tuned time series model. The fine-tuned time series model predicts a blood pressure sequence of a specific length. The blood pressure value at each time point in the predicted sequence is checked according to the standard definition of hypotension, and the prediction of whether hypotension will occur is output.
[0082] The present invention also provides a personalized intraoperative hypotension prediction system based on continuous learning, the system comprising:
[0083] The timing model parameter update module is used to fine-tune the timing model based on personalized instance multi-tasks. Specifically, it is used to update the pre-trained timing model parameters by using the patient's blood pressure time series data before surgery;
[0084] The time series model parameter fine-tuning module is used for online continuous learning. Specifically, before intraoperative hypotension prediction, the time series model after updating the model parameters continues to use the reconstruction self-supervised learning algorithm and uses real-time blood pressure data to further fine-tune the time series model parameters;
[0085] The intraoperative hypotension prediction module is used to predict intraoperative hypotension. Specifically, it is used to input the blood pressure sequence to be predicted into the fine-tuned timing model. The fine-tuned timing model outputs the predicted blood pressure sequence. Then, the blood pressure sequence output by the fine-tuned timing model is checked according to the standard definition of hypotension to obtain a prediction of whether hypotension will occur.
[0086] The present invention also 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 personalized intraoperative hypotension prediction method based on continuous learning when executing the program.
[0087] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for predicting personalized intraoperative hypotension based on continuous learning is implemented.
[0088] The present invention also provides a computer program product, comprising a computer program, which, when executed by a processor, implements the personalized intraoperative hypotension prediction method based on continuous learning.
[0089] To verify the effectiveness of this invention, a unified time series model (Units) was used on the VitalDB dataset. The Units model converts raw time series data into three types of tokens, enabling unified processing of various tasks: sequence tokens, prompt tokens, and task tokens.
[0090] The token types are described as follows:
[0091] Sequence tokens correspond to the original time series data. The original sequence is divided into multiple patches using non-overlapping windows. Each patch is mapped through a multi-layer perceptron (MLP) to generate an embedded representation of the sequence token; the dimensions are formally expressed as follows:
[0092] Z x ∈R s×v×d
[0093] Among them, s is the number of segments after being divided into patches, v is the number of time series variables, and d is the dimension of the embedding vector.
[0094] Prompttoken is similar to the softprompt concept in natural language processing. It generates several learnable prompt embeddings and concatenates them with sequence tokens. It is mainly used to absorb contextual information in the sequence based on the characteristics of the task. Each task corresponds to its own unique prompt embedding; the dimensions are formally expressed as follows:
[0095] Z p ∈R p×v×d
[0096] Among them, p is the number of prompttokens, v is the number of time series variables, and d is the dimension of the embedding vector.
[0097] Tasktokens are used to unify different tasks, including Gentokens and CLS tokens. Gentokens is used to unify time series prediction and time series filling. The results generated by Gen are prediction results. CLS tokens are used for tasks related to time series classification. The dimensions are formally expressed as follows:
[0098] Z m ∈R 1×v×d
[0099] For tasks involving prediction, the GEN token Z is copied according to the desired prediction length m f times, get Z m ∈R f ×v×d ; These features are then connected with the sample and prompt tags and input into the Units model for time series prediction. The data dimensions input into the Units model are formally expressed as follows:
[0100] Z fore =Unit(CA(Z p ,Z x ,Zm ,))∈R (p+s+f)×v×d .
[0101] In order to verify the prediction performance of the method of the present invention, a comparative experiment was conducted. Figure 3 As shown, Figure 3 In the figure, lr = learning rate, wd = weight-decay, dp = dropout, ep = epoch, and the superscript z indicates zero-shot. CT-Units represents cross-testing (ablation experiments), where the model is trained on the current case and then randomly tested on test data from other cases. The result is the average of 10 cross-tests.
[0102] from Figure 3 It can be seen that:
[0103] Personalized prediction performance based on online continuous learning is superior;
[0104] In terms of time series prediction, the present invention (i.e. Figure 3 After using the Units-TTT method in
[15] , Units performs better in terms of MAE (mean absolute error) and MSE (mean square error) (the smaller the MAE and MSE, the better the performance of the model).
[0105] The Units model after using the method of the present invention (i.e. Figure 3 The Units-TTT in
[15] improves the Zero-shot performance of Units by 18.22% and 26.28% respectively. Compared with Units with some parameter fine-tuning, the online continuous learning version of Units improves MAE and MSE by 6.36% and 18.47%.
[0106] From the perspective of classification tasks, the Units (i.e. Figure 3 The performance of Units-TTT in the
[15] has also been greatly improved. For example, compared with the Zero-shot version, F1 has increased by nearly 3 times, and Recall has increased by nearly 6 times. Compared with the version with fine-tuning some parameters, F1 has increased by 9.9% and Recall has increased by 22.64%.
[0107] The application areas of the present invention include:
[0108] (1) Real-time prediction of intraoperative hypotension: The present invention is mainly used for early prediction and real-time warning of hypotension during surgery, providing real-time decision support for anesthesiologists.
[0109] (2) Intelligent anesthesia management system: It can be integrated into the intelligent anesthesia monitoring system to monitor the patient's physiological signals and predict the risk of hypotension to ensure the safety of surgery.
[0110] The specific embodiments of the present invention are described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the knowledge of ordinary technicians in this field without departing from the scope of the present invention.
Claims
1. A personalized intraoperative hypotension prediction method based on continuous learning, characterized by: The method comprises: Step 1: Fine-tune the time series model based on personalized instance multi-task: Update the pre-trained time series model parameters by using the patient's blood pressure time series data before surgery; Step 2, online continuous learning: Before intraoperative hypotension prediction, the time series model after updating the model parameters continues to use the reconstruction self-supervised learning algorithm to further fine-tune the time series model parameters using real-time blood pressure data; Step 3. Prediction of intraoperative hypotension: The blood pressure sequence to be predicted is input into the fine-tuned time series model. The fine-tuned time series model outputs the predicted blood pressure sequence. The blood pressure sequence output by the fine-tuned time series model is then checked according to the standard definition of hypotension to obtain a prediction of whether hypotension will occur.
2. The method for predicting intraoperative hypotension based on continuous learning according to claim 1, characterized in that: Step 1 includes: Step 1.1, transfer learning: use the pre-trained time series model Units model for fine-tuning; Step 1.2, Model fine-tuning: Use the patient's previous blood pressure time series data to fine-tune the parameters of the pre-trained Units model to ensure that the model adapts to individual physiological characteristics and time series data distribution; Step 1.
3. Introduce self-supervised learning tasks: During the fine-tuning process, combine it with the mask reconstruction task, that is, randomly mask the input data and let the Units model learn to reconstruct the masked data.
3. The method for predicting intraoperative hypotension based on continuous learning according to claim 2, characterized in that: In Step 1.3, each training sample (x, y) of the self-supervised learning task contains two training losses: Supervised learning loss: The Units model predicts the future time series based on the input data x, obtains the predicted value y′, and calculates the mean squared error between the predicted value y′ and the true value y; Self-supervised reconstruction loss: Mask the input data x, reconstruct it using the Units model, and calculate the mean square error between the reconstructed data and the original data; Taking into account the losses of the time series prediction task and the mask reconstruction task, the loss of fine-tuning the Units model based on personalized instance multi-task is formally expressed as follows: Among them, θ units Represents the backbone network of the Units model; θ fore ,θ imp Represents the learnable prompt tokens for time series prediction and mask reconstruction tasks respectively; N refers to N training samples; θ fore,units Represents the model parameters used for time series prediction tasks; θ imp,units Represents the model parameters used for mask reconstruction tasks; θ fore,units (x) represents the process of time series prediction by the model; θ imp,units (x') represents the process of mask reconstruction task; MSE is the mean square error, which calculates the square of the error between the true value and the predicted value at each moment and then takes the average. The expression of MSE is: where Y i It represents the true value at the i-th moment, Y i ' is the predicted value at the i-th moment.
4. The method for predicting intraoperative hypotension based on continuous learning according to claim 1, characterized in that: Step 2 includes: The input data x is copied, which is the data of a single observation window, to the number of a batch size, and then randomly masking the data of the entire batch size, the backbone network θ units Update the weights on a batch of data to get the updated backbone network θ' units , the learning loss is as follows: Updated backbone network θ' units +θ fore Make the final low blood pressure prediction to achieve online continuous learning.
5. The method for predicting intraoperative hypotension based on continuous learning according to claim 1, characterized in that: In Step 3, the intraoperative hypotension prediction includes: The data of the blood pressure observation window of a predefined length is input into the fine-tuned time series model. The fine-tuned time series model predicts a blood pressure sequence of a specific length. The blood pressure value at each time point in the predicted sequence is checked according to the standard definition of hypotension, and the prediction of whether hypotension will occur is output.
6. The method for predicting intraoperative hypotension based on continuous learning according to claim 1, characterized in that: In Step 3, the fine-tuned time series model processes the blood pressure sequence to be predicted by: The time series model uses the Units model. The Units model converts raw time series data into three types of tokens to achieve unified processing of various tasks: sequence token, prompt token, and task token. The token types are described as follows: Sequence tokens correspond to the original time series data. The original sequence is divided into multiple patches by using non-overlapping windows. Each patch is mapped through a multi-layer perceptron (MLP) to generate an embedded representation of the sequence token. The dimensionality is formally expressed as follows: WITH x ∈R s×v×d Among them, s is the number of segments after being divided into patches, v is the number of time series variables, and d is the dimension of the embedding vector; The prompt token generates several learnable prompt embeddings and concatenates them with the sequence token to absorb contextual information in the sequence based on the task characteristics. Each task corresponds to its own unique prompt embedding. The dimensionality is formally expressed as follows: WITH p ∈R p×v×d Where p is the number of prompttokens, v is the number of time series variables, and d is the dimension of the embedding vector; Tasktoken is used to unify different tasks, including Gen tokens and CLS tokens. The former is used to unify time series prediction and time series filling. The results generated by Gen are prediction results. CLS tokens are used for tasks related to time series classification. The dimension is formally expressed as follows: Z m ∈R 1×v×d ; For tasks involving prediction, the GEN token Z is copied according to the desired prediction length m f times, get Z m ∈R f×v×d ; These features are then connected with the sample and prompt tags and input into the Units model for time series prediction; the data dimension input into the Units model is formally expressed as follows: Z fore =Unit(CA(Z p ,Z x ,Z m ,))∈R (p+s+f)×v×d .
7. A personalized intraoperative hypotension prediction system based on continuous learning, characterized by: The system includes: a module for executing the personalized intraoperative hypotension prediction method based on continuous learning according to any one of claims 1 to 6.
8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the method for predicting personalized intraoperative hypotension based on continuous learning as described in any one of claims 1 to 6 is implemented.
9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the personalized intraoperative hypotension prediction method based on continuous learning as described in any one of claims 1 to 6.
10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, it implements the personalized intraoperative hypotension prediction method based on continuous learning as described in any one of claims 1 to 6.