Hemodialysis hypotension prediction system based on machine learning

Abstract

The invention provides a hemodialysis hypotension prediction system based on machine learning, which relates to the technical field of medical health, and comprises a real-time data stream processing module, a dynamic modeling and prediction module, a real-time data stream processing module, a real-time data stream processing module, a real-time data stream processing module, a real-time data stream processing module and a real-time data stream processing module, the risk prediction module is used for dynamically modeling and predicting the hypotension risk of a patient according to physiological parameter data input in real time, and the personalized adaptation module is used for automatically adjusting a risk prediction model. According to the hemodialysis hypotension prediction system based on machine learning, by introducing the real-time data flow processing module, physiological parameter data in the hemodialysis process can be collected and processed in real time. By adopting the dynamic modeling and predicting module, the hypotension risk can be dynamically modeled and predicted based on the physiological parameter data input in real time, and the prediction model is automatically adjusted through the personalized adaptation module, so that the risk prediction result of each patient is more accurate.

Description

Technical Field

[0001] This invention relates to the field of medical and health technology, specifically to a machine learning-based system for predicting low blood pressure during hemodialysis. Background Technology

[0002] Hemodialysis is one of the most common clinical treatments for end-stage renal disease. The dialysis process continuously removes metabolic waste and excess water from the body while maintaining the patient's fluid and electrolyte balance. Clinically, dialysis can cause numerous complications, with hypotension being one of the most common. Hypotension can occur during hemodialysis due to decreased circulating blood volume, weakened vascular tone regulation, and insufficient cardiac reserve. Once this happens, it can cause dizziness, nausea, and confusion; in severe cases, it can even lead to insufficient blood supply to the heart and brain, endangering life. Healthcare professionals typically rely on real-time monitoring of blood pressure, heart rate, and blood volume, and use experience to assess a patient's risk of hypotension. In recent years, some dialysis devices have integrated basic hemodynamic monitoring functions, allowing for dynamic acquisition of certain parameters. Simultaneously, some hospitals are attempting to combine big data and information systems to store and analyze historical dialysis data, providing basic references for healthcare professionals. However, these methods primarily rely on static data and lack in-depth analysis of multi-source dynamic data.

[0003] While existing technologies have achieved data acquisition and preliminary analysis to some extent, real-time performance remains insufficient in the practical application of hypotension prediction during dialysis. Many studies rely primarily on offline analysis, resulting in delayed data processing and an inability to rapidly model and predict newly acquired physiological parameters during dialysis. This forces healthcare professionals to rely on experience or reactive measures, failing to achieve truly dynamic early warning of hypotension risk. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a machine learning-based system for predicting low blood pressure during hemodialysis. The technical problem this invention aims to solve is: how to achieve accurate prediction and real-time early warning of low blood pressure risk through a feedback mechanism that integrates real-time data acquisition, dynamic modeling and prediction, personalized adaptation, and automatic adjustment.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a machine learning-based system for predicting low blood pressure during hemodialysis, comprising:

[0006] The real-time data stream processing module is used to collect and process patients' physiological parameter data in real time during hemodialysis.

[0007] The dynamic modeling and prediction module is used to dynamically model and predict the patient's risk of hypotension based on real-time input physiological parameter data.

[0008] Personalized adaptation module for automatically adjusting risk prediction models;

[0009] The real-time early warning and reminder module is used to calculate the low blood pressure risk value in real time.

[0010] Preferably, the physiological parameters include: blood pressure, heart rate, body temperature, and blood oxygen saturation.

[0011] Preferably, the dynamic modeling and prediction module models and predicts the patient's risk of hypotension by including the following steps:

[0012] S3.1. A feature selection module is adopted, which uses dynamic feature selection based on information gain algorithm to automatically select the features that have the most influence on the prediction of low blood pressure risk, so as to reduce redundant and noisy data;

[0013] S3.2. After feature selection, the model is pre-trained based on a large-scale health dataset to predict the risk of hypotension, thereby improving its adaptability to different patient groups and dialysis environments.

[0014] S3.3. A feedback mechanism that automatically adjusts risk thresholds and alarms based on real-time clinical feedback improves the accuracy and responsiveness of hypotension risk prediction.

[0015] Preferably, the feature selection module employs an L1 regularization-based feature selection algorithm and uses an L1 penalty term to sparsify the features, automatically selecting features related to low blood pressure risk prediction and reducing redundant features.

[0016] Preferably, when the feedback mechanism receives a clinical feedback signal, it rewards or punishes the system based on the strength and accuracy of the feedback signal to optimize the system's strategy for adjusting the risk threshold. The formula for the feedback mechanism is as follows:

[0017] R a =R c +β·f

[0018] Among them, R a R is the adjusted reward value. c denoted as the current reward value, β as the adjustment coefficient for feedback learning, and f as the accuracy of the clinical feedback.

[0019] Preferably, the automatic adjustment of the risk prediction model includes the following steps:

[0020] S6.1. Establish a risk prediction model:

[0021] W a =W b ×(1+α·X p )

[0022] Among them, Wa W represents the adjusted weighting coefficients. b X is the initial weighting coefficient, α is the adjustment factor, and X is the weighting factor. p For each patient's individual feature vector;

[0023] S6.2. Dynamically adjust the parameters of the hypotension risk prediction model based on real-time data input and the adjusted weighting coefficients;

[0024] S6.3. When the patient's condition changes beyond the safety threshold, the process of retraining the model is automatically triggered to ensure that the model maintains the latest predictive capabilities.

[0025] Preferably, the safety thresholds are: blood pressure fluctuation: exceeding 20 mmHg, heart rate fluctuation: exceeding 15 beats / minute, blood oxygen saturation exceeding 5%, and body temperature fluctuation: exceeding 1°C.

[0026] Preferably, the real-time early warning and reminder module pushes alerts to medical staff's mobile devices or workstations via sound alarms, visual alarms, and real-time notifications.

[0027] This invention provides a machine learning-based system for predicting low blood pressure during hemodialysis. It offers the following advantages:

[0028] This machine learning-based hypotension prediction system for hemodialysis, through the introduction of a real-time data stream processing module, can collect and process physiological parameter data during hemodialysis in real time. Employing a dynamic modeling and prediction module, it can dynamically model and predict the risk of hypotension based on real-time input physiological parameter data, and automatically adjust the prediction model through a personalized adaptation module, making the risk prediction results more accurate for each patient.

[0029] The hypotension risk prediction model is dynamically adjusted based on each patient's individual characteristics. Through a feature selection module, the most influential features are selected using an information gain algorithm, and L1 regularization is applied to sparsify these features, enabling the prediction model to adapt to different patients' physiological states. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the structure for realizing an invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] like Figure 1 As shown, this embodiment of the invention provides a machine learning-based system for predicting low blood pressure during hemodialysis, including a real-time data stream processing module for real-time acquisition and processing of physiological parameter data of patients during hemodialysis, including: blood pressure, heart rate, body temperature, and blood oxygen saturation.

[0033] The dynamic modeling and prediction module is used to dynamically model and predict the patient's risk of hypotension based on real-time input physiological parameter data. The dynamic modeling and prediction module for modeling and predicting the patient's risk of hypotension includes the following steps:

[0034] S3.1. A feature selection module is adopted, which uses dynamic feature selection based on information gain algorithm to automatically select the features most influential on the prediction of hypotension risk, so as to reduce redundant and noisy data. The feature selection module adopts a feature selection algorithm based on L1 regularization and uses L1 penalty term to sparsify the features, automatically filtering out the features related to the prediction of hypotension risk and reducing redundant features.

[0035] S3.2. After feature selection, the model is pre-trained based on a large-scale health dataset to predict the risk of hypotension, thereby improving its adaptability to different patient groups and dialysis environments.

[0036] S3.3. A feedback mechanism based on real-time clinical feedback automatically adjusts risk thresholds and alarms, improving the accuracy and responsiveness of hypotension risk prediction. Upon receiving a clinical feedback signal, the mechanism rewards or penalizes the system based on the signal's strength and accuracy, optimizing the system's risk threshold adjustment strategy. The formula for the feedback mechanism is as follows:

[0037] R a =R c +β·f

[0038] Among them, R a R is the adjusted reward value. c denoted as the current reward value, β as the adjustment coefficient for feedback learning, and f as the accuracy of the clinical feedback.

[0039] The personalized adaptation module is used to automatically adjust the risk prediction model. The automatic adjustment of the risk prediction model includes the following steps:

[0040] S6.1. Establish a risk prediction model:

[0041] W a =W b ×(1+α·X p )

[0042] Among them, W a W represents the adjusted weighting coefficients. bX is the initial weighting coefficient, α is the adjustment factor, and X is the weighting factor. p For each patient's individual feature vector;

[0043] S6.2. Dynamically adjust the parameters of the hypotension risk prediction model based on real-time data input and the adjusted weighting coefficients;

[0044] S6.3. When the patient's condition changes beyond the safety threshold, the process of retraining the model is automatically triggered to ensure that the model maintains the latest predictive ability. The safety thresholds are: blood pressure fluctuation: more than 20 mmHg, heart rate fluctuation: more than 15 beats / minute, blood oxygen saturation: more than 5%, body temperature fluctuation: more than 1℃.

[0045] The real-time warning and reminder module is used to push the calculated low blood pressure risk value to the mobile devices or workstations of medical staff through sound alarms, visual alarms and real-time notifications.

[0046] Example 2

[0047] This example demonstrates how to improve the accuracy and real-time performance of hypotension risk prediction by using feature selection, model pre-training, and risk threshold adjustment based on real-time clinical feedback.

[0048] 1. Feature Selection Module

[0049] 1.1 Information Gain Algorithm

[0050] First, the relationship between each physiological parameter and the risk of hypotension is calculated using the information gain algorithm. For example, consider the following dataset:

[0051]

[0052] By calculating information gain, the correlation between various features and the risk of low blood pressure can be determined. For example, the calculated information gain values ​​are as follows:

[0053] Blood pressure: 0.25

[0054] Heart rate: 0.35

[0055] Body temperature: 0.10

[0056] Blood oxygen saturation: 0.40

[0057] Based on the calculation results, blood oxygen saturation and heart rate are selected as the features most influential in predicting the risk of low blood pressure because they have the highest information gain values.

[0058] 1.2L1 Regularization

[0059] After feature selection, L1 regularization is used for sparsity reduction. A simple linear regression model is used to predict the probability of hypotension, and the loss function of L1 regularization is:

[0060]

[0061] In this model, w i Let λ be the weight of the feature and λ be the regularization coefficient. L1 regularization can reduce the weights of some redundant or irrelevant features to zero. For example, after L1 regularization, the weight of body temperature might be reduced to zero, indicating its small contribution to predicting low blood pressure.

[0062] 2. Model pre-training

[0063] 2.1 Dataset Selection

[0064] When selecting a dataset, a large health dataset containing multiple patient groups was used. A clinical dataset containing 1000 patients was used, which included information such as age, sex, blood pressure, heart rate, body temperature, and blood oxygen saturation.

[0065] 2.2 Pre-training process

[0066] A deep neural network was pre-trained on the dataset to predict the probability of hypotension. During pre-training, a neural network with three hidden layers and 256 neurons per layer was selected. Through training, the model's weights were optimized, ultimately reducing the prediction error (e.g., mean squared error) to 0.02 and achieving a training accuracy of 95%.

[0067] 3. Adjustment of risk thresholds based on clinical feedback

[0068] 3.1 Reception of Clinical Feedback Signals

[0069] During real-time dialysis, medical staff input the following clinical feedback:

[0070] Patient 1: Blood pressure dropped from 120 / 80 to 90 / 60, heart rate increased from 72 to 85, and dizziness occurred.

[0071] Patient 2: Blood oxygen saturation dropped from 98% to 95%, with no obvious symptoms.

[0072] The system will receive these feedback signals and process them through the following steps.

[0073] 3.2 Reward and Punishment Mechanism

[0074] Based on the accuracy and strength of the clinical feedback signals, the system will reward or penalize according to the following formula:

[0075] Ra =R c +γ·f

[0076] For patient 1, the accuracy of the feedback signal is high (dizziness symptoms), γ = 1.0, f = 1.2 (given higher weights based on the severity of the patient's symptoms).

[0077] For patient 2, the accuracy of the feedback signal was moderate (decreased blood oxygenation, but no symptoms), γ = 0.8, f = 0.8.

[0078] Therefore, the reward value is adjusted as follows:

[0079] Patient 1: R a =R c +1.0 × 1.2 = R c +1.2

[0080] Patient 2: R a =R c +0.8 × 0.8 = R c +0.64

[0081] 3.3 Automatic adjustment of risk threshold

[0082] The system automatically adjusts the risk threshold based on the strength of the feedback signal. If Patient 1's feedback signal indicates an increased risk of hypotension, the system lowers the risk threshold and issues an early warning. For example, if the initial risk threshold is 0.7 (an alert is triggered when the predicted risk of hypotension is greater than 0.7), after Patient 1's clinical feedback, the system lowers the threshold to 0.6. The adjustment formula is as follows:

[0083] T a =T b ×(1+α·F f )

[0084] Patient 1's feedback signal makes α = 0.1 and F f =1.2, then:

[0085] T a =0.7×(1+0.1×1.2)=0.7×1.12=0.784

[0086] This means that the system will issue an alert when patient 1's predicted risk of low blood pressure is greater than 0.784.

[0087] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A machine learning-based system for predicting hypotension during hemodialysis, characterized in that, include: The real-time data stream processing module is used to collect and process patients' physiological parameter data in real time during hemodialysis. The dynamic modeling and prediction module is used to dynamically model and predict the patient's risk of hypotension based on real-time input physiological parameter data. Personalized adaptation module for automatically adjusting risk prediction models; The real-time early warning and reminder module is used to calculate the low blood pressure risk value in real time.

2. The machine learning-based hypotension prediction system for hemodialysis according to claim 1, characterized in that: The physiological parameters include: blood pressure, heart rate, body temperature, and blood oxygen saturation.

3. The machine learning-based hypotension prediction system for hemodialysis according to claim 1, characterized in that: The dynamic modeling and prediction module models and predicts the patient's risk of hypotension, including the following steps: S3.

1. A feature selection module is adopted, which automatically selects the features with the most influence on the prediction of low blood pressure risk by performing dynamic feature selection based on the information gain algorithm; S3.

2. After feature selection, the model is pre-trained based on a large-scale health dataset to predict the risk of low blood pressure. S3.

3. Feedback mechanism for automatically adjusting risk thresholds and alarms based on real-time clinical feedback.

4. The machine learning-based hypotension prediction system for hemodialysis according to claim 3, characterized in that: The feature selection module employs an L1 regularization-based feature selection algorithm and uses an L1 penalty term to sparsify the features.

5. A machine learning-based hypotension prediction system for hemodialysis according to claim 4, characterized in that: When the feedback mechanism receives a clinical feedback signal, it rewards or punishes the user based on the strength and accuracy of the feedback signal. The formula for the feedback mechanism is as follows: R a =R c +β·f Among them, R a R is the adjusted reward value. c denoted as the current reward value, β as the adjustment coefficient for feedback learning, and f as the accuracy of the clinical feedback.

6. A machine learning-based hypotension prediction system for hemodialysis according to claim 3, characterized in that: The automatic adjustment of the risk prediction model includes the following steps: S6.

1. Establish a risk prediction model: W a =W b ×(1+α·X p ) Among them, W a W represents the adjusted weighting coefficients. b X is the initial weighting coefficient, α is the adjustment factor, and X is the weighting coefficient. p For each patient's individual feature vector; S6.

2. Dynamically adjust the parameters of the hypotension risk prediction model based on real-time data input and the adjusted weighting coefficients; S6.

3. When the patient's condition changes beyond the safety threshold, the process of retraining the model is automatically triggered.

7. A machine learning-based hypotension prediction system for hemodialysis according to claim 6, characterized in that: The safety thresholds are: blood pressure fluctuation: exceeding 20 mmHg, heart rate fluctuation: exceeding 15 beats / minute, blood oxygen saturation exceeding 5%, and body temperature fluctuation: exceeding 1°C.

8. A machine learning-based hypotension prediction system for hemodialysis according to claim 7, characterized in that: The real-time early warning and reminder module pushes alerts to medical staff's mobile devices or workstations via sound alarms, visual alarms, and real-time notifications.

Images