Prediction and early warning method and system for medical instrument related pressure damage

Abstract

The invention belongs to the field of medical instruments, and particularly relates to a medical instrument related pressure damage prediction and early warning method and system. The method mainly comprises the following steps that a multi-sensor network is adopted, a key area where a patient makes contact with a medical instrument is covered, key area data are collected, preprocessing, filtering and denoising are conducted on the collected original data of the key area so that abnormal values and noise interference can be eliminated, and a solid foundation is laid for subsequent data analysis; and carrying out data fusion, feature extraction and model construction on the preprocessed data, predicting the MDRPI risk by using an MDRPI risk prediction model, carrying out risk assessment in combination with the specific condition of a patient, and giving an alarm in time when the risk is high. According to the MDRPI risk prediction method, accurate prediction of the MDRPI risk is realized through the MDRPI risk prediction model, false report and missing report are reduced, continuous optimization of the model is promoted through an intelligent feedback mechanism, and the prediction accuracy and the system applicability are improved.

Description

Technical Field

[0001] The present invention belongs to the field of medical devices, and specifically relates to a method and system for predicting and warning of medical device-related pressure injuries. Background Art

[0002] Almost all hospitalized patients require at least one medical device for care and treatment, which puts them at risk of injury. This is especially true for patients in the ICU, where a variety of medical devices or equipment (MDs) are commonly used to maintain the lives of critically ill patients. Therefore, the risk of medical device-related pressure injuries (MDRPIs) is inevitable.

[0003] Studies have shown that MDRPIs are an important type of pressure injury (PI) in critically ill adult patients, with a combined incidence of approximately 0.4% to 44.7% of PI. However, due to differences in the populations surveyed, injury sites, and medical devices used in various studies, the incidence of MDRPIs in the research results also varies greatly (0.4% to 44.7%). At present, due to insufficient clinical attention and difficulties in early identification, the evaluation and reporting data of MDRPIs are relatively scarce, and the clinical understanding of related issues such as the prevention, occurrence, and care of MDRPIs is not comprehensive. Therefore, the present invention provides a more efficient and stable MDRPIs prediction and early warning method and system to achieve efficient and accurate prediction and early warning of MDRPIs. Summary of the invention

[0004] The present invention provides a medical device-related pressure injury prediction and early warning method and system, which are used to solve the technical problems of inaccurate prediction of existing MDRPIs and failure to timely feedback abnormalities.

[0005] In order to achieve the above-mentioned purpose, the technical solution adopted by the present invention is as follows: A method for predicting and warning medical device-related pressure injuries, comprising the following steps:

[0006] (1) Data collection: A multi-sensor network is used to cover the key areas where the patient contacts the medical device and collect key area data, including pressure distribution, temperature, and humidity environmental parameters of the skin contact surface, as well as the patient's heart rate, blood pressure, and respiratory rate vital sign parameters;

[0007] (2) Data preprocessing: Preprocess the raw data collected in the key areas, filter and denoise them to remove outliers and noise interference, laying the foundation for subsequent data analysis;

[0008] (3) Data analysis: The preprocessed data is subjected to three processes: data fusion, feature extraction, and model construction. The data fusion adopts the normalization processing method to ensure that the data with different parameters are compatible with each other. The feature extraction uses the linear discriminant analysis (LDA) algorithm to automatically identify and extract feature data reflecting the patient's physiological state, medical device usage, and environmental factors in three dimensions. After the feature data is extracted, the feature data is fused and learned using a deep learning model to construct an MDRPI s risk prediction model - a hybrid model structure of a convolutional neural network (CNN) combined with a recurrent neural network (RNN).

[0009] (4) Risk assessment and early warning: After the MDRP I s risk prediction model predicts the MDRP I s risk, it conducts risk assessment based on the patient's specific situation. When the assessment result shows a high risk, an early warning signal is automatically triggered. The early warning information also includes the patient's basic information and risk assessment results, so that medical staff can quickly understand the situation and take appropriate treatment measures;

[0010] (5) Intelligent feedback and optimization: By recording each warning and subsequent processing, a closed-loop feedback mechanism is formed, and the prediction model is continuously optimized using reinforcement learning algorithms to improve prediction accuracy and warning timeliness.

[0011] Furthermore, the filtering adopts exponential smoothing filtering: t =αx t +(1-α)y t-1 , where α is the smoothing coefficient (0<α<1), x t is the current data point, y t is the current filtered data point, y t-1 is the previous filtered data point.

[0012] Furthermore, the moving average denoising algorithm is used to filter the data y t Averaging is performed to remove high frequency noise.

[0013] Furthermore, the Z-score method is used to identify outliers.

[0014] Furthermore, the data fusion adopts a normalization processing method to collect the pressure distribution P of the skin contact surface of the key area data. (t) , Temperature T (t) 、Humidity M (t) , heart rate H (t) , blood pressure B (t) , respiratory rate F (t) Perform time weighting and normalization to form the data sequence X (t) .

[0015] Furthermore, the specific process of the linear discriminant analysis LDA algorithm is as follows:

[0016] Input: d-dimensional dataset y i ∈{1,2,...,N}, the dimension to be reduced to d';

[0017] Output: dataset X' after dimension reduction;

[0018] (1) Calculate the intra-class scatter matrix S w ;

[0019] (2) Calculate the inter-class scatter matrix S b ;

[0020] (3) Calculation matrix

[0021] (4) Yes Perform singular value decomposition to obtain the singular value λ i and its corresponding eigenvector w i ,i=1,2,...,N-1;

[0022] (5) Take the eigenvector group or projection matrix W corresponding to the singular value with the largest d' before;

[0023] (6) Calculate each sample x in the sample set i The projection z in the new low-dimensional space i =W T x i ;

[0024] (7) Obtain the sample set after dimensionality reduction That is, characteristic data related to risk categories.

[0025] Furthermore, a hybrid model structure of convolutional neural network (CNN) and recurrent neural network (RNN) is adopted. CNN can effectively extract the spatial characteristics of the data, while RNN can capture the time series characteristics of the data. Through the combination of these two models, a comprehensive and accurate prediction of MDRP I s risk can be achieved.

[0026] Furthermore, the reinforcement learning algorithm uses a policy gradient algorithm to update the policy parameters by maximizing the expectation of the cumulative reward. The core is the gradient ascent method, which is used to maximize the objective function J(θ).

[0027] The gradient of J(θ) with respect to θ is used to update θ along the gradient direction.

[0028] A system including a method for predicting and warning medical device-related pressure injuries includes a data acquisition module, a data preprocessing module, a data analysis module, a risk assessment and warning module, an intelligent feedback and optimization module, and a control system, wherein the data acquisition module uses a multi-sensor network to cover key areas where patients contact medical devices and collect key area data, and the key area data includes pressure distribution, temperature, and humidity environmental parameters of the skin contact surface, and the patient's heart rate, blood pressure, and respiratory rate vital sign parameters; the data preprocessing module preprocesses the original data collected in the key area, filters, and denoises to eliminate abnormal values ​​and noise interference, laying a solid foundation for subsequent data analysis; the data analysis module performs three processes of data fusion, feature extraction, and model construction on the preprocessed data, and the data fusion uses a normalization processing method to ensure that data with different parameters are compatible with each other; the feature extraction uses a linear discriminant analysis LDA algorithm to automatically identify and extract feature data in three dimensions reflecting the patient's physiological state, medical device usage, and environmental factors. After the feature data is extracted, the feature data is fused and learned using a deep learning model to construct an MDRPI s risk prediction model - a hybrid model structure of convolutional neural network CNN combined with recurrent neural network RNN; the risk assessment and early warning module performs risk assessment based on the patient's specific situation. When the assessment result shows a high risk, an early warning signal is automatically triggered. The early warning information also includes the patient's basic information and risk assessment results, so that medical staff can quickly understand the situation and take corresponding treatment measures; the intelligent feedback and optimization module records each early warning and subsequent processing to form a closed-loop feedback mechanism, and uses reinforcement learning algorithms to continuously optimize the prediction model to improve the prediction accuracy and early warning timeliness.

[0029] Furthermore, the data acquisition module, data preprocessing module, data analysis module, risk assessment and early warning module, and intelligent feedback and optimization module are controlled by the control system.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] (1) The present invention achieves accurate prediction of MDRPIs risk and reduces false positives and negative negatives through multi-source data fusion and a hybrid model structure of the MDRPIs risk prediction model - a convolutional neural network (CNN) combined with a recurrent neural network (RNN).

[0032] (2) The present invention can immediately notify medical staff, buy valuable time for early intervention, effectively reduce the degree of injury, and provide customized nursing advice based on individual differences of patients, thereby improving the quality of nursing care.

[0033] (3) The present invention promotes continuous optimization of the model through an intelligent feedback mechanism, thereby improving prediction accuracy and system applicability. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is a flow chart of the medical device-related pressure injury prediction and early warning method of the present invention. DETAILED DESCRIPTION

[0035] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0036] like Figure 1 As shown, a method for predicting and warning medical device-related pressure injuries includes the following steps:

[0037] (1) Data collection: A multi-sensor network is used to cover the key areas where the patient contacts the medical device and collect key area data, including pressure distribution, temperature, and humidity environmental parameters of the skin contact surface, as well as the patient's heart rate, blood pressure, and respiratory rate vital sign parameters; the sensor is connected to the data preprocessing terminal via wireless or wired means to achieve real-time transmission and storage of data.

[0038] (2) Data preprocessing: Preprocess the raw data collected in key areas, filter and denoise them to remove outliers and noise interference, laying a solid foundation for subsequent data analysis;

[0039] The filtering adopts exponential smoothing filter: t =αx t +(1-α)y t-1

[0040] Among them, α is the smoothing coefficient (0<α<1), x t is the current data point, y t is the current filtered data point, y t-1 is the previous filtered data point.

[0041] Denoising algorithm: Use the moving average denoising algorithm to filter the data y t Averaging is performed to eliminate high frequency noise. Where n represents the number of data points, X t Represents the denoised data.

[0042] Outlier detection and elimination: The Z-score method is used to identify outliers. The outlier detection formula is: Among them, x t is the denoised data point, μ is the data mean, and σ is the data standard deviation. Usually, |Zt Data points with |>3 are considered outliers.

[0043] (3) Data analysis: The preprocessed data is subjected to three processes: data fusion, feature extraction, and model construction. The data fusion adopts the normalization processing method to ensure that the data with different parameters are compatible with each other. The feature extraction uses the linear discriminant analysis (LDA) algorithm to automatically identify and extract feature data of three dimensions reflecting the patient's physiological state, medical device usage, and environmental factors. After the feature data is extracted, the feature data is fused and learned using a deep learning model to construct an MDRPIs risk prediction model - a hybrid model structure of a convolutional neural network (CNN) combined with a recurrent neural network (RNN).

[0044] Data fusion: Since the collected data comes from different sensors, there may be differences in their formats, dimensions, sampling rates, etc. Therefore, in the data fusion stage, the data is processed in a unified manner, including data format conversion, dimension unification, and sampling rate matching, to ensure that the data from different data sources are compatible with each other and can be used for subsequent analysis.

[0045] Pressure distribution P on the skin contact surface (t) , Temperature T (t) 、Humidity M (t) , heart rate H (t) , blood pressure B (t) , respiratory rate F (t) The above data is time-weighted and normalized to form the data sequence X (t) .X (t) =ω p ·P (t) +ω T ·T (t) +ω M ·M (t) +ω H ·H (t) +ω B ·B (t) +ω F ·F (t) , where ω i =∑i t / t,i∈{P,T,M,H,B,F}.

[0046] Feature extraction: In order to extract key features related to MDRPIs risk from massive amounts of data, we use advanced feature extraction algorithms. These algorithms can automatically identify and extract multiple dimensional features that reflect the patient's physiological state, medical device usage, and environmental factors, such as skin contact surface pressure distribution characteristics, temperature change characteristics, and patient vital sign change characteristics.

[0047] Feature extraction algorithm: Linear discriminant analysis (LDA) algorithm is used. LDA is a supervised learning dimensionality reduction technique that attempts to find a linear combination to maximize the ratio of between-class variance to within-class variance. In MDRPIs risk analysis, LDA can be used to extract the pressure distribution P of the patient's skin contact surface. (t) , Temperature T (t) 、Humidity M (t) , heart rate H (t) , blood pressure B (t) , respiratory rate F (t) and extracting risk-category-related features from medical device usage data.

[0048] The specific process of the linear discriminant analysis LDA algorithm is as follows:

[0049] Input: d-dimensional dataset y i ∈{1,2,...,N}, the dimension to be reduced to d';

[0050] Output: dataset X' after dimension reduction;

[0051] (1) Calculate the intra-class scatter matrix S w :

[0052] (2) Calculate the inter-class scatter matrix S b :

[0053] Where m i is the number of samples in the i-th category, μ is the mean vector of all samples;

[0054] (3) Calculation matrix

[0055] (4) Yes Perform singular value decomposition to obtain the singular value λ i and its corresponding eigenvector w i ,i=1,2,...,N-1;

[0056] (5) Take the eigenvector group or projection matrix W corresponding to the singular value with the largest d' before;

[0057] (6) Calculate each sample x in the sample set i The projection z in the new low-dimensional space i =W T x i ;

[0058] (7) Obtain the sample set after dimensionality reduction That is, characteristic data related to risk categories.

[0059] Model construction: After extracting the feature data, the deep learning model is used to fuse and learn the features to build the MDRPIs risk prediction model. Specifically, the CNN network architecture: A convolutional neural network is mainly composed of the following 5 layers:

[0060] Input layer (I nput layer), convolution layer (CONV layer, including activation function, ReLU, sigmo id, tanh), pooling layer (Pooling layer), fully connected layer (FC layer), output layer (Ouput layer).

[0061] RNN network architecture: Instead of rigidly memorizing all fixed-length sequences, RNN stores information from previous time steps through hidden states. RNN has a loop structure. A typical RNN network architecture contains an input, an output, and a neural network unit. The difference from ordinary feedforward neural networks is that the neural network unit of RNN is not only connected to the input and output, but also has a loop / circuit / loop / return (loop) itself. This loop allows information to be passed from one step in the network to the next step.

[0062] (4) Risk assessment and early warning: After the MDRP I s risk prediction model predicts the MDRP I s risk, it conducts risk assessment based on the patient's specific situation. When the assessment result shows a high risk, an early warning signal is automatically triggered. The early warning information also includes the patient's basic information and risk assessment results, so that medical staff can quickly understand the situation and take appropriate treatment measures;

[0063] Risk assessment: After the model predicts the risk of MDRPIs, a risk assessment is conducted based on the patient's specific situation. This includes considering individual differences such as the patient's age, gender, and physical condition, as well as external factors such as the type of medical device, how it is used, and how long it is used. By comprehensively considering these factors, a more accurate risk assessment result can be obtained.

[0064] Warning signal triggering: When the assessment result shows high risk, the system automatically triggers the warning signal. These warning signals can be conveyed to medical staff in a variety of ways, such as sound and light alarms, mobile device push, etc. At the same time, the warning information will also include key content such as the patient's basic information, risk assessment results and recommended intervention measures, so that medical staff can quickly understand the situation and take appropriate measures.

[0065] (5) Intelligent feedback and optimization: By recording each warning and subsequent processing, a closed-loop feedback mechanism is formed, and the prediction model is continuously optimized using reinforcement learning algorithms to improve prediction accuracy and warning timeliness.

[0066] Closed-loop feedback mechanism: Each warning and subsequent processing will be automatically recorded by the system and stored in the database. Through the analysis of these data, the shortcomings of the warning system are found, such as the changes in indicators such as false alarm rate and missed alarm rate.

[0067] Reinforcement learning algorithm: In order to further improve the accuracy and adaptability of the prediction model, the reinforcement learning algorithm is introduced. The algorithm can automatically adjust the model parameters to maximize the prediction accuracy or minimize the false alarm rate and other objective functions by simulating the prediction and decision-making process in different scenarios. In this way, the continuous optimization and performance improvement of the prediction model can be achieved.

[0068] Personalized care recommendations: Based on historical data and individual patient information, the system can also provide personalized care recommendations to medical staff

[0069] Reinforcement learning algorithm uses policy gradient algorithm: Policy gradient algorithm is a policy-based reinforcement learning method that directly optimizes the policy function π(a|s), that is, the probability distribution of selecting action a given state s. The goal of the policy gradient algorithm is to update the policy parameter θ by maximizing the expectation of the cumulative reward, so that the intelligent feedback and optimization system can learn the optimal policy.

[0070] The core of the policy gradient algorithm is the gradient ascent method, which is used to maximize the objective function J(θ) (usually the expectation of the cumulative reward). The objective function can be expressed as:

[0071]

[0072] Among them, τ represents a trajectory from the initial state to the terminal state, which consists of a series of states, actions and rewards; πθ represents the strategy defined by the parameter θ; r(s t ,a t ) represents the immediate reward after executing action at in state st; T is the length of the trajectory; γ is the discount factor used to balance the importance of immediate rewards and future rewards. In order to maximize J(θ), we need to calculate the gradient of J(θ) with respect to θ and update θ along the gradient direction. The gradient can be estimated by the following formula: Where N is the number of sampled trajectories; is the state-action value function An estimate of Next action After that, the expectation of the future cumulative rewards that can be obtained by following the strategy π; is the logarithm of the gradient of the policy function, which is used to guide the update direction of the parameter θ.

[0073] In practical applications, It is usually estimated through methods such as Monte Carlo methods or temporal difference learning.

[0074] In the intelligent feedback and optimization module, the policy gradient algorithm can be used to optimize the parameters of the prediction model. Specifically, the parameters of the prediction model are regarded as the parameters θ of the policy function, and the prediction accuracy or false alarm rate is regarded as part of the reward function. By continuously allowing the prediction model to interact with real data (i.e., making predictions and receiving feedback on real results) and using the policy gradient algorithm to adjust the model parameters, the prediction model can gradually learn how to predict more accurately and reduce the false alarm rate.

[0075] In addition, in combination with the need for personalized nursing advice, the relevant parameters in the prediction model are dynamically adjusted according to the patient's historical data and individual difference information (such as through feature weight adjustment, model structure selection, etc.) to provide personalized nursing advice that is more in line with the patient's actual situation. In this way, the accuracy of the prediction model can be improved, and more accurate and effective nursing guidance can be provided to medical staff.

[0076] A medical device-related pressure injury prediction and early warning system includes a data acquisition module, a data preprocessing module, a data analysis module, a risk assessment and early warning module, an intelligent feedback and optimization module, and a control system, wherein the data acquisition module uses a multi-sensor network to cover the key areas where the patient contacts the medical device and collects key area data, and the key area data includes the pressure distribution, temperature, and humidity environmental parameters of the skin contact surface, and the patient's heart rate, blood pressure, and respiratory rate vital sign parameters; the data preprocessing module preprocesses the original data collected in the key area, filters, and denoises to eliminate abnormal values ​​and noise interference, laying a solid foundation for subsequent data analysis; the data analysis module performs three processes of data fusion, feature extraction, and model construction on the preprocessed data, and the data fusion adopts a normalization processing method to ensure that data with different parameters are compatible with each other; the feature extraction adopts a linear discriminant analysis LDA algorithm to automatically identify and extract feature data of three dimensions reflecting the patient's physiological state, medical device usage, and environmental factors. After the feature data is extracted, the feature data is fused and learned using a deep learning model to construct an MDRPI s risk prediction model - a hybrid model structure of convolutional neural network CNN combined with recurrent neural network RNN; the risk assessment and early warning module conducts risk assessment in combination with the specific situation of the patient. When the assessment result shows high risk, the early warning signal is automatically triggered. The early warning information also contains the patient's basic information and risk assessment results, so that medical staff can quickly understand the situation and take corresponding treatment measures; the intelligent feedback and optimization module records each early warning and subsequent processing to form a closed-loop feedback mechanism, and uses the reinforcement learning algorithm to continuously optimize the prediction model to improve the prediction accuracy and early warning timeliness. The data acquisition module, data preprocessing module, data analysis module, risk assessment and early warning module, and intelligent feedback and optimization module are controlled by the control system.

[0077] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A method for predicting and warning medical device-related pressure injuries, characterized in that: The following steps are involved: (1) Data collection: A multi-sensor network is used to cover the key areas where the patient contacts the medical device and collect key area data, including pressure distribution, temperature, and humidity environmental parameters of the skin contact surface, as well as the patient's heart rate, blood pressure, and respiratory rate vital sign parameters; (2) Data preprocessing: Preprocess the raw data collected in the key areas, filter and denoise them to remove outliers and noise interference, laying the foundation for subsequent data analysis; (3) Data analysis: The preprocessed data is subjected to three processes: data fusion, feature extraction, and model construction. The data fusion adopts the normalization processing method to ensure that the data with different parameters are compatible with each other. The feature extraction uses the linear discriminant analysis (LDA) algorithm to automatically identify and extract feature data of three dimensions reflecting the patient's physiological state, medical device usage, and environmental factors. After the feature data is extracted, the feature data is fused and learned using a deep learning model to construct an MDRPIs risk prediction model - a hybrid model structure of a convolutional neural network (CNN) combined with a recurrent neural network (RNN). (4) Risk assessment and early warning: After the MDRPIs risk prediction model predicts the MDRPIs risk, it conducts risk assessment based on the patient's specific situation. When the assessment result shows a high risk, an early warning signal is automatically triggered. The early warning information also includes the patient's basic information and risk assessment results, so that medical staff can quickly understand the situation and take appropriate treatment measures; (5) Intelligent feedback and optimization: By recording each warning and subsequent processing, a closed-loop feedback mechanism is formed, and the prediction model is continuously optimized using reinforcement learning algorithms to improve prediction accuracy and warning timeliness.

2. A method for predicting and warning medical device-related pressure injuries according to claim 1, characterized in that: Exponential smoothing filtering is used for filtering.

3. A method for predicting and warning medical device-related pressure injuries according to claim 1, characterized in that: The moving average denoising algorithm is used to filter the data y t Averaging is performed to remove high frequency noise.

4. A method for predicting and warning medical device-related pressure injuries according to claim 1, characterized in that: The Z-score method is used to identify outliers and outlier detection.

5. A method for predicting and warning medical device-related pressure injuries according to claim 1, characterized in that: The data fusion adopts a normalization processing method to collect the pressure distribution P of the skin contact surface of the key area data (t) , Temperature T (t) 、Humidity (t) , heart rate H (t) , blood pressure B (t) , respiratory rate F (t) Perform time weighting and normalization to form the data sequence X (t) .

6. A method for predicting and warning medical device-related pressure injuries according to claim 1, characterized in that: The specific process of the linear discriminant analysis LDA algorithm is as follows: Input: d-dimensional dataset The dimension to be reduced to d'; Output: dataset X' after dimension reduction; (1) Calculate the intra-class scatter matrix S w ; (2) Calculate the inter-class scatter matrix S b ; Where m i is the number of samples in the i-th category, μ is the mean vector of all samples; (3) Calculation matrix (4) Yes Perform singular value decomposition to obtain the singular value λ i and its corresponding eigenvector w i ,i=1,2,...,N-1; (5) Take the eigenvector group or projection matrix W corresponding to the singular value with the largest d' before; (6) Calculate each sample x in the sample set i The projection z in the new low-dimensional space i =W T x i ; (7) Obtain the sample set after dimensionality reduction That is, characteristic data related to risk categories.

7. A method for predicting and warning medical device-related pressure injuries according to claim 1, characterized in that: A hybrid model structure of convolutional neural network (CNN) and recurrent neural network (RNN) is adopted. CNN can effectively extract the spatial characteristics of data; RNN can capture the time series characteristics of data. Through the combination of these two models, a comprehensive and accurate prediction of MDRPIs risks can be achieved.

8. A method for predicting and warning medical device-related pressure injuries according to claim 1, characterized in that: The reinforcement learning algorithm uses the policy gradient algorithm to update the policy parameters by maximizing the expectation of the cumulative reward. The core is the gradient ascent method, which is used to maximize the objective function J(θ). The gradient of J(θ) with respect to θ, and updates θ along the gradient direction; Where N is the number of sampled trajectories; is the state-action value function An estimate of Next action After that, the expectation of the future cumulative rewards that can be obtained by following the strategy π; It is the logarithm of the gradient of the policy function, which guides the update direction of the parameter θ.

9. A system for implementing a method for predicting and warning medical device-related pressure injuries as described in any one of claims 1 to 8, characterized in that: It includes a data acquisition module, a data preprocessing module, a data analysis module, a risk assessment and early warning module, an intelligent feedback and optimization module, and a control system. The data acquisition module adopts a multi-sensor network to cover the key areas where the patient contacts the medical device and collects data in the key areas. The data in the key areas include the pressure distribution, temperature, and humidity environmental parameters of the skin contact surface, and the patient's heart rate, blood pressure, and respiratory rate vital sign parameters. The data preprocessing module preprocesses the raw data collected in the key areas, filters, and denoises them to eliminate abnormal values ​​and noise interference, laying a solid foundation for subsequent data analysis. The data analysis module performs three processes of data fusion, feature extraction, and model construction on the preprocessed data. The data fusion adopts a normalization processing method to ensure that data with different parameters are compatible with each other. The feature extraction adopts a linear The discriminant analysis LDA algorithm automatically identifies and extracts feature data in three dimensions reflecting the patient's physiological state, medical device usage and environmental factors. After extracting the feature data, the deep learning model is used to fuse and learn the feature data to construct an MDRPIs risk prediction model - a hybrid model structure of a convolutional neural network CNN combined with a recurrent neural network RNN; the risk assessment and early warning module conducts risk assessment based on the patient's specific situation. When the assessment result shows a high risk, an early warning signal is automatically triggered. The early warning information also includes the patient's basic information and risk assessment results, so that medical staff can quickly understand the situation and take appropriate measures; the intelligent feedback and optimization module records each early warning and subsequent processing to form a closed-loop feedback mechanism, and uses a reinforcement learning algorithm to continuously optimize the prediction model to improve the prediction accuracy and early warning timeliness.

10. A system for predicting and warning medical device-related pressure injuries according to claim 9, characterized in that: The data acquisition module, data preprocessing module, data analysis module, risk assessment and early warning module, and intelligent feedback and optimization module are controlled by the control system.

Images