Personalized urethral catheterization scheme optimization method for nephrology department patient based on deep learning

Through deep learning technology, using methods such as GAN, CNN, LSTM and VAE, a personalized catheterization plan for nephrology patients was constructed, which solved the shortcomings of catheter quality assessment and dynamic physiological changes monitoring in traditional methods, and achieved the goal of nephrology patients. The precise evaluation and real-time adjustment of personalized catheterization regimens improve treatment effect and patient comfort.

CN119027728AInactive Publication Date: 2024-11-26ZHENGZHOU CENT HOSPITAL
View PDF 0 Cites 3 Cited by

Patent Information

Application Number
CN202411085616.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-08
Publication Date
2024-11-26
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing personalized catheterization program for nephrology patients has shortcomings in evaluating catheter quality and dynamic physiological changes. It lacks the ability to comprehensively analyze multiple data sources. It is difficult to achieve real-time adjustment of traditional methods, which increases the risk of treatment and predict concurrency. inaccuracy of the disease.

Method used

Using a deep learning-based method, a diverse training data is generated by generating adversarial networks (GANs), a joint feature extraction network is constructed, and image and time series data features are extracted using convolutional neural networks (CNNs) and long and short-term memory networks (LSTMs). Design a multi-task learning framework, introduce attention mechanisms, generate personalized catheterization schemes, and combine variable autoencoder (VAE) to generate personalized catheterization schemes.

Benefits of technology

Improve the accuracy and personalization of the assessment of the catheterization regimen, enhance the generalization ability and robustness of the model, realize comprehensive assessment and real-time adjustment of the patient's physiological status, reduce the risk of treatment, improve the treatment effect and patient Comfort.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119027728A_ABST
    Figure CN119027728A_ABST
Patent Text Reader

Abstract

The invention relates to a personalized urethral catheterization scheme optimization method for nephrology department patients based on deep learning. The method comprises the following steps: firstly, collecting medical history, medicine use records, image data, physiological parameters and laboratory detection data of patients through a plurality of hospitals, and performing data enhancement through a generative adversarial network (GAN) to generate diversified training data; then constructing a joint feature extraction network, extracting image data features by using a convolutional neural network CNN, and extracting time sequence data features by using a long short-term memory network LSTM; a multi-task learning framework is designed, and a plurality of related indexes such as the urine flow rate, the bladder pressure and the complication risk are predicted at the same time; an attention mechanism is introduced, so that the personalized urethral catheterization scheme model focuses on key features influencing the urethral catheterization scheme of the individual patient; and finally, encoding the extracted features through an encoder to generate a patient feature vector, and generating a personalized urethral catheterization scheme by adopting a variational auto-encoder (VAE).
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a method for optimizing a personalized urinary catheterization program for patients in nephrology, and specifically to a method for optimizing a personalized urinary catheterization program for patients in nephrology based on deep learning. Background Art

[0002] In nephrology treatment, the use of catheters is a common medical operation, and the formulation of personalized catheterization plans lies in how to select and adjust the use of catheters according to the specific conditions of patients. However, there are still some deficiencies and drawbacks in the optimization methods of personalized catheterization plans for patients in nephrology, which affect the accuracy and personalization level of catheterization plans. Now, combined with the Chinese invention "A method and system for intelligent evaluation of catheter quality" with patent number 202210765097.X, the defects of the current personalized catheterization plan optimization method are analyzed in detail: At present, the formulation of personalized catheterization plans for patients in nephrology mostly relies on clinical experience and traditional medical evaluation methods. Doctors usually choose the type and frequency of catheterization based on the patient's medical history, clinical symptoms and basic physiological parameters (urinary flow rate and bladder pressure). The main defect of this method is the lack of accurate data support and personalized analysis methods, which leads to the formulation of plans that are often rough and difficult to fully consider the individual differences and dynamic physiological changes of patients. First of all, the traditional method has obvious deficiencies in evaluating the quality of catheters. Although patent 202210765097.X provides an intelligent catheter quality assessment method that can improve the accuracy and precision of quality assessment by analyzing the catheter information and its quality parameters (structure, material and functional parameters), this assessment process still faces challenges in actual application. The quality assessment of the catheter needs to comprehensively consider its matching degree with the patient's physiological structure, the material's physiological response to the patient, and the actual performance of the function in clinical practice, and these factors are difficult to fully reflect through static parameters.

[0003] Secondly, existing personalized catheterization programs usually lack real-time monitoring and analysis of dynamic physiological parameters. The patient's physiological state (bladder capacity, urethral resistance, and urodynamics) will change over time and with the progress of treatment, and these changes are often not captured by a single physiological measurement. This leads to deviations in the implementation of the catheterization program, and it cannot be adjusted in time according to the patient's actual situation. In addition, traditional programs lack accuracy in predicting complications during catheterization, which increases the risk of treatment. Thirdly, traditional personalized catheterization programs are also insufficient in integrating multiple data sources. Existing methods often rely only on the patient's medical history and basic physiological parameters, and lack the ability to comprehensively analyze multiple data sources (imaging data, laboratory test data, and time series physiological data). Although the intelligent evaluation method mentioned in patent 202210765097.X can improve the accuracy of catheter quality assessment, in the actual application of personalized programs, these multidimensional data have not yet been fully utilized for personalized analysis and program optimization. In addition, the optimization of personalized catheterization programs is also limited by algorithms and models. Traditional decision-making models are mainly based on rules and experience and lack adaptive learning capabilities. This means that when faced with complex individual differences among patients, the model cannot automatically adjust its strategy to cope with different clinical scenarios. Although emerging technologies based on deep learning have powerful learning and prediction capabilities, their application in nephrology catheterization programs is still in its infancy and lacks verification and support from large-scale clinical data. Summary of the invention

[0004] The purpose of the present invention is to provide a method for optimizing personalized urinary catheterization programs for nephrology patients based on deep learning, thereby solving some of the drawbacks and shortcomings pointed out in the background technology.

[0005] The present invention solves the above-mentioned technical problems by adopting the following technical solutions:

[0006] S1. First, collect patients’ medical history, medication records, imaging data, physiological parameters, and laboratory test data from multiple hospitals, and generate diverse training data through data enhancement using the Generative Adversarial Network (GAN).

[0007] S2. Then, a joint feature extraction network is constructed, using convolutional neural network (CNN) to extract features of image data and long short-term memory (LSTM) to extract features of time series data; and a multi-task learning framework is designed to simultaneously predict multiple related indicators such as urine flow rate, bladder pressure, and complication risk;

[0008] S3. Introduce an attention mechanism to enable the personalized catheterization plan model to focus on the key features that affect the catheterization plan for individual patients.

[0009] S4. Finally, the extracted features are encoded through the encoder to generate the patient feature vector, and the variational autoencoder VAE is used to generate a personalized catheterization plan.

[0010] Furthermore, the process of generating diversified training data includes: first, collecting the medical history, drug use records, imaging data, physiological parameters and laboratory test data of patients in the Department of Nephrology from multiple hospitals, and the data collection formula is defined as:

[0011]

[0012] Where x represents a continuous variable, which is any physiological measurement or test result of the patient, a(x) is the weight function of the data type, and μ and σ are the expected and standard deviations, respectively;

[0013] Secondly, the collected data is preprocessed, including standardization and denoising, and the preprocessing formula is applied:

[0014]

[0015] Where k is the sensitivity parameter for adjusting data smoothing, y represents the data points after preprocessing, and α i is the weight coefficient specific to the data type, and n is the number of data points;

[0016] Next, we use the generative adversarial network GAN to enhance the data. The enhancement formula is:

[0017] f3(z)=z·(sin(az 2 )+cos(bz 2 ))

[0018] Where a and b are parameters used to adjust the degree of variation of the synthetic data, and z is the random noise drawn from a normal distribution.

[0019] Furthermore, the joint feature extraction network construction process is as follows:

[0020] S1. First, use medical imaging data, including CT or MRI images of the bladder and urethra, develop a deep convolutional neural network (CNN) architecture to extract features from the imaging data, and use image enhancement techniques, including contrast adjustment and sharpening, to improve the recognizability of key structures and pathological changes in the imaging data.

[0021] S2, then processes the physiological time series data directly related to catheterization, including urine flow rate and bladder pressure, through the long short-term memory network LSTM;

[0022] S3. Construct a multi-task learning framework to share the features extracted by CNN and LSTM to predict urine flow rate, bladder pressure and the risk of potential complications respectively, and adopt loss functions adapted to different tasks. The urine flow rate and bladder pressure use mean square error, while the risk of complications uses cross entropy loss.

[0023] Furthermore, the method of formulating a deep convolutional neural network (CNN) architecture extracts the image data features of the bladder and urethra through the function:

[0024]

[0025] Where x and y represent the spatial coordinates of the image data, α i is the adjustment parameter, n is the number of network layers;

[0026] Next, the image data is enhanced, including contrast adjustment and sharpening, using the function to process each pixel:

[0027]

[0028] Where z represents the intensity value or gray level of each pixel in the image;

[0029] Finally, we use the deep learning model to extract and analyze features using the function:

[0030]

[0031] To evaluate the statistical importance of features, u and v represent the coordinates in the multidimensional data space, p and q represent the extracted feature dimensions, and β is the normalization parameter.

[0032] Furthermore, the long short-term memory network LSTM construction process is as follows:

[0033] S1. First, the physiological time series data, including the collected urine flow rate and bladder pressure, are preprocessed and the normalization processing formula is applied:

[0034]

[0035] Transform the data t, where μ and σ represent the mean and standard deviation of the data set respectively;

[0036] S2. Next, the time series data is converted into a training sequence through a complex sliding window technique, and the function is applied:

[0037]

[0038] Where s(u) is the data value at time point u, δ is the window length, and λ is the time decay coefficient;

[0039] S3. Then design and implement the long short-term memory network LSTM model, which processes data through a dynamically adjusted network formula:

[0040] L(S)=tanh(W h h t-1 +W x x t +W c ·(c t-1 ⊙σ f )+b)

[0041] where h t-1 and c t-1 They are the hidden state and cell state at the previous moment, x t is the current input, W h ,W x ,W c and b are network parameters, σ f is the activation function of the forget gate, ⊙ represents element-wise multiplication, tanh and σ are activation functions that simulate and predict the dynamic changes of physiological data.

[0042] Furthermore, the method for constructing a multi-task learning framework includes: firstly using a composite convolutional neural network CNN and a long short-term memory network LSTM framework, through the formula:

[0043]

[0044] To fuse the image features f extracted by CNN and the time series features g extracted by LSTM, where σ represents the adjustment factor of feature fusion;

[0045] Next, we build a multi-task learning framework using a variational method:

[0046]

[0047] Processing and optimization features x, t is the integration variable;

[0048] For the prediction tasks of urine flow rate and bladder pressure, the mean square error loss function is applied:

[0049]

[0050] where y(t) and Represent the functional forms of the true value and the predicted value respectively, t is the integral variable,

[0051] For the prediction of complication risk, the cross entropy loss function is used:

[0052]

[0053] Where p(x) and q(x) represent the density functions of the continuous probability distribution, and x represents a continuous variable.

[0054] Furthermore, the attention mechanism construction method is as follows:

[0055] S1. First, a hybrid neural network framework is used to combine the convolutional neural network (CNN) and the long short-term memory (LSTM) network, and a self-attention mechanism is applied to enhance the model's feature understanding ability. The attention mechanism is expressed by the formula:

[0056]

[0057] Calculate the weight of each feature f, where β is a learnable parameter that represents the adjustment coefficient of feature importance;

[0058] S2. Subsequently, the features extracted by CNN and LSTM are fused, and the importance of each feature is dynamically evaluated using the integral formula:

[0059]

[0060] To optimize feature focusing, x represents the features extracted from the data.

[0061] Furthermore, the process of generating a personalized catheterization plan is as follows: First, by combining the convolutional neural network CNN and the long short-term memory network LSTM, the key features of the medical images and physiological time series data of the bladder and urethra are extracted, and the encoding formula is adopted:

[0062]

[0063] The feature v i Converted to a coded vector, where α i and β i To adjust the parameters;

[0064] Next, use the variational autoencoder VAE to generate the formula:

[0065]

[0066] The personalized catheterization plan is sampled and reconstructed from the latent space, where z represents the latent variable sampled from the normal distribution and γ is the coefficient to adjust the reconstruction error.

[0067] Beneficial effects of the present invention:

[0068] First, the present invention can extract key features from multi-source data through deep learning models, convolutional neural networks (CNN) and long short-term memory networks (LSTM), so that the model can effectively analyze medical imaging data (CT or MRI images of the bladder and urethra) and physiological time series data (urinary flow rate and bladder pressure changes), thereby achieving a comprehensive and accurate assessment of the patient's physiological state. Compared with traditional manual assessment methods, this data-driven analysis method can reduce human errors and improve the accuracy and consistency of assessments.

[0069] Secondly, the present invention uses generative models such as generative adversarial networks (GANs) and variational autoencoders (VAEs) to improve the generalization ability of the model through data enhancement and feature reconstruction. By generating diverse synthetic data, the model can learn more patient characteristics and pathological changes during training, which not only improves the robustness of the model in practical applications, but also maintains a high prediction accuracy when the amount of data is insufficient. In addition, by generating personalized catheterization plans from the latent space through VAE, the system can customize the plan according to the characteristics of different patients to meet the individual needs of patients.

[0070] In addition, the present invention emphasizes the real-time adjustment and feedback mechanism of personalized programs. By combining with the Internet of Things technology, the system can monitor the changes in the patient's physiological data in real time and dynamically adjust the catheterization program according to the changes. This real-time adaptability not only improves the effectiveness of treatment, but also significantly improves the patient's comfort and satisfaction. Through timely feedback and adjustment, the patient's subjective experience and treatment compliance are improved, thereby improving the overall treatment effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 This is a flow chart of the method for optimizing personalized urinary catheterization plans for nephrology patients based on deep learning in the present invention.

[0072] Figure 2 Generate a diverse training data module graph for the present invention.

[0073] Figure 3 A flow chart for constructing a joint feature extraction network of the present invention.

[0074] Figure 4 This is a diagram of the long short-term memory network LSTM building module of the present invention. DETAILED DESCRIPTION

[0075] The specific implementation modes of the present invention will be described in detail below in conjunction with the accompanying drawings.

[0076] Combined with Figure 1, a method for optimizing personalized catheterization programs for patients in the Department of Nephrology based on deep learning, first collects data from multiple hospitals, including medical history, drug use records, imaging data CT and MRI, heart rate and blood pressure, and various laboratory test data, and uses the generative adversarial network GAN to enhance these data and generate diverse training data;

[0077] Then, a convolutional neural network (CNN) was used to extract key spatial features from the image data, and a long short-term memory (LSTM) network was used to process the time series data of physiological parameters.

[0078] To achieve accurate prediction, a multi-task learning framework was constructed, which used features to predict urine flow rate, bladder pressure, and potential complication risk. This method enables the model to handle multiple output tasks simultaneously, improving the efficiency and accuracy of prediction.

[0079] Introducing an attention mechanism to strengthen the model's focus on key features that affect catheterization decisions, including abnormal structures in images or abnormal fluctuations in time series, thereby ensuring the relevance and practicality of the model output;

[0080] Finally, all key features are converted into patient feature vectors through the encoder, and then a variational autoencoder (VAE) is used to generate personalized catheterization plans based on these feature vectors.

[0081] Embodiment 1:

[0082] View attached Figure 2 In the deep learning-based optimization method for personalized catheterization of nephrology patients, blood pressure data of patients are collected and processed from multiple hospitals to evaluate the catheterization treatment of nephrology patients. Blood pressure data is affected by many factors, including the patient's age, gender, and underlying diseases. The following formula is used for data enhancement and normalization:

[0083]

[0084] in:

[0085] x represents the blood pressure measurement;

[0086] a(x) is the weight function of the data type, set to The weight used to balance the data;

[0087] μ (expected value) and σ (standard deviation) were set to 120 (average level of normal systolic blood pressure) and 15 (taking into account the variability of blood pressure in general adults), respectively.

[0088] Through the Gaussian function:

[0089]

[0090] Weighting each blood pressure measurement x for data augmentation helps normalize the blood pressure data, making it more suitable for input into subsequent CNN and LSTM networks.

[0091] Assuming that a patient's blood pressure measurement is 135 mmHg, use the above formula to calculate its weighted value:

[0092] Compute the normalized expression of the x value (135) with respect to μ (120) and σ (15):

[0093]

[0094] Apply a Gaussian function and weights:

[0095]

[0096] F1(135)≈1×0.0263=0.0263

[0097] The embodiment further preprocesses the collected blood pressure physiological data, including standardization and denoising, and uses the following formula to perform data smoothing and normalization processing:

[0098]

[0099] In this formula:

[0100] y i represents the standardized blood pressure measurement;

[0101] α i It is the weight coefficient for different types of data. Blood pressure data is more important, so α is set i =0.6; for data such as body temperature that do not directly affect the catheterization decision, set α i =0.4;

[0102] k is an adjustment factor, which is used to adjust the sensitivity of data smoothing. It is set according to the volatility of actual data, and k is set to 5;

[0103] n is the total number of data points involved in the calculation. Setting n=3 means there are three indicators: blood pressure, body temperature and heart rate.

[0104] Suppose there are data points that have been processed by Gaussian: blood pressure y BP =0.0263; body temperature y temp =0.015; heart rate y HR =0.020. Applying formula F2(y), first calculate the weighted sum:

[0105]

[0106] Then substitute this value into F2(y):

[0107]

[0108] After preprocessing, the comprehensive feature values ​​are converted into a form more suitable for input into the deep learning model.

[0109] Next, the embodiment uses the generative adversarial network GAN to further enhance the data by generating more synthetic data to improve the model's generalization ability and ability to handle the complexity of the real world. The specific data enhancement formula is:

[0110] F3(z)=z·(sin(az 2 )+cos(bz 2 ))

[0111] In this formula:

[0112] z is random noise drawn from a normal distribution N(0,1).

[0113] a and b are parameters that control the variability of synthetic data, and the value range is set to a = 0.5 to 1.5 and b = 0.5 to 1.5, ensuring the diversity of generated data without deviating too much from the characteristics of actual data.

[0114] A random value z=0.8 is drawn from a normal distribution, and a=1.2 and b=0.8 are set. The synthetic data values ​​calculated using the enhancement formula are as follows:

[0115] Calculate z 2 :

[0116] z 2 =0.8 2 =0.64

[0117] Substituting into F3(z):

[0118] F3(z)=0.8·(sin(1.2×0.64)+cos(0.8×0.64))

[0119] sin(0.768)≈0.694

[0120] cos(0.512)≈0.876

[0121] F3(z)=0.8·(0.694+0.876)=0.8·1.57=1.256

[0122] This calculation result F3(z)=1.256 represents the enhanced data points generated by GAN. This data can be used to train deep learning models to improve the accuracy and robustness of the model when parsing real-world data.

[0123] Embodiment 2:

[0124] Combined with Figure 3 In the embodiment, in the method for optimizing personalized urinary catheterization scheme for patients in nephrology based on deep learning, a deep convolutional neural network (CNN) architecture is used to extract the image data features of the bladder and urethra, and a feature extraction function is used:

[0125]

[0126] in:

[0127] x and y represent the spatial coordinates of the image data;

[0128] α i It is a tuning parameter used to control the network’s sensitivity to features at different levels and the granularity of extraction. The value range is set between 0.1 and 1.0 to adjust the degree of attention to detail.

[0129] n is the number of network layers, ranging from 5 to 20 layers, depending on the complexity of the model and the level of detail that needs to be parsed.

[0130] Then a set of medical images are processed to extract key features from these images to determine whether there are bladder or urethral abnormalities that affect catheterization. For the image point (x, y) = (0.5, 0.5), α = 0.5 and the number of network layers n = 10 are set.

[0131] Calculate the feature extraction contribution of each layer:

[0132]

[0133] Add up the contributions of all layers:

[0134]

[0135] Under the given network configuration and parameters, the features of specific points are successfully extracted and represented significantly in the sum.

[0136] Next, the embodiment performs image enhancement processing to improve the recognizability of key structures and pathological changes in the image data by adjusting contrast and sharpening, thereby optimizing the accuracy and personalization level of the catheterization plan. The function used in the image enhancement processing is as follows:

[0137]

[0138] z represents the intensity value or gray level of each pixel in the image. It is the derivative of the logistic regression function. For image processing, this derivative acts on each pixel, which can enhance the edge effect of the pixel and make the details of the image clearer.

[0139] Set the original intensity value z of a pixel to 150 (in the grayscale range of 0 to 255) and calculate the derivative of the logistic regression function:

[0140] First, ask The value of Then the derivative of logu is

[0141] calculate because but

[0142] so:

[0143]

[0144] Applied to the original pixel value z=150:

[0145] e -150 is almost 0, so

[0146] G(150)=150·0=0, indicating that the enhancement effect is not significant at high pixel values.

[0147] The final example uses a deep learning model for feature extraction and analysis to assess the statistical importance of features using the following formula:

[0148]

[0149] u and v are coordinates in the feature space, representing the extracted image and physiological data features, p and q are the corresponding feature dimensions, and β = 0.5 is set as the normalization parameter to adjust the smoothness and importance of the features in statistical analysis.

[0150] Analyze the medical data of a nephrology patient, including two characteristics: bladder wall thickness and urine flow rate. The bladder wall thickness p is 2.0 mm, and the urine flow rate q is 15.0 ml / s.

[0151] First, calculate the numerator, which represents the cumulative probability within the feature dimension:

[0152]

[0153] Using Gaussian integration method, the node and weight table are as follows:

[0154] For 1-dimensional integral Select 5 nodes and weights:

[0155] Node x i :{-2.020,-0.958,0,0.958,2.020}

[0156] Weight w i :{0.0199,0.3936,0.9453,0.3936,0.0199}

[0157] Apply the Gaussian integral in two dimensions:

[0158]

[0159] Calculate the integral value of Gaussian nodes and weights at u∈[-∞,2.0] and v∈[-∞,15.0]:

[0160] u i ={-2.020,-0.958,0,0.958,2.0}

[0161] v j ={-15.0,-7.5,0,7.5,15.0}

[0162] Calculate the product of each node and weight:

[0163] For u = 2.0 and v = 15.0, taking β = 0.5, we calculate:

[0164]

[0165] The denominator represents the total probability of the entire feature space and is used for normalization:

[0166]

[0167] This integral is a standard Gaussian integral, and the result is

[0168]

[0169] Substitute the above results into the feature evaluation formula:

[0170]

[0171] This result H(2.0,15.0)=0.14 indicates statistical significance within the specific feature range.

[0172] Embodiment 3:

[0173] In the deep learning-based optimization method for personalized urinary catheterization of nephrology patients, physiological time series data of a nephrology patient, including urine flow rate and bladder pressure, are processed, and these data are preprocessed, and the normalization processing formula is applied:

[0174]

[0175] Here, t is a data point, μ and σ represent the mean and standard deviation of the data set, respectively.

[0176] The following urine flow rate data are set: {20, 25, 22, 30, 28} (unit: ml / s), and the mean μ and standard deviation σ are μ=25 and σ=4 respectively.

[0177] Normalize these data:

[0178] (Since negative numbers cannot be logarithmic, the formula is adjusted to handle negative values ​​or values ​​less than the average):

[0179]

[0180] Do the same for the other data points:

[0181]

[0182] P(25)=log(1)=0

[0183]

[0184] P(22)=log(0.75)≈-0.287

[0185]

[0186] After the above preprocessing, the normalized urine flow rate data is obtained as follows:

[0187] {0.223,0,-0.287,0.223,0.287}

[0188] The preprocessed data will be input into the LSTM model. LSTM can capture the long-term dependencies of time series data through its gating mechanism, thereby accurately predicting changes in urine flow rate and bladder pressure.

[0189] Combined with Figure 4 , using the normalized data above, we build and train an LSTM model. The input layer of the model receives the preprocessed time series data, and the hidden layer contains LSTM units, which can learn the time dependencies in the data by adjusting the internal state and output weights.

[0190] The trained LSTM model can be used to predict the future urine flow rate and bladder pressure. The model output is as follows when predicting the urine flow rate data within the next 5 seconds:

[0191] {0.200,0.150,-0.100,0.100,0.250}

[0192] The predicted values ​​are converted back to actual values ​​by inverse normalization (i.e. applying the inverse of the original formula):

[0193] t=σ·exp(P(t))+μ

[0194] For the first prediction value P(t) = 0.200:

[0195] t=4·exp(0.200)+25≈29.3

[0196] The reverse normalization process of other data points is similar to obtain the predicted urine flow rate data:

[0197] {29.3,27.9,23.1,26.5,30.4}

[0198] Next, the embodiment uses the following formula to convert the time series data into a training sequence:

[0199]

[0200] In this formula:

[0201] s(u) is the data value at time point u;

[0202] δ is the length of the sliding window, which is used to determine the time span of each integration;

[0203] λ is the time decay coefficient, ranging from 0.1 to 1.0, which is used to control the decay speed.

[0204] Suppose there is a set of urine flow rate time series data as follows:

[0205] Time point (seconds): {0,1,2,3,4,5};

[0206] Corresponding urine flow rate (ml / s): {20, 22, 25, 23, 21, 24};

[0207] The sliding window length δ is set to 2 seconds and the time decay coefficient λ is set to 0.5.

[0208] The time series data is transformed by sliding window technology. Taking the time point t = 0 as an example, W(0,δ) is calculated:

[0209] The sliding window range is selected from time point t=0 to t+δ=2.

[0210] Compute the weighted sum within a sliding window:

[0211]

[0212] Select discrete time point u as {0,1,2}.

[0213] For each u, calculate e -0.5(0-u) s(u):

[0214] u=0:e 0 20 = 20

[0215] u=1:e 0.5 ·22≈1.6487·22≈36.27

[0216] u=2:e 1.0 ·25≈2.7183·25≈67.96

[0217] Sliding window results:

[0218] W(0,2)=20+36.27+67.96=124.23

[0219] The final embodiment designs and implements a long short-term memory network LSTM model, and processes data through the following dynamically adjusted network formula:

[0220] L(S)=tanh(W h h t-1 +W x x t +W c ·(c t-1 ⊙σ f )+b)

[0221] In this formula:

[0222] h t-1 and c t-1 are the hidden state and cell state at the previous moment respectively;

[0223] x t is the current input, representing the time series data at the current moment;

[0224] W h ,W x ,W c , and b are network parameters used to adjust the weights and biases of the model;

[0225] σ f is the activation function of the forget gate, using the Sigmoid function, ranging from 0 to 1;

[0226] ⊙ represents element-by-element multiplication;

[0227] Tanh is an activation function used to clamp the output between -1 and 1.

[0228] Suppose that a set of urine flow rate data is being processed after preprocessing and sliding window processing:

[0229] Input sequence at time point t=3: {0.223, 0, -0.287};

[0230] Set the parameters of the LSTM model: W h =0.5,W x =0.3,W c =0.8; b = 0.1;

[0231] Forget the activation function of the gate:

[0232] Current input x t =-0.287;

[0233] Set h t-1 =0.2,c t-1 =0.5.

[0234] Calculate the forget gate:

[0235]

[0236] Update cell status:

[0237] c t =σ f ·c t-1 +W c ·(c t-1 ⊙σ f )

[0238] c t =0.523·0.5+0.8·(0.5·0.523)

[0239] c t =0.2615+0.2092≈0.4707

[0240] Compute the hidden state:

[0241] h t =tanh(W h ·h t-1 +W x ·x t +W c ·c t +b)

[0242] h t =tanh(0.5·0.2+0.3·-0.287+0.8·0.4707+0.1)

[0243] h t =tanh(0.1-0.0861+0.3766+0.1)

[0244] ht =tanh(0.4905)≈0.454

[0245] At this time point t=3, the LSTM model outputs the hidden state h t = 0.454. This hidden state is used as the input for the next moment and can also be used to predict the trend and change of urine flow rate.

[0246] Embodiment 4:

[0247] In the deep learning-based optimization method for personalized urinary catheterization of patients in the Department of Nephrology, the following formula is used to fuse the image feature f extracted by CNN and the time series feature g extracted by LSTM:

[0248]

[0249] In this formula:

[0250] f represents the image features extracted by CNN;

[0251] g represents the time series features extracted by LSTM;

[0252] σ is the adjustment factor of feature fusion, which ranges from 0.1 to 1.0 and is used to control the smoothness of fusion and the importance of features.

[0253] There is a set of image features and time series features as follows:

[0254] Image feature f = 1.5 (feature value of bladder wall thickness extracted by CNN);

[0255] Time series feature g = {1.2, 1.4, 1.6, 1.8} (feature values ​​of urine flow rate extracted by LSTM);

[0256] The adjustment factor of feature fusion is set to σ = 0.5.

[0257] The fusion result is calculated by formula C(f,g):

[0258] Feature fusion formula:

[0259]

[0260] For each g, calculate

[0261] When g=1.2:

[0262]

[0263] When g=1.4:

[0264]

[0265] When g=1.6:

[0266]

[0267] When g=1.8:

[0268]

[0269] Calculate the fusion result:

[0270] Sum all the calculation results:

[0271]

[0272] Substituting into the calculation, we get:

[0273] C(f,g)≈0.676+0.781+0.781+0.676=2.914

[0274] The fused feature C(f,g) calculated by the feature fusion formula can be used as the input of the multi-task learning framework.

[0275] The next step of the embodiment is to build a multi-task learning framework and use the following variational method formula to process and optimize the feature x:

[0276]

[0277] Where x is the feature to be optimized and t is the integration variable ranging from 0 to 1.

[0278] It is assumed that a set of fused feature values ​​are being processed, one of the important features being the fused feature x=2.914 of urine flow rate (which is the result of feature fusion in the previous step).

[0279] Calculate feature optimization results:

[0280]

[0281] Calculation using Simpson's integral method The approximate value of the integral interval t is divided into several small intervals, and the integral point t is selected. i For {0, 0.25, 0.5, 0.75, 1}, calculate the value of each integration point

[0282] t=0:

[0283] 2.914 0 e -2.914·0 =1×1=1

[0284] t=0.25:

[0285] 2.914 0.25 e -2.914·0.25 ≈1.241·e -0.7285 ≈1.241·0.483≈0.599

[0286] t=0.5:

[0287] 2.914 0.5 e -2.914·0.5 ≈1.707·e -1.457 ≈1.707·0.233≈0.398

[0288] t=0.75:

[0289] 2.914 0.75 e -2.914·0.75 ≈2.347 e -2.1855 ≈2.347·0.113≈0.265

[0290] t=1:

[0291] 2.914 1 e -2.914·1 ≈2.914·e -2.914 ≈2.914·0.054≈0.157

[0292] Use Simpson's formula or the trapezoidal formula to sum:

[0293]

[0294] The optimized features M(x) calculated by the variational method can be used as the input of the multi-task learning framework.

[0295] The next step of the embodiment uses the mean square error loss function to quantify the difference between the predicted value and the true value, which is defined as follows:

[0296]

[0297] y(t) is the true value at time t; is the predicted value at time t; t is an integrating variable ranging from 0 to 1.

[0298] Suppose there is a set of true and predicted values ​​of urine flow rate data:

[0299] True value y(t): {25, 28, 22, 30, 27} (unit: ml / s);

[0300] Predicted value {26,27,24,29,26} (unit: ml / s).

[0301] These data were normalized to a function of t in the range 0 to 1 for integration.

[0302] Normalize the true and predicted values:

[0303] True value y(t) = {0.5, 0.56, 0.44, 0.6, 0.54};

[0304] Predicted value

[0305] Use polynomial fitting or spline interpolation to convert discrete points into continuous function form, and the function fitting is obtained:

[0306] y(t)=0.5+0.1t-0.2t 2 +0.05t 3

[0307]

[0308] Calculate the loss function:

[0309]

[0310] Use Simpson's integration method to perform numerical integration and choose the integration point t i For {0, 0.25, 0.5, 0.75, 1}, calculate the square of the difference at each integration point

[0311] t=0:

[0312]

[0313] t=0.25:

[0314]

[0315] t=0.5:

[0316]

[0317] t=0.75:

[0318]

[0319] t=1:

[0320]

[0321] Use Simpson's formula or the trapezoidal formula to sum:

[0322]

[0323] The error value calculated by the mean square error loss function is used to optimize the parameters of the model and adjust the difference between the predicted value and the true value.

[0324] The last embodiment uses the cross entropy loss function to quantify the difference between the true probability distribution and the predicted probability distribution, which is defined as follows:

[0325]

[0326] p(x) is the true probability density function; q(x) is the predicted probability density function; x is a continuous variable.

[0327] Suppose there is a set of true and predicted probability distributions of complication risk. Use Gaussian distribution to describe these probability density functions:

[0328] True distribution p(x): Gaussian distribution, mean μ = 0.5, standard deviation σ = 0.1

[0329] Prediction distribution q(x): Gaussian distribution, mean μ = 0.55, standard deviation σ = 0.15

[0330] Define the probability density function:

[0331] True distribution

[0332] Predicted Distribution

[0333] Calculate the cross entropy:

[0334]

[0335] Use Simpson's integral method to calculate, divide the integral interval x into several small intervals, select the integral point x i For {-0.5,-0.25,0,0.25,0.5,0.75,1}, calculate the value of each integration point:

[0336] Calculate p(x i ) and q(x i ):

[0337] For x = 0.5:

[0338]

[0339] For x = 0.75:

[0340]

[0341]

[0342] Evaluate and sum over all integration points:

[0343]

[0344] The calculation results are:

[0345] L ce (p,q)≈-(3.989·log(1.541)+0.176·log(0.882)+…)≈0.115

[0346] The loss value calculated by the cross entropy loss function is used to optimize the parameters of the model and adjust the difference between the predicted distribution and the true distribution.

[0347] Embodiment 5:

[0348] In the deep learning-based optimization method for personalized urinary catheterization of patients in the Department of Nephrology, the attention mechanism is introduced to improve the model feature understanding ability, and the following formula is used to calculate each feature f i Weight:

[0349]

[0350] In this formula:

[0351] f i represents the value of feature i;

[0352] β is a learnable parameter used to adjust the importance of features, ranging from 0.1 to 1.0.

[0353] The assumption is that we are processing a set of features that are extracted through CNN and LSTM and combined to evaluate the catheterization plan for patients in the Department of Nephrology. The following is a list of features and their corresponding initial values:

[0354] Imaging features (bladder wall thickness) f1 = 2.5;

[0355] Time series characteristics (urinary flow rate changes) f2 = 3.0;

[0356] Physiological parameter characteristics (blood pressure) f3 = 1.8;

[0357] Set the learnable parameter β = 0.5.

[0358] Feature f1:

[0359]

[0360] Feature f2:

[0361]

[0362]

[0363] Feature f3:

[0364]

[0365] By calculating the attention weights, the model can focus on important features more effectively. The weight of the image feature f1 is 0.335, the weight of the time series feature f2 is 0.429, and the weight of the physiological parameter feature f3 is 0.236.

[0366] The embodiment further optimizes feature focusing by fusing features extracted by a convolutional neural network (CNN) and a long short-term memory (LSTM) network and dynamically evaluating the importance of each feature:

[0367]

[0368] In this formula:

[0369] x represents the features extracted from the data;

[0370] A(x) is the feature weight calculated by the attention mechanism.

[0371] Assume there is a set of fused feature values, based on the previously calculated attention weights:

[0372] 1. Image feature (bladder wall thickness) f1 = 2.5 weight A(f1) = 0.335;

[0373] 2. Time series characteristics (urinary flow rate changes) f2 = 3.0 weight A(f2) = 0.429;

[0374] 3. Physiological parameter characteristics (blood pressure) f3 = 1.8 weight A (f3) = 0.236.

[0375] Approximate the integral as a discrete sum and choose the integration point x i ={1.0,1.5,2.0,2.5,3.0}, set the Gaussian function to be mainly concentrated near these points, and calculate the value of each integral point

[0376] x=1.0:

[0377]

[0378] x=1.5:

[0379]

[0380] x=2.0:

[0381]

[0382] x=2.5:

[0383]

[0384] x=3.0:

[0385]

[0386] Sum all the calculation results:

[0387] F(x)≈0.086+0.053+0.015+0.002+0.000158≈0.156

[0388] The optimized feature F(x) calculated by the feature optimization formula reflects the comprehensive influence of each feature. The model can use this optimized feature to improve the prediction accuracy of multiple physiological indicators.

[0389] Embodiment 6:

[0390] The embodiment combines the convolutional neural network CNN and the long short-term memory network LSTM to extract the key features of the patient's bladder and urethra imaging data and physiological time series data, and uses the following encoding formula to encode the feature v i Convert to an encoded vector:

[0391]

[0392] In this formula:

[0393] v i are key features extracted from medical images and physiological data;

[0394] α i and β i It is a tuning parameter used to control the contribution and nonlinear response of the feature, and its value range is 0.1 to 1.0;

[0395] n is the number of features.

[0396] Suppose that we are processing data for a patient, including the following characteristics:

[0397] Bladder wall thickness (imaging feature) v1 = 3.2 mm;

[0398] Urethral resistance (imaging feature) v2 = 1.8;

[0399] Changes in urine flow rate (time series characteristics) v3 = 2.5;

[0400] Blood pressure fluctuation (physiological parameter) v4=2.0.

[0401] Set the tuning parameter α i = 0.5 and βi = 0.3 for all features.

[0402] Compute the encoded value for each feature:

[0403] Features v1:

[0404] E(v1)=0.5·tanh(0.3·3.2)=0.5·tanh(0.96)≈0.5·0.744≈0.372

[0405] Features v2:

[0406] E(v2)=0.5·tanh(0.3·1.8)=0.5·tanh(0.54)≈0.5·0.494≈0.247

[0407] Features v3:

[0408] E(v3)=0.5·tanh(0.3·2.5)=0.5·tanh(0.75)≈0.5·0.635≈0.318

[0409] Features v4:

[0410] E(v4)=0.5·tanh(0.3·2.0)=0.5·tanh(0.6)≈0.5·0.537≈0.268

[0411] The sum is then used to get the final encoding vector:

[0412] Total encoding vector E(v):

[0413]

[0414] The encoding vector E(v)=1.205 obtained by calculation will be used as input to the subsequent personalized solution generation model.

[0415] The last example generates a personalized catheterization plan through a variational autoencoder VAE, using the formula:

[0416]

[0417] in:

[0418] z is a latent variable sampled from a normal distribution N(0,1);

[0419] γ is the coefficient for adjusting the reconstruction error, which is set to γ ​​= 0.4 here;

[0420] Sampling from N(0,1) yields z=0.6.

[0421] First calculate the integral:

[0422]

[0423] z 2 -γz is converted to The integral formula then becomes:

[0424]

[0425] This integral can be obtained using the Gaussian integral formula:

[0426]

[0427] Because the result of the Gaussian integral formula is a constant, the overall integral is weighted:

[0428]

[0429] For γ = 0.4, calculate:

[0430]

[0431] Therefore, the integral result is approximately

[0432] Calculate G(z):

[0433] G(0.6)=0.6×1.844≈1.106

[0434] The result G(z)=1.106 is obtained by calculation. After sampling from the latent space, this result can be used to adjust the specific parameters of the personalized catheterization plan.

[0435] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A method for optimizing personalized urinary catheterization scheme for patients in nephrology based on deep learning, characterized by The following steps are involved: S1. First, collect patients’ medical history, medication records, imaging data, physiological parameters, and laboratory test data from multiple hospitals, and generate diverse training data through data enhancement using the Generative Adversarial Network (GAN). S2. Then, a joint feature extraction network is constructed, using convolutional neural network (CNN) to extract features of image data and long short-term memory (LSTM) to extract features of time series data; and a multi-task learning framework is designed to simultaneously predict multiple related indicators such as urine flow rate, bladder pressure, and complication risk; S3. Introduce an attention mechanism to enable the personalized catheterization plan model to focus on the key features that affect the catheterization plan for individual patients. S4. Finally, the extracted features are encoded through the encoder to generate the patient feature vector, and the variational autoencoder VAE is used to generate a personalized catheterization plan.

2. The method for optimizing personalized urinary catheterization scheme for nephrology patients based on deep learning according to claim 1, characterized in that The process of generating diverse training data includes: first, collecting the medical history, medication use records, imaging data, physiological parameters and laboratory test data of patients in the Department of Nephrology from multiple hospitals. The data collection formula is defined as: Where x represents a continuous variable, which is any physiological measurement or test result of the patient, a(x) is the weight function of the data type, and μ and σ are the expected and standard deviations, respectively; Secondly, the collected data is preprocessed, including standardization and denoising, and the preprocessing formula is applied: Where k is the sensitivity parameter for adjusting data smoothing, y represents the data points after preprocessing, and α i is the weight coefficient specific to the data type, and n is the number of data points; Next, we use the generative adversarial network GAN to enhance the data. The enhancement formula is: f3(z)=z·(sin(az) 2 )+cos(bz 2 )) Where a and b are parameters used to adjust the degree of variation of the synthetic data, and z is the random noise drawn from a normal distribution.

3. The method for optimizing personalized urinary catheterization scheme for nephrology patients based on deep learning according to claim 1, characterized in that The joint feature extraction network construction process is as follows: S1. First, use medical imaging data, including CT or MRI images of the bladder and urethra, develop a deep convolutional neural network (CNN) architecture to extract features from the imaging data, and use image enhancement techniques, including contrast adjustment and sharpening, to improve the recognizability of key structures and pathological changes in the imaging data. S2, then processes the physiological time series data directly related to catheterization, including urine flow rate and bladder pressure, through the long short-term memory network LSTM; S3. Construct a multi-task learning framework to share the features extracted by CNN and LSTM to predict urine flow rate, bladder pressure and the risk of potential complications respectively, and adopt loss functions adapted to different tasks. The urine flow rate and bladder pressure use mean square error, while the risk of complications uses cross entropy loss.

4. The method for optimizing personalized urinary catheterization scheme for nephrology patients based on deep learning according to claim 3, characterized in that The method of developing a deep convolutional neural network (CNN) architecture extracts the image data features of the bladder and urethra through the function: Where x and y represent the spatial coordinates of the image data, α i is the adjustment parameter, n is the number of network layers; Next, the image data is enhanced, including contrast adjustment and sharpening, using the function to process each pixel: Where z represents the intensity value or gray level of each pixel in the image; Finally, we use the deep learning model to extract and analyze features using the function: To evaluate the statistical importance of features, u and v represent the coordinates in the multidimensional data space, p and q represent the extracted feature dimensions, and β is the normalization parameter.

5. The method for optimizing personalized urinary catheterization scheme for nephrology patients based on deep learning according to claim 3, characterized in that The long short-term memory network LSTM construction process is as follows: S1. First, the physiological time series data, including the collected urine flow rate and bladder pressure, are preprocessed and the normalization processing formula is applied: Transform the data t, where μ and σ represent the mean and standard deviation of the data set respectively; S2. Next, the time series data is converted into a training sequence through a complex sliding window technique, and the function is applied: Where s(u) is the data value at time point u, δ is the window length, and λ is the time decay coefficient; S3. Then design and implement the long short-term memory network LSTM model, which processes data through a dynamically adjusted network formula: L(S)=tanh(W h h t-1 +W x x t +W c ·(c t-1 ⊙σ f )+b) where h t-1 and c t-1 They are the hidden state and cell state at the previous moment, x t is the current input, W h ,W x ,W c and b are network parameters, σ f is the activation function of the forget gate, ⊙ represents element-wise multiplication, tanh and σ are activation functions that simulate and predict the dynamic changes of physiological data.

6. The method for optimizing personalized urinary catheterization scheme for nephrology patients based on deep learning according to claim 3, characterized in that The method for constructing a multi-task learning framework includes: firstly, using a composite convolutional neural network (CNN) and a long short-term memory (LSTM) network framework, by formula: To fuse the image features f extracted by CNN and the time series features g extracted by LSTM, where σ represents the adjustment factor of feature fusion; Next, we build a multi-task learning framework using a variational method: Processing and optimization features x, t is the integration variable; For the prediction tasks of urine flow rate and bladder pressure, the mean square error loss function is applied: where y(t) and Represent the functional forms of the true value and the predicted value respectively, t is the integral variable, For the prediction of complication risk, the cross entropy loss function is used: Where p(x) and q(x) represent the density functions of the continuous probability distribution, and x represents a continuous variable.

7. The method for optimizing personalized urinary catheterization scheme for nephrology patients based on deep learning according to claim 1, characterized in that The attention mechanism construction method is as follows: S1. First, a hybrid neural network framework is used to combine the convolutional neural network (CNN) and the long short-term memory (LSTM) network, and a self-attention mechanism is applied to enhance the model's feature understanding ability. The attention mechanism is expressed by the formula: Calculate each feature f i The weight of , where β is a learnable parameter, representing the adjustment coefficient of feature importance; S2. Subsequently, the features extracted by CNN and LSTM are fused, and the importance of each feature is dynamically evaluated using the integral formula: To optimize feature focusing, x represents the features extracted from the data.

8. The method for optimizing personalized urinary catheterization scheme for nephrology patients based on deep learning according to claim 1, characterized in that The process of generating a personalized catheterization plan is as follows: First, the key features of the medical images and physiological time series data of the bladder and urethra are extracted by combining the convolutional neural network CNN and the long short-term memory network LSTM, and the encoding formula is used: The feature v i Converted to a coding vector, where α i and β i To adjust the parameters; Next, use the variational autoencoder VAE to generate the formula: The personalized catheterization plan is sampled and reconstructed from the latent space, where z represents the latent variable sampled from the normal distribution and γ is the coefficient to adjust the reconstruction error.

Citation Information

Cited By

  • Method for predicting life of battery of humanoid robot and storage medium

    CN119783538A

  • Construction method of human liver deformation parameter prediction model for individualized virtual laparoscope system

    CN120108743A

  • Methods for constructing a human liver deformation parameter prediction model for personalized virtual laparoscopic systems

    CN120108743B