AI-assisted hemodialysis patient nutrition algorithm

Abstract

The invention provides an AI-assisted hemodialysis patient nutrition algorithm, which constructs a hemodialysis patient nutrition algorithm calculation model through a convolutional neural network, and calculates the nutrition demand and potential risk of a patient by inputting related parameter data of the patient into the calculation model. Personalized customization schemes of diet structure suggestions, specific nutrient supplement suggestions and diet taboo are made for the patient, physiological indexes and diet habit changes of the patient are continuously monitored in real time through a wearable device and an intelligent sensor, a nutrition algorithm calculation model can respond in time, and a nutrition scheme is dynamically adjusted; and when the physiological indexes, dietary habits or psychological states of the patient change, the calculation model is automatically triggered to re-evaluate and update the nutrition scheme, so that the nutrition scheme is always matched with the actual demand of the patient, and the hemodialysis patient can be favorably recovered.

Description

Technical Field

[0001] The present invention belongs to the technical field of nutrition, and specifically relates to an AI-assisted nutrition algorithm for hemodialysis patients. Background Art

[0002] While some existing technologies are relevant to nutritional management of hemodialysis patients, they still have some shortcomings. Existing technologies often fail to fully account for individual patient differences, resulting in inaccurate nutritional algorithms that struggle to meet the specific needs of diverse patients. Furthermore, low integration with other healthcare processes hinders comprehensive health management.

[0003] Traditional nutritional assessment methods for hemodialysis patients rely primarily on self-reporting, a process significantly impacted by memory bias, subjective perceptions, and compliance. Patients may omit information about their food intake or miscalculate portion sizes, resulting in inaccurate data that fails to truly reflect their actual nutritional intake. Furthermore, frequent inquiries can be burdensome for patients and affect their compliance.

[0004] Some existing artificial intelligence technologies have a low degree of integration with the hospital's medical system, making it impossible to achieve data sharing and collaborative management. This requires doctors and nutritionists to spend a lot of time and energy collecting and organizing data when obtaining patients' nutritional information, affecting the efficiency and quality of treatment.

[0005] Therefore, it is necessary to study an AI-assisted nutrition algorithm for hemodialysis patients, which can not only solve the problem of inaccurate nutritional assessment, but also realize the generation of personalized nutrition plans and provide a dynamic adjustment mechanism for hemodialysis patients. Summary of the Invention

[0006] In response to the shortcomings of the existing technology, the present invention provides an AI-assisted nutrition algorithm for hemodialysis patients to solve the problems in the existing technology of inaccurate nutrition assessment, inability to perform dynamic adjustments, and inability to form personalized nutrition plans.

[0007] The technical solution of the present invention is: an AI-assisted hemodialysis patient nutrition algorithm, comprising the following steps:

[0008] Step 1: The structural composition of the nutrition calculation model, including:

[0009] The basic physiological index collection module uses high-precision sensors built into wearable devices to monitor the basic physiological indicators of hemodialysis patients, such as blood pressure, heart rate, and weight, in real time. These wearable devices automatically transmit data to mobile terminals or directly upload data to cloud servers via Bluetooth 4.0 or above or Wi-Fi wireless networks. The data transmission frequency of the basic physiological index collection module can be set to store data once every fixed time period;

[0010] The blood test index collection module takes samples from hemodialysis patients and uses the hospital's advanced blood testing equipment to obtain relevant indicators of the patient's hemoglobin, serum albumin, and electrolyte levels. The patient's test results are then imported into the blood test index collection module through the hospital's testing and inspection system via a safe and reliable interface program. The data update frequency of the blood test index collection module is set to a fixed period based on the patient's dialysis cycle and the doctor's advice.

[0011] The dialysis treatment parameter collection module collects relevant parameters of hemodialysis patients, such as dialysis frequency, dialysis time, and ultrafiltration volume, through the hospital's hemodialysis management system. The hospital's hemodialysis management system automatically transmits the relevant parameters to the dialysis treatment parameter collection module immediately after each dialysis treatment of the hemodialysis patient.

[0012] Patient food and diet collection module: Hemodialysis patients fill out a detailed diet record questionnaire, recording the type, amount, and time of food intake each day. This questionnaire must be completed at least once a day, and the completed data is uploaded to the patient food and diet collection module in real time;

[0013] In the psychological status assessment module, hemodialysis patients fill out a professionally validated psychological assessment scale in detail. The patient's anxiety and depression status are assessed at least once a week, and the assessment data is uploaded to the psychological status assessment module in real time.

[0014] Step 2: Data preprocessing;

[0015] Screen and preprocess the collected multi-dimensional data, remove obvious outliers and obvious erroneous data, standardize different data, set the original data as X, and the standardized data as X std , for a certain data feature i, its standardization formula is where μ i is the mean of the feature in the training data, σ i is the standard deviation of the feature in the training data;

[0016] Step 3: Use convolutional neural network to build a computational model;

[0017] Input layer: Receives preprocessed multi-dimensional data. If the batch size is m and the number of features is n, the input shape is (m, n). This layer is mainly responsible for inputting data into the network.

[0018] Convolution layer: Use multiple convolution kernels to perform feature extraction operations, where the convolution kernel size is set to (K1, K2), and slide the window with a step size S on the feature dimension of the data. The input shape (m, n) is convolved, and the output feature map size is calculated according to the formula Calculate, where p is the convolutional layer filling value, here we use VALID filling, that is, p = 0, then the output feature map size is

[0019] The number of convolution kernels is set to C. Each convolution kernel is convolved with the input data and then transformed nonlinearly using the ReLU activation function g(x)=max(0,X) to obtain multiple feature maps and enhance the network's ability to perceive data features.

[0020] Pooling layer: The maximum pooling method is used, the pooling specification is (K3, K4), and the dimension of the convolution layer output map is reduced with a step size of S1. The input feature map (m, o) is pooled and the output feature map size is Where p1 is the filling value of the pooling layer, which is usually 0, effectively reducing the amount of data and retaining key features, reducing the risk of overfitting;

[0021] Fully connected layer: First, flatten the pooling layer output into a one-dimensional vector. Set the pooling layer output shape to (p2, q), then the length after flattening is p2×q;

[0022] The fully connected layer contains r neurons, the weight matrix W2 is of dimension (p2×q, r), and the bias vector b2 is of length r. The input vector is multiplied by W2 and added to b2, and then activated by ReLU to obtain the intermediate result H2;

[0023] The final fully connected layer output layer is used to predict the patient's nutritional needs assessment results and personalized nutrition plan recommendations;

[0024] Depending on the output type, the number of neurons and activation function are different. When predicting whether a patient is deficient in a certain nutrient, the number of neurons in the output layer, r, is 2, and the activation function can be the softmax function. When predicting the amount of nutrient deficiency a patient needs, the number of neurons in the output layer is determined by the type of nutrient to be predicted, and the activation function can be a linear function.

[0025] Set the output layer weight matrix to W3 and the bias vector to b3, then the model's final output is Y = f(W3H2+b3), where H2 is the intermediate result of the fully connected layer;

[0026] Parameter initialization: The convolutional kernel weights of the convolutional layer are initialized to random values ​​ranging from -0.1 to 0.1, and the bias is initialized to an all-zero vector. The weight matrices W2 and W3 of the fully connected layer are also initialized to random values ​​ranging from -0.1 to 0.1, and the bias vector b2 and b3 are initialized to all-zero vectors. This initialization method helps break the symmetry and enables the network to learn different features in the early stages of training.

[0027] Step 4: Compute model training and optimization layer;

[0028] The computational model is trained using a large amount of multi-dimensional data from historical hemodialysis patients as a training set. Through multiple iterations of training, the weights and bias parameters in the computational model are continuously adjusted to improve the accuracy and generalization ability of the computational model. The computational model is evaluated regularly to analyze its accuracy in predicting nutritional needs and formulating nutrition.

[0029] For a training sample (X i , Y i ), the model prediction output is The loss function L selects an appropriate function according to the task type. If it is a classification task, the cross entropy loss function is selected. Where C is the number of categories, y ij is the true label, is the predicted probability; if the mean square error loss function is selected for the regression task Where m is the number of samples; this is used to measure the difference between the predicted output and the true label;

[0030] Use the gradient descent algorithm and adopt the learning rate decay strategy according to the training situation to update the parameters. The update formula is the convolution layer weight Fully connected layer weights The bias update formula is similar to ( and are the gradients of the loss function with respect to b1, b2, and b3 respectively;

[0031] Repeat the above steps until the accuracy of the calculated model on the training set reaches the predetermined threshold or the number of training rounds reaches the set value;

[0032] Step 5: Computational model testing and optimization layer;

[0033] Patient data different from the training set is used as the test set to test and evaluate the trained and optimized calculation model. The evaluation indicators are calculation accuracy and Where TP is a true positive example, TN is a true negative example, FP is a false positive example, and FN is a false negative example; recall rate The harmonic mean of precision and recall

[0034] Based on the evaluation results, if the calculation model does not achieve the expected results, optimization operations are performed;

[0035] Step 6: Generating personalized nutrition plan for patients;

[0036] For new hemodialysis patients, their relevant data parameters are collected and pre-processed before being input into a trained and optimized computational model. The computational model analyzes the input data to calculate the nutritional needs and potential risks of hemodialysis patients. Combining medical nutrition guidelines and nutrition experts, a personalized nutrition plan is generated for the patient. The personalized nutrition plan includes dietary structure, specific supplements of required nutrients, dietary taboos, and is combined with the patient's taste preferences and lifestyle habits.

[0037] Step 7: Patient’s personalized nutrition plan adjustment stage;

[0038] The wearable device continuously monitors the patient's physiological parameters through high-precision sensors built into the device. When the monitored data exceeds the preset range or the patient's feedback changes, the calculation model is automatically triggered to re-evaluate the patient's nutritional status and the patient's nutritional plan is promptly adjusted and updated based on the new data.

[0039] In the first step, the wearable device is a smart bracelet or a smart watch, the data transmission frequency of the basic physiological index acquisition module is set to 15 minutes, and the data update frequency of the blood test index acquisition module is set to one week.

[0040] In the fourth step, the computational model is iteratively trained no less than 1,000 times, the evaluation period for weight and bias parameters in the computational model is set to one month, and the accuracy threshold of the computational model on the training set is set to 90%.

[0041] Beneficial effects of the present invention:

[0042] The present invention constructs a nutritional algorithm calculation model for hemodialysis patients through a convolutional neural network. By inputting the patient's relevant parameter data into the calculation model, the patient's nutritional needs and potential risks are calculated, and a personalized customized plan with dietary structure recommendations, specific nutrient supplement recommendations, and dietary taboos is formulated for the patient. In addition, the patient's physiological indicators and eating habits are continuously monitored in real time through wearable devices and smart sensors. The nutritional algorithm calculation model can respond in a timely manner and dynamically adjust the nutritional plan. When the patient's physiological indicators, eating habits or psychological state change, the calculation model is automatically triggered to re-evaluate and update the nutritional plan to ensure that the nutritional plan always matches the patient's actual needs, thereby facilitating the recovery of hemodialysis patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a schematic diagram of the principle flow of the present invention. DETAILED DESCRIPTION

[0044] In order to make the objectives, technical solutions and beneficial effects of the present invention more clear, the embodiments of the present invention are further described in detail below with reference to the accompanying drawings.

[0045] An AI-assisted nutrition algorithm for hemodialysis patients includes the following steps:

[0046] Step 1: The structural composition of the nutrition calculation model, including:

[0047] The basic physiological index acquisition module uses the high-precision sensors built into wearable devices to monitor the basic physiological indicators of blood pressure, heart rate, and weight of hemodialysis patients in real time. Wearable devices usually use smart bracelets and smart watches. The measurement accuracy of wearable devices can reach . These wearable devices automatically transmit data to mobile terminals or directly upload them to cloud servers via Bluetooth 4.0 and above or Wi-Fi (802.11n and above standards) wireless networks. The data transmission frequency of the basic physiological index acquisition module can be set to store once every 15 minutes.

[0048] The blood test index collection module takes samples from hemodialysis patients and uses the hospital's advanced blood testing equipment to obtain the patient's hemoglobin, serum albumin, and electrolyte level related indicators. The patient's test results are then imported into the blood test index collection module through a safe and reliable interface program through the hospital's testing and inspection system. The data update frequency of the blood test index collection module is set to a fixed cycle based on the patient's dialysis cycle and the doctor's recommendations.

[0049] The dialysis treatment parameter acquisition module collects the relevant parameters of hemodialysis patients, such as dialysis frequency, dialysis time, and ultrafiltration volume, through the hospital's hemodialysis management system. The hospital's hemodialysis management system automatically transmits the relevant parameters to the dialysis treatment parameter acquisition module immediately after each dialysis treatment of the hemodialysis patient. The dialysis frequency accurately records the time interval between each dialysis of the patient, and the dialysis time must be accurate to the minute.

[0050] Patient food and diet collection module: Hemodialysis patients fill out a detailed diet record questionnaire to record the types, amounts, and meal time parameters of daily food intake. The questionnaire must be filled out at least once a day, and the completed data will be uploaded to the patient food and diet collection module in real time. The types of food must be accurate to the specific ingredient names, the amount of food intake must be accurate to , and the meal time must be accurate to the minute.

[0051] In the psychological state assessment module, hemodialysis patients fill out a professionally validated psychological assessment scale in detail. The patient's anxiety and depression status are assessed at least once a week, and the assessment data are uploaded to the psychological state assessment module in real time. The psychological assessment scale usually includes the Self-Rating Anxiety Scale (SAS) and the Self-Rating Depression Scale (SDS). Both the Self-Rating Anxiety Scale (SAS) and the Self-Rating Depression Scale (SDS) are calculated using standard scoring rules. The score range of the Self-Rating Anxiety Scale (SAS) is 20-80 points, and the score range of the Self-Rating Depression Scale (SDS) is 25-100 points.

[0052] Step 2: Data preprocessing;

[0053] Screen and preprocess the collected multi-dimensional data, remove obvious outliers and obvious erroneous data, standardize different data, set the original data as X, and the standardized data as X std , for a certain data feature i, its standardization formula is where μ i is the mean of the feature in the training data, σ i It is the standard deviation of the feature in the training data; obvious outliers such as blood pressure higher than 200 / 120mmHg or lower than 80 / 50mmHg are beyond the normal physiological range; obvious erroneous data refer to those with incorrect format or obviously unreasonable data.

[0054] For example: For weight data, if the mean weight in the training set is Standard deviation σ w =10kg, then the standardized value of a patient's weight W = 70kg is

[0055] Step 3: Use convolutional neural network to build a computational model;

[0056] Input layer: Receives preprocessed multi-dimensional data. Set the batch size to m, m is 32, the number of features to n, n is 15, and the input shape is (m, n). This layer is mainly responsible for inputting data into the network.

[0057] Convolution layer: Use multiple convolution kernels to perform feature extraction operations, where the convolution kernel size is set to (K1, K2), and slide the window with a step size S on the feature dimension of the data. The input shape (m, n) is convolved, and the output feature map size is calculated according to the formula Calculate, where p is the convolutional layer filling value, here we use VALID filling, that is, p = 0, then the output feature map size is This embodiment inputs (32, 15) and outputs (32, 15).

[0058] Set the number of convolution kernels to C, C is 32, each convolution kernel is convolved with the input data and then transformed nonlinearly using the ReLU activation function g(x)=max(0,X) to obtain multiple feature maps and enhance the network's ability to perceive data features;

[0059] Pooling layer: The maximum pooling method is used, the pooling specification is (K3, K4), and the dimension of the convolution layer output map is reduced with a step size of S1. The input feature map (m, o) is pooled and the output feature map size is Where p1 is the filling value of the pooling layer, which is usually 0, effectively reducing the amount of data and retaining key features, reducing the risk of overfitting; in this embodiment, K3 is 2, K4 is 1, S3 is 2, the input is (32, 15), and the output becomes (16, 8).

[0060] Fully connected layer: First, flatten the pooling layer output into a one-dimensional vector, and set the pooling layer output shape to (p2, q). The length after flattening is p2×q. In this embodiment, p2 is 16 and q is 8, so the length after flattening is p2×q=108.

[0061] The fully connected layer contains r neurons, and the weight matrix W2 has dimensions (p2×q, r). In this embodiment, r is 64, so the weight matrix W2 has dimensions (108, 64), and the bias vector b2 has length r. The input vector is multiplied by W2 and added to b2, and then activated by ReLU to obtain the intermediate result H2;

[0062] The final fully connected layer output layer is used to predict the patient's nutritional needs assessment results and personalized nutrition plan recommendations;

[0063] Depending on the output type, the number of neurons and activation function are different. When predicting whether a patient is deficient in a certain nutrient, the number of neurons in the output layer, r, is 2, and the activation function can be the softmax function. When predicting the amount of nutrient deficiency a patient needs, the number of neurons in the output layer is determined by the type of nutrient to be predicted, and the activation function can be a linear function.

[0064] Set the output layer weight matrix to W3 and the bias vector to b3, then the model's final output is Y = f(W3H2+b3), where H2 is the intermediate result of the fully connected layer;

[0065] Parameter initialization: The convolutional kernel weights of the convolutional layer are initialized to random values ​​ranging from -0.1 to 0.1, and the bias is initialized to an all-zero vector. The weight matrices W2 and W3 of the fully connected layer are also initialized to random values ​​ranging from -0.1 to 0.1, and the bias vector b2 and b3 are initialized to all-zero vectors. This initialization method helps break the symmetry and enables the network to learn different features in the early stages of training.

[0066] Step 4: Compute model training and optimization layer;

[0067] A large amount of multi-dimensional data of historical hemodialysis patients is used as a training set. In this embodiment, data of no less than 1000 patients in the past five years are used as a training set <(X1, Y1), (X2, Y2), ..., (X n , Y n)>, training the computational model, through multiple iterative training, the number of training rounds in this embodiment is 1000 rounds, continuously adjusting the weights and bias parameters in the computational model to improve the accuracy and generalization ability of the computational model, regularly evaluating the computational model to analyze the accuracy of the computational model in predicting nutritional needs and formulating nutrition;

[0068] For a training sample (X i , Y i ), the model prediction output is The loss function L selects an appropriate function according to the task type. If it is a classification task, the cross entropy loss function is selected. Where C is the number of categories, y ij is the true label, is the predicted probability; if the mean square error loss function is selected for the regression task Where m is the number of samples; this is used to measure the difference between the predicted output and the true label;

[0069] Use the gradient descent algorithm and adopt the learning rate decay strategy according to the training situation to update the parameters. The update formula is the convolution layer weight Fully connected layer weights The bias update formula is similar to ( and are the gradients of the loss function with respect to b1, b2, and b3, respectively. In this embodiment, stochastic gradient descent (SGD) is adopted, with a learning rate of α = 0.001, and a learning rate decay strategy can be adopted according to the training situation, such as decaying the parameters by 0.9 every 100 rounds.

[0070] Repeat the above steps until the accuracy of the calculation model on the training set reaches a predetermined threshold or the number of training rounds reaches a set value; in this embodiment, the predetermined threshold is 90% and the set value of the number of training rounds is 1000 rounds.

[0071] Step 5: Computational model testing and optimization layer;

[0072] Patient data different from the training set is used as the test set to test and evaluate the trained and optimized calculation model. The evaluation indicators are calculation accuracy and Where TP is a true positive example, TN is a true negative example, FP is a false positive example, and FN is a false negative example; recall rate The harmonic mean of precision and recall

[0073] Based on the evaluation results, if the calculation model does not achieve the expected results, optimization operations are performed;

[0074] In this embodiment, the learning rate α is adjusted to 0.0005, the regularization parameter λ is adjusted to 0.01, etc. The regularization term can be added to the loss function, such as To prevent overfitting and improve the model structure, in this embodiment, the hidden layer depth is increased to 4 layers and the number of neurons is adjusted to 128-64-32-16.

[0075] Step 6: Generating personalized nutrition plan for patients;

[0076] For new hemodialysis patients, their relevant data parameters are collected and pre-processed before being input into a trained and optimized computational model. The computational model analyzes the input data to calculate the nutritional needs and potential risks of hemodialysis patients. Combining medical nutrition guidelines and nutrition experts, a personalized nutrition plan is generated for the patient. The personalized nutrition plan includes dietary structure, specific supplements of required nutrients, dietary taboos, and is combined with the patient's taste preferences and lifestyle habits.

[0077] In this embodiment, the patient's blood test indicators, dialysis treatment parameters and dietary habits are analyzed to determine whether the patient has protein deficiency P. If Hb std <110g / L and Alb std <35g / L, then P=1, otherwise P=0); determine whether the patient has electrolyte imbalance E, if K std <3.5mmol / L or K std >5.5mmol / L, etc., then E=1, otherwise E=0.

[0078] In this example, the recommended food types and proportions for intake, if the patient is protein deficient, the recommended amount of protein source food The recommended daily protein intake is 1.2 g / kg, which is determined based on the patient's condition. protein is the protein content of common protein sources; if the patient is vitamin D deficient, that is, Y vtcamtnD <20ng / ml, it is recommended to supplement vitamin D; if the patient’s blood potassium is high, that is, K std >5.5mmol / L, high potassium foods are contraindicated.

[0079] Step 7: Patient’s personalized nutrition plan adjustment stage;

[0080] The wearable device continuously monitors the patient's physiological parameters through high-precision sensors built into the device. When the monitored data exceeds the preset range or the patient's feedback changes, the calculation model is automatically triggered to re-evaluate the patient's nutritional status and the patient's nutritional plan is promptly adjusted and updated based on the new data.

[0081] The above is a preferred embodiment of the present invention. Those skilled in the art should understand that the embodiments of the present invention are not limited to the above content. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the scope of protection of the present invention.

Claims

1. An AI-assisted nutrition algorithm for hemodialysis patients, characterized by: The following steps are involved: Step 1: The structural composition of the nutrition calculation model, including: The basic physiological index collection module uses high-precision sensors built into wearable devices to monitor the basic physiological indicators of hemodialysis patients, such as blood pressure, heart rate, and weight, in real time. These wearable devices automatically transmit data to mobile terminals or directly upload data to cloud servers via Bluetooth 4.0 or above or Wi-Fi wireless networks. The data transmission frequency of the basic physiological index collection module can be set to store data once every fixed time period; The blood test index collection module takes samples from hemodialysis patients and uses the hospital's advanced blood testing equipment to obtain relevant indicators of the patient's hemoglobin, serum albumin, and electrolyte levels. The patient's test results are then imported into the blood test index collection module through the hospital's testing and inspection system via a safe and reliable interface program. The data update frequency of the blood test index collection module is set to a fixed period based on the patient's dialysis cycle and the doctor's advice. The dialysis treatment parameter collection module collects relevant parameters of hemodialysis patients, such as dialysis frequency, dialysis time, and ultrafiltration volume, through the hospital's hemodialysis management system. The hospital's hemodialysis management system automatically transmits the relevant parameters to the dialysis treatment parameter collection module immediately after each dialysis treatment of the hemodialysis patient. Patient food and diet collection module: Hemodialysis patients fill out a detailed diet record questionnaire, recording the type, amount, and time of food intake each day. This questionnaire must be completed at least once a day, and the completed data is uploaded to the patient food and diet collection module in real time; In the psychological status assessment module, hemodialysis patients fill out a professionally validated psychological assessment scale in detail. The patient's anxiety and depression status are assessed at least once a week, and the assessment data is uploaded to the psychological status assessment module in real time. Step 2: Data preprocessing; Screen and preprocess the collected multi-dimensional data, remove obvious outliers and obvious erroneous data, standardize different data, set the original data as X, and the standardized data as X std , for a certain data feature i, its standardization formula is where μ i is the mean of the feature in the training data, σ i is the standard deviation of the feature in the training data; Step 3: Use convolutional neural network to build a computational model; Input layer: Receives preprocessed multi-dimensional data. If the batch size is m and the number of features is n, the input shape is (m, n). This layer is mainly responsible for inputting data into the network. Convolution layer: Use multiple convolution kernels to perform feature extraction operations, where the convolution kernel size is set to (K1, K2), and slide the window with a step size S on the feature dimension of the data. The input shape (m, n) is convolved, and the output feature map size is calculated according to the formula Calculate, where p is the convolutional layer filling value, here we use VALID filling, that is, p = 0, then the output feature map size is Set the number of convolution kernels to C, convolve each convolution kernel with the input data and perform nonlinear transformation using the ReLU activation function g(x)=max(0,X) to obtain multiple feature maps and enhance the network's ability to perceive data features. Pooling layer: The maximum pooling method is used, the pooling specification is (K3, K4), and the dimension of the convolution layer output map is reduced with a step size of S1. The input feature map (m, o) is pooled and the output feature map size is Where p1 is the filling value of the pooling layer, which is usually 0, effectively reducing the amount of data and retaining key features, reducing the risk of overfitting; Fully connected layer: First, flatten the pooling layer output into a one-dimensional vector. Set the pooling layer output shape to (p2, q), then the length after flattening is p2×q; The fully connected layer contains r neurons, the weight matrix W2 is of dimension (p2×q, r), and the bias vector 2 is of length r. The input vector is multiplied by W2 and added with b2, and then activated by ReLU to obtain the intermediate result H2; The final fully connected layer output layer is used to predict the patient's nutritional needs assessment results and personalized nutrition plan recommendations; Depending on the output type, the number of neurons and activation function are different. When predicting whether a patient is deficient in a certain nutrient, the number of neurons in the output layer, r, is 2, and the activation function can be the softmax function. When predicting the amount of nutrient deficiency a patient needs, the number of neurons in the output layer is determined by the type of nutrient to be predicted, and the activation function can be a linear function. Set the output layer weight matrix to W3 and the bias vector to b3, then the model's final output is Y = f(W3H2+b3), where H2 is the intermediate result of the fully connected layer; Parameter initialization: The convolutional kernel weights of the convolutional layer are initialized to random values ​​ranging from -0.1 to 0.1, and the bias is initialized to an all-zero vector. The weight matrices W2 and W3 of the fully connected layer are also initialized to random values ​​ranging from -0.1 to 0.1, and the bias vector b2 and b3 are initialized to all-zero vectors. This initialization method helps break the symmetry and enables the network to learn different features in the early stages of training. Step 4: Compute model training and optimization layer; The computational model is trained using a large amount of multi-dimensional data from historical hemodialysis patients as a training set. Through multiple iterations of training, the weights and bias parameters in the computational model are continuously adjusted to improve the accuracy and generalization ability of the computational model. The computational model is evaluated regularly to analyze its accuracy in predicting nutritional needs and formulating nutrition. For a training sample (X i , Y i ), the model prediction output is The loss function L selects an appropriate function according to the task type. If it is a classification task, the cross entropy loss function is selected. Where C is the number of categories, y ij is the true label, is the predicted probability; if the mean square error loss function is selected for the regression task Where m is the number of samples; this is used to measure the difference between the predicted output and the true label; Use the gradient descent algorithm and adopt the learning rate decay strategy according to the training situation to update the parameters. The update formula is the convolution layer weight Fully connected layer weights The bias update formula is similar to ( and are the gradients of the loss function with respect to b1, b2, and b3 respectively; Repeat the above steps until the accuracy of the calculated model on the training set reaches the predetermined threshold or the number of training rounds reaches the set value; Step 5: Computational model testing and optimization layer; Patient data different from the training set is used as the test set to test and evaluate the trained and optimized calculation model. The evaluation indicators are calculation accuracy and Where TP is a true positive example, TN is a true negative example, FP is a false positive example, and FN is a false negative example; recall rate The harmonic mean of precision and recall Based on the evaluation results, if the calculation model does not achieve the expected results, optimization operations are performed; Step 6: Generating personalized nutrition plan for patients; For new hemodialysis patients, their relevant data parameters are collected and pre-processed before being input into a trained and optimized computational model. The computational model analyzes the input data to calculate the nutritional needs and potential risks of hemodialysis patients. Combining medical nutrition guidelines and nutrition experts, a personalized nutrition plan is generated for the patient. The personalized nutrition plan includes dietary structure, specific supplements of required nutrients, dietary taboos, and is combined with the patient's taste preferences and lifestyle habits. Step 7: Patient’s personalized nutrition plan adjustment stage; The wearable device continuously monitors the patient's physiological parameters through high-precision sensors built into the device. When the monitored data exceeds the preset range or the patient's feedback changes, the calculation model is automatically triggered to re-evaluate the patient's nutritional status and the patient's nutritional plan is promptly adjusted and updated based on the new data.

2. The AI-assisted hemodialysis patient nutrition algorithm according to claim 1, characterized in that: In the first step, the wearable device is a smart bracelet or a smart watch, the data transmission frequency of the basic physiological index acquisition module is set to 15 minutes, and the data update frequency of the blood test index acquisition module is set to one week.

3. The AI-assisted hemodialysis patient nutrition algorithm according to claim 1, characterized in that: In the fourth step, the computational model is iteratively trained no less than 1,000 times, the evaluation period for weight and bias parameters in the computational model is set to one month, and the accuracy threshold of the computational model on the training set is set to 90%.

Images