Pediatric nursing quality prediction and optimization method based on transfer learning

Abstract

The invention discloses a pediatric nursing quality prediction and optimization method based on transfer learning, and the method comprises the following steps: S1, collecting target pediatric nursing data, and carrying out the preprocessing; s2, selecting source data, and extracting source data features; s3, constructing and training a nursing quality prediction model; s4, carrying out migration and fine tuning on the nursing quality prediction model by adopting an improved fine tuning method; s5, inputting target pediatric nursing data to the fine-adjusted nursing quality prediction model, calculating a nursing quality index, and evaluating a prediction result; s6, generating a nursing scheme based on the prediction result, and optimizing the nursing scheme according to the execution effect; and S7, applying the optimized nursing scheme to actual nursing work, and establishing an updating mechanism, so that the nursing quality prediction model is combined with new data to carry out iterative optimization. The pediatric nursing quality prediction and optimization are realized by adopting transfer learning and deep learning methods, and the pediatric nursing quality prediction and optimization method has the advantages of high data adaptability, high prediction accuracy and intelligent optimization of the nursing scheme.

Description

Technical Field

[0001] The present invention relates to the field of medical nursing technology, and in particular to a pediatric nursing quality prediction and optimization method based on transfer learning. Background Art

[0002] With the rapid development of medical informatization, data such as electronic medical records, nursing records, and patient rehabilitation status have gradually been digitized, providing possibilities for nursing quality prediction and optimization. In the medical field, nursing quality directly affects the patient's rehabilitation effect and hospitalization experience. Therefore, accurately evaluating nursing quality and providing personalized optimization solutions have become an important direction of modern medical management. At present, artificial intelligence and data-driven methods are widely used in medical prediction tasks, such as disease prediction, clinical risk assessment, and intelligent nursing systems. These technologies can improve nursing efficiency, optimize medical resource allocation, and reduce nursing errors caused by human factors. However, existing technologies still have many problems in nursing quality prediction and optimization, especially in the field of pediatric nursing. Due to the scarcity and complexity of nursing data, the generalization ability of existing models is poor and it is difficult to adapt to the nursing environment of different medical institutions.

[0003] Existing nursing quality prediction methods are mainly based on traditional statistical methods and machine learning methods. In early studies, medical institutions mainly used traditional machine learning methods such as regression analysis, decision trees and support vector machines to analyze nursing data to evaluate nursing effects and risks. Although these methods can provide nursing quality predictions to a certain extent, they have the following defects: First, traditional methods are highly dependent on data and usually require a large amount of structured data, while nursing data often contains a large amount of unstructured text and time series information, which limits the performance of traditional models; second, existing statistical and traditional machine learning methods have limited performance in complex nursing scenarios, especially when nursing quality is affected by multiple factors (such as patient characteristics, nursing staff experience, nursing environment, etc.), the model is difficult to capture complex feature interactions; third, due to the relatively small amount of pediatric nursing data, the problem of data imbalance is particularly prominent, and the traditional model has insufficient learning ability in a small sample environment, resulting in low accuracy of prediction results.

[0004] In recent years, deep learning and neural networks have made significant progress in the medical field. Neural network architectures such as convolutional neural networks, long short-term memory networks, and gated recurrent units have been applied to medical prediction tasks. However, in terms of pediatric nursing quality prediction, existing deep learning models still face challenges. First, deep neural networks usually rely on a large amount of labeled data for training, and pediatric nursing data is often difficult to collect and label due to ethical restrictions, privacy protection and other factors, resulting in insufficient model training data, making it difficult for deep learning models to fully learn effective features. Secondly, medical data varies greatly across institutions. The nursing standards, equipment conditions, and patient groups of different medical institutions may be significantly different, making it difficult for models trained in a single medical institution to generalize to other institutions. In addition, the nursing quality prediction task involves not only the patient's physiological indicators, but also unstructured information in the nursing process, such as subjective feedback from nursing staff, nursing text records, etc. Existing neural network models still have certain limitations in processing multimodal data, and it is difficult to fully mine the key features in nursing data.

[0005] In terms of nursing optimization, existing methods mainly rely on rule-making and experience guidance, that is, based on the experience of nursing staff, standardized nursing processes are formulated and regularly optimized. This approach has improved the quality of nursing to a certain extent, but there are several key problems. First, the nursing plan lacks dynamic adaptability. Existing nursing optimization plans are usually based on historical experience and cannot be personalized according to the patient's real-time data. For example, different patients have different recovery speeds, but nursing plans are usually fixed and lack optimization mechanisms for individual situations. Secondly, it is difficult to optimize the allocation of nursing resources intelligently. In medical institutions, nursing resources (such as nursing staff, nursing time, and nursing equipment) are limited. How to optimally allocate them according to patient needs is a key challenge. Existing methods mainly rely on the experience of nursing managers to allocate resources, which makes it difficult to achieve precise optimization. In addition, the measurement of nursing quality is usually based on macro indicators such as nursing success rate and patient recovery status, and how to combine these indicators with the optimization of specific nursing plans still lacks a systematic solution.

[0006] Therefore, how to provide a pediatric nursing quality prediction and optimization method based on transfer learning is an urgent problem that technicians in this field need to solve. Summary of the invention

[0007] One purpose of the present invention is to propose a pediatric nursing quality prediction and optimization method based on transfer learning. The present invention utilizes transfer learning technology and improves the fine-tuning method to enable the nursing quality prediction model to adapt to the nursing environment of the target medical institution. The adaptive attention mechanism is adopted to enable the nursing quality prediction model to automatically focus on key factors in the nursing process and improve the sensitivity of the prediction model to important features.

[0008] A pediatric nursing quality prediction and optimization method based on transfer learning according to an embodiment of the present invention comprises the following steps:

[0009] S1. Collect target pediatric nursing data and preprocess the target pediatric nursing data;

[0010] S2, select source data and extract source data features;

[0011] S3, constructing a nursing quality prediction model, training the nursing quality prediction model based on source data features, enabling the nursing quality prediction model to learn nursing features, and storing training parameters of the nursing quality prediction model;

[0012] S4. Based on the preprocessed target pediatric nursing data and the training parameters of the nursing quality prediction model, the nursing quality prediction model is migrated and fine-tuned using an improved fine-tuning method, and the nursing quality prediction model and nursing quality prediction model parameters are saved;

[0013] S5. Input the target pediatric nursing data into the fine-tuned nursing quality prediction model, calculate the nursing quality indicators, and evaluate the prediction results;

[0014] S6. Generate nursing plans based on prediction results and optimize nursing plans according to implementation effects;

[0015] S7. Apply the optimized nursing plan to actual nursing work and establish an update mechanism to iteratively optimize the nursing quality prediction model in combination with new data.

[0016] Optionally, the target pediatric nursing data includes electronic medical records, nursing records, patient recovery status and nursing staff feedback from medical institutions, and the preprocessing includes data cleaning, standardization, missing value filling and outlier detection.

[0017] Optionally, the source data includes pediatric nursing data and historical nursing data from other medical institutions, and the source data characteristics include basic patient information, nursing process information, nursing quality indicators and nursing environment information.

[0018] Optionally, the S3 specifically includes:

[0019] S31. Constructing a nursing quality prediction model, wherein the nursing quality prediction model includes:

[0020] The data input layer receives source data features and performs data format conversion and normalization;

[0021] Feature extraction layer: Encode the input data and extract nursing features. Fully connected neural network is used to extract features for numerical data, category mapping matrix is ​​used to vectorize category data, and improved gated recurrent unit network is used to extract time dependencies for time series data.

[0022] Improve the attention fusion layer. Based on the adaptive attention mechanism, dynamically adjust the importance weights in combination with the extracted nursing features to increase the focus on key nursing factors, and introduce a gating mechanism to further optimize the screening of nursing features.

[0023] The prediction output layer uses a fully connected neural network to predict and output nursing quality indicators based on the final nursing features after fusion. The nursing quality indicators include patient recovery status, nursing success rate and nursing risk;

[0024] S32, processing source data features, formatting numerical data, categorical data, and time series data, normalizing numerical data, and converting categorical data using a one-hot encoding method;

[0025] S33. Use fully connected neural network to process numerical data:

[0026]

[0027]

[0028] in, represents the output of the first layer of fully connected neural network, represents the output of the second layer of fully connected neural network, σ represents the ReLU activation function, and denote the weight matrices of the first and second layers respectively, and Denote the bias terms of the first and second layers respectively, X num Represents the input matrix, dimension is m×d num , the dimension of the input matrix m represents the number of samples, d num Represents the numerical feature dimension;

[0029] S34, use the category mapping matrix to convert the categorical data into a low-dimensional vector representation. Suppose the category variable C has n different categories (c 1 ,c 2 ,...,c n ), define the category mapping matrix:

[0030]

[0031] Among them, M cat represents the category mapping matrix, dembed Represents the embedding dimension of the category feature, each category c i Corresponding to a vector n represents the number of categories, R represents the set of real numbers, and ∈ represents "belongs to";

[0032] Perform feature mapping:

[0033] H cat =M cat C;

[0034] Among them, H cat represents the category feature representation after mapping, and C represents the one-hot encoding representation of the category feature;

[0035] S35, using an improved gated recurrent unit network to process time series data, wherein the improved gated recurrent unit network introduces a time attenuation gate γ t , the hidden state update of the modified gated recurrent unit network:

[0036] γ t =exp(-λΔt);

[0037] Among them, γ t represents the time decay gate, exp represents the natural exponential function, λ represents the time decay rate, and Δt represents the time interval between the current nursing event and the previous nursing event;

[0038] When Δt is large, that is, the nursing event interval is long, γ t Become smaller, so that the impact of longer-term care events is reduced;

[0039] When Δt is small, that is, nursing events occur frequently, γ t becomes larger, and the weight of recent events becomes higher;

[0040] Perform hidden state update of improved gated recurrent unit network:

[0041] z t =σ(W z X seq +U z h t-1 +b z );

[0042] r t =σ(W r X seq +U r h t-1 +b r );

[0043] h t =(1-z t )h t-1 +zt γ t tanh(W h X seq +U h (r t ⊙h t-1 )+b h );

[0044] Among them, z t represents the update gate, r t represents the reset gate, h t represents the current hidden state, W z , W r , W h , U z , U r and U h represents the weight matrix, b z 、b r and b h represents the bias, h t-1 represents the hidden state of the previous time step, X seq Represents the input time series data matrix, with dimensions of m×T×d seq , the dimension of the input time series data matrix is ​​m, which represents the number of samples, T, which represents the time step, and d seq represents the number of nursing features contained in each time step, ⊙ represents element-by-element multiplication, which is used to control the influence of the hidden state of the previous time step on the current hidden state, t represents the time step, tanh represents the hyperbolic tangent activation function, and σ represents the ReLU activation function;

[0045] S36. Adopt an adaptive attention mechanism to calculate the weight of the nursing feature, and calculate the attention-weighted nursing feature vector:

[0046]

[0047] Among them, α i represents the importance weight of the i-th nursing feature, exp represents the natural exponential function, and W i , W j 、b i and b j represents an adaptive parameter, which is automatically adjusted through training and is used to give higher weights to important features that affect the quality of care. i and h j represents the nursing feature vector, k represents the total number of input features, and h θ represents the fused nursing feature vector;

[0048] S37. After attention calculation, a gating mechanism is introduced to optimize feature screening and calculate the final nursing features:

[0049] g=Sigmoid(W g h θ +b g );

[0050] h o =g·h θ ;

[0051] Among them, g represents the global feature gating value, Sigmoid represents the activation function, and W g and b g represents the training parameters, h θ represents the nursing characteristics after fusion, h o represents the final nursing characteristics;

[0052] S38. Train the nursing quality prediction model. Suppose the training data set is x in the training dataset i represents the input feature vector, y i Represents the nursing quality score and defines the loss function:

[0053]

[0054] Among them, L represents the loss function, f(x i ,θ) represents the prediction output, θ represents the training parameters of the nursing quality prediction model, and y i represents the nursing quality score, and N represents the number of data samples;

[0055] Gradient descent is used to optimize the training parameters of the nursing quality prediction model:

[0056]

[0057] Among them, θ (t+1) represents the training parameters of the nursing quality prediction model in the t+1th iteration, θ (t) represents the training parameters of the nursing quality prediction model in the tth iteration, η represents the learning rate, represents the gradient of the loss function with respect to the training parameters of the nursing quality prediction model;

[0058] S39. Storing the training parameters of the trained nursing quality prediction model.

[0059] Optionally, the S4 specifically includes:

[0060] S41. Perform feature matching on the target pediatric nursing data, calculate the feature similarity between the source data and the target pediatric nursing data, and use the Pearson correlation coefficient to select highly correlated features:

[0061]

[0062] Among them, r represents the feature correlation coefficient, x i Represents the source data characteristics, y i represents the target pediatric nursing data characteristics, and represent the means of the source data characteristics and the target pediatric nursing data characteristics, respectively;

[0063] Set a feature screening threshold and retain all target pediatric nursing data features whose feature correlation coefficients are greater than the feature screening threshold;

[0064] S42. The nursing quality prediction model obtained by training with the source data is used as the source model:

[0065] θ s = {W s ,b s};

[0066] Among them, θ s represents the parameters of the source model, W s and b s Represent the weight and bias parameters of the source model respectively;

[0067] S43. Input the target pediatric nursing data into the source model and calculate the initial prediction error, using the mean square error as the loss function:

[0068]

[0069] Among them, L in represents the mean square error loss function, f s (x i ,θ s ) represents the source model f s The predicted output, Y i represents the nursing quality score, and N represents the number of data samples;

[0070] If the initial prediction error L in If the preset threshold is exceeded, migration and fine-tuning are performed;

[0071] S44, adjusting the source model parameters by using an improved fine-tuning method, wherein the improved fine-tuning method comprises:

[0072] In the first stage, the feature extraction layer of the source model is frozen, and only the prediction output layer of the source model is fine-tuned:

[0073]

[0074] in, Indicates the prediction output layer parameters that need to be fine-tuned in the c+1 round of training, Indicates the feature layer parameters frozen in the c+1 round of training, which are not updated in the first stage. Indicates the prediction output layer parameters that need to be fine-tuned in the cth round of training, represents the feature layer parameters frozen in the cth round of training, η c represents the learning rate of the cth round of training, L in represents the mean square error loss function, represents the gradient of the mean square error loss function, η c represents the learning rate of the cth round of training, L in represents the mean square error loss function, represents the gradient of the mean square error loss function;

[0075] In the second stage, an adaptive unfreezing strategy is used to gradually unfreeze the hidden layers close to the output layer and gradually adjust the weights of the entire source model:

[0076]

[0077] in, represents the parameters of the first layer in the c+1th round of training, β l Represents the adjustment weight of layer 1, with the initial value set to 1. represents the parameters of layer 1 in round c of training, represents the new parameters of Tier 1 on the target pediatric care data;

[0078] In the third stage, after the second stage is stable, all source model layers are unfrozen and the whole model is fine-tuned;

[0079]

[0080] in, represents the fine-tuned parameters;

[0081] S45. During fine-tuning, adaptively adjust the learning rate:

[0082]

[0083] Among them, η c represents the learning rate of the cth round of training, η 0 represents the initial learning rate, ρ represents the learning rate attenuation coefficient, which is used to reduce the learning rate in the later stage of training;

[0084] S46, in the mean square error loss function L in Add L2 regularization and calculate the final loss:

[0085] L final =L in +L 2 ;

[0086] Among them, L in represents the mean square error loss function, fs (x i ,θ s ) represents the source model f s The predicted output, Y i represents the nursing quality score, and N represents the number of data samples;

[0087] If L final If L is less than the error threshold, the optimization is stopped. final If it is greater than the error threshold, continue to optimize the parameters;

[0088] S47. Store the nursing quality prediction model and nursing quality prediction model parameters after fine-tuning and optimization.

[0089] Optionally, the nursing quality indicators include patient recovery status, nursing success rate and nursing risks.

[0090] The beneficial effects of the present invention are:

[0091] First, in the construction of the nursing quality prediction model, the present invention adopts an improved deep learning network architecture, and realizes efficient analysis of nursing data through the combination of data input layer, feature extraction layer, improved attention fusion layer and prediction output layer. Among them, the feature extraction layer combines the fully connected neural network, the category mapping matrix and the improved gated recurrent unit network to effectively process numerical data, category data and time series data, so that the nursing model can understand the patient status and nursing process more comprehensively. In particular, the adaptive attention mechanism proposed in the present invention can dynamically allocate the weights of nursing features, highlight key factors, improve the model's attention to factors affecting nursing quality, and make the prediction results more accurate. Compared with the traditional nursing quality prediction model that only predicts based on fixed rules or static features, the present invention uses an adaptive attention mechanism to optimize feature selection, which improves the prediction ability of the model in complex nursing scenarios.

[0092] Secondly, compared with the existing nursing quality prediction methods, the present invention introduces an improved fine-tuning strategy in the training and optimization process, adopts a staged fine-tuning and parameter importance screening method, avoids the high computational cost and overfitting problems caused by full model training, and improves the adaptability of the model to the target data. The present invention also combines an adaptive learning rate adjustment strategy to optimize gradient updates, making the training process more stable, and improving the model convergence speed and prediction accuracy. In addition, the present invention adds L2 regularization in the fine-tuning process to suppress model overfitting, so that the model has better generalization ability when processing pediatric nursing data.

[0093] Finally, the present invention adopts feature similarity screening and feature distribution matching methods in the migration process to ensure that the data features of the target medical institution are highly consistent with the source data features, thereby improving the effectiveness of transfer learning. In traditional transfer learning methods, due to the heterogeneity of nursing data in different medical institutions, directly applying the model trained with source data to the target medical institution may lead to prediction bias and affect the accuracy of nursing quality assessment. The present invention combines feature matching distribution optimization strategy to optimize the migration process, so that the nursing quality prediction model can adapt to the target data environment more accurately and improve the generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0094] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0095] Figure 1 This is a flow chart of a pediatric nursing quality prediction and optimization method based on transfer learning proposed by the present invention;

[0096] Figure 2 This is a flow chart of the transfer learning method for the pediatric nursing quality prediction and optimization method based on transfer learning proposed in the present invention. DETAILED DESCRIPTION

[0097] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0098] refer to Figure 1 and Figure 2 , a pediatric nursing quality prediction and optimization method based on transfer learning, comprising the following steps:

[0099] S1. Collect target pediatric nursing data and preprocess the target pediatric nursing data;

[0100] S2, select source data and extract source data features;

[0101] S3, constructing a nursing quality prediction model, training the nursing quality prediction model based on source data features, enabling the nursing quality prediction model to learn nursing features, and storing training parameters of the nursing quality prediction model;

[0102] S4. Based on the preprocessed target pediatric nursing data and the training parameters of the nursing quality prediction model, the nursing quality prediction model is migrated and fine-tuned using an improved fine-tuning method, and the nursing quality prediction model and nursing quality prediction model parameters are saved;

[0103] S5. Input the target pediatric nursing data into the fine-tuned nursing quality prediction model, calculate the nursing quality indicators, and evaluate the prediction results;

[0104] S6. Generate nursing plans based on prediction results and optimize nursing plans according to implementation effects;

[0105] S7. Apply the optimized nursing plan to actual nursing work and establish an update mechanism to iteratively optimize the nursing quality prediction model in combination with new data.

[0106] In this embodiment, the target pediatric nursing data includes electronic medical records, nursing records, patient recovery status and nursing staff feedback from medical institutions, and the preprocessing includes data cleaning, standardization, missing value filling and outlier detection.

[0107] In this embodiment, the source data includes pediatric nursing data and historical nursing data from other medical institutions, and the source data features include basic patient information, nursing process information, nursing quality indicators and nursing environment information.

[0108] In this implementation, S3 specifically includes:

[0109] S31. Constructing a nursing quality prediction model, wherein the nursing quality prediction model includes:

[0110] The data input layer receives source data features and performs data format conversion and normalization;

[0111] Feature extraction layer: Encode the input data and extract nursing features. Fully connected neural network is used to extract features for numerical data, category mapping matrix is ​​used to vectorize category data, and improved gated recurrent unit network is used to extract time dependencies for time series data.

[0112] Improve the attention fusion layer. Based on the adaptive attention mechanism, dynamically adjust the importance weights in combination with the extracted nursing features to increase the focus on key nursing factors, and introduce a gating mechanism to further optimize the screening of nursing features.

[0113] The prediction output layer uses a fully connected neural network to predict and output nursing quality indicators based on the final nursing features after fusion. The nursing quality indicators include patient recovery status, nursing success rate and nursing risk;

[0114] S32, processing source data features, formatting numerical data, categorical data, and time series data, normalizing numerical data, and converting categorical data using a one-hot encoding method;

[0115] S33. Use fully connected neural network to process numerical data:

[0116]

[0117] in, represents the output of the first layer of fully connected neural network, represents the output of the second layer of fully connected neural network, σ represents the ReLU activation function, and denote the weight matrices of the first and second layers respectively, and Denote the bias terms of the first and second layers respectively, X num Represents the input matrix, dimension is m×d num , the dimension of the input matrix m represents the number of samples, d num Represents the numerical feature dimension;

[0118] S34, use the category mapping matrix to convert the categorical data into a low-dimensional vector representation. Suppose the category variable C has n different categories (c 1 ,c 2 ,...,c n ), define the category mapping matrix:

[0119]

[0120] Among them, M cat represents the category mapping matrix, d embed Represents the embedding dimension of the category feature, each category c i Corresponding to a vector n represents the number of categories, R represents the set of real numbers, and ∈ represents "belongs to";

[0121] Perform feature mapping:

[0122] H cat =M cat C;

[0123] Among them, H cat represents the category feature representation after mapping, and C represents the one-hot encoding representation of the category feature;

[0124] S35, using an improved gated recurrent unit network to process time series data, wherein the improved gated recurrent unit network introduces a time attenuation gate γ t , the hidden state update of the modified gated recurrent unit network:

[0125] γ t =exp(-λΔt);

[0126] Among them, γ trepresents the time decay gate, exp represents the natural exponential function, λ represents the time decay rate, and Δt represents the time interval between the current nursing event and the previous nursing event;

[0127] When Δt is large, that is, the nursing event interval is long, γ t Become smaller, so that the impact of longer-term care events is reduced;

[0128] When Δt is small, that is, nursing events occur frequently, γ t becomes larger, and the weight of recent events becomes higher;

[0129] Perform hidden state update of improved gated recurrent unit network:

[0130] z t =σ(W z X seq +U z h t-1 +b z );

[0131] r t =σ(W r X seq +U r h t-1 +b r );

[0132] h t =(1-z t )h t-1 +z t γ t tanh(W h X seq +U h (r t ⊙h t-1 )+b h );

[0133] Among them, z t represents the update gate, r t represents the reset gate, h t represents the current hidden state, W z , W r , W h , U z , U r and U h represents the weight matrix, b z 、b r and b h represents the bias, h t-1 represents the hidden state of the previous time step, X seq Represents the input time series data matrix, with dimensions of m×T×d seq, the dimension of the input time series data matrix is ​​m, which represents the number of samples, T, which represents the time step, and d seq represents the number of nursing features contained in each time step, ⊙ represents element-by-element multiplication, which is used to control the influence of the hidden state of the previous time step on the current hidden state, t represents the time step, tanh represents the hyperbolic tangent activation function, and σ represents the ReLU activation function;

[0134] S36. Adopt an adaptive attention mechanism to calculate the weight of the nursing feature, and calculate the attention-weighted nursing feature vector:

[0135]

[0136] Among them, α i represents the importance weight of the i-th nursing feature, exp represents the natural exponential function, and W i , W j 、b i and b j represents an adaptive parameter, which is automatically adjusted through training and is used to give higher weights to important features that affect the quality of care. i and h j represents the nursing feature vector, k represents the total number of input features, and h θ represents the fused nursing feature vector;

[0137] S37. After attention calculation, a gating mechanism is introduced to optimize feature screening and calculate the final nursing features:

[0138] g=Sigmoid(W g h θ +b g );

[0139] h o =g·h θ ;

[0140] Among them, g represents the global feature gating value, Sigmoid represents the activation function, and W g and b g represents the training parameters, h θ represents the nursing characteristics after fusion, h o represents the final nursing characteristics;

[0141] S38. Train the nursing quality prediction model. Suppose the training data set is x in the training dataset i represents the input feature vector, y i Represents the nursing quality score and defines the loss function:

[0142]

[0143] Among them, L represents the loss function, f(x i ,θ) represents the prediction output, θ represents the training parameters of the nursing quality prediction model, and y i represents the nursing quality score, and N represents the number of data samples;

[0144] Gradient descent is used to optimize the training parameters of the nursing quality prediction model:

[0145]

[0146] Among them, θ (t+1) represents the training parameters of the nursing quality prediction model in the t+1th iteration, θ (t) represents the training parameters of the nursing quality prediction model in the tth iteration, η represents the learning rate, represents the gradient of the loss function with respect to the training parameters of the nursing quality prediction model;

[0147] S39. Storing the training parameters of the trained nursing quality prediction model.

[0148] In this implementation manner, the S4 specifically includes:

[0149] S41. Perform feature matching on the target pediatric nursing data, calculate the feature similarity between the source data and the target pediatric nursing data, and use the Pearson correlation coefficient to select highly correlated features:

[0150]

[0151] Among them, r represents the feature correlation coefficient, x i Represents the source data characteristics, y i represents the target pediatric nursing data characteristics, and represent the means of the source data characteristics and the target pediatric nursing data characteristics, respectively;

[0152] Set a feature screening threshold and retain all target pediatric nursing data features whose feature correlation coefficients are greater than the feature screening threshold;

[0153] S42. The nursing quality prediction model obtained by training with the source data is used as the source model:

[0154] θ s = {W s ,b s};

[0155] Among them, θ s represents the parameters of the source model, W s and b s Represent the weight and bias parameters of the source model respectively;

[0156] S43. Input the target pediatric nursing data into the source model and calculate the initial prediction error, using the mean square error as the loss function:

[0157]

[0158] Among them, L in represents the mean square error loss function, f s (x i ,θ s ) represents the source model f s The predicted output, Y i represents the nursing quality score, and N represents the number of data samples;

[0159] If the initial prediction error L in If the preset threshold is exceeded, migration and fine-tuning are performed;

[0160] S44, adjusting the source model parameters by using an improved fine-tuning method, wherein the improved fine-tuning method comprises:

[0161] In the first stage, the feature extraction layer of the source model is frozen, and only the prediction output layer of the source model is fine-tuned:

[0162]

[0163] in, Indicates the prediction output layer parameters that need to be fine-tuned in the c+1 round of training, Indicates the feature layer parameters frozen in the c+1 round of training, which are not updated in the first stage. Indicates the prediction output layer parameters that need to be fine-tuned in the cth round of training, represents the feature layer parameters frozen in the cth round of training, η c represents the learning rate of the cth round of training, L in represents the mean square error loss function, represents the gradient of the mean square error loss function, η c represents the learning rate of the cth round of training, L in represents the mean square error loss function, represents the gradient of the mean square error loss function;

[0164] In the second stage, an adaptive unfreezing strategy is used to gradually unfreeze the hidden layers close to the output layer and gradually adjust the weights of the entire source model:

[0165]

[0166] in, represents the parameters of the first layer in the c+1th round of training, β l Represents the adjustment weight of layer 1, with the initial value set to 1. represents the parameters of layer 1 in round c of training, represents the new parameters of Tier 1 on the target pediatric care data;

[0167] In the third stage, after the second stage is stable, all source model layers are unfrozen and the whole model is fine-tuned;

[0168]

[0169] in, represents the fine-tuned parameters;

[0170] S45. During fine-tuning, adaptively adjust the learning rate:

[0171]

[0172] Among them, η c represents the learning rate of the cth round of training, η 0 represents the initial learning rate, ρ represents the learning rate attenuation coefficient, which is used to reduce the learning rate in the later stage of training;

[0173] S46, in the mean square error loss function L in Add L2 regularization and calculate the final loss:

[0174] L final =L in +L 2 ;

[0175] Among them, L in represents the mean square error loss function, f s (x i ,θ s ) represents the source model f s The predicted output, Y i represents the nursing quality score, and N represents the number of data samples;

[0176] If L final If L is less than the error threshold, the optimization is stopped. final If it is greater than the error threshold, continue to optimize the parameters;

[0177] S47. Store the nursing quality prediction model and nursing quality prediction model parameters after fine-tuning and optimization.

[0178] In this embodiment, the nursing quality indicators include patient recovery status, nursing success rate and nursing risks.

[0179] Embodiment 1:

[0180] In order to verify the feasibility of the present invention in implementation, the present invention was applied to the pediatric ward of a tertiary-level Class A children's hospital, and a six-month experimental study was conducted. The hospital faces many challenges in the process of pediatric care, such as the heavy workload of nursing staff, the quality of care is greatly affected by individual differences and fluctuations in the nursing process, and the traditional nursing quality assessment method has low efficiency in data utilization, making it difficult to achieve accurate prediction and intelligent optimization. The main goal of this experiment is to verify the practical application value of the present invention in terms of nursing quality prediction accuracy, nursing plan optimization effect, nursing resource utilization efficiency, etc., and to compare the differences between the method of the present invention and the traditional nursing quality management method to evaluate its applicability in clinical care.

[0181] The method of the present invention was compared in a pediatric ward of a hospital, involving a total of 300 hospitalized children, of which 150 patients were treated with traditional nursing methods and 150 patients were treated with the optimized intelligent nursing method of the present invention. The basic conditions of the two groups of patients were similar, covering different age groups, diseases and nursing needs. The experiment lasted for six months, with the first two months used for model training and fine-tuning, and the last three months used for actual testing and effect evaluation.

[0182] In the model training stage, the method of the present invention first collects the pediatric nursing data of the hospital, including basic patient information, nursing process data, nursing quality indicators, nursing environment information, etc., and introduces pediatric nursing data from other medical institutions, and uses transfer learning methods to perform feature matching and distribution optimization on the data to ensure the applicability of cross-institutional data. Then, the nursing time series data is modeled using an improved gated recurrent unit network, and an adaptive attention mechanism is introduced to optimize the feature extraction process and improve the accuracy of nursing quality prediction. In the model fine-tuning stage, the nursing quality prediction model is optimized by calculating the mean square error to make it more accurately adapt to the pediatric nursing environment of the hospital.

[0183] During the testing phase, hospital nursing staff formulate personalized nursing plans based on the nursing quality prediction results of the present invention, and evaluate the nursing quality during the actual nursing process. The test mainly examines the following core indicators:

[0184] The accuracy of nursing quality prediction, that is, the ability of the method of the present invention to predict the nursing success rate, nursing risk, and patient recovery status;

[0185] The effect of nursing plan optimization, that is, the shortening of patients’ recovery time and the improvement of nursing success rate;

[0186] The utilization of nursing resources, namely, the reduction of nursing staff’s workload and the optimization of nursing resources allocation.

[0187] Table 1 Experimental data comparison table

[0188]

[0189] In terms of nursing quality prediction, the method of the present invention significantly improves the prediction accuracy of nursing success rate, nursing risk and recovery time. Among them, the prediction accuracy of nursing success rate increased from 73.2% to 89.6%, the prediction accuracy of nursing risk increased from 68.7% to 85.2%, and the prediction accuracy of recovery time increased from 70.4% to 88.1%, with an overall improvement of 22.4% to 25.2%. This shows that with the help of the method of the present invention, hospital nursing staff can more accurately predict nursing effectiveness and take intervention measures in advance, thereby improving the quality of patient care.

[0190] In terms of nursing plan optimization, the personalized nursing optimization mechanism of the present invention significantly shortens the patient's recovery time and improves the nursing success rate. Experimental data show that the average recovery time of patients is shortened from 6.8 days of traditional methods to 5.2 days, a reduction of 23.5%, indicating that the intelligent nursing optimization method of the present invention can effectively shorten the hospitalization time and improve the patient's recovery efficiency. At the same time, the nursing success rate increased from 82.5% to 91.3%, and the nursing risk was reduced by 21.8%, indicating that the method of the present invention can dynamically adjust the nursing plan according to the prediction results, reduce the possibility of nursing failure, and provide patients with safer and more effective nursing plans.

[0191] In terms of nursing resource utilization, the method of the present invention reduces the workload of nursing staff and improves the utilization rate of nursing resources. Experimental data show that the average daily nursing time of each nursing staff is reduced from 7.5 hours to 6.2 hours, and the workload is reduced by 17.3%, which effectively improves the work efficiency of nursing staff. At the same time, the utilization rate of nursing resources is increased by 14.6%, indicating that the nursing optimization method of the present invention can allocate nursing resources more reasonably, ensure a more accurate match between nursing staff and patients, and improve overall nursing efficiency.

[0192] In summary, the intelligent nursing quality prediction and optimization method of the present invention can not only provide more accurate nursing quality prediction, but also dynamically adjust the nursing plan according to the prediction results and optimize the allocation of nursing resources. Compared with traditional nursing methods, the method of the present invention improves the nursing success rate, reduces nursing risks, shortens the patient's recovery time, and improves the workload of nursing staff.

[0193] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A pediatric nursing quality prediction and optimization method based on transfer learning, characterized in that: The steps include: S1. Collect target pediatric nursing data and preprocess the target pediatric nursing data; S2, select source data and extract source data features; S3, constructing a nursing quality prediction model, training the nursing quality prediction model based on source data features, enabling the nursing quality prediction model to learn nursing features, and storing training parameters of the nursing quality prediction model; S4. Based on the preprocessed target pediatric nursing data and the training parameters of the nursing quality prediction model, the nursing quality prediction model is migrated and fine-tuned using an improved fine-tuning method, and the nursing quality prediction model and nursing quality prediction model parameters are saved; S5. Input the target pediatric nursing data into the fine-tuned nursing quality prediction model, calculate the nursing quality indicators, and evaluate the prediction results; S6. Generate nursing plans based on prediction results and optimize nursing plans according to implementation effects; S7. Apply the optimized nursing plan to actual nursing work and establish an update mechanism to iteratively optimize the nursing quality prediction model in combination with new data.

2. A pediatric nursing quality prediction and optimization method based on transfer learning according to claim 1, characterized in that: The target pediatric nursing data includes electronic medical records, nursing records, patient recovery status and nursing staff feedback from medical institutions, and the preprocessing includes data cleaning, standardization, missing value filling and outlier detection.

3. A pediatric nursing quality prediction and optimization method based on transfer learning according to claim 1, characterized in that: The source data includes pediatric nursing data and historical nursing data from other medical institutions, and the source data characteristics include basic patient information, nursing process information, nursing quality indicators and nursing environment information.

4. The method for predicting and optimizing pediatric nursing quality based on transfer learning according to claim 1, characterized in that: The S3 specifically includes: S31. Constructing a nursing quality prediction model, wherein the nursing quality prediction model includes: The data input layer receives source data features and performs data format conversion and normalization; Feature extraction layer: Encode the input data and extract nursing features. Fully connected neural network is used to extract features for numerical data, category mapping matrix is ​​used to vectorize category data, and improved gated recurrent unit network is used to extract time dependencies for time series data. Improve the attention fusion layer. Based on the adaptive attention mechanism, dynamically adjust the importance weights in combination with the extracted nursing features to increase the focus on key nursing factors, and introduce a gating mechanism to further optimize the screening of nursing features. The prediction output layer uses a fully connected neural network to predict and output nursing quality indicators based on the final nursing features after fusion. The nursing quality indicators include patient recovery status, nursing success rate and nursing risk; S32, processing source data features, formatting numerical data, categorical data, and time series data, normalizing numerical data, and converting categorical data using a one-hot encoding method; S33. Use fully connected neural network to process numerical data: in, represents the output of the first layer of fully connected neural network, represents the output of the second layer of fully connected neural network, σ represents the ReLU activation function, and denote the weight matrices of the first and second layers respectively, and Denote the bias terms of the first and second layers respectively, X num Represents the input matrix, dimension is m×d num , the dimension of the input matrix m represents the number of samples, d num Represents the numerical feature dimension; S34, use the category mapping matrix to convert the categorical data into a low-dimensional vector representation. Suppose the category variable C has n different categories (c1, c2, ..., c n ), define the category mapping matrix: Among them, M cat represents the category mapping matrix, d embed Represents the embedding dimension of the category feature, each category c i Corresponding to a vector n represents the number of categories, R represents the set of real numbers, and ∈ represents "belongs to"; Perform feature mapping: H cat =M cat C; Among them, H cat represents the category feature representation after mapping, and C represents the unique hot encoding representation of the category feature; S35, using an improved gated recurrent unit network to process time series data, wherein the improved gated recurrent unit network introduces a time attenuation gate γ t , the hidden state update of the modified gated recurrent unit network: c t =exp(-λΔt); Among them, γ t represents the time decay gate, exp represents the natural exponential function, λ represents the time decay rate, and Δt represents the time interval between the current nursing event and the previous nursing event; When Δt is large, that is, the nursing event interval is long, γ t Become smaller, so that the impact of longer-term care events is reduced; When Δt is small, that is, nursing events occur frequently, γ t becomes larger, and the weight of recent events becomes higher; Perform hidden state update of improved gated recurrent unit network: z t =σ(W z X seq +U z h t-1 +b z ); r t =σ(W r X seq +U r h t-1 +b r ); h t =(1-z t )h t-1 +z t γ t tanh(W h X seq +U h (r t ⊙h t-1 )+b h ); Among them, z t represents the update gate, r t represents the reset gate, h t represents the current hidden state, W z , W r , W h , U z , U r and U h represents the weight matrix, b z , b r and b h represents the bias, h t-1 represents the hidden state of the previous time step, X seq Represents the input time series data matrix, with dimensions of m×T×d seq , the dimension of the input time series data matrix is ​​m, which represents the number of samples, T, which represents the time step, and d seq represents the number of nursing features contained in each time step, ⊙ represents element-by-element multiplication, which is used to control the influence of the hidden state of the previous time step on the current hidden state, t represents the time step, tanh represents the hyperbolic tangent activation function, and σ represents the ReLU activation function; S36. Adopt an adaptive attention mechanism to calculate the weight of the nursing feature, and calculate the attention-weighted nursing feature vector: Among them, α i represents the importance weight of the i-th nursing feature, exp represents the natural exponential function, and W i , W j , b i and b j represents an adaptive parameter, which is automatically adjusted through training and is used to give higher weights to important features that affect the quality of care. i and h j represents the nursing feature vector, k represents the total number of input features, and h θ represents the fused nursing feature vector; S37. After attention calculation, a gating mechanism is introduced to optimize feature screening and calculate the final nursing features: g=Sigmoid(W g h θ +b g ); h o =g·h θ ; Among them, g represents the global feature gating value, Sigmoid represents the activation function, and W g and b g represents the training parameters, h θ represents the nursing characteristics after fusion, h o represents the final nursing characteristics; S38. Train the nursing quality prediction model. Suppose the training data set is x in the training dataset i represents the input feature vector, y i Represents the nursing quality score and defines the loss function: Among them, L represents the loss function, f(x i ,θ) represents the prediction output, θ represents the training parameters of the nursing quality prediction model, and y i represents the nursing quality score, and N represents the number of data samples; Gradient descent is used to optimize the training parameters of the nursing quality prediction model: Among them, θ (t+1) represents the training parameters of the nursing quality prediction model in the t+1th iteration, θ (t) represents the training parameters of the nursing quality prediction model in the tth iteration, η represents the learning rate, represents the gradient of the loss function with respect to the training parameters of the nursing quality prediction model; S39. Storing the training parameters of the trained nursing quality prediction model.

5. The method for predicting and optimizing pediatric nursing quality based on transfer learning according to claim 1, characterized in that: The S4 specifically includes: S41. Perform feature matching on the target pediatric nursing data, calculate the feature similarity between the source data and the target pediatric nursing data, and use the Pearson correlation coefficient to select highly correlated features: Among them, r represents the feature correlation coefficient, x i Represents the source data characteristics, y i represents the target pediatric nursing data characteristics, and represent the means of the source data characteristics and the target pediatric nursing data characteristics, respectively; Set a feature screening threshold and retain all target pediatric nursing data features whose feature correlation coefficients are greater than the feature screening threshold; S42. The nursing quality prediction model obtained by training with the source data is used as the source model: i s ={W s ,b s }; Among them, θ s represents the parameters of the source model, W s and b s Represent the weight and bias parameters of the source model respectively; S43. Input the target pediatric nursing data into the source model and calculate the initial prediction error, using the mean square error as the loss function: Among them, L in represents the mean square error loss function, f s (x i ,θ s ) represents the source model f s The predicted output, Y i represents the nursing quality score, and N represents the number of data samples; If the initial prediction error L in If the preset threshold is exceeded, migration and fine-tuning are performed; S44, adjusting the source model parameters using an improved fine-tuning method, wherein the improved fine-tuning method comprises: In the first stage, the feature extraction layer of the source model is frozen, and only the prediction output layer of the source model is fine-tuned: in, Indicates the prediction output layer parameters that need to be fine-tuned in the c+1 round of training, Indicates the feature layer parameters frozen in the c+1 round of training, which are not updated in the first stage. Indicates the prediction output layer parameters that need to be fine-tuned in the cth round of training, represents the feature layer parameters frozen in the cth round of training, η c represents the learning rate of the cth round of training, L in represents the mean square error loss function, represents the gradient of the mean square error loss function, η c represents the learning rate of the cth round of training, L in represents the mean square error loss function, represents the gradient of the mean square error loss function; In the second stage, an adaptive unfreezing strategy is used to gradually unfreeze the hidden layers close to the output layer and gradually adjust the weights of the entire source model: in, represents the parameters of the first layer in the c+1th round of training, β l Represents the adjustment weight of layer 1, with the initial value set to 1. represents the parameters of layer 1 in round c of training, represents the new parameters of Tier 1 on the target pediatric care data; In the third stage, after the second stage is stable, all source model layers are unfrozen and the whole model is fine-tuned; in, represents the fine-tuned parameters; S45. During fine-tuning, adaptively adjust the learning rate: Among them, η c represents the learning rate of the cth round of training, η0 represents the initial learning rate, and ρ represents the learning rate attenuation coefficient, which is used to reduce the learning rate in the later stage of training; S46, in the mean square error loss function L in Add L2 regularization and calculate the final loss: L final =L in +L2; Among them, L in represents the mean square error loss function, f s (x i ,θ s ) represents the source model f s The predicted output, Y i represents the nursing quality score, and N represents the number of data samples; If L final If L is less than the error threshold, the optimization is stopped. final If it is greater than the error threshold, continue to optimize the parameters; S47. Store the nursing quality prediction model and nursing quality prediction model parameters after fine-tuning and optimization.

6. The method for predicting and optimizing pediatric nursing quality based on transfer learning according to claim 1, characterized in that: The nursing quality indicators include patient recovery status, nursing success rate and nursing risks.

Images