Clinical pathway construction method based on deep learning and data enhancement

Abstract

The present application aims at the clinical path optimization problem existing in the medical system, and provides a clinical path construction method based on deep learning and data enhancement, a clinical path optimization framework constructed by using deep learning and data enhancement method, patient diagnosis and treatment log data is preprocessed by using multi-label feature coding, key diagnosis and treatment theme mode in diagnosis and treatment day is extracted from the hospitalization log by using latent Dirichlet allocation topic model, and the diversity and complexity of medical text data are increased by using data enhancement through high-consistency theme composed of diagnosis and treatment mode theme combination set, the generalization ability and adaptability in processing personalized and dynamic clinical path are improved; cooperate with bidirectional long short time memory network to capture the time sequence characteristics of diagnosis and treatment day, memory and analyze the time sequence characteristics of diagnosis and treatment day, use the full connection layer to process the output of bidirectional long short time memory network layer, ensure the comprehensive understanding of each time step data, and output the diagnosis and treatment project of each diagnosis and treatment day after the first day of hospitalization.

Description

Technical Field

[0001] This invention relates to deep learning technology, and in particular to a deep learning-based clinical pathway construction technology in medical systems. Background Technology

[0002] In the current healthcare system, the management and optimization of clinical pathways rely primarily on the experience and intuition of medical experts. This traditional approach faces significant limitations when dealing with personalized treatment plans and large-scale patient data. In particular, existing technologies have many shortcomings in data analysis, pathway adaptability, and real-time response. A clinical pathway refers to a standardized treatment model and process established for a specific disease. It is a comprehensive model of clinical diagnosis and treatment, guided by evidence-based medicine and guidelines, to promote the organization of medical care and disease management, ultimately playing a role in standardizing medical behavior, reducing variability, lowering costs, and improving quality. Clinical pathways typically consist of a series of diagnostic and treatment items, including specific examinations, surgical procedures, and medication administration. However, while rule-based systems can manage clinical pathways to some extent, they struggle to adapt to rapidly changing healthcare environments and personalized needs.

[0003] Graph mining methods are used to analyze the processes and patterns of clinical pathways, but they often lack the ability to handle complex and variable data. Furthermore, although machine learning algorithms (such as decision trees and random forests) have been applied to data analysis, their capabilities in processing large-scale time series data and providing high-accuracy predictions are limited. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a clinical pathway construction method that performs better in terms of data processing complexity, pathway adaptability, and real-time dynamic optimization, addressing the challenges of clinical pathway optimization in medical systems.

[0005] The technical solution adopted by this invention to solve the above-mentioned technical problems is a clinical pathway construction method based on deep learning and data augmentation, comprising the following steps:

[0006] Data collection and preprocessing steps: Collect medical orders from medical databases in chronological order for specific individuals; after data filtering and structural optimization of the medical orders, obtain medical log data.

[0007] The steps for mining treatment topics are as follows: The latent Dirichlet assignment topic model receives the input treatment log data. For each treatment day, it first generates a treatment day-topic distribution and a topic-treatment item distribution based on the Dirichlet distribution prior. Based on the treatment day-topic distribution, a topic set is obtained. Then, based on the topic-treatment item distribution, the set of treatment items corresponding to each treatment day's topic is obtained and output. The treatment day-topic distribution represents the set of topics used on the corresponding treatment day; the topic-treatment item distribution represents the set of treatment items used under the corresponding topic.

[0008] Data augmentation steps: Using the generated diagnosis and treatment topics generated by the Latent Dirichlet Assignment Topic Model, the consistency score for each diagnosis and treatment topic is output to complete the selection and combination of multiple diagnosis and treatment topics, and output the set of diagnosis and treatment item combinations corresponding to the combination of diagnosis and treatment topics;

[0009] Medical pathway prediction steps: First, a multi-label binary classification method is used to construct multi-label feature codes for each treatment day in the received treatment topic combination, and these codes are input into a bidirectional long short-term memory network. The bidirectional long short-term memory network outputs the temporal features of each treatment day to a fully connected layer. The fully connected layer predicts the set of treatment items for each treatment day, thereby obtaining the predicted clinical pathway for the treatment day.

[0010] This invention utilizes a clinical pathway optimization framework constructed using deep learning and data augmentation methods. Patient treatment log data is preprocessed using multi-label feature encoding. A Latent Dirichlet Allocation (LDA) topic model is combined with a Bidirectional Long Short-Term Memory (LSTM) network. The LDA topic model extracts key treatment topic patterns from the hospitalization logs for each treatment day. Data augmentation is then used to increase the diversity and complexity of the medical text data by creating a set of treatment pattern topic combinations with high consistency, improving generalization and adaptability in handling personalized and dynamic clinical pathways. The Bidirectional LSTM network captures the time-series features of each treatment day, memorizing and analyzing these features. Fully connected layers process the output of the Bidirectional LSTM network layers, ensuring a comprehensive understanding of the data at each time step and outputting the treatment items for each treatment day after the first day of hospitalization.

[0011] The beneficial effects of this invention are that it demonstrates outstanding effectiveness in mining treatment patterns, effectively extracting key features from large amounts of clinical data. The deep learning model, composed of a latent Dirichlet assignment topic model and a bidirectional long short-term memory network model architecture, exhibits excellent treatment pattern memorization capabilities during the training phase, particularly showing significant improvements in key performance indicators such as accuracy, precision, recall, and F1 score. This demonstrates its superior learning and generalization abilities, providing not only more accurate predictions but also rapid adaptation to new situations. It maintains high accuracy when handling unknown data during the validation phase, exhibiting good adaptability. It maintains accurate and stable predictive capabilities on diverse treatment process data, effectively improving the efficiency and quality of medical services while reducing costs. Attached Figure Description

[0012] Figure 1 Build flowcharts for data- and deep learning-driven clinical pathways;

[0013] Figure 2 A schematic diagram of clinical pathway construction. Detailed Implementation

[0014] This invention constructs clinical pathways on a daily basis, defining the specific treatment items to be implemented each day. To achieve this method, a novel model for clinical pathway construction is proposed. This model integrates a latent Dirichlet assignment topic model and a bidirectional long short-term memory network.

[0015] This example describes a method for constructing a clinical pathway for patients with acute ST-segment elevation myocardial infarction (STEMI). The specific implementation steps are as follows: Figure 1 As stated above.

[0016] S1. Data collection and preprocessing steps:

[0017] We first obtained electronic medical records of patients with acute ST-segment elevation myocardial infarction from a medical database. From these records, we extracted medical text data detailing the patient's treatment process. Then, we extracted key data related to the patient's treatment, primarily including medical orders. During preprocessing, we screened the data, removing information irrelevant to the treatment. To facilitate time series analysis, we focused on converting the execution time of medical orders into specific days relative to the start of the patient's hospitalization, ensuring the accuracy and continuity of the data over time. Furthermore, to improve data processing efficiency, we optimized the structure of the medical order data, including data merging and deduplication, to simplify the dataset structure and improve the efficiency of subsequent analysis. After data screening and structural optimization, we obtained the treatment log data.

[0018] S2. Steps for identifying treatment models (treatment topics):

[0019] Extracting high-quality and representative features from clinical log data is crucial for model building, especially when the amount of original patient data is limited and the daily clinical data has high dimensionality. To address this challenge, this invention employs a Latent Dirichlet Assignment Topic Model to first extract treatment patterns, using these patterns as prior knowledge features of clinical pathways. Treatment patterns are essential for ensuring the accuracy and robustness of subsequent deep learning models.

[0020] The Latent Dirichlet Topic Allocation Model is based on the assumption that a set of implicit topics exists within multiple document datasets, each topic consisting of a set of related words. The model aims to infer these implicit topics and their distribution within the documents. In this invention, the words in the Latent Dirichlet Topic Allocation Model represent the diagnostic and treatment items that constitute a medical pathway, and the documents represent diagnostic and treatment log data. That is, the Latent Dirichlet Topic Allocation Model in this embodiment needs to infer a set of implicit topics as diagnostic and treatment pattern topics (diagnostic and treatment topics) from the diagnostic and treatment log data. Then, a set of related diagnostic and treatment items can be obtained from the diagnostic and treatment topics, thereby revealing the potential medical pathway. The implicit diagnostic and treatment topics are not artificially defined diagnostic and treatment pattern topics with traditional medical significance, but rather implicit diagnostic and treatment pattern topics obtained by the model through the potential correlations between data, which are beneficial for subsequent data processing.

[0021] After the clinical log data is input into the latent Dirichlet assignment topic model, for each clinical day, a clinical day-topic distribution and a topic-treatment item distribution are generated based on the Dirichlet distribution prior. The clinical day-topic distribution represents the set of topics used on the corresponding clinical day; the topic-treatment item distribution represents the set of treatment items used under the corresponding topic. After obtaining the topic set based on the clinical day-topic distribution, the treatment item set for each topic is obtained based on the topic-treatment item distribution. Summarizing the treatment item sets for all topics yields the set of treatment items to be used on that clinical day.

[0022] The implementation example employs a variational Bayesian inference-based approach to realize a latent Dirichlet assignment topic model. Variational parameters δ and γ are introduced to approximate the true posterior distribution, and the model is optimized by adjusting the Dirichlet hyperparameters to discover latent topics in patient logs. During the variational expectation maximization process, the focus is on maximizing the lower bound of evidence to ensure the model can accurately extract the optimal patient log-topic distribution and topic-treatment item distribution from the data.

[0023] The process of finding the optimal solution employs a variational expectation maximization optimization process based on the latent Dirichlet assignment topic model. The specific implementation maximizes the lower bound of evidence by adjusting the Dirichlet hyperparameters. The detailed steps are as follows:

[0024] a) Variational expectation step: Optimize the variational parameter δ to obtain the topic-treatment item distribution, where each component δ in the variational parameter δ is δ d,j,k This represents the probability that treatment item j is assigned to topic k on treatment day d;

[0025] b) Variational maximization step: Optimize the variational parameter γ to obtain the treatment day-topic distribution, where each component γ in the variational parameter γ d,k This represents the variational parameter associated with topic k in the treatment day d.

[0026] Through iterative parameter updates in these steps, the lower bound of evidence is maximized based on the approximation between the topic-treatment item distribution and the treatment day-topic distribution. This determines whether the latent Dirichlet allocation topic model has converged, allowing the invention to gradually approximate a stable variational distribution, roughly approximating the target posterior distribution. This enables effective analysis of treatment log data for patients with acute ST-segment elevation myocardial infarction and accurate extraction of implicit treatment topics. The identification of these implicit topics provides rich prior knowledge features for subsequent bidirectional long short-term memory network model analysis, enabling the invention to more accurately predict and optimize clinical pathways. In the topic generation training for each treatment day, the latent Dirichlet allocation topic model applies the same dynamically adjusted range of topic quantity (2-30 in this example) to determine the optimal number of topics for each treatment day. Key treatment topic patterns are dynamically extracted from the treatment day, and the dynamic changes in the data are analyzed from the extracted static text data features. The determination of the optimal topics utilizes the consistency and perplexity scores of the latent Dirichlet allocation topic model for each topic. Consistency assesses the stability of topic content, while perplexity is used to evaluate the model's accuracy in predicting prior probabilities of words on the test set in the original data, reflecting the model's generalization performance. During evaluation, optimal consistency is prioritized. Within a small range of variation with the highest consistency score (0.005 in this example), the number of topics with the lowest perplexity score is selected to determine the topic model, thus obtaining the optimal topics.

[0027] For example, a topic model with 2 topics yields the consistency and perplexity of each of the two topics. The average of the consistency and perplexity is then calculated as the consistency and perplexity score of the topic model under this parameter.

[0028] S3. Theme Combination Selection (Data Augmentation) Steps:

[0029] The data augmentation strategy in this invention relies on multiple treatment topics output by the Latent Dirichlet Allocation Topic Model. When the Latent Dirichlet Allocation Topic Model outputs a treatment topic, it also outputs a consistency score for that topic. The core of the data augmentation strategy is to select and combine multiple treatment topics to re-represent the textual data of patient treatment items for each treatment day, transforming it into a series of topic combinations, i.e., a set of treatment item combinations.

[0030] Consistency scores measure the semantic relevance of words within a topic, reflecting the topic's quality and stability. In the medical field, topic interpretability is a crucial factor; topics with high consistency scores are generally easier for medical professionals to understand and apply. Topics with high consistency scores exhibit strong correlations between diagnostic and treatment items, meaning that in practical medical settings, this topic better represents specific diagnostic and treatment processes and medical behaviors. In this invention, we assume that a high-quality topic provides a stable, specific, and personalized description of diagnostic and treatment processes. Furthermore, using consistency scores as a selection criterion helps avoid introducing noise from low-quality or irrelevant topics, which is crucial for improving the clinical pathway performance of the model.

[0031] For each treatment day, the consistency score is selected as the criterion for choosing high-quality topics. Using the top n topics with the highest consistency scores, and considering the n preferred topics for each treatment day, all possible combinations of treatment topics are constructed. In this example, n is set to 2. This step is equivalent to resampling data points in a high-dimensional feature space to generate a sequence of treatment items representing each treatment day, i.e., a new set of feature vectors for each treatment day.

[0032] The use of highly consistent thematic sets of diagnosis and treatment processes significantly increases the diversity and complexity of medical text data, including more detailed and multi-faceted personalized descriptions of the diagnosis and treatment process. This not only provides richer information on the diagnosis and treatment process for model training but also helps reduce overfitting of the model to specific samples, providing a more accurate and personalized foundation for predicting subsequent treatment plans.

[0033] S4. Medical Pathway Prediction Steps:

[0034] Step S3 combines the time-series information of the treatment day into multiple treatment item sequences to further understand the dynamic changes of the treatment topics in the time series, and thereby introduces a deep learning model to analyze the time-series data in the treatment process.

[0035] In the deep learning model implemented in step S4, the example uses a model based on a bidirectional long short-term memory network and a temporally distributed fully connected layer. This model is designed to process sequential data and classify the output at each time step. Since the diagnosis and treatment process of patient hospitalization logs is long-term and continuous, the bidirectional long short-term memory network can memorize the temporal characteristics of the treatment days, while the bidirectional memory network understands the changes in the treatment pattern status of adjacent treatment days by remembering the complete diagnosis and treatment process in both forward and reverse directions.

[0036] For each treatment item in the treatment topic combination obtained after data augmentation, a binary classification multi-label feature code is established. Each treatment item in the treatment topic combination is assigned a label. When a treatment item is needed on a treatment day, the value of the label corresponding to that treatment item is yes; otherwise, the value of the label is no, thus forming a set of treatment item labels for the treatment day.

[0037] A bidirectional long short-term memory (LSTM) network stores the temporal features of treatment days and connects them to fully connected layers, enabling the processing of the set of treatment item labels for treatment days within a single time step. Based on the data characteristics of clinical pathways—locally disordered but globally ordered—we established a bidirectional sequence-annotated LSM network model for clinical pathway construction, combining multi-label features of treatment items for local treatment days with temporal features of treatment days in the global treatment process.

[0038] 4-1. Multi-label feature encoding steps:

[0039] Before a bidirectional sequence-annotated long short-term memory (LSTM) network model can understand a diagnosis and treatment process, feature encoding is a crucial step in the data preprocessing, especially when dealing with text data. Transforming words into a fixed numerical format that the bidirectional LSTM network model can understand is essential. To allow the network to learn complex properties between different labels, we transform unstructured text data into structured numerical features suitable for deep learning models. In multi-label classification problems, each instance of a diagnosis and treatment process may contain multiple labels (diagnosis and treatment items). We use a multi-label binary classification method to convert the list of target labels into a binary array. Specifically, this involves two main steps: First, identifying all possible labels and establishing a mapping L = Labels{I1, I2, ..., I...} to index positions. N}=

[0040] {l1,l2,…,l N}={1,…,N}, where Labels represents the set of label index positions, I N Indicates the Nth label, l NThe first step represents the index position of the Nth label, where N is the index number. The second step converts the actual label set into a binary vector, where each element corresponds to a label. An element is 1, indicating that the corresponding treatment item has been used; otherwise, the element is 0. Finally, the multi-label binary classification method outputs a vector of length N, consisting of elements of 0 or 1, as a multi-label feature code for a treatment day. The multi-label feature code for each treatment day serves as the input to the bidirectional long short-term memory network model.

[0041] 4-2. Processing steps of bidirectional long short-term memory network:

[0042] A bidirectional long short-term memory network model is used to output the temporal characteristics of each treatment day.

[0043] Recurrent Neural Networks (RNNs) are neural networks that process sequential data, capable of handling sequences of arbitrary length, where information can be passed from one unit to the next over time. Standard RNNs face the vanishing and exploding gradient problems, limiting their learning ability on long sequences. Long Short-Term Memory (LSTM) networks, by introducing three gating structures to regulate information flow, effectively preserve long-term dependencies and avoid gradient problems. Bidirectional LSM networks combine forward and backward LSM architectures, considering both the context and background information of the sequence data. In bidirectional LSM networks, the output at each time step is determined by the hidden states in both directions, providing richer contextual information. Bidirectional LSM networks are an extension of LSM networks, consisting of two LSM layers, processing the forward and backward sequences of data respectively. This allows the network to capture the forward and backward dependencies of the sequence, with the output at each time step being a concatenation of the hidden states from the forward and backward LSM units. Bidirectional LSM networks use a series of ingenious gating mechanisms to regulate information flow, solving the vanishing gradient problem that traditional RNNs often encounter when processing long sequences. Each time step of a Long Short-Term Memory (LSTM) network unit contains four core operations: forget gate, input gate, unit state, and output gate. This effectively maintains long-term information flow and improves the performance of time series prediction and label classification tasks.

[0044] In these gating mechanisms, the forget gate determines how much information from the previous time step's cell state is retained. The output gate determines how much new information is updated into the cell state. Candidate memory cells generate new information candidates. During state updates, the new cell state is updated based on the outputs of the forget and input gates, incorporating information forgotten from the previous cell state and adding new information. The tanh function ensures the output remains between -1 and 1. The final hidden state is calculated based on the output gate and the current cell state, and this determines the information content of the output. These operations ensure that the hidden state contains information from the current time step and will serve as input for the next time step.

[0045] The bidirectional long short-term memory network model not only enhances the understanding of the diagnosis and treatment subject but also improves the accuracy and efficiency of treatment pathway prediction. Through this method, the present invention ensures high accuracy and robustness in the analysis of the diagnosis and treatment process, further providing patients with personalized and optimized treatment plans.

[0046] Bidirectional Long Short-Term Memory (BSSM) networks are used to effectively capture and analyze time-series data from patient care processes. The unique feature of BSSM networks lies in their bidirectional structure, enabling the network to simultaneously process and remember forward and backward dependencies in care logs. Specifically, when analyzing data from a specific care day, a BSSM network considers not only the care patterns prior to that day but also those subsequent. This comprehensive consideration of context allows BSSM networks to grasp the overall contextual relationships within the care process more comprehensively and accurately when handling clinical pathway analysis.

[0047] During implementation, the bidirectional long short-term memory network (LSTM) effectively constructed comprehensive temporal characteristics for each treatment day through the joint operation of its forward and backward LSM layers. This characteristic construction is not based solely on treatment information at a single point in time, but rather on comprehensive information from multiple consecutive time points, thus providing a richer and more in-depth analysis of the treatment process. Through this analysis, the bidirectional LSM can accurately predict key turning points and potential risks in the treatment process, providing healthcare professionals with crucial insights into patient treatment pathways, thereby optimizing treatment plans and improving patient outcomes.

[0048] In summary, bidirectional long short-term memory networks, through their efficient time-series data processing capabilities, provide strong technical support for in-depth analysis of patient diagnosis and treatment processes in electronic medical records, further improving the accuracy and efficiency of clinical pathway analysis.

[0049] 4-3. Processing steps for time-distributed fully connected layers

[0050] The time-distributed fully connected layer is used to output the predicted treatment items for each treatment day, thus completing the medical pathway construction.

[0051] The design of the temporally distributed fully connected layer aims to enhance the processing power of the deep learning model, ensuring that it can independently output the corresponding attribute label at each time step of the sequence. Each time step corresponds to information about each treatment day. This design enables the model to make accurate predictions at each time point; that is, the bidirectional long short-term memory layer captures the forward and backward information of each time step, providing an integration of contextual information from the previous and next treatment days. The fully connected layer utilizes this information to ensure a comprehensive understanding of the entire treatment process sequence. It independently processes the output of each time step of the bidirectional LSTM layer. Information from each time step is passed to the fully connected layer and undergoes the same processing flow (applying the sigmoid activation function for binary classification of each label), making an accurate prediction at each time step. The fully connected layer uses the same weight matrix to process each time step of the sequence in the bidirectional long short-term memory network layer, thus ensuring that the model processes information consistently and in parallel throughout the entire sequence. The advantage of this architecture lies in improving parameter efficiency and ensuring a comprehensive understanding of the data at each time step. Especially when characterizing complex diagnostic and treatment processes, the fully connected layer can capture and retain complete information about the diagnostic and treatment topics throughout the day.

[0052] To effectively handle multi-label classification problems in the output layer, the sigmoid activation function was employed. This activation function adjusts the opening degree of the output gate at each time step, generating a probability distribution for the occurrence of each label in the label set. This probability distribution reflects the likelihood of each treatment procedure occurring on a specific treatment day. In the model's output layer, we set a probability threshold to determine whether a treatment item label will be used; enabled labels are then identified as treatment procedures to be performed on the treatment day.

[0053] By combining a time-distributed fully connected layer and the sigmoid activation function, our proposed model can meticulously capture and predict label information in time series data. This method offers significant advantages for time series analysis in the medical field and has important practical significance for deeply understanding and learning from disease-specific treatment data. Through this invention, we can provide a more accurate and effective tool to help healthcare professionals optimize treatment pathways and provide better treatment options for patients.

[0054] like Figure 2As shown, the model for clinical pathway construction that implements the above steps includes a latent Dirichlet assignment topic model, data augmentation strategies, a deep learning model containing multi-label feature encoding, a bidirectional long short-term memory network, and a fully connected layer for event distribution. When the clinical log data of historical treatment days is input into the clinical pathway construction model, after the above steps, the model ultimately outputs the predicted clinical pathways for subsequent treatment days.

[0055] This embodiment uses clinical log data from patients with acute ST-segment elevation myocardial infarction (STEMI) at a municipal tertiary hospital, and performs several preprocessing steps to ensure data accuracy and relevance. Next, a Latent Dirichlet Allocation Topic Model (LDPA) is applied to mine topic content from the data of each hospitalization day, revealing key treatment patterns in the treatment process. The LPA establishes document-vocabulary relationships between treatment days and treatment items to extract implicit topics from the treatment process. Then, a bidirectional Long Short-Term Memory (LSTM) network model is used to analyze time-series data in the treatment process to understand and predict treatment pathways. The bidirectional LSTM model provides richer contextual information by considering the contextual information of the sequence data. Furthermore, a data augmentation strategy is employed, relying on the LPA to generate new treatment process datasets, increasing the diversity and complexity of medical text data through data re-representation and topic combination. This combination of model architectures enables the proposed framework not only to identify key treatment patterns but also to understand how these patterns change over time, thereby providing more accurate and personalized treatment plans.

[0056] The implementation scheme underwent a series of experimental verifications, demonstrating its significant feasibility and advantages in optimizing clinical pathways. Experimental results show that the model exhibits high efficiency and effectiveness in mining treatment patterns, particularly demonstrating its superior ability to extract key features from clinical data through a comprehensive consideration of consistency and perplexity scores. During the deep learning model training phase, the latent Dirichlet assignment topic model combined with the bidirectional long short-term memory network model architecture showcased its powerful treatment pattern memorization ability, especially in terms of significant improvements in key performance indicators such as accuracy, precision, recall, and F1 score, thus reflecting its superior learning and generalization capabilities.

[0057] During the validation phase, the model maintained high accuracy on new and unknown clinical pathway datasets, demonstrating its good adaptability and generalization ability. During the testing phase, the model maintained a high level of accuracy and stability even when dealing with diverse clinical pathway data. These results indicate that our model can not only accurately predict clinical pathways but also maintain its predictive stability and accuracy across different types of clinical data.

[0058] In summary, the experimental results of the Latent Dirichlet Assignment Topic Model and the Bidirectional Long Short-Term Memory Network Model strongly demonstrate the application value of this invention in the field of clinical pathway optimization. It is not only theoretically innovative but also exhibits significant effects and advantages in practical applications. These advantages make this invention an efficient and reliable tool in medical practice, helping to improve the quality and efficiency of medical services, thus making a significant contribution to clinical pathway management in the medical field.

Claims

1. A clinical pathway construction method based on deep learning and data augmentation, characterized in that, Including the following steps: Data collection and preprocessing steps: Collect medical orders from the medical database in chronological order for patients with acute ST-segment elevation myocardial infarction. After data filtering and structural optimization of the medical orders, obtain the medical log data. The steps for mining treatment topics are as follows: The latent Dirichlet assignment topic model receives the input treatment log data. For each treatment day, it first generates a treatment day-topic distribution and a topic-treatment item distribution based on the Dirichlet distribution prior. Based on the treatment day-topic distribution, a topic set is obtained. Then, based on the topic-treatment item distribution, a set of treatment items corresponding to each treatment day is obtained and output. The treatment day-topic distribution is the topic set used on the corresponding treatment day. The topic-treatment item distribution is the set of treatment items used under the corresponding topic; Data augmentation steps: Using the generated diagnosis and treatment topics generated by the Latent Dirichlet Assignment Topic Model, the consistency score for each diagnosis and treatment topic is output to complete the selection and combination of multiple diagnosis and treatment topics, and output the set of diagnosis and treatment item combinations corresponding to the combination of diagnosis and treatment topics; Medical pathway prediction steps: First, a multi-label binary classification method is used to construct multi-label feature codes for each treatment day in the received treatment topic combination and input them into a bidirectional long short-term memory network; the bidirectional long short-term memory network outputs the temporal features of each treatment day to a fully connected layer, and the fully connected layer predicts the set of treatment items for the treatment day, thereby obtaining the predicted clinical pathway for the treatment day; In the data augmentation step, for each treatment day, a set number of topics with the highest consistency scores are selected to construct all possible combinations of treatment topics as the output treatment topic combinations.

2. The method as described in claim 1, characterized in that, Data screening in the data collection and preprocessing steps involves removing information irrelevant to treatment; structure optimization involves data merging and deduplication.

3. The method as described in claim 1, characterized in that, In the treatment topic mining step, the variational Bayesian inference method is used to introduce variational parameters to approximate the true posterior distribution during the optimization process of the variational expectation maximization of the latent Dirichlet assignment topic model. The lower bound of evidence is maximized by adjusting the Dirichlet hyperparameters to obtain the optimal treatment day-topic distribution and topic-treatment item distribution.

4. The method as described in claim 1, characterized in that, In the medical pathway prediction step, the length of the multi-label feature encoding is N, where N is the number of treatment items under the corresponding treatment topic. Each treatment item corresponds to one element. When the element value is 1, it means that the treatment item is used, and when the element value is 0, it means that the treatment item is not used.

5. The method as described in claim 1, characterized in that, In the medical pathway prediction step, the fully connected layer uses the same weight matrix to process each time step of the temporal features in the bidirectional long short-term memory network layer, and finally outputs the probability distribution of each label appearing on the predicted treatment day, thereby obtaining the set of treatment items for the predicted treatment day.

6. The method as described in claim 5, characterized in that, The fully connected layer uses the sigmoid activation function to adjust the opening degree of the output gate at each time step, generating the probability distribution of each label.

Images