Case feature label-based medical path accurate classification model generation method and system

Abstract

The application discloses a medical path accurate classification model generation method and system based on case feature labels, first extracts medical logs into structured logs, and collects treatment event sequences of each case; then, similar cases are aggregated through trajectory clustering, process mining is carried out on each clustering cluster, a process tree model is generated and a cycle node is optimized, and a medical path is extracted from the process tree model; a trained neural network model cluster is used to determine the case feature labels of the extracted medical path; finally, the case feature label determination and the corresponding medical path are integrated to form a usable medical path accurate classification model. The application combines trajectory clustering, process discovery, neural network and other technologies, fully gives play to the advantages of each technology, and the performance of the obtained model is much higher than that of a model generated by a single technology, so that the purpose of accurately recommending a medical path according to case features can be achieved.

Description

Technical Field

[0001] This invention belongs to the field of process mining and deep learning fusion in medical applications, and specifically relates to the construction of medical pathway models for complex cases and complications. Background Technology

[0002] A clinical pathway refers to a standardized treatment model and procedure established for a specific disease. It is a comprehensive model of clinical treatment that uses evidence-based medicine and guidelines to improve treatment organization and disease management, ultimately aiming to standardize medical practices, reduce variability, lower costs, and improve quality. It is an important way for medical institutions to provide "appropriate, efficient, low-cost, and high-quality" medical services under limited healthcare resources.

[0003] However, current clinical pathways are mainly used for the diagnosis of single diseases. When encountering complex cases, such as patients with other diseases that may lead to comorbidities and complications, blindly adopting the treatment plan of the clinical pathway may cause adverse consequences. The attending physician needs to adjust the treatment plan based on experience.

[0004] Generating reliable medical pathways for complex cases using existing massive medical logs has become a current research hotspot. Process mining techniques, which have emerged in recent years, have played a significant role in the field of medical pathways. They can discover features of treatment processes from vast medical logs, thereby mining medical process models to validate and guide physicians in determining treatment plans. However, traditional process discovery techniques fail to achieve satisfactory results when dealing with medical logs containing complex cases. This is because even for the same disease, treatment pathways can vary significantly due to different case characteristics. Traditional process discovery techniques do not differentiate between these differences, integrating them into a single process model, resulting in a complex and overgeneralized model. Furthermore, the medical pathways extracted from process models lack classification methods, making it difficult to match them with corresponding complex cases. Summary of the Invention

[0005] Purpose of the invention: To address the problems existing in the above-mentioned methods, the purpose of this invention is to provide a method and system for generating a precise medical pathway classification model based on case feature tags. This method can effectively utilize existing medical logs to provide corresponding treatment plans for cases with different characteristics, and provides a reliable approach for the application of complex cases.

[0006] Technical Solution: The technical solution to achieve the above-mentioned invention objective is: a method for generating a precise medical pathway classification model based on case feature tags, comprising the following steps:

[0007] Step 1: Extract the medical logs into a structured log; the structured log summarizes the treatment events of each collected case in chronological order.

[0008] Step 2: The set of treatment paths in the medical logs is divided into multiple subsets according to the similarity of the treatment event sequences by trajectory clustering, and each subset corresponds to a cluster.

[0009] Step 3: Each cluster is mined using process discovery techniques to obtain multiple process tree models;

[0010] Step 4: Extract medical pathways from each process tree model, extracting at least one medical pathway from each process tree model.

[0011] Step 5: Input the medical pathway extracted in Step 4 into the trained neural network model cluster for case feature label determination; the case feature labels include case attributes associated with complications, and each neural network model in the neural network model cluster is trained based on the training dataset consisting of the treatment event sequence processed in Step 1 and the selected case feature labels.

[0012] Step 6: Integrate case feature label determination and corresponding medical pathways to generate a precise classification model of medical pathways based on case feature labels.

[0013] Furthermore, the specific process of extracting medical logs in step 1 includes:

[0014] Step 1-1, Statistical analysis by case: Summarize the complete medical process of a patient's single medical visit from the medical log;

[0015] Steps 1-2: Sort by time: Sort all treatment methods for a case by time.

[0016] Steps 1-3: Summarize according to the XES format: The XES file has a hierarchical relationship of Log-Trace-Event tags. All medical logs correspond to the Log tag, a specific case corresponds to the Trace tag, and the Event corresponds to each treatment event, resulting in structured logs.

[0017] Furthermore, the specific process of trajectory clustering in step 2 includes:

[0018] Step 2-1: Using Levenshtein distance as an indicator to measure the similarity between case treatment pathways, calculate the Levenshtein distance between each pair of case treatment pathways;

[0019] Step 2-2: Use the Markov clustering algorithm to cluster the case treatment paths according to process similarity.

[0020] Furthermore, the specific process of process discovery in step 3 includes:

[0021] Step 3-1: Perform process discovery on each cluster from Step 2;

[0022] Step 3-2: Using the inductive mining algorithm, the direct following graph is segmented using four segmentation methods: "exclusive", "sequential", "parallel", and "cyclic" until it cannot be further decomposed.

[0023] Step 3-3: Represent the decomposition process in the form of a tree to generate a process tree model.

[0024] Furthermore, the specific process of extracting medical pathways in step 4 includes:

[0025] Step 4-1: Replace the cyclic nodes in each process tree model obtained in Step 3 with parallel nodes;

[0026] Step 4-2: Utilizing the characteristics of trees, under the condition that the process tree's nodes are "exclusive," "sequential," and "parallel," a set number or proportion of medical paths are obtained through random walks. The paths obtained through random walks are then filtered by the range of the number of medical events in the set paths to obtain medical paths that meet the conditions.

[0027] Furthermore, in step 5, the neural network model adopts the GRU-Attention model. The model uses GRU as the neural network unit and uses the Attention mechanism to further extract features of the treatment process. The extracted medical path is input into the trained GRU-Attention model for label determination, represented as follows:

[0028]

[0029] Where σ1, σ2, …, σ m Here, m represents the number of extracted medical pathways, and t1, t2, ..., t3 represents the number of extracted medical pathways. m This is the corresponding case feature label result, e i,j .attention represents the attention value of the j-th treatment event in the i-th medical path, and n is the maximum number of treatment events in the medical path.

[0030] Furthermore, the specific process of generating the medical pathway model in step 6 includes:

[0031] Step 6-1: After the extracted medical path is classified by the GRU-Attention model corresponding to multiple case feature labels, multiple case feature labels and corresponding medical event attention tables can be obtained. These are then combined and concatenated to form a triple d = (ct, p, a). m×k ), where ct is the case feature label group, p is the medical pathway, and a m×kIt is an attention table of event pairs in the medical pathway, where m is the number of case feature labels in ct and k is the number of medical events in p;

[0032] Step 6-2: Put all the extracted medical paths into the set of triples obtained by the model to form a medical path accurate classification model based on case feature labels.

[0033] Furthermore, one or more case feature labels are input into the medical pathway accurate classification model, and the model outputs one or more recommended medical pathways and the attention value of each treatment event in the medical pathway to each label.

[0034] Based on the same inventive concept, the present invention discloses a medical pathway accurate classification model generation system based on case feature tags, comprising:

[0035] The preprocessing module is used to extract medical logs into structured logs; the structured logs summarize the treatment events of each collected case in chronological order.

[0036] The clustering module is used to divide the set of treatment paths in medical logs into multiple subsets based on the similarity of treatment event sequences through trajectory clustering. Each subset corresponds to a cluster.

[0037] The process discovery module is used to mine each cluster using process discovery techniques to obtain multiple process tree models;

[0038] The path extraction module is used to extract medical paths from each process tree model, with at least one medical path extracted from each process tree model.

[0039] The label determination module is used to input the medical path extracted by the path extraction module into the trained neural network model cluster for case feature label determination; the case feature labels include case attributes associated with complications, and each neural network model in the neural network model cluster is trained based on the training dataset composed of the treatment event sequence obtained by the preprocessing module and the selected case feature labels.

[0040] It also includes a model generation module, which integrates case feature label determination and corresponding medical pathways to generate a precise classification model of medical pathways based on case feature labels.

[0041] Furthermore, the system also includes a medical pathway recommendation module, which is used to input one or more case feature labels into the medical pathway accurate classification model, and the model outputs one or more recommended medical pathways and the attention value of each treatment event in the medical pathway to each label.

[0042] Beneficial Effects: This invention effectively utilizes existing medical logs to provide corresponding treatment plans for cases with different characteristics. First, trajectory clustering divides the treatment path set in the medical logs into multiple subsets with similar behaviors. Then, process discovery technology is used to mine multiple simplified process models. Reasonable medical paths are extracted from each process model, and a trained neural network model cluster is used to determine case feature labels and obtain the attention of events in the medical paths. Finally, the results are integrated to generate a precise classification model of medical paths based on case feature labels, providing a reliable approach for complex cases. Compared with existing technologies, this invention has the following significant advantages: 1) The trajectory clustering proposed in this invention divides the treatment path set in the logs into multiple subsets with similar processes. Process discovery is performed on each subset, and the resulting process models can exhibit unique treatment methods for different complications or different case characteristics; 2) This invention extracts medical paths from process models and finally uses a neural network model to classify the medical paths by label, determining the applicable case characteristics for each medical path; 3) The process discovery algorithm and neural network model used in this invention are mature and stable, achieving a good balance between speed and performance. Attached Figure Description

[0043] Figure 1 This is a flowchart illustrating an embodiment of the present invention.

[0044] Figure 2 This is a schematic diagram illustrating the model of an embodiment of the present invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention. It should be understood that the specific embodiments described herein are only used to explain this invention and are not intended to limit this invention.

[0046] Example 1

[0047] like Figure 1 As shown in the figure, an embodiment of the present invention discloses a method for generating a precise medical pathway classification model based on case feature labels, which includes the following steps:

[0048] Step 1: Extract the medical logs into structured logs;

[0049] Step 2: The set of treatment paths in the medical logs is divided into multiple subsets according to the similarity of the treatment event sequences by trajectory clustering, and each subset corresponds to a cluster.

[0050] Step 3: Each cluster is mined using process discovery techniques to obtain multiple process tree models;

[0051] Step 4: Extract medical pathways from each process tree model;

[0052] Step 5: Input the extracted medical pathways into the trained neural network model cluster for case feature label determination;

[0053] Step 6: Integrate case feature label determination and corresponding medical pathways to generate a precise classification model of medical pathways based on case feature labels.

[0054] Specifically, step 1 involves extracting the medical logs into XES format log files. The specific process includes:

[0055] Step 1-1, statistical analysis by case: summarize the complete medical process of a patient's one visit in the medical log across different information systems;

[0056] Steps 1-2: Sort by time: Sort all treatment methods for a case by time.

[0057] Steps 1-3: Summarize according to the XES format: The XES file has a tag hierarchy of Log-Trace-Event. All medical logs correspond to the Log tag, a specific case corresponds to the Trace tag, and the Event corresponds to each treatment event. Transform medical logs in text and other formats into structured logs.

[0058] Using the approach in this embodiment, we extracted treatment pathways from sepsis medical logs and constructed a medical log dataset containing 1000 sepsis treatment cases. Using a real-world dataset allows for thorough testing of the method's performance.

[0059] The specific process of trajectory clustering in step 2 includes:

[0060] Step 2-1: Using Levenshtein distance as an indicator to measure the similarity between case treatment pathways, calculate the Levenshtein distance between each pair of case treatment pathways;

[0061] Step 2-2: Use the Markov clustering algorithm to cluster the case treatment paths according to process similarity.

[0062] Using the scheme in this embodiment, we divided the logs into 7 clusters, which have significant differences in process between each other and high similarity within each cluster.

[0063] Step 3, the process discovery process specifically includes:

[0064] Step 3-1: Perform process discovery on each cluster from Step 2;

[0065] Step 3-2: Using the Inductive Miner algorithm, the direct following graph is segmented using four segmentation methods: "exclusive", "sequential", "parallel", and "cyclic" until it cannot be further decomposed.

[0066] Step 3-3: Represent the decomposition process in the form of a tree to generate a process tree model.

[0067] Using the scheme in this embodiment, by summarizing the characteristics of mining, we can effectively mine treatment models for several medical log clusters, and then obtain treatment paths for several sepsis cases with different characteristics through tree traversal algorithms.

[0068] Step 4, the specific process of extracting medical pathways, includes:

[0069] Step 4-1: Replace the cyclic nodes in the process tree models obtained in Step 3 with parallel nodes;

[0070] Step 4-2: Utilizing the characteristics of trees, a log generation algorithm is used to obtain several effective medical paths by random walks under the condition that the nodes meet the characteristics of "exclusive", "sequential", and "parallel" nodes of the process tree. (The number or proportion of paths can be set, such as the number of paths in the original log / 10 + 1). Finally, an interval of the number of events in the path is set to filter out unreasonable paths.

[0071] The approach used in this embodiment, by optimizing the loop nodes, allows for further extraction of treatment plan features. In actual treatment, some treatment methods may be implemented multiple times depending on the efficacy and symptoms, but they only appear once in the plan. Ultimately, 11 medical pathways are extracted.

[0072] In this embodiment, the neural network model in step 5 adopts the GRU-Attention model, and the specific process includes:

[0073] Step 5-1, use GRU as the neural network unit

[0074] z t =σ g (W z ·[h t-1 x t ])

[0075] r t =σ g (W r ·[h t-1 x t ])

[0076]

[0077]

[0078] Where x t h represents the input vector. t Represents the output vector, z t It updates the gate vector, r t It is the reset gate vector. This represents the current memory content, where t is the current time unit, and t-1 is the previous time unit. W z W r , These are different weight matrices. σ g The sigmoid function outputs 0-1 as a gate signal. The tanh function outputs values ​​from -1 to 1.

[0079] Step 5-2: Use the Attention mechanism to further extract features of the treatment process. The new hidden state vector s is the weight coefficient α of each hidden state in the new hidden state. i With respect to the hidden state h of the initial input i The sum of products.

[0080]

[0081]

[0082] e i =a(h i ) = tanh(w i h i +b i )

[0083] Among them, e i It is h i The hidden layer representation, w i It is the weight, b i It is a bias, α i These are the normalized weight coefficients after processing with the softmax function, u w It is the attention weight matrix, which is trained along with the model, and n is the number of hidden states;

[0084] Step 5-3: Based on the output of the Attention layer, calculate y = softmax(ws + b) using the softmax function to classify the case paths. The objective function used for training is the cross-entropy loss function. Where y is the label value. is the predicted value, w is the weight, and b is the bias. The dataset used for training consists of treatment event sequences and case feature labels obtained through preprocessing of medical logs.

[0085] Step 5-4: Use the trained model to determine the labels. The GRU-Attention model can be represented as:

[0086]

[0087] Where σ1, σ2, …, σ m Here, m represents the number of extracted medical pathways, and t1, t2, ..., t3 represents the number of extracted medical pathways. m This is the corresponding case feature label result, ei, j .attention represents the attention value of the j-th treatment event in the i-th medical path, and n is the maximum number of treatment events in the medical path, with zeros padded if the number is insufficient.

[0088] Using the scheme in this embodiment, 10 classifier models are trained based on 10 attributes related to complications in the log as labels to determine the labels of the extracted medical pathways.

[0089] Step 6, generating the medical pathway model, specifically includes:

[0090] Step 6-1: After the extracted medical path is classified by the GRU-Attention model corresponding to multiple case feature labels, multiple case feature labels and corresponding medical event attention tables can be obtained. After combination and concatenation, the triple d = (ct, p, a) can be formed as defined. m×k ); where ct is the case feature label group, p is the medical pathway, and a m×k This is an attention table for event pairs in the medical pathway, where m is the number of case feature labels in ct and k is the number of medical events in p.

[0091] Step 6-2: All extracted medical pathways are processed by the model to obtain triples, which are then placed into a set to form a precise medical pathway classification model based on case feature labels. For example... Figure 2 As shown.

[0092] When using the specific model, one or more case feature labels are input into the model, and the model can output one or more recommended medical paths and the attention value of each treatment event in the medical path to each label.

[0093] Using the scheme in this embodiment, a medical pathway model for sepsis can be obtained, and corresponding treatment plans can be derived based on the different physical conditions of patients, such as hypertension.

[0094] Example 2

[0095] Based on the same inventive concept, this invention discloses a medical path accurate classification model generation system based on case feature labels, comprising: a preprocessing module for extracting medical logs into structured logs; the structured logs summarize the treatment events of each collected case in chronological order; a clustering module for dividing the set of treatment paths in the medical logs into multiple subsets according to the similarity of the treatment event sequences through trajectory clustering, each subset corresponding to a cluster; a process discovery module for mining each cluster using process discovery technology to obtain multiple process tree models; a path extraction module for extracting medical paths from each process tree model, with each process tree model extracting at least one medical path; a label determination module for inputting the medical paths extracted by the path extraction module into a trained neural network model cluster for case feature label determination; the case feature labels include case attributes associated with complications, and each neural network model in the neural network model cluster is trained based on a training dataset consisting of the treatment event sequences obtained by the preprocessing module and the selected case feature labels; and a model generation module for integrating the case feature label determination and the corresponding medical paths to generate a medical path accurate classification model based on case feature labels. Furthermore, the system also includes a medical pathway recommendation module, which is used to input one or more case feature labels into the medical pathway accurate classification model, and the model outputs one or more recommended medical pathways and the attention value of each treatment event in the medical pathway to each label.

[0096] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of each module described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. The division of modules is merely a logical functional division; in actual implementation, there may be other division methods, such as multiple modules being combined or integrated into another system.

[0097] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

Claims

1. A method for generating a precise classification model of medical pathways based on case feature labels, characterized in that, Includes the following steps: Step 1: Extract the medical logs into a structured log; the structured log summarizes the treatment events of each collected case in chronological order. Step 2: The set of treatment paths in the medical logs is divided into multiple subsets according to the similarity of the treatment event sequences by trajectory clustering, and each subset corresponds to a cluster. Step 3: Each cluster is mined using process discovery techniques to obtain multiple process tree models; Step 4: Extract medical pathways from each process tree model, extracting at least one medical pathway from each process tree model. Step 5: Input the medical pathways extracted in Step 4 into the trained neural network model cluster for case feature label determination; the case feature labels include case attributes associated with complications, and each neural network model in the neural network model cluster is trained based on the training dataset consisting of the treatment event sequence processed in Step 1 and the selected case feature labels; the neural network model adopts the GRU-Attention model, which uses GRU as the neural network unit and uses the Attention mechanism to further extract treatment process features. The extracted medical pathways are input into the trained GRU-Attention model for label determination as follows: ； in This represents the extracted medical pathways, where m is the number of extracted medical pathways. These are the corresponding case feature label results. Let represent the attention value of the j-th treatment event in the i-th medical path, and n be the maximum number of treatment events in the medical path; Step 6: Integrate case feature label determination and corresponding medical pathways to generate a precise classification model of medical pathways based on case feature labels.

2. The method for generating a precise medical pathway classification model based on case feature labels according to claim 1, characterized in that, Step 1, the specific process of extracting medical logs, includes: Step 1-1, Statistical analysis by case: Summarize the complete medical process of a patient's single medical visit from the medical log; Steps 1-2: Sort by time: Sort all treatment methods for a case by time. Steps 1-3: Summarize according to the XES format: The XES file has a hierarchical relationship of Log-Trace-Event tags. All medical logs correspond to the Log tag, a specific case corresponds to the Trace tag, and the Event corresponds to each treatment event, resulting in structured logs.

3. The method for generating a precise medical pathway classification model based on case feature labels according to claim 1, characterized in that, The specific process of trajectory clustering in step 2 includes: Step 2-1: Using Levenshtein distance as an indicator to measure the similarity between case treatment pathways, calculate the Levenshtein distance between each pair of case treatment pathways; Step 2-2: Use the Markov clustering algorithm to cluster the case treatment paths according to process similarity.

4. The method for generating a precise medical pathway classification model based on case feature labels according to claim 1, characterized in that, Step 3, the process discovery process specifically includes: Step 3-1: Perform process discovery on each cluster from Step 2; Step 3-2: Using the inductive mining algorithm, the direct following graph is segmented using four segmentation methods: "exclusive", "sequential", "parallel", and "cyclic" until it cannot be further decomposed. Step 3-3: Represent the decomposition process in the form of a tree to generate a process tree model.

5. The method for generating a precise medical pathway classification model based on case feature labels according to claim 1, characterized in that, Step 4, the specific process of extracting medical pathways, includes: Step 4-1: Replace the cyclic nodes in each process tree model obtained in Step 3 with parallel nodes; Step 4-2: Utilizing the characteristics of trees, under the conditions of meeting the "exclusive", "sequential", and "parallel" node characteristics of a process tree, a set number or a set proportion of medical paths are obtained through random walks. The paths obtained through random walks are then filtered by the range of the number of medical events in the set paths to obtain medical paths that meet the conditions.

6. The method for generating a precise medical pathway classification model based on case feature labels according to claim 1, characterized in that, Step 6, generating the medical pathway model, specifically includes: Step 6-1: After the extracted medical paths are classified by the GRU-Attention model corresponding to multiple case feature labels, multiple case feature labels and corresponding medical event attention tables can be obtained. These are then combined and concatenated to form triples. Where ct is the case feature label group, and p is the medical pathway. It is an attention table of event pairs in the medical pathway, where m is the number of case feature labels in ct and k is the number of medical events in p; Step 6-2: Put all the extracted medical paths into the set of triples obtained by the model to form a medical path accurate classification model based on case feature labels.

7. The method for generating a precise medical pathway classification model based on case feature labels according to claim 6, characterized in that, Step 6 also includes inputting one or more case feature labels into the medical pathway accurate classification model, and the model outputting one or more recommended medical pathways and the attention value of each treatment event in the medical pathway to each label.

8. A system for generating a precise medical pathway classification model based on case feature labels, characterized in that, include: The preprocessing module is used to extract medical logs into structured logs; the structured logs summarize the treatment events of each collected case in chronological order. The clustering module is used to divide the set of treatment paths in medical logs into multiple subsets based on the similarity of treatment event sequences through trajectory clustering. Each subset corresponds to a cluster. The process discovery module is used to mine each cluster using process discovery techniques to obtain multiple process tree models; The path extraction module is used to extract medical paths from each process tree model, with at least one medical path extracted from each process tree model. The label determination module is used to input the medical paths extracted by the path extraction module into a trained neural network model cluster for case feature label determination. The case feature labels include case attributes associated with complications. Each neural network model in the cluster is trained on a training dataset consisting of treatment event sequences obtained from the preprocessing module and the selected case feature labels. The neural network model employs a GRU-Attention model, using GRUs as neural network units and employing an attention mechanism to further extract treatment process features. The extracted medical paths are input into the trained GRU-Attention model for label determination as follows: ； in This represents the extracted medical pathways, where m is the number of extracted medical pathways. These are the corresponding case feature label results. Let represent the attention value of the j-th treatment event in the i-th medical path, and n be the maximum number of treatment events in the medical path; It also includes a model generation module, which integrates case feature label determination and corresponding medical pathways to generate a precise classification model of medical pathways based on case feature labels.

9. The medical pathway accurate classification model generation system based on case feature tags according to claim 8, characterized in that, It also includes a medical pathway recommendation module, which is used to input one or more case feature labels into the medical pathway accurate classification model, and the model outputs one or more recommended medical pathways and the attention value of each treatment event in the medical pathway to each label; The process of generating a precise medical pathway classification model includes: after the extracted medical pathways are classified by a GRU-Attention model corresponding to multiple case feature labels, multiple case feature labels and corresponding medical event attention tables are obtained, which are then combined and concatenated to form triples. Where ct is the case feature label group, and p is the medical pathway. It is an attention table of event pairs in the medical pathway, where m is the number of case feature labels in ct and k is the number of medical events in p; the triples obtained from all extracted medical pathways through the model are put into the set to form a medical pathway accurate classification model based on case feature labels.

Images