Research On Process Mining Method And Technology For Stroke Diagnosis And Treatment Optimization

With the development and progress of medical technology,the complexity of disease diagnosis and treatment process is increasing.The effectiveness of these activities in diagnosis and treatment processes often determines the quality of medical services.In the era of big data,optimizing the diagnosis and treatment process through data analysis is of great significance to improve service quality and reduce medical costs.Recently,process mining methods and techniques have become a research hotspot in medical informatics.In the widely used medical information system,a large number of detailed log data of clinical activities are recorded,including activity name,timestamp,executor,etc.By analyzing these log data,process mining can realize model discovery,conformance check,and healthcare improvement.However,the current research results still have some limitations,which are challenging to meet the requirements of standardization and flexibility in clinical practice.In recent years,with the development of social economy and the change of people’s lifestyle,the risk factors of cerebrovascular disease like stroke are largely exposed.Stroke has the characteristics of high incidence rate,high disability rate,high mortality rate,and high recurrence rate.It has led to a rapid increase in population and heavy disease burden,becoming an increasingly serious public health problem.To this end,our country is currently promoting the establishment of five emergency centers(stroke center,chest pain center,etc.)in hospitals,to improve the effectiveness of emergency treatment for these diseases.However,since the diagnosis and treatment process of those diseases is a complex task with high risk,multi-department participation,and urgent time,the existing process mining methods are difficult to adapt to these application scenarios.There exists some practical problems,such as the low standardization degree of process models,the low accuracy of thrombolysis decision-making by primary doctors,the difficulty in estimating value of medical interventions,and the lack of follow-up data from discharged patients.In this paper,taking stroke as an example,we optimize and integrate the diagnosis and treatment based on process mining method.Below is a summary of the key research findings:1.To solve the problems of insufficient standardization,weak expansibility,and incomplete representation in clinical process models,a new multi-perspective method based on declarative modeling is proposed.Referring to the archetype specification of openEHR international standard,the attributes of clinical activities are classified by business,and the constraints between events are represented by the relation matrix.It is convenient for understanding and extension by clinical experts.Moreover,in order to support the comprehensive modeling and representation for medical process,the constraint relationships in data flow such as time and organization are added.Finally,a conformance checking algorithm is designed.The feasibility and effectiveness of this representation method are verified on a real-life data set.This method combines the two specifications of openEHR and process mining,representing a medical process model that conforms to international standards and clinical practice habits.2.To solve the problem of inconvenient access to event logs for process mining,a tool for data extraction and event log conversion based on openEHR archetype query language(AQL)is developed.First,the corresponding relationship in metadata is determined by parsing the archetype and template files of openEHR.Second,according to users’ query keywords,the possible archetypes and corresponding data items are listed after query extension.Third,the AQL scripts are automatically generated based on the data item chosen by users and retrieval conditions.Finally,AQL statements are executed on the EHR server to extract the data and convert it to standard event log format(XES)used in process mining.The tool was used in the national stroke registry program,and the feasibility of the method was verified.It provides an easy-to-use tool for semantic interoperability between information systems and for clinical researchers to obtain event logs.It also provides a new way for clinical researchers to query and obtain event logs through a domain knowledge model instead of the actual database structure,to realize semantic interoperability across information systems.3.In order to solve the problem that clinical guidelines cannot cover all detailed business scenarios,we propose a clinical decision support method using predictive process monitoring.The thrombolytic process of acute ischemic stroke is tested in experimental verification to make up for clinical guidelines.At present,the mainstream methods of building clinical assistant decision-making models can be divided into two categories.One is the machine learning method based on big data to discover or generate new knowledge,but the interpretability of this kind of model often faces the problem of insufficient interpretability.The other is expert system method based on logical rules,expressing medical knowledge such as clinical guidelines,but it is difficult to adapt to a complex application environment.In this paper,by manually constructing a labeled data set,the data-driven and rule-driven methods are combined to identify the best decision for each case and whether the actual activities conform to the clinical guidelines.After the prefix extraction and filtering in control flow,the data flow is encoded to train corresponding prediction models.The experimental results showed that the performance for intravenous thrombolysis and arterial thrombolysis was significantly improved.Therefore,this method can supply clinical guidelines and provide doctors with more accurate decision-making services from different perspectives,exploring the medical AI driven by data and knowledge.4.Aiming at the problem that it is difficult to assess the value of intervention activities in diagnosis and treatment,this paper proposes a relative value evaluation method for clinical intervention based on process mining.Firstly,the clinical pathway of cerebral infarction published by the National Health Commission is expressed in the BPMN model.Secondly,the event log is compared with the process model to detect the variations of clinical pathway.Finally,for the variations that are not required to be done,the relative value of each activity is calculated based on cost-utility ratio.The experimental group and control group are matched by propensity score,and the results are compared with the results of the model.The experimental results showed that neuroprotective drugs were the most common variation activities on the data set of ischemic stroke patients in the cooperative hospital.Both the relative value of variation and recommended activities in clinical pathways are calculated respectively.It provides a new way to evaluate the value of intervention activities.This method combines the effect and cost of medical intervention,providing a reference for evaluating the value of intervention activities and realizing value-based healthcare.5.In order to overcome the difficulties of traditional methods in building recurrence models on discharged patients,a recurrence risk prediction method based on medical process discovery and transfer learning is proposed.Firstly,process models are found from clinical guidelines to analyze the similarity of data collected by different medical institutions,using the control flow variables as additional characteristics of patients.Moreover,by adjusting the weight of target domain data and source domain data,the recurrence prediction models are trained using instance-based transfer learning method.Finally,taking the recurrence of ischemic cerebrovascular events(ICE)as an example to verify the effect of this method,205 discharged cases from a tertiary hospital(target domain)and 2954 cases from the national stroke screening cohort(source domain)were tested.The results show this framework can effectively improve the performance of three instance-based migration algorithms,and is superior to the clinical risk scoring tools commonly used,alleviating the limitation of insufficient labeled follow-up data in hospitals.The importance of model features is calculated by two methods to overcome the problem of insufficient interpretability.It proved that representing process variables as new features can improve the model performance.Integrating process discovery and transfer learning,this framework expands the application scope of process mining and enhances the performance of transfer learning model.It provides a new solution for the key technical problems of machine learning,such as insufficient samples and variable selection.In summary,although we take stroke as a case study in this paper,the process mining technology framework proposed has certain universality,which can be extended to clinical process mining of acute and critical illnesses like cardiovascular disease.It can improve the secondary utilization effect of medical data,providing method and tool support to optimize diagnosis and treatment process.