On Treatment Patterns for Modeling Medical Treatment Processes in Clinical Practice Guidelines

Liqin Yang,Liang Zhang

1School of Computer Science,Fudan University,Shanghai 200438,China

2Shanghai Key Laboratory of Data Science,Shanghai 200438,China

3Institute of Intelligent Electronics and Systems,Shanghai 200438,China

4Computer Teaching and Research Section,Shanghai University of Traditional Chinese Medicine,Shanghai 201203,China

Abstract:Clinical practice guidelines(CPGs)contain evidence-based and economically reasonable medical treatment processes.Executable medical treatment processes in healthcare information systems can assist the treatment processes.To this end,business process modeling technologies have been exploited to model medical treatment processes.However,medical treatment processes are usually flexible and knowledge-intensive.To reduce the effort in modeling,we summarize several treatment patterns(i.e.,frequent behaviors in medical treatment processes in CPGs),and represent them by three process modeling languages(i.e.,BPMN,DMN,and CMMN).Based on the summarized treatment patterns,we propose a pattern-based integrated framework for modeling medical treatment processes.A modeling platform is implemented to support the use of treatment patterns,by which the feasibility of our approach is validated.An empirical analysis is discussed based on the coverage rates of treatment patterns.Feedback from interviewed physicians in a Chinese hospital shows that executable medical treatment processes of CPGs provide a convenient way to obtain guidance,thus assisting daily work for medical workers.

Keywords:clinical practice guidelines;treatment patterns;process modeling;BPMN;DMN;CMMN

I.INTRODUCTION

Clinical practice guidelines(CPGs)are systematically developed statements that contain evidence-based and economically reasonable medical treatment processes[1].Using CPGs can improve the quality of care,limit unjustified practice variations and reduce healthcare costs[2,3].However,it takes much effort to consult text form CPGs for medical workers.It is necessary to develop executable medical treatment processes in CPGs,which can assist medical workers to make treatment plans more effectively.Business process modeling has been exploited to represent medical treatment processes[4–7].Business processes in medical domain are usually flexible(i.e.,unstructured)[8]and knowledge-intensive[9](i.e.,including quite a lot of diagnosis decision logic and time constraint medical activities).These characteristics result in complex medical treatment process models.Thus modeling medical treatment processes becomes a difficult and demanding task.

To reduce the effort in modeling,pattern-based approaches have been proposed to help modelers to generate process models efficiently.As described in[10,11],a pattern“is the abstraction from a concrete form which keeps recurring in specific non-arbitrary contexts”.There is wide agreement that patterns can accelerate the process of designing a solution and reduce modeling time[12,13].Patterns enable participants of a community to communicate more effectively,with greater conciseness and less ambiguity.Russell et al.[14]have presented 43 workflow patterns derived from general business processes,but many of them are not used in medical domain,and some common medical process patterns are not included.Works in[15–17]defined linguistic patterns which can be used to formally represent the knowledge contained in CPG documents.Peleg and Tu[18]defined solution patterns and developed templates restricted to domain of screening and immunization for CPG modeling.However,these patterns are either applied as a preprocessing step in developing CPG models or restricted to limited clinical circumstances.To complement process patterns needed for treatment process modeling,we summarize seven treatment patterns from medical treatment process models of CPGs.A treatment pattern is a generic process fragment of frequent behavior in medical treatment processes.Compared to previous research on patterns,our treatment patterns are process-oriented so that they can be directly realized with existing process modeling languages.Besides,the treatment patterns are frequently used in CPGs for all kinds of diseases,i.e.,not restricted to certain clinical circumstances.Modelers will apply treatment patterns as templates and con figure them as process fragments conveniently rather than starting from scratch.

Considering the characteristics of medical treatment processes,the treatment patterns are realized with a combination of BPMN[19],DMN[20]and CMMN[21].Object Management Group(OMG)released these three modeling standards for publishing,sharing,and applying models and specification documents that describe clinical pathways.Works in[22–24]agreed that medical treatment process models represented with the combination of BPMN,DMN and CMMN are easy to understand and maintain.Besides,Camunda[25]provides an integrated modeling platform and execution engine supporting these three languages simultaneously.Therefore,we select them as the realization languages of our treatment patterns.In this research,we propose a set of process-oriented treatment patterns that can be applied to various clinical circumstances,and implement a modeling platform to support the use of treatment patterns.The contributions of this study are as follows:

·To reduce the effort in modeling medical treatment processes,we summarize a set of treatment patterns and represent them with BPMN,DMN,and CMMN.

·Based on treatment patterns,we propose a treatment pattern based framework to model medical treatment processes in CPGs.

·A platform is implemented to support process modeling with treatment patterns to validate the feasibility of our pattern-based approach.

·Our work is evaluated by an empirical analysis based on coverage rate of each pattern,and feedback under effectiveness from interviewed physicians after using the executable CPGs.

The rest of this paper is organized as follows.Section II introduces a treatment pattern based framework for modeling medical treatment processes given a new CPG document.The detailed treatment pattern descriptions and their realizations are presented in Section III.Section IV reports the implementation of our modeling platform for supporting the use of treatment patterns and an example of how to use the proposed patterns by modeling a medical treatment process of non-secondary hypertension.The evaluation of our approach is also reported in this section.The related work is presented in Section V,which is followed by the conclusion and future word in Section VI.

II.TREATMENT PATTERN BASED FRAMEWORK FOR MODELING MEDICAL TREATMENT PROCESSES

Before modeling medical treatment processes,treatment patterns are extracted from Chinese CPG documents of four kinds of diseases(see Table 1).To ensure the universality of the proposed patterns,the CPG documents involve various types,such as surgical operation,nursing,prevention and treatment of chronic diseases and emergency treatment.The associated medical treatment processes included in each CPG document are modeled manually with BPMN,DMN and CMMN.During the modeling process,we interviewed some domain experts in Putuo hospital to understand medical knowledge better.As a result,96 medical treatment process models are generated.Among them,seven treatment patterns are extracted according to their coverage rates in these models.Some of them can also be found in empirical evidence from literature[22–24,30].The details of these treatment patterns will be described in Section III.An empirical analysis based on the patterns’coverage rates will be discussed in Section IV.

Table 1.Four Chinese CPGs used as sources of treatment patterns.

Figure 1.Treatment pattern based framework for modeling medical treatment processes.

Next,we present the treatment pattern based framework for modeling medical treatment processes,as shown in Figure 1.Based on the summarized treatment patterns,the associated medical treatment processes of a new CPG document can be modeled with the following steps.Step 1:By analyzing the CPG document,descriptive information of medical treatment processes can be identified.As for a disease,there may be many treatment plans for different cases.One treatment plan corresponds to a medical treatment process.Step 2:For a medical treatment process,the descriptive information of its included treatment patterns can be identified from the process descriptive information.Step 3:For a treatment pattern,it can be instantiated by using our implementation platform and con figured to form a process model fragment according to its descriptive information.Step 4:As for the other process elements or structures,they can be modeled with general modeling notations or work flow patterns[14],and then combined to generate a complete medical treatment process model.Even though all the steps are conducted manually,a process model fragment can be efficiently produced by configuring the patterns with our modeling platform.An example of using the treatment patterns to model medical treatment processes in a real-world CPG document will be represented in Section IV.

III.TREATMENT PATTERNS

In this section, we present the description for treatment patterns.These patterns are classified into three categories based on business operations.Category I is used to describe activities with time constraints,category II is used to describe diagnosis decision making,and category III is used to describe unstructured process fragments in medical treatment processes.The treatment patterns are realized with process modeling languages,i.e.,BPMN,DMN or CMMN.Although a complete treatment process is represented as a BPMN model,the included Task and Sub-Process elements are modeled with BPMN,DMN or CMMN.Specifically,a DMN model is invoked by a Business Rule Task[19],a CMMN model is invoked by a Call Task[19],and a BPMN model can be embedded in a Sub-Process[19]element.

3.1 Category I:Patterns for Activities with Time Constraints

Pattern 1(Periodic Activity):An activity is repeatedly performed at set time intervals.It will be interrupted by expiration of time or some interrupting conditions.For example,a patient takes drugs two times one day and the treatment will last for one course.

Figure 2.The generic model of pattern 1(periodic activity).

Figure 3.The generic model of pattern 2(reaching goal within a valid period).

The generic model of this pattern is represented with BPMN.As shown in Figure 2,once the medical task starts at the Start Time,it will be performed repeatedly at Time Intervals.The interrupting event is triggered by Interrupting Timer Intermediate Event[19](see Figure 2a)or by Interrupting Condition Intermediate Event[19](see Figure 2b)on the boundary of the sub-process.

Pattern 2(Reaching Goal within a Valid Period):An activity is performed to achieve some goal within a particular valid period.For example,blood pressure is slowly reduced to 160/100mmHg within 24-48 hours.

The generic model of this pattern is represented with BPMN.As shown in Figure 3,the medical task in the sub-process is performed continually,until the goal is achieved.The valid period is represented by an Earliest Time and a Latest Time.The normal completion of the sub-process implies the medical goal is reached within the valid period or before the earliest time.Process variable Tag equals to False after the execution of sub-process indicates that the medical goal is reached in advance.The interruption from the sub-process means that the medical goal is not reached before the Latest Time.

Figure 4.The generic DMN model of pattern 3(counting algorithm).

3.2 Category II:Patterns for Diagnosis Decision Making

Pattern 3(Counting Algorithm):Get the number of conditions patients meet.The number is usually used for disease diagnosis.For example,as described in the guidelines of diagnosis and treatment of chronic constipation,if three or more corresponding conditions are met within three months,functional constipation will be diagnosed.

The generic model of this pattern is represented with DMN.Its generic Decision Requirements Diagram(DRD)[20]is shown in Figure 4a.It displays a set of conditions required for the Counting Algorithm decision.Figure 4b shows the generic Decision Table(DT)[20]of the counting algorithm decision.The decision table uses the Collect hit policy[20]with the Count operator(represented by C#).It does not matter what values are in the output column.The output of the decision table is the number of rows whose rules evaluate to be True.

Figure 5.The generic model of pattern 4(diagnose-action).

Figure 6.The generic DMN model of pattern 5(scorerecommend).

Pattern 4(Diagnose-Action):Determine the next plan carried out for a patient according to the results of the diagnosis decision.

The generic model of this pattern is realized as a BPMN process fragment,shown in Figure 5a.This pattern extracts “diagnosis decision”aspects,represented as a DMN model,from a BPMN model.The Exclusive Gateway[19]expresses branching logic based on the results of the Diagnose task(i.e.Business Rule Task).As shown in Figure 5a,the DMN model is embedded in the Diagnose task.The generic DT of the diagnose decision is shown in Figure 5b.The decision table uses Collect hit policy with the Unique operator(represented by U).According to an input entry that is True,a specific output value should apply.

Figure 7.The generic CMMN model of pattern 6(dynamic response).

Pattern 5(Score-Recommend):Determine the most suitable treatment plans for a patient among several alternatives by assessing their applicability.The pattern follows these steps:(1)a set of indicators are taken out to evaluate the alternative plans.(2)Each indicator for an alternative plan is associated with a value.Scoring algorithm works through adding together the values of all of the indicators.(3)Recommend the most suitable plans with the highest scores.

The generic model of this pattern is represented with DMN.The generic form of the Score-Recommend DRD is shown in Figure 6a.Each Score decision is carried out for an alternative plan to score it.The generic decision table in each Score decision is depicted in Figure 6b.The output of each indicator for an alternative plan is associated with a value.The table uses the collection hit policy with the Sum operator(represented by C+).For each row whose rule evaluates to True,the output value will be added to the sum representing the total score for the alternative plan.The results of each Score decision(i.e.score of each alternative plan)are then used as inputs for Recommend decision.The recommend algorithm is coded by the Literal Expression in the Recommend decision.The pseudo-code of the recommend algorithm is shown in Algorithm 1.The input is a set of alternative plans and the number of recommended plans.The output is a list of recommended plans.Line 3-5 collects all the alternative plans and their scores at first.Line 6 sorts these plans by their scores in descending order.Line 7 recommends k plans with the highest scores.Assuming there are N alternative plans,the time complexity of Algorithm 1 is O(N).

Figure 8.The generic CMMN model of pattern 7(activities of limited freedom).

3.3 Category III:Patterns for Unstructured Process Fragments

Pattern 6(Dynamic Response):Based on the feedback of the patients,doctors continuously decide what to do for the patients dynamically.This pattern often occurs in perioperative fields of surgery or in emergency treatment.For example,a patient is adjusted to operational status through various treatments according to his conditions.

The generic model of this pattern is shown in Figure 7,represented with a CasePlanModel[21](i.e.,a case)of CMMN.In the CasePlanModel,each treatment task has an entry criterion(i.e.,a sentry,represented by♢).After the Decide Further Steps task being finished,the result will be verified on the IfPart[21]of the sentries of each treatment task to decide if it should be started.As shown in Figure 7,all the tasks(marked by hashtag #)are represented as repeatable.This means that once any patient data changes,the decision task will be executed.The DMN model embedded in the Decide Further Steps task allows making complex decisions at one point.

Pattern 7(Activities of Limited Freedom):It describes medical activities with limited freedom.Medical activities are prescribed at the point of care according to the doctor’s individual experiences and skills.They also should be restrained by rules in CPGs.For example,among the laboratory examinations,some items are essential,and some items may be prescribed at discretionary times based on the patients’condition according to CPGs.

The generic model of this pattern is represented with a CasePlanModel of CMMN.As shown in Figure 8,tasks in this pattern are divided into two groups(i.e.,Stages[21]):basic task group and optional task group.Tasks in basic tasks group are further divided into three categories.The first category is Basic Tasks[21]which should be performed on all patients of a specific disease,represented by solid line frames.The second category is Conditioned Basic Tasks[21]which should be performed on patients with special characteristics,represented by solid line frames with sentries.The third category is Discretionary Basic Tasks[21]which should be planned at discretionary times based on the outcome of previous tasks,represented by dashed line frames with sentries.Tasks in the optional tasks group are Discretionary Tasks[21]which would be planned by doctors according to their experiences.

IV.IMPLEMENTATION AND EVALUATIONS

In this section,a modeling platform is implemented for supporting the use of treatment patterns,by which the feasibility of the pattern-based approach is validated.With the platform,a modeling example shows how these treatment patterns can be used.Then we evaluate our work from two aspects.On one side,we evaluate our approach by calculating the coverage rate of each treatment pattern in four Chinese CPGs.On the other side,the feedback from interviewed physicians who used the executable hypertension treatment processes shows the effectiveness of developed executable CPGs.

4.1 Implementation

We extend the open-source process modeling platform Camunda Modeler[25]with our proposed treatment patterns as integrated elements on the palette.Camunda Modeler includes BPMN,DMN,and CMMN modeling components in one platform.As shown in Figure 9,the widgets(i.e.,P1,P2,and P4)annotated with a solid line frame on the palette represent pattern 1(periodic activity),pattern 2(reaching goal within a valid period)and pattern 4(diagnoseaction)respectively,which can be modeled just with BPMN.It should be noted that other treatment patterns are not appeared in Figure 9 because they are modeled with DMN or CMMN modeling language.For example,pattern 3(counting algorithm)and pattern 5(score-recommend)are in the DMN component;pattern 6(dynamic response)and pattern 7(activities of limited freedom)are in the CMMN component of Cumanda modeler.Modelers can drag a treatment pattern from the palette to the main area to quickly obtain its instance.And then a process model fragment is generated by configuring the instance with basic process modeling elements.The implementation is coded in JavaScript programming language,and the source codes can be downloaded from Github(https://github.com/cocoylq/develop).

We take the medical treatment process of nonsecondary hypertension as an example to show how these treatment patterns can be used for modeling.By analyzing the description of the treatment process for non-secondary hypertension[31],five treatment patterns are identified.The identified fragments and the corresponding treatment patterns are listed in Table 2.We apply the treatment patterns as templates and configure them as process fragments according to the corresponding description.The complete medical treatment process model of non-secondary hypertension is generated by composing these process fragments,as shown in Figure 9.The used treatment patterns are annotated with dashed line frames,which are con figured as follows.

Configuration of Pattern 7:The process fragment of diagnosis assessment(usage of pattern 7)is shown in Figure 10.Inquiry about medical history,blood pressure(BP)test,blood biochemical,electrocardiogram,urine test and complete blood count(CBC)are basic items that should be performed on all patients with hypertension.Quantitative analysis of urinary albumin is conditioned basic item,compulsory for patient with diabetes mellitus.A postprandial sugar level test is recommended for patient whose fasting blood sugar(FBS)is greater than 6.1mmol/L.A quantitative urinary protein test is recommended for proteinpositive person.These two are discretionary basic items.For patients with suspected secondary hypertension,optional items can be selected,such as TC,MRI,NMN,brain function,cardiac function,etc.

Configuration of Pattern 4:The process fragment of stratification of cardiovascular risk(SCR)has been illustrated in Figure 9,annotated with a dashed line frame.The SCR DRD is shown in Figure 11.Target organ damage(TOD),complications,BP class and the number of cardiovascular risk factors are required as inputs for SCR decision.

Configuration of Pattern 3:The DMN model of CCRF(usage of Pattern 3)is shown in Figure 12.CCRF DRD is shown in Figure 12a.Age(over 65 years old),smoking,obesity,dyslipidemia,homocysteine,impaired glucose tolerance(IGT)and family histories(FH)are all cardiovascular risk factors.The decision table of CCRF is shown in Figure 12b.The result of CCRF decision is used as an input of SCR decision(see Figure 11).

Configuration of Pattern 5:The Recommend Drugs DRD(usage of Pattern 5)is shown in Figure 13a.According to the guidelines,five kinds of common used antihypertensive drugs are angiotension receptor blockers(ARB),angiotension-converting enzyme inhibitor(ACEI),dihydropyridine calcium chan-nel blockers(D-CCB),diuretics and β-blocker.Every kind of drugs has compelling or relative indications and contraindications.Given the limited space available,Figure 13b just displays the decision table of scoring for ACEI.Decision tables of other drugs are similar to that of ACEI.

Figure 9.The implementation screenshot of our modeling platform and a treatment process model for non-secondary hypertension.

Table 2.Detected content and the corresponding treatment patterns in the treatment process of non-secondary hypertension.

Configuration of Pattern 1:The drug treatment fragment is modeled by configuring Pattern 1,as shown in Figure 9(annotated with a dashed line frame).Patients take drugs at regular intervals,and visit doctors every three months.

Pattern 2 and Pattern 6 are not used in the medical treatment process of non-secondary hypertension.In fact,pattern 2 is used in the treatment process of hypertensive urgencies,and Pattern 6 is used in processes about perioperative fields of surgery or emergency treatment.As the limited space available,the usage of Pattern 2 and Pattern 6 are not written here.

4.2 Evaluations

Empirical Analysis:As mentioned in Section II,we select the most frequent process fragments as treatment patterns from 96 medical treatment process models of four Chinese CPGs.Here,frequency is evaluated by the coverage rate of each candidate pattern.The coverage rate for each candidate pattern can be expressed as

Figure 10.The model of diagnostic assessments(usage of pattern 7).

Figure 11.Stratify Cardiovascular Risk DRD.

Figure 12.The DMN model of CCRF(usage of pattern 3).

where F(P)is the number of medical treatment process models including Pattern P,and N is the total number of process models.Note that if Pattern P appears more than once in the same medical treatment process model,account one to F(P).

Figure 13.The DMN model of recommend drugs(usage of pattern 5).

Figure 14.Coverage rates of treatment patterns.

The coverage rate for each pattern is shown in Figure 14.P1 to P7 respectively correspond to Pattern 1 to Pattern 7 introduced in Section III.The most frequently used pattern is Pattern 4,which can be found in 64.6% medical treatment process models.And Pattern 7 can be found in 61.5% medical treatment process models.Pattern 4 is often used in diagnosis decisions,and Pattern 7 is often used in examination activities,so they are common in medical treatment processes.The coverage rates of Pattern 1 and 5 are close,and they can be found in 47.9% and 46.9% medical treatment process models respectively.Pattern 1 is often used in courses of treatment and pattern 5 is often used to select the most suitable treatment plans from alternatives.The coverage rate of Pattern 3 is 29.2%.Pattern 3 usually being used in Pattern 4,is often used to get the number of diagnostic indexes of patients.Pattern 2 and 6 have lower coverage rates with 15.6% and 13.5% respectively.In fact,they are often used for modeling unstructured process fragments in perioperative fields of surgery or in emergency treatment.Patterns 1～7 are most frequently used in practice,while the other process fragments(i.e.,Pattern 8 and Pattern 9)are much less used,so we finally select Pattern 1～7 as our treatment patterns.

Although the sources of treatment patterns are 96 medical treatment process models of four Chinese CPGs,to ensure the universality of the proposed patterns,the selected CPGs documents involve various types,such as surgical operation,nursing,prevention and treatment of chronic diseases,and emergency treatment.We believe that using these treatment patterns can reduce the time cost for modeling medical treatment processes.For example,when modeling a process fragment of drug recommendation,rather than starting out from scratch,Pattern 5 enables the modelers to start with a Score-Recommend pattern at hand.

Findings:Using the modeling platform,we generate 18 medical treatment process models from Chinese hypertension CPGs[31]based on the proposed treatment patterns.These process models are running on the Camunda execution engine.After the diseaserelated information(e.g.symptoms and signs related indicators,etc.)is input through the provided interface,the recommended treatment activities will be displayed on a physician’s task list and the complete process model can be browsed.We invited 12 physicians working in Putuo hospital to use the executable medical treatment processes.To improve the reliability of the evaluations,we randomly selected the physicians working in the department of cardiology or geriatrics,because physicians in both departments treat many patients with hypertension every day.Besides,the physicians are with different professional qualifications,from primary post to senior post.The profiles of these physicians are listed in Table 3.Most of the invited physicians are qualified with primary technical posts,because they need more clinical guidance from CPGs in their daily work.Then,they were asked to answer a questionnaire.The questionnaire is designed to subjectively evaluate the system according to five measures:Usefulness,Functionality,Usability,Efficiency and Experience on use.Questions designed for these five measures are listed in Table 4.Question 1 to Question 4 are answered on four-point Likert scales,from 1(completely disagree)to 4(completely agree).Question 5 is answered by descriptive texts,used to capture physicians’experiences of using the system.

Table 3.Profiles of the interviewed physicians.

Table 5 presents the percentage frequencies of responses for all options on question 1 to question 4.If the percentage frequencies of responses for“completely agree”and “agree”options are added up,100% of the interviewed physicians considered that the system was convenient to use(Usability)and can improve the efficiency for querying CPG(efficiency).91.7% of the interviewed physicians considered that the recommended treatment activities were useful for them to make treatment plans(Usefulness).Physicians answering “generally disagree”on this question considered that the description of recommended treatment activities was a little bit simple,and suggested that it would be better if the recommending reason had been added in the system.83.3% of the physicians thought that the functions provided by the system could meet the needs of current diagnosis and treatment work(Functionality).Physicians answering“generally disagree”on this question thought that the functions provided by the system were not so comprehensive enough.This suggests that new functions should be added to better assist the diagnosis and treatment work,such as the function of predicting the treatment effect.

The answers of question 5 are collected.They are summarized as follows:

-Capturing guidelines through computer system is more convenient than reading text form CPGs.

-It is easy to understand the treatment process of a disease through graphical models of medical treatment processes.

-Users can get targeted guidelines by information system which the medial treatment processes are running on.This saves a lot of troubles to search for the whole CPG documents.

Table 4.Questionnaire of using executable medical treatment processes.

Table 5.Percentage frequencies of responses for all options.

-Recommended activities according to the input are useful for making treatment plans for medical workers.

-The user interface is concise,thus it is easy to get started to use the system for getting guidelines.

-It seems more convenient to obtain guidance,which is much helpful to make treatment plans.

Findings from the feedbacks suggest that the development of executable medical treatment processes in CPGs is of significance for assisting the daily work of medical workers.Some suggestions provided by the interviewed physicians will be realized in the application system based on medical treatment processes.

V.RELATED WORK

Business process modeling has been exploited to represent medical treatment processes[4–7].BPMN[19],DMN[20]and CMMN[21]are three standard process modeling languages,and process models modeled with these languages are easy to understand and maintain[22–24].Thus,they are selected to realize treatment patterns in CPGs in this paper.At present,there are some works related to CPGs modeling.These works can be divided into two classes:ontology based CPGs modeling approach,and pattern based CPGs modeling approach.

5.1 Ontology Based CPGs Modeling Approach

Computer interpretable guidelines(CIGs)have been proposed to facilitate the transition from narrative guidelines to formal representations,such as GLIF,ASBRU,PROforma,GUIDE,and EON.Peleg et al.[32]reviewed these CIGs modeling approaches and established a consensus on their common structure.In all CIGs formalisms,the focus is to represent the clinical knowledge that specifies the decision logic and the ordering of clinical actions that helps a healthcare professional to manage a single patient.However,the main disadvantage is that most CIGs require an execution engine that is not freely available.In addition,complex work flow patterns cannot be easily modeled in this way.To the best of our knowledge,none of the CIGs formalisms has been widely adopted and applied nowadays.

5.2 Pattern Based CPGs Modeling Approach

In regard to the use of patterns,work flow patterns have been suggested for representing different workflow aspects:control flow[14],resources[33],data[34]and exception handling[35].However,some common medical process patterns needed for medical domain are not included.For supporting the modeling of CPGs,Serban et al.[15]defined linguistic patterns,which can be used to formally represent the knowledge about medical actions contained in CPG texts.Kaiser et al.[16]defined syntactic and semantic patterns that are used to develop extraction rules to identify and extract actions and processes out of CPG texts.This approach was further developed by the definition of semantic relation patterns to automatically formalize actions in a CPG language[17].However,the pat-terns in these works are at the text level,which are applied as a preprocessing step in the manual development of CPG models.Peleg and Tu[18]defined design patterns and developed visual templates restricted to the domain of screening and immunization.Different from the patterns introduced above,our treatment patterns in this study are process-oriented,and are commonly detected in CPGs regardless of disease addressed by the guidelines.

VI.CONCLUSION AND FUTURE WORK

This study presents seven treatment patterns i.e.,frequently occurring behaviors in four CPG documents of various types.They are realized by BPMN,DMN,and CMMN.Based on the treatment patterns,we present a treatment pattern based framework to model medical treatment processes for a new CPG document.A modeling platform is implemented to support the use of our treatment patterns.We believe that using these patterns will improve the efficiency of medical treatment process modeling.However,it must be mentioned that there might be other treatment patterns not included in this paper because the sources of the proposed patterns are limited.Besides,compared to starting from scratch, it is more efficient to produce process fragments by configuring treatment patterns,but the configuration and the modeling of a complete medical process is still a manual procedure.Moreover,feedbacks from 12 interviewed physicians might be not strong to validate the effectiveness of developed executable CPGs.

In the future,we will further model more medical processes in other new CPG documents,which give the possibility to find new treatment patterns.And we will try to find an automatic or semi-automatic method to con figure treatment patterns according to the descriptive information of process fragments.Besides,We will invite more physicians from different hospitals to use the executable medical treatment processes and give the feedbacks.

ACKNOWLEDGEMENT

The work is supported by Chinese National Key Research and Development Program (No.2017YFB1400604).

The authors are thankful for the positive support received from the Putuo hospital affiliated to Shanghai University of Traditional Chinese Medicine as well as to all medical staff involved.