The interoperability of clinical information is a major issue that hinders data exchange and knowledge reuse between clinical institutions [1]. An important type of clinical information is the representation of institution-specific care protocols, known as clinical pathways (CPs), defined as “task orientated care plans which detail essential steps in the care of patients with a specific clinical problem and describe the patient’s expected clinical course” [2]. This definition has been reviewed in the studies by De Bleser et al [3] and Vanhaecht et al [4] and further refined in the study by Kinsman et al [5] through five criteria: “(1) a structured multidisciplinary plan of care; (2) used to channel the translation of guidelines or evidence into local structures; (3) detailed the steps in a course of treatment or care in a plan, pathway, algorithm, guideline, protocol or other inventory of actions; (4) had timeframes or criteria-based progression (that is, steps were taken if designated criteria were met); and (5) aimed to standardise care for a specific clinical problem, procedure or episode of healthcare in a specific population.” CPs often rely on medical evidence expressed in clinical guidelines (CGs), which standardize abstract best practices for the diagnosis and treatment of specific medical conditions based on both medical evidence and expert consensus to ultimately improve the quality and uniformity of care, describing the strict temporal order in which clinical work needs to be carried out. For this purpose, many clinical task-oriented tools have been used over the years, such as Asbru, GLIF, GLARE, PROforma, EON, or GUIDE [6], but none of them seem to have reached the popularity of general-purpose process management tools such as the business process model and notation (BPMN) from the Object Management Group (OMG) [7], probably because of their complexity and narrow use limited to high-technology institutions with enough financial and technical means. In the last years, BPMN, which is widely used in other industry domains, has also emerged in clinical domains as a process management standard. Many studies describe the use of BPMN to represent clinical processes to improve efficiency or serve as a basis for the development of clinical decision support systems, which require seamless integration between electronic health record (EHR) data, decisions, and the specific workflow of a medical institution [8-12]. Although BPMN has proven to be effective in representing clinical processes, the transition to process execution in real clinical institutions is still scarce [13]. Indeed, although BPMN can effectively help clinicians visually understand clinical processes and detect possible inefficiencies, the implementation of otherwise complex, stepwise clinical workflows in an ordered manner is not easy to accomplish as temporal interrelations between tasks are crucial. In the last years, new clinical-oriented process representation standards have emerged that address the specific needs of clinical workflows, such as openEHR Task Planning (TP) [14]. Our work is based on the hypothesis that the clinical-oriented business process management (BPM) tool TP provides a more adequate way to represent CPs, bridging the gap between abstract CG logic and clinical workflow and offering essential tools for the representation of complex health care processes such as infection treatment. In our research, we analyze how TP can represent complex infection CPs taking as a case study the catheter-related bloodstream infection (CR-BSI) CGs developed by the Johns Hopkins Hospital (JHH) Antimicrobial Stewardship Program, an international gold standard that includes recommendations aiming to standardize clinical practice around antibiotic prescription, thereby helping minimize antibiotic resistance [15].

The structure of this paper is as follows: in the Introduction section, we explain the background; in the Methods section, we illustrate the methodology used; in the Results section, we describe our experiments; and, in the Discussion section, we discuss the relevance and limitations of the results.

The best up-to-date evidence-based clinical knowledge is usually expressed in CGs, defined as “systematically developed statements to assist practitioner and patient decisions about appropriate healthcare for specific clinical circumstances” [16], often used in a paper-based mode. Many studies have successfully represented CG knowledge using different notations, such as the procedural medical-oriented Arden Syntax, rule-based systems such as Drools, guideline definition languages (Graphic Language for Interactive Design [GLIDE] and the openEHR Guideline Definition Language), or Semantic Web rule systems such as SPARQL Inferencing Notation (SPIN) or Shapes Constraint Language (SHACL) [17]. However, knowledge representation technologies do not suffice to represent a specific patient evolution over time, which has led to the development of medical-oriented task-based systems such as PROforma [18], Asbru [19], or Prodigy [20], which have been successfully used in many clinical scenarios but are effectively limited to a few medical institutions owing to the high costs and efforts associated with their implementation. This has favored the introduction in the last years of popular, easy-to-use, general-purpose BPM standards in the health care landscape as a working alternative to complex, medical task-oriented knowledge systems. However, modeling specific constraints in clinical processes remains a challenge because of the intrinsic domain complexity. Much research has been conducted over the years to eliminate or compensate for the temporal shortcomings of generic BPM notations, such as with Petri nets (a graphical workflow formalism for conceptual modeling of distributed systems [21,22]), Process Mining for Healthcare [23], or improved integration with data or decision support [9,10,24].

CR-BSI is a highly recurring infection and a major cause of morbidity and costs in hospitals. The leading cause of CR-BSI is gram-positive bacteria present in intravascular catheters, especially the coagulase-negative Staphylococcus species, which must be treated with a multidisciplinary approach based on catheter removal or catheter salvage combined with an antimicrobial lock therapy. Only in the United States, >150 million intravascular catheters are purchased by hospitals each year [25,26]. In addition, >250,000 intravascular CR-BSI cases happen each year, with an attributed mortality rate of 12% to 25% [27]. Of these, approximately 80,000 CR-BSI cases take place in intensive care units, where “more than 15 million central vascular catheter (CVC) days occur each year” [28]. This happens in a context of insufficient research on new antibiotics and limited supply of the existing ones, which narrows down the possible therapies for highly recurring bacterial infections, many of which are still susceptible to generic antibiotics with lower toxicity levels and lower risk of resistance development [29]. The treatment options for CR-BSI depend on the microorganism causing the infection, which determines the type of antibiotic and the time, frequency, and dosage of its intake. Furthermore, the type of catheter, the way it is handled, the duration of its placement, and the patient conditions codetermine the risk of a hospital-acquired CR-BSI, increasing the length of hospital stay and mortality figures especially in patients who are critically ill and whose catheter is not removed [30]. Therefore, specific CGs have been developed for health care staff to prevent the development of CR-BSI and issue recommendations regarding the most suitable course of treatment. The JHH CGs, based on current literature, Infectious Diseases Society of America national guidelines [31], and JHH medical evidence, cover the entire infection course from diagnosis through antibiotic treatment and assist clinicians in the selection of the optimal antibiotic therapy in an attempt to fight antibiotic resistance and still remain effective [15].

BPMN is a mature, general-purpose BPM graphical representation and ISO standard developed by OMG based on an unstructured graph-oriented language combined with features from other workflow languages that can be represented using Petri nets [32]. In BPMN, a process is a free sequence of activities or events ordered in a sequence flow and connected through split or merge gateways that redirect the flow into one or multiple paths. This standard has been widely used by business process managers in many different application domains owing to its simplicity. Despite not being specifically designed for clinical processes, BPMN has proven its value in the health care domain to represent CPs, allowing for an easy-to-understand representation of CP recommendations [12,13,33,34]. However, its use in clinical domains is still scarce owing to the complexity of clinical processes and organizations, the critical relevance of care protocols, the difficulties in handling temporal constraints, and the uncertainty surrounding patient evolution and treatment effectiveness [13]. Furthermore, the integration of BPMN process models with EHR data is usually modeled using Unified Modeling Language [35] to address the BPMN shortcomings in that regard [24]. In addition, BPMN provides a limited set of capabilities for modeling temporal constraints [34,36] as the standard lacks out-of-the-box temporal semantics. Therefore, many extensions of the BPMN metadata model have been proposed over the years to address the specific demands of clinical processes [11,34,36-44].

openEHR is an initiative of the openEHR Foundation working closely with the European Committee for Standardization (CEN), the International Organization for Standardization (ISO), Health Level 7 (HL7), OMG, and other organizations on EHR and clinical modeling standards that provides an archetype-based standard data model for EHR systems designed with interoperability in mind [45]. In 2016, the openEHR Foundation released a new clinical-oriented workflow management standard, the TP specification, that includes a Visual Modeling Language (TP-Visual Modeling Language [TP-VML]) backed up by formal semantics, which ultimately would allow for the automatic translation of graphical workflow models into executable models [14,46]. TP complements openEHR by allowing modeling orders and actions in the future, structured as work or task plans, as well as their eventual execution in distributed environments, improving features partially present in other workflow languages (BPMN, Yet Another Workflow Language [YAWL] [47], Case Management Model and Notation [CMMN] [48], or Decision Model and Notation [DMN] [49]). The TP engine executes both TP work plans and openEHR Decision Language (DL) rules [50]. Moreover, TP allows for the specification of complex times in medication orders (eg, “3 times a day before meals” [51]), and it is conceived as a clinical process navigator, empowering users to perform ad hoc modifications at the time of execution to reflect real-time changes and supporting auditing, reporting, and billing to analyze task performance, execution, and eventual deviations from the original plan. Although openEHR has a wide coverage as a research area [52-54], the TP specification has been used in fewer studies [55-57] and is still under development. The contributions of our work to support our hypothesis of the native suitability of clinical-oriented BPM formalisms for infection CPs are as follows: (1) identification of key features required for modeling infection CPs, as laid out in the literature, the JHH CGs, and the successive BPMN extensions; (2) theoretical analysis of how TP addresses the identified key features; (3) a case study to empirically show the native TP adequacy by modeling typical CR-BSI infection patterns using TP; and (4) analysis and proposal of potential extensions of the current version of TP for improved modeling of infection CPs and possibly of other complex CPs.

Our research aimed to explore and identify key features to represent infection CPs to then analyze how they are present in the TP standard. The analysis can lead to a potential list of benefits and possible improvements for the representation of infection CPs using TP. For this purpose, in this study we adopt the methodology illustrated in Figure 1.

We identified key features required for modeling infection CPs by reviewing scientific literature, BPMN extensions in clinical settings, JHH CGs, and an extended BPMN-based model of JHH CR-BSI. The identified key features formed a basic set of requirements that served as a starting point to analyze the theoretical adequacy of TP for the representation of infection CPs.

We conducted a theoretical analysis of the presence of the previously identified key features in the TP specification by performing an in-depth evaluation of the current version of TP to understand how these capabilities were addressed by the standard. When possible, a one-to-one comparison of both standards’ capabilities was performed, which could help identify possible improvements. In addition, when new, potentially interesting TP capabilities were detected, we discussed their relevance.

We conducted an empirical test of the suitability of the current version of TP for conceptual modeling of infection CPs. To that end, we modeled typical infection treatment patterns from the JHH CR-BSI CGs using the same logic used in the extended BPMN version [44]. When opportunities for improvements were detected, we proposed alternative TP models.

Taking the output of steps B and C, we proceeded with the analysis of benefits and potential enhancements to the current version of the TP standard and made a proposal of possible extensions to improve the representation of infection CPs. For this purpose, we followed the steps of the methodology for BPMN extensions laid out in the study by Braun and Schlieter [58], which we consider adequate for TP.

BPMN has been extended over the years with domain-specific concepts to meet the main requirements of process representations in clinical domains. Therefore, an analysis of the BPMN missing capabilities through its extensions in clinical settings is essential to identify key features that any CP representation should fulfill. Furthermore, any language for modeling CPs should contain (1) concepts for medical business process modeling (patient state, treatment step, decision, or process flow) plus the ability to integrate information objects and responsibilities, (2) indefinite order relations as well as compulsory parallel relations between treatment steps and iterating treatment steps, (3) the evidence class of any recommendation and decision and a link to the source of the evidence, and (4) temporal dependencies and explicit time events [58].

The key features proposed in this study were derived from an analysis of existing literature selected based on quality and content regarding (1) BPM extensions in clinical domains affecting the representation of, among others, time, resources, and coordination of CPs by multidisciplinary teams working together in a care process [8,36-39,41,58-63]; (2) review studies of BPMN extensions both in clinical and nonclinical domains [13]; and (3) analysis of JHH CG features for infection management [15,43,44], such as parallel work, synchronization with EHR data, or uncertainty in the course of an infection treatment. Although broadly inclusive, our criteria were proposed as a set of minimum requirements to represent infection CPs. First, the well-structuredness of a process conceptual model helps increase its legibility and understanding by clinical staff, thereby enabling the detection and analysis of possible process improvements. In addition, it is a requirement to avoid deadlocks and inconsistencies in and between processes, thereby facilitating the future maintenance and evolution of the process representation. Process modularity allows for the decomposition of large, complex clinical processes with many involved parties in their constituent subprocesses, thereby allowing for increased readability, process reuse, and faster model building and verification of soundness across the organization. Events allow for conceptual modeling of facts that take place at the time of execution, the occurrence of which is possibly unknown at the time of design. Events deal with uncertainty and are always present in the treatment of infections, whose course depends on factors unknown at the time of design, such as the patients’ eventual response to medication. Parallel execution of tasks optimizes execution time to promptly react to clinical happenings as clinical institutions consist of complex multilayered organizational units, possibly externalized and interacting with each other. Task duration allows for the expression of a constraint, a deadline, or simply information to all process actors. For example, a blood test can take a minimum of 1 day and a maximum of 2 days to complete. If purely informational, workflow execution should not wait for this task to complete but, if meant as a constraint, the model should include semantics to hold the execution of subsequent tasks. The time distance between nonconsecutive tasks can be specified using relative time constraints such as edge duration between the starting or ending instant of a predecessor task A and the starting or ending instant of a successor task B. These constraints are determined by best practices or the institution’s average figures to define a minimum or maximum waiting time when dispatching work to other organizational units. Synchronization between subprocesses is ultimately reduced to the synchronization between their inner tasks. Use of resources is relevant in infection CPs as catheters are a main cause of infection in hospitalized patients, requiring both catheter lock therapy (CLT) and systemic therapy (ST) applied repeatedly and alternatively but not simultaneously through the existing catheter, which must therefore be used in exclusive mode by the executing task. Multiple instances of a task might be required to unfold repeatable tasks such as therapy or follow-up tasks, although, in many cases, their multiplicity level is unknown at the time of design and can only be determined at the time of execution (either by the user or by the process logic) as it depends on the clinical evolution of a patient or the adequacy of the prescribed treatment. Delays between iterations of looping activities, executed as long as a loop condition evaluates to true, allow for the expression of things such as “Vancomycin every 8 hours during 7 days” [7]. Furthermore, the integration of clinical data is useful for reviewing the availability of process-required data, their location, or the detection of inconsistencies or potential improvements. Data are used for input or output to make decisions, initiate a process, coordinate tasks, or inform process actors, whereas data stores allow for data persistence beyond the execution of a process model. In addition, in infection CPs, a close relationship between EHR information and model metadata is required to be able to dynamically adapt to clinical happenings. Finally, overrides at execution time, either by the user or the system’s logic, are required to deal with unexpected happenings during the course of an infection, which can unleash relevant changes in the care process.

The TP specification is natively oriented toward the conceptual modeling of clinical processes, thus theoretically providing out-of-the-box capabilities that are present in BPMN only after successive extensions. In this section, we analyze how TP supports the key features identified in the previous section. First, well-structuredness is a built-in feature in the TP standard semantics, which impose implicit restrictions on a process model. Process modularity is achieved in TP by the intrinsic grouping of reusable pieces of work in work plans, task plans, and, optionally, subplans for fine-grained work details. TP events regulate the temporal behavior of TP tasks or task groups by means of task-waits that prevent task execution until facts occur that satisfy a wait condition, such as the occurrence of a task transition or changes in global tracked variables. Parallel execution of tasks can be achieved in TP using the execution type attribute of the plan item top-level class. In addition, conditional decision structures exist to represent if-elseif-else, case, or event-based conditions, and the optional concurrency-mode attribute defines 4 possible states of a task group during and after the execution of its conditional paths. However, task duration, either as constraint or merely informative, does not exist in the current version of TP. When duration is meant as a constraint, TP implements this behavior through task-waits associated with events instead. Relative time constraints between tasks are managed in TP through task-waits for both deterministic and nondeterministic events combined with the use of dispatchable tasks. Resources are minimally represented in TP through Resource_Participation objects required for task execution. The allocation and tracking of workers to a task or task group is done at the time of execution as part of the materialized model (ie, the M* classes, which are only minimally specified in the current TP version). Multiple instances of a TP task can be dynamically generated at the time of execution if they are marked as repeatable in the process model by the optional repeat-spec attribute, which specifies the minimum and maximum number of iterations, and an optional terminate condition to exit the repeat loop. Each iteration is unrolled into literal sequential copies in the materialized image of the work plan. Delays between iterations of looping tasks or task groups are possible with the optional period attribute of a repeat-spec associated with a repeatable task or task group to be able to express things such as daily every 8 hours, acting as a spacer between the execution of the literal copies of each iteration. Furthermore, clinical data integration is indirectly achieved through the capture data set attribute, a data set template or form via which data can be introduced during task execution. As TP is part of the openEHR standard, data references are encapsulated using archetypes, so the template-id attribute specifies the archetype human-readable identifier (HRID) to be used, triggering the population of a data set [64]. Finally, overrides at execution time are built-in in TP, which by default assumes that, at the time of execution, users can have information that is unknown to the system for any number of reasons. The designed process model is performed in an advisory, adaptive way, allowing for logical deletion and addition of tasks at runtime and override of plan parameters such as task execution time or preconditions [64]. Table 1 shows the list of the identified key features and their coverage in extended BPMN and TP.

The results of our experiments show, on the one hand, the native suitability of TP for infection management and possibly for other complex clinical scenarios, confirming our initial hypothesis, and, on the other hand, the potential of the TP standard to simplify complex CPs by extending some of its native capabilities following the methodology for BPMN extensions laid out in the study by Braun and Schlieter [58], which we consider applicable to TP. The proposed TP extensions are mainly focused on increased synchronization between tasks, supporting Allen temporal relations, new dynamic behavior of repeat constructs, and enhancements in the resource perspective to emphasize the relevance of medical-associated devices, such as implanted catheters, in the course of infectious diseases. When applicable, we also proposed the use of complementary specifications such as DL.

In pattern 1, we saw how the rich set of nondeterministic TP events is better suited to represent uncertainty about task duration than the deterministic events used in the corresponding BPMN model, allowing for a dynamic response to unknown clinical happenings and, as a result, faster process execution. Nevertheless, we proposed to enhance the TP abstract syntax with new task duration attributes to allow for fine-grained fixed temporal constraints that could be of use in specific scenarios.

In the case of pattern 2, we initially modeled it in the depicted pattern 2 model as a cognitive decision-making task expressed as a multiple decision tree type of task, as done in BPMN. However, we proposed an alternative pattern 2* TP model to encapsulate the complexity of the cognitive decision-making task in a single rule instead. To this end, the TP standard is complemented by the DL specification, allowing for the evocation of DL rules from within a task or task group, which is not possible using BPM alone, resulting in simpler workflows and improved maintenance and evolution of task-related knowledge.

Pattern 3 highlights the need for a more advanced adaptive behavior of the TP execution engine to address new facts or clinical happenings that cannot be foreseen in advance in static BPM models, allowing for dynamic changes in the process logic at the time of execution. The pattern 3 TP model mimics the corresponding BPMN model, with 2 distinct cyclic therapy tasks. However, in the alternative pattern 3* model, we proposed extending TP to enhance the override capabilities of the repeat loop, which is widely used in infection CPs to represent, for example, cyclic medication tasks. Dynamic overrides were not available in BPMN, resulting in an unnecessary increase in the workflow complexity.

Finally, pattern 4 shows the synchronization between 2 tasks that cannot be executed simultaneously through the same device; in our case, a catheter. This pattern could not be represented in BPMN owing to its complexity. We could represent it easily in the pattern 4 TP depicted model using out-of-the-box TP constructs by modeling 2 instrumental tasks for synchronization purposes complemented with the rich set of nondeterministic TP events. However, in the alternative pattern 4* model, we showed how enhancing the TP standard with a resource extension could greatly simplify the pattern 4 representation by adding new attributes to the Resource_Participation class and proposing both a new nondeterministic resource_transition event and a new state machine in the TP engine. The proposed TP extension could be generally used to represent medically relevant devices such as catheters, mechanical ventilators, or any other disease-associated devices, thus improving the TP native resource perspective.

Furthermore, we theoretically analyzed other possible extensions that could be of use in infection management and other clinical scenarios, such as adding complex event processing capabilities to the standard, an enhanced task-wait condition logical expression using the newly proposed event-list-relation and offset attributes, and a new Resume_type in a task synchronous dispatch. Finally, we examined the TP temporal restrictions between tasks from a theoretical point of view, analyzing how the TP standard addresses the 13 Allen temporal relations and proposing extensions to support 3 of the 6 symmetric Allen relations plus the equals temporal operator. However, the proposed temporal extensions must be weighed against their implementation costs as they require an enhanced task-wait logic for a task to transition not only to the available state but also to the completed state, as well as new time reference attributes to be able to express more sophisticated time constraints between tasks.

The proposed extensions provide a more consistent approach to dealing with uncertainty, thereby improving overall workflow efficiency and duration by adding dynamic response capabilities during workflow execution. They also acknowledge the importance of clinically relevant resources in the course of infection treatment, improving their visibility and clinical role.

A possible limitation is that many of the proposed extensions require changes in the abstract syntax and the behavior of the TP engine; for example, in pattern 3* to dynamically monitor and reset the repeat loop count. These changes in the standard should be weighed against the expected benefits. In addition, the use of rules within CPs to represent knowledge-intensive tasks requires further definition of the DL specification, which is still under development. Furthermore, the use of resources in CPs should be synchronized with the resource information defined in the EHR through specific mechanisms. Finally, the overall suitability of the standard for other complex clinical domains should be further assessed in future studies addressing the specific domain constraints.

Future work will focus on analyzing the DL specification [50] in relation to TP, a new openEHR formalism for evoking rules from TP expressed using the openEHR Expression Language [68] and based on the openEHR Basic Meta Model [51]. We will also analyze the new OMG BPM+ Health initiative aiming to promote the use of the 3 OMG standards—BPMN, Case Management Model and Notation, and Decision Model and Notation—in health care domains [69].