32nd Annual International Conference of the IEEE EMBS Buenos Aires, Argentina, August 31 - September 4, 2010
Activity-Based Process Mining for Clinical Pathways Computer Aided Design Carlos Fern´andez-Llatas, Teresa Meneu, Jose Miguel Bened´ı and Vicente Traver Abstract— Current trends in health management improvement demand the standardization of care protocols to achieve better quality and efficiency. The use of Clinical Pathways is an emerging solution for that problem. However, current Clinical Pathways are big manuals written in natural language and highly affected by human subjectivity. These problems make the deployment and dissemination of them extremely difficult in real practice environments. In this work, a complete computer based architecture to help the representation and execution of Clinical Pathways is suggested. Furthermore, the difficulties inherent to the design of formal Clinical Pathways in this way requires new specific design tools to help making the system useful. Process Mining techniques can help to automatically infer processes definition from execution samples. Yet, the classical Process Mining paradigm is not totally compatible with the Clinical Pathways paradigm. In this paper, a pattern recognition algorithm based in an evolution of the Process Mining classical paradigm is presented and evaluated as a solution to this situation. The proposed algorithm is able to infer Clinical Pathways from execution logs to support the design of Clinical Pathways.
Nevertheless, the hand-made design of Workflows that precisely describes the actions involved in a care process is a hard task. In that way, an iterative design process can be the solution for Clinical Pathways deployment. But, in that case, it is required to include the use of tools that allow analyzing how the process is working in order to evaluate the status of the deployment in each iteration. Process Mining techniques can offer a non-subjective view about the implementation of those standardized processes. Process Mining algorithms allow experts automatically infer Workflows explaining real action flows using the activity logs. Unfortunately, traditional approaches of Process Mining are no fully compatible with the Clinical Pathways inferring problem. In this paper, based in a closed-loop methodology for designing Clinical Pathways using Workflow technology, a better paradigm for Process Mining devoted to the Clinical Pathways problem and a Process Mining algorithm dealing with this is presented and evaluated.
I. INTRODUCTION The design of Clinical Pathways represents a challenge in the scheduling of medical care, specially for complex healthcare processes. The description of the flows of action implicated in the care processes with the required level of quality and in an efficient way its not an easy task. The use of natural language to describe Clinical Pathways results in ambiguous process descriptions and usually its automation is severely compromised. In this line in order to define these processes with an acceptable quality, the use of formal models is required to avoid the ambiguity and ensure the automation allowing professionals to describe, share and use Clinical Pathways. Workflow technology is a useful approach to describe those processes [7]. The use of formally defined Workflows allows the creation of non-ambiguous languages to be executed automatically in computerized systems. In addition, this framework has a battery of representation, simulation, execution and data mining tools available. Those tools can be used to empower health professionals to create high quality Clinical Pathways that facilitate the understanding, deployment, evaluation and improvement of those complex processes. This work was supported by TSB group at ITACA institute at Universidad Polit´ecnica de Valencia Carlos Fern´andez-Llatas, Teresa Meneu and Vicente Traver is with TSB group at ITACA institute, Universidad Polit´ecnica de Valencia, Spain
{cfllatas, tmeneu, vtraver}@itaca.upv.es Jos´e Miguel Bened´ı is with ITI institute, Universidad Polit´ecnica de Valencia, Spain
[email protected]
978-1-4244-4124-2/10/$25.00 ©2010 IEEE
II. CLINICAL PATHWAYS FOR HEALTH CARE STANDARDIZATION Currently, the standardization is considered one of the best practices to improve the way in which the processes are executed. Clinical Pathways are multidisciplinary standardized actuation protocols based on medical knowledge where the clinical activities, such as the medical support and management are sequentially detailed. Clinical Pathways are thought to improve the quality of patient care and make a more efficient usage of health resources [8]. In literature there are lots of documentation repositories where lots of Clinical Pathways are shown like PubMed [13] or the Cochrane [3] Clinical Pathways are expected to complement the practitioners expertise. This approach has some advantages which can improve the quality and efficiency of medicine science. Those standards can be used as a guideline to facilitate the praxis for health professionals. This can help not only health experts but also for general medicine professionals and students. In addition, Clinical Pathways force experts to unify criteria. This will allow avoiding the variability of daily clinical practice and improving the quality of service. Other advantage of those standards is their capability to enable the administrative management of care. This will allow to better foresee and account the costs of treating each specific patient, enabling the improvement of the management of health units Nevertheless, the deployment of those standards has some difficulties in real scenarios. Clinical Pathways are defined by using big documents created in complex iterative processes
6178
by consensus of expert groups. Those documents have ambiguities and are too complex to be used in a practical way. In addition to this, those documents are affected by subjectivity of the creation members and must be adapted to each health unit before been deployed. Also, the coordination of Clinical Pathways stakeholders is critical to make an efficient use of resources. This requires an agile communication system among health actors to be useful. In this way, to allow the implantation of Clinical Pathways in real environments it is necessary to use a framework that supports the agile description and deployment of the health processes. III. WORKFLOW TECHNOLOGY On the other side, the use of ICT (Information and Communication Technologies) is increasingly growing in health centres and these technologies are key for the automating of the process execution. In that way, the formal definition of processes enables its automatic execution in computer-based system. This paradigm is covered by Workflow technology. The use of those technologies can be a solution for the coordination among the stakeholders of Clinical Pathways. Workflow Technology is a research field focused on the creation of process specification languages and its dynamic execution. A Workflow is defined as the automation of a business process, in whole or part, during which documents, information or tasks are passed from one participant to another for action, according to a set of procedural rules [20]. A Workflow is a formal description of a process designed to be automated. A Workflow language is a formalism that allows the definition of Workflows. A formal Workflow can be automatically executed in a computerized system by using a Workflow Engine. This will allow the automatic execution of standardized protocols and the guidance of processes in computerized environments. Clinical Pathways are one of the most demanding fields in terms of process complexity. For that, the Workflow language selected is crucial for the successful deployment of a Workflow system. Clinical Pathways represent very complex processes and the language must be enough expressive to represent them. In addition, Clinical Pathways must be understood by non-computer experts humans, and for that purpose they need a high understandability. In previous works the authors proposed a Workflow language that is able to represent and execute Clinical Pathways called Life Activity Protocols (LAPs) [7]. This formalism has a very high expressivity and was be used by non-computer experts to define complex processes in PIPS project [2]. Despite the use of expressive and understandable languages, the design problem is not solved. The design of formal Clinical Pathways requires a big quantity of resources to be specified. There is a great difficulty to manually design those guidelines for all specific patients because their high variability (i.e. pluripatological patients). In addition to that, the subjective load and design errors on Clinical Pathways suppose critical implantation problems that can make useless the system. To debug those systems, an iterative design is
required. But, to debug the system it is needed to gather and process all the information about the execution exceptions of the standardized process. This work is normally unaffordable manually because of the large quantity of resources required. In that way, the use of tools to help this iterative design of Clinical Pathways is mandatory if a successful design and deployment is desired. A. Process Mining When a Clinical Pathway is firstly deployed in a real environment a large account of exceptions can be expected. Most patients do not follow the pattern described by the flow due to its specific characteristics. This fact would cause modifications on the Clinical Pathway instance for each specific patient. If these modifications are stored and further analyzed, it will be possible to create new pathways that incorporate those exceptions or that correct the errors made in previous design iterations. This technique, that is normally not possibly performed manually, can be approached by using Process Mining (also Workflow Mining) techniques. The Process Mining [18] idea is based in the automatic learning of Workflows for business process inference. Process Mining algorithms use the execution samples to infer the Workflow that describe the real process. Using this technique the logs of care protocols actions applied to each patient can be used to learn Workflows that formally represent those Clinical Pathways. This helps Clinical Pathways designers to modify previous iteration processes according to the real implantation incorporating the new exceptions and correcting design errors. Most of the works in literature are based on the use of transactional logs as samples for Workflow inferences [4] [15] [16]. These models use logs from general systems of Workflow management as input samples to infer models that explain the whole system. These models are usually based on the Event-based approach. The Event-based Process Mining approach [16] learns Workflows using as input samples of the events occurred on a determinate process. This information only includes the information of actions (the action name and the starting time) forgetting the action results. Clinical Pathways usually base their decisions on the results of activities (also indicators). For instance, the action Take Temperature can generate the indicators Fever or Not Fever. In this way, depending on the indicator of this action the flow that the patient follows in the Workflow will be different. Thus, on one hand if the result of the action was Fever, the next action might be Take Antipyretic. On the other hand, if the result of the action was Not Fever, the next action might be Do Nothing. Unfortunately, the Event-Based Process Mining approach cant make that distinction. In this framework, a new paradigm and algorithms to incorporate indicators to Process Mining technology is required. IV. ACTIVITY-BASED PROCESS MINING FOR CLINICAL PATHWAYS AUTOMATIC LEARNING As it was shown previously, there are many problems that cant be solved using the Event-based approach. In complex
6179
cases, the information available in Event-Based learning corpora is not enough to infer the flow transitions rules based on action results, which is the case of Clinical Pathways. In order to solve that, the authors propose Activity-Based Process Mining to approach this problem. Activity-based process mining proposes to enrich the Event-based training logs with information about the indicators. Using this approach it will be possible to learn the flows of the processes in accordance to their actions’ results and, by these means, improve dramatically the Clinical Pathways design through the usage of Process Mining technologies. This approach has been used with success in other research fields as in human behaviour modelling [10]. In this section, an Activity-Based Process Mining algorithm (PALIA) is used evaluate the accuracy of those models using a controlled Clinical Pathway inference experiment. A. PALIA Algorithm The Parallel Activity-based Log Inference Algorithm(PALIA) [9] is an Activity-Based Process mining algorithm. This algorithm require samples with the actions, the starting and the finish time of those actions as well as the actions’ indicators. PALIA Algorithm infers TPAs (Timed Parallel Automaton) [9] that is the mathematical basis of LAP. PALIA algorithm is divided into five stages. • The Parallel Acceptor Tree Algorithm stage builds a graph tree with the samples, taking into account the beginning and the end of the processes and their parallelism. At the end of the algorithm, a basic TPA that only accepts the entry sample is built. • In the Onward Merge stage, all the branches that are equivalent are fused. For each node, the algorithm verifies if their posterior branches are equivalent; if so, the algorithm fuses the nodes and transitions. Two branches are equivalent if all the nodes and transitions use the same tokens for the same processes. • The Parallel Merge algorithm stage merges the nodes that are sequentially together and can be fused. Two nodes can be fused if they represent the same event. The merging process fuses the nodes and corrects the arcs. At the end of the algorithm, a corrected TPA is returned. • In the last two stages of the PALIA Algorithm, Delete Repeated Transitions and Delete Unused Nodes, repeated transitions and unused nodes are deleted. The result of this algorithm is a TPA that models the system. The PALIA algorithm uses syntactical pattern recognition techniques to learn TPAs. The stages of the algorithm are inspired in the OSTIA algorithm [14] that is used to infer transducers. In our case, the PALIA algorithm allows to infer parallel structures representing Workflows to solve ActivityBased Process Mining problems. B. Experimental results To test PALIA algorithm an Activity-Based Process Mining was designed. For this experiment a free simplifica-
tion of a Heart Failure Clinical Pathway from FISTERRA database [12] was used. This simplification includes a set of complex Workflow Patterns that are used in Clinical Pathways specifically prepared to test the accuracy of process mining algorithms. Workflow Patterns [17] are typical Workflow structures that define the flow rules in a process. Fig. 1 represent the Workflow of the experiment. This experiment shows a treatment for a cardiac illness depending of its seriousness. In the example, there are three paths possible: a simple sequence, a complex sequence and a very complex parallel sequence with protected sections that can be atomically executed. This Clinical Pathway was simulated to gather 510 samples that represent execution instances. Using these samples the Process Mining algorithm infers a Workflow that is compared with the original one in order to evaluate the accuracy of the algorithm. Classical Pattern recognition techniques cant be directly applied to Process Mining evaluation because most of the algorithms are not thought to accept or classify samples. Those algorithms are expected to identify flow patterns. To do so, the accuracy of the algorithm is evaluated using two indicators, the Efficacy (Eff) and the Relative Understandability coefficient (CfRU ) [11]. The Efficacy shows the capability of the algorithm to identify Workflow Patterns. This coefficient is the ratio between the number of patterns correctly identified and the total of patterns existing in the original Workflow: Ef f =
P atternsIdentif ied P atternsT otal
The Relative Understandability coefficient (CfRU ) shows the overhead of structures used to identify the patterns. The lower is the number, the less structures are used to represent each identified pattern. This offers a measure of the easiness of visual understanding of the Workflow. This coefficient is the ratio between number of Nodes by Arc of inferred Workflow needed divided by identified patterns and the number of Nodes By Arc of original Workflow divided by original patterns: CfRU =
NInf erred XAInf erred P atternsIdentif ied NOriginal XAOriginal P atternsOriginal
The CfRU of the original Workflow is 1. The more near to 1 this value is the more efficient the algorithm is. If the value is lower that 1, usually is because the algorithm has sacrificed some patterns to better define others. In addition to evaluate the inference capability of ActivityBased systems, PALIA was be compared with existing Process Mining Algorithms to evaluate PALIA flow identification accuracy. Those Process Mining algorithms are Heuristic Miner(HM) [19], Genetic Process Miner(GPM) [6], α [16] and α++ [5]. Unfortunately, all algorithms selected were Event-Based because there are no other Activity-Based algorithms available in literature.
6180
Original PALIA Heuristic Miner Genetic Process Mining α
α++
S/F S F S/F S F S/F S F S/F S F
N 15 15 30 16 14 30 16 14 30 16 14 16 14
A 25 35 59 29 27 54 29 26 49 5 3 9 7
P atterns 15 14 11 10 9 10 10 8 13 2 1 8 7
Ef f 0.93 0.73 0.67 0.60 0.67 0.67 0.53 0.87 0.13 0.07 0.53 0.47
CfRU 1.50 6.44 1.86 1.68 6.48 1.86 1.82 4.52 1.60 1.68 0.72 0.56
approach. Based in that approach, PALIA has been tested in a controlled experiment of Clinical Pathways inference problems. PALIA has achieved better results inferring the flow of actions than classical Process Mining algorithms, plus identifying the transitions causes. Currently, the authors are involved in a European Project called Heart Cycle [1] that is incorporating those technology in real scenarios to help the design of cardiac Clinical Pathways. R EFERENCES
TABLE I E XPERIMENT R ESULTS
Table I shows the result of the experiments 1 2 3 . PALIA is the algorithm with the best efficiency keeping a very good Relative Understandability. Note that PALIA is the only one of those algorithms able to infer Activity-Based corpora. This supposes a handicap for PALIA, because, for PALIA it is required the correct inference of not only the pattern flow, but also the indicator that fires that transition. For the Event-Based algorithms it is only required the correct inference of the flow because they do not have indicator inference capabilities.
Fig. 1.
Workflow of the experiment
V. CONCLUSIONS AND FUTURE WORKS The use of this Workflow approach to the Clinical Pathways problem will ensure a better standardization without incomplete and ambiguously defined processes. In addition, Workflow-Based Clinical Pathways can be automated and simulated due to their formal definition. Activity-based Process mining approach can help Clinical Pathways design better than event-based traditional 1 S/F:
Inference performed using Start and Finish events Inference performed using Start events 3 F: Inference performed using Finish events 2 S:
[1] VII Framework Program IST Project 216695. Heart cycle project:compliance and effectiveness in hf and chd closed-loop management, 2008-2011. [2] VI Framework Program IST Project 507019. Pips project. personalised information platform for life and health services. [3] The Cochrane Collaboration. Cochrane library: http://www.cochrane.org/index.htm. [4] Jonathan Cook and Zhidian Du. Discovery thread interactions in a concurrent system. Journal of Systems and Software, 7:285–297, 2005. [5] A. K. Alves de Medeiros, B. F. Dongen, W. M. P. van der Aalst, and A. J. M. M. Weijters. Process mining extending the alpha algorithm to mine short loops. Technical report, WP113 Beta Paper Series Eindhoven University of Technology, 2004. [6] Ana Karla A. de Medeiros, A. J. M. M. Weijters, and W.M.P. van der Aalst. Genetic process mining: A basic approach and its challenges. In Business Process Management Workshops, pages 203–215, 2005. [7] David Dominguez, Carlos Fernandez, Teresa Meneu, Juan Bautista Mocholi, and Riccardo Seraffin. Medical guidelines for the patient: Introducing the life assistance protocols. In Computer-based Medical Guidelines and Protocols: A Primer and Current Trends, volume 139, pages 282–293. IOS Press, 2008. [8] N.R. Every, J. Hochman, R. Becker, S. Kopecky, and C.P. Cannon. Critical pathways. a review. Circulation, 101:461–465, 2000. [9] Carlos Fernandez and Jose Miguel Bened. Timed parallel automaton learning in workflow mining problems. In Ciencia y Tecnologa en la Frontera, number special 2008, pages 181–187, 2008. [10] Carlos Fern´andez, Juan Pablo L´azaro, and Jose Miguel Bened´ı. Workflow mining application to ambient intelligence behavior modeling. In Universal Access in Human-Computer Interaction, volume 5615 of Lecture Notes in Computer Science, pages 160–167. Springer, 2009. [11] Carlos Fernandez-Llatas. Representacion, Interpretacion y Aprendizaje de flujos de trabajo orientados a actividades para la estandarizacin de Vias Clnicas. PhD thesis, Universidad Policnica de Valencia, 2009. [12] Fisterra. Biblioteca de vias clnicas fisterra: http://www.fisterra.com/fisterrae/. [13] National Library of Medicine and The National Institutes of Health. Pubmed library: http://www.pubmed.gov. [14] Jose Oncina, Pedro Garcia, and Enrique Vidal. Learning subsequential transducers for pattern recognition interpretation tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 448–458, 1993. [15] Guido Schimm. Process miner - a tool for mining process schemes from event-based data. In JELIA ’02: Proceedings of the European Conference on Logics in Artificial Intelligence, pages 525–528, London, UK, 2002. Springer-Verlag. [16] W. van der Aalst, A. Weijters, and L. Maruster. Workflow mining: Discovering process models from event logs. IEEE Transactions on Knowledge and Data Engineering, 16:1128 – 1142, 2004. [17] Wil M. P. van der Aalst, Alistair P. Barros, Arthur H. M. ter Hofstede, and Bartek Kiepuszewski. Workflow patterns. Distributed and Parallel Databases, page 70, 2003. [18] Wil M. P. van der Aalst, B. F. van Dongen, J. Herbst, L. Maruster, G. Schimm, and A. J. M. M. Weijters. Workflow mining: A survey of issues and aproaches. Data and Knowledge Engineering 47 2003, pages 237–267, 2003. [19] A. J. M. M. Weijters, W. M. P. van der Aalst, and A. K. Alves de Medeiros. Process mining with the heuristics miner algorithm. Technical report, WP166 Beta Paper Series Eindhoven University of Technology, 2006. [20] WfMC. Workflow Management Coalition Terminology Glossary. WFMC-TC-1011, Document Status Issue 3.0, 1999.
6181
