Many information systems nowadays record data about the process instances executed at the organization in the form of traces in an event log. In this paper we present a framework able to convert actions found in the traces into higher level concepts, on the basis of domain knowledge. Abstracted traces are then provided as an input to semantic process mining. The approach has been tested in the medical domain of stroke care, where we show how the abstraction mechanism allows the user to mine process models that are easier to interpret, since unnecessary details are hidden, but key behaviors are clearly visible.
Most commercial information systems, including those adopted by many health care organizations, record information about the executed process instances in the form of an event log [15]. The event log stores the sequences (traces [9] henceforth) of actions that have been executed at the organization, typically together with key execution parameters, such as times, cost and resources. Event logs can be provided in input to process mining [15, 10] algorithms, a family of a-posteriori analysis techniques able to extract non-trivial knowledge from these historic data; within process mining, process model discovery algorithms, in particular, take as input the log traces and build a process model, focusing on its control ow constructs. Classical process mining algorithms, however, provide a purely syntactical analysis, where actions in the traces are processed only referring to their names. Action names are strings without any semantics, so that identical actions, labeled by synonyms, will be considered as di erent, or actions that are special cases of other actions will be processed as unrelated.
On the other hand, the capability of relating semantic structures such as ontologies to actions in the log can enable trace comparison and process mining techniques to work at di erent levels of abstraction (i.e., at the level of instances and/or concepts) and, therefore, to mask irrelevant details, to promote reuse, and, in general, to make process analysis much more exible and reliable.
In fact, it has been observed that human readers are limited in their cognitive
capabilities to make sense of large and complex process models [
Interestingly, semantic process mining, de ned as the integration of semantic processing capabilities into classical process mining techniques, has been recently proposed in the literature (see Section 5). However, while more work has been done in the eld of semantic conformance checking (another branch of process mining) [8, 11], to the best of our knowledge semantic process model discovery needs to be further investigated.
In this paper, we present a knowledge-based abstraction mechanism (see Section 2), able to operate on event log traces. In our approach: { actions in the log are mapped to the ground terms of an ontology; { a rule base is exploited, in order to identify which of the multiple ancestors of an action should be considered for abstracting the action itself. Medical knowledge and contextual information are resorted to in this step; { when a set of consecutive actions on the trace abstract as the same ancestor, they are merged into the same abstracted macro-action, labeled as the common ancestor at hand. This step requires a proper treatment of delays and/or actions in-between that descend from a di erent ancestor.
Our abstraction mechanism is then provided as an input to semantic
process mining (see Section 3). In particular, we rely on classical process model
discovery algorithms embedded in the open source framework ProM [
We also describe our experimental work (see Section 4) in the eld of stroke care, where the application of the abstraction mechanism on log traces has allowed us to mine simpler and more understandable process models. 2
In our framework, trace abstraction has been realized as a multi-step mechanism. The following subsections describe the various steps. 2.1
As a rst step, every action in the trace to be abstracted is mapped to a ground term of an ontology, formalized resorting to domain knowledge.
In our current implementation, we have de ned an ontology related to the eld of stroke management, where ground terms are patient management actions, while abstracted terms represent medical goals. Figure 1 shows an excerpt of the stroke domain ontology, formalized resorting to the Protege editor. Legend • CAT is Computer Assisted Tomography • MRI is Magnetic Resonance Imaging • MRI_with_CE is Contrast Enhanced Magnetic Resonance Imaging • MRI_with_DWI is Diffusion-weighted Magnetic Resonance Imaging • TPA is Tissue Plasminogen Activator
In particular, a set of classes, representing the main goals in stroke management, have been identi ed, namely: \Administrative Actions", \Brain Damage Reduction",\Causes Identi cation", \Pathogenetic Mechanism Identi cation", \Prevention", and \Other". These main goals can be further specialized into subclasses, according to more speci c goals (e.g., \Parenchima Examination" is a subgoal of \Pathogenetic Mechanism Identi cation", while \Early Relapse Prevention" is a subgoal of \Prevention"), down to the ground actions, that will implement the goal itself.
Some actions in the ontology can be performed to implement di erent goals. For instance, a Computer Assisted Tomography (CAT) can be used to check therapy e cacy in \Early Relapse Prevention", or to perform \Parenchima Examination" (see gure 1).
The proper goal to be used in the abstraction phase will be selected on the basis of the context of execution, as formalized in the rule base, described in the following subsection. 2.2
As a second step in the trace abstraction mechanism, a rule base is exploited to identify which of the multiple ancestors of an action in the ontology should be considered for abstracting the action itself. The rule base encodes medical knowledge. Contextual information (i.e., the actions that have been already executed on the patient at hand, and/or her/his speci c clinical conditions) is used to activate the correct rules. The rule base has been formalized in Drools [17].
As an example, referring to the CAT action mentioned in the previous subsection, the following rule states that, if intra-venous (ev tPA) or intra-arterial (ia tPA) anti-thrombotic therapies have been administered, then CAT implements the \Early Relapse Prevention" goal. rule "CAT" when (groundActionIsBefore("ev_tPA") || groundActionIsBefore("ia_tPA")) then end
macroAction.setAncestorName("Early_Relapse_Prevention"); where \groundActionIsBefore" is a function that, given the name of a ground action, returns true if this action precedes CAT in the trace, false otherwise.
On the contrary, if the context is di erent (i.e., anti-thrombotic therapy was not administered), CAT has to be intended as a means for \Parenchima Examination" (see gure 1).
More complex situations, where it is necessary to activate a chain of multiple rules - not described here due to space constraints - can also be managed by our system. 2.3
Once the correct ancestor of every action has been identi ed, trace abstraction can be completed.
In this last step, when a set of consecutive actions on the trace abstract as the same ancestor, they have to be merged into the same abstracted macro-action, labeled as the common ancestor at hand. This procedure requires a proper treatment of delays, and of actions in-between that descend from a di erent ancestor (interleaved actions henceforth).
Trace abstraction has been realized by means of the procedure described in Algorithm 1 below.
The function abstraction takes in input an event log trace, the domain ontology onto, and the level in the ontology chosen for the abstraction (e.g., level = 1 corresponds to the choice of abstracting the actions up to the sons of the ontology root). It also takes in input three thresholds (delay th, n inter th and inter th). These threshold values have to be set by the domain expert in order to limit the total admissible delay time within a macro-action, the total number of interleaved actions, and the total duration of interleaved actions, respectively. In fact, it would be hard to justify that two ground actions share the same goal (and can thus be abstracted to the same macro-action), if they are separated by very long delays, or if they are interleaved by many/long di erent ground actions, meant to ful ll di erent goals.
The function outputs an abstracted trace.
For every action i in trace, an iteration is executed (lines 3-27). First, a macro-action mi, initially containing just i, and sharing its starting and ending times, is created. mi is labeled referring to the ancestor of i (the one identi ed by the rule based reasoning procedure) at the abstraction level provided as an input. Accumulators for this macro-action (total-delay, num-inter and totalinter, commented below) are initialized to 0 (lines 4-10). Then, a nested cycle is executed (lines 11-25): it considers every element j following i in the trace, where a trace element can be an action, or a delay between a pair of consecutive actions. Di erent scenarios can occur: { if j is a delay, total delay is updated by summing the length of j (lines 12-14). { if j is an action, and j shares the same ancestor of i at the input abstraction level, then j is incorporated into the macro-action mi. This operation is always performed, provided that total delay, number inter and total inter do not exceed the threshold passed as an input (lines 15-19). j is then removed from the actions in trace that could start a new macro-action, since it has already been incorporated into an existing one (line 18). This kind of situation is described in Figure 2 (a). { if j is an action, but does not share the same ancestor of i, then it is treated as an interleaved action. In this case, num inter is increased by 1, and total inter is updated by summing the length of j (lines 20-23). This situation, in the end, may generate di erent types of temporal constraints end end num inter = num inter + 1; total inter = total inter + j:length; 1 abs trace = abs algorithm(trace; onto; level; delay th; n inter th; inter th); 2 abs trace = ;; 3 for every i 2 activities in trace do 4 if (i:startF lag = yes) then 5 create : mi as ancestor(i; level); 6 mi:start = i:start; 7 mi:end = i:end; 8 total delay = 0; 9 num inter = 0; 10 total inter = 0; 11 for (every j 2 elements in trace) do 12 if (j is a delay) then 13 total delay = total delay + j:length; 14 else 15 16 if (ancestor(j; level)=ancestor(i; level)) then if (total delay < delay th ^ num inter < n inter th ^ total inter < inter th) then mi:end = max(mi:end; j:end); j.startFlag = no; 17 18 19 20 21 22 23 24 end 25 26 27 end 28 return abs trace;
end
append mi to abs trace;
Algorithm 1: Multi-level abstraction algorithm
between macro-actions, as the ones described in Figure 2 (b) (Allen's during
[
Finally, the macro-action mi is appended to abs trace, that, in the end, will contain the list of all the macro-actions that have been created by the procedure (line 26).
Complexity. The cost of abstracting a trace is O(actions elements), where actions is the number of actions in the input trace, and elements is the number of elements (i.e., actions + delay intervals) in the input trace. 3
In our approach, process mining, made semantic by the exploitation of the
abstraction mechanism illustrated above, is implemented resorting to the
wellknown process mining tool ProM, extensively described in [
For the work described in this paper, we have exploited ProM's Heuristic
Miner [
Currently, we have chosen to rely on Heuristics Miner, because it is known to be tolerant to noise, a problem that may a ect medical event logs (e.g., sometimes the logging may be incomplete). Anyway, testing of other mining algorithms available in ProM 6 is foreseen in our future work. 4
In this section, we describe the experimental work we have conducted, in the application domain of stroke care.
The available event log is composed of more than 15000 traces, collected at the 40 Stroke Unit Network (SUN) collaborating centers of the Lombardia region, Italy. Traces are composed of 13 actions on average. The 40 Stroke Units (SUs) are not all equipped with the same human and instrumental resources: in particular, according to resource availability, they can be divided into 3 classes. Class-3 SUs are top class centers, able to deal with particularly complex stroke cases; class-1 SUs, on the contrary, are the more generalist centers, where only standard cases can be managed.
We have tested whether our capability to abstract the event log traces on the basis of their semantic goals allowed to obtained process models where unnecessary details are hidden, but key behaviors are clear. Indeed, if this hypothesis holds, in our application domain it becomes easier to compare process models of di erent SUs, highlighting the presence/absence of common paths, regardless of minor action changes (e.g., di erent ground actions that share the same goal) or irrelevant di erent action ordering or interleaving (e.g., sets of ground actions, all sharing a common goal, that could be executed in any order).
Figure 3 compares the process models of two di erent SUs (SU-A and SUB), mined by resorting to Heuristic Miner, operating on ground traces. Figure 4, on the other hand, compares the process models of the same SUs as gure 3, again mined by resorting to Heuristic Miner, but operating on traces abstracted according to the goals of the ontology in gure 1. In particular, abstraction was conducted up to level 2 in the ontology (where level 0 is the root, i.e.. \Goal").
Generally speaking, a visual inspection of the two graphs in gure 3 is very di cult. Indeed, these two ground processes are \spaghetti-like" [9], and the extremely large number of nodes and edges makes it hard to identify commonalities in the two models.
The abstract models in gure 4, on the other hand, are much more compact, and it is possible for a medical expert to analyze them.
In particular, the two graphs in gure 4 are not identical, but in both of them it is easy to a identify the macro-actions which corresponds to the treatment of a typical stroke patient.
However, the model for SU-A at the top of gure 4 exhibits a more complex control ow (with the presence of loops), and shows three additional macroactions with respect to the model of SU-B, namely \Extracranial Vessel Inspecem_neuro_TAC_senza_mdc
Start em_Terapia_anticoagulante_orale em_ecgperformed ric_test ranscranicdop ler valutazione_neurologica
Antidiabetici_se_diabete
Adeguamento_dietetico ric_testcoagulscre ning
Cons_Ematologo ric_rehabtherapy_FKT_entro_48_ore
Terapia_anticomiziale_se_crisi_epilet ica ric_consultvascular em_Terapia_eparinica_ev_sc ric_clinantiag regating em_neuro_AngioRMN_vasi_intra_ed_extracra ric_testencephalicnmrdwi Cons_Cardiologo_se_FA_IMA_o_aritmia_ventricolare
Terapia_anti pertensiva_se_crisi_ipertensiva ric_compltherantihypert ric_compltherinsulin r_Stop_fumo_se_fumatori ric_theradvicesexercise em_hematocperformed em_Antiag reganti_se_NO_TPA em_dop tsaperformed em_dop tcranperformed Cons_Fisiatra_precedente_a_FKT TPA_ev ric_testdop lersovratrunk ric_testecg ric_test ransesophecg ric_testencephalic atcont TrombosiSeniVenosi em_neuro_AngioTAC_vasi_intra_ed_extracra dimis ione End SU-B ric_test ranschestecg ric_testvenecocolorlower ric_clinanticoagulating ric_theradvicesantihypert Terapia_O2_con_ipos ia Cons_Rianimatore_se_ipert_endocr_o_aritmia_ventr
ric_testencephalicnmrcont
Start ric_testecg Terapia_anticomiziale_se_crisi_epilet ica tion", \Intracranial Vessel Inspection" and \Recanalization". This nding can be explained, since SU-A is a class-2 SU, where di erent kinds of patients, including some atypical/more critical ones, can be managed, thanks to the availability of di erent skills and instrumental resources. These patients may require the additional macro-actions reported in the model, and/or the repetition of some procedures, in order to better characterize and manage the patient's situation.
On the other hand, SU-B is a class-1 SU, i.e., a more generalist one, where very speci c human knowledge or technical resources are missing. As a consequence, the overall model control ow is simpler, and some activities are not executed at all.
Interestingly, our abstraction mechanism, while hiding irrelevant details, allows to still appreciate these di erences. 5
The use of semantics in business process management, with the aim of operating at di erent levels of abstractions in process discovery and/or analysis, is a relatively young area of research, where much is still unexplored.
One of the rst contributions in this eld was proposed in [
Hepp et al. [13] propose a framework able to merge semantic web,
semantic web services, and business process management techniques to build semantic
business process management, and use ontologies to provide machine-processable
semantics in business processes [14]. The work in [
The topic was studied in the SUPER project [
Ontologies, references from elements in logs to concepts in ontologies, and
ontology reasoners (able to derive, e.g., concept equivalence), are described as
the three essential building blocks for semantic process mining in [8]. This paper
also shows how to use these building blocks to extend ProM's LTL Checker [
The work in [
Start Causes_Identification
SU-B CardioEmbolic_Mechanism
Other Early_Relapse_Prevention
Coagulation_Screening
NeuroProtection In-Hospital_Disability_Reduction
Parenchima_Examination
Long_Term_Relapse_Prevention Administrative_Actions
End corrective actions. Detection, diagnosis and resolution present interesting challenges that, on the authors' opinion, can strongly bene t from knowledge-based techniques.
In [
In [7] an example of process discovery at di erent levels of abstractions is presented. It is however a very simple example, where a couple of ground actions are abstracted according to their common ancestor. However, the management of interleaved actions or delays is not addressed, and multiple inheritance is not considered. A more recent work [16] introduces a methodology that combines domain and company-speci c ontologies and databases to obtain multiple levels of abstraction for process mining. In this paper data in databases become instances of concepts at the bottom level of a taxonomy tree structure. If consecutive tasks in the discovered model abstract as the same concepts, those tasks are aggregated. However, also in this work we could nd neither a clear description of the abstraction algorithm, nor the management of interleaved actions or delays.
Other interesting contributions can be found in [
However, most of the papers cited above (including [8, 7]) present theoretical frameworks, and not yet a detailed technical architecture nor a concrete implementation of all their ideas.
Referring to medical applications, the work in [11] proposes an approach, based on semantic process mining, to verify the compliance of a Computer Interpretable Guideline with medical recommendations. In this case, semantic process mining refers to conformance checking rather than to process discovery (as it is also the case in [8]). These works are thus only loosely related to our contribution.
In conclusion, in the current research panorama, our work appears to be very innovative, for several reasons: { many approaches, illustrating very interesting and sometimes ambitious ideas, just provide pure theoretical frameworks, which can be very important to inspire more engineering-style work. However, concrete implementations of algorithms and complete architectures of systems are often missing, leaving open research opportunities for contributions like the one we have presented; { in semantic process mining, more work has been done in the eld of conformance checking (also in medical applications), while process discovery still deserves attention (also because many approaches are still at the theoretical level, as commented above); { as regards trace abstraction, it is often proposed as a very powerful means to obtain better process discovery and analysis results, but technical details of the abstraction mechanism are usually not provided, or are illustrated through very simple examples, where the issues related to the management of interleaved actions or delays do not emerge.
In this paper, we have presented a framework for knowledge-based abstraction of event log traces. In our approach, abstracted traces are then provided as an input to semantic process mining. Semantic process mining relies on ProM algorithms; indeed, the overall integration of our approach within ProM is foreseen in our future work.
The rst experimental results in the eld of stroke management suggest that the capability of abstracting the event log traces on the basis of their semantic goal may allow to mine clearer process models, where unnecessary details are hidden, but key behaviors are clear.
In the future, we plan to conduct a validation study, by quantitatively comparing di erent process models (of di erent SUs) obtained from abstracted traces. Comparison will resort to knowledge-intensive process similarity metrics, such as the one we described in [18]. We will also extensively test the approach in di erent application domains.
Finally, an abstraction mechanism directly operating on process models (i.e., on the graph, instead of the event log), may be considered, and abstraction results will be compared to the ones currently enabled by our framework. Academy Doctoral Consortium, MONET, OnToContent, ORM, PerSys, PPN, RDDS, SSWS, and SWWS 2007, Vilamoura, Portugal, November 25-30, 2007, Proceedings, Part II, volume 4806 of Lecture Notes in Computer Science, pages 1244{1255. Springer, 2007. 7. A. K. Alves de Medeiros and W. M. P. van der Aalst. Process mining towards semantics. In T. S. Dillon, E. Chang, R. Meersman, and K. P. Sycara, editors, Advances in Web Semantics I - Ontologies, Web Services and Applied Semantic Web, volume 4891 of Lecture Notes in Computer Science, pages 35{80. Springer, 2009. 8. A. K. Alves de Medeiros, W. M. P. van der Aalst, and C. Pedrinaci. Semantic process mining tools: Core building blocks. In W. Golden, T. Acton, K. Conboy, H. van der Heijden, and V. K. Tuunainen, editors, 16th European Conference on Information Systems, ECIS 2008, Galway, Ireland, 2008, pages 1953{1964, 2008. 9. W. Van der Aalst. Process Mining. Discovery, Conformance and Enhancement of
Business Processes. Springer, 2011. 10. W. Van der Aalst, B. van Dongen, J. Herbst, L. Maruster, G. Schimm, and A. Weijters. Work ow mining: a survey of issues and approaches. Data and Knowledge Engineering, 47:237{267, 2003. 11. M. A. Grando, M. H. Schonenberg, and W. M. P. van der Aalst. Semantic process mining for the veri cation of medical recommendations. In V. Traver, A. L. N. Fred, J. Filipe, and H. Gamboa, editors, HEALTHINF 2011 - Proceedings of the International Conference on Health Informatics, Rome, Italy, 26-29 January, 2011, pages 5{16. SciTePress, 2011. 12. D. Grigori, F. Casati, M. Castellanos, U. Dayal, M. Sayal, and M.C. Shan. Business process intelligence. Computers in Industry, 53(3):321{343, 2004. 13. M. Hepp, F. Leymann, J. Domingue, A. Wahler, and D. Fensel. Semantic business process management: A vision towards using semantic web services for business process management. In F. C. M. Lau, H. Lei, X. Meng, and M. Wang, editors, 2005 IEEE International Conference on e-Business Engineering (ICEBE 2005), 18-21 October 2005, Beijing, China, pages 535{540. IEEE Computer Society, 2005. 14. M. Hepp and D. Roman. An ontology framework for semantic business process management. In A. Oberweis, C. Weinhardt, H. Gimpel, A. Koschmider, V. Pankratius, and B. Schnizler, editors, eOrganisation: Service-, Prozess-, MarketEngineering: 8. Internationale Tagung Wirtschaftsinformatik - Band 1, WI 2007, Karlsruhe, Germany, February 28 - March 2, 2007, pages 423{440. Universitaetsverlag Karlsruhe, 2007. 15. http : ==www:win:tue:nl=ieeetf pm=doku:php?id = shared : processminingmanif esto. IEEE Taskforce on Process Mining: Process Mining Manifesto (last accessed on 22/11/2017). 16. W. Jareevongpiboon and P. Janecek. Ontological approach to enhance results of business process mining and analysis. Business Proc. Manag. Journal, 19(3):459{ 476, 2013. 17. M. N. De Maio M. Salatino, E. Aliverti. Mastering JBoss Drools 6 for Developers.
Packt Publishing, 2016. 18. S. Montani, G. Leonardi, S. Quaglini, A. Cavallini, and G. Micieli. A knowledgeintensive approach to process similarity calculation. Expert Syst. Appl., 42(9):4207{ 4215, 2015. 19. C. Pedrinaci and J. Domingue. Towards an ontology for process monitoring and mining. In M. Hepp, K. Hinkelmann, D. Karagiannis, R. Klein, and N. Stojanovic, editors, Proceedings of the Workshop on Semantic Business Process and Product