Business Process Monitoring through a Case-Based Approach

1. A Case-based Approach to Business Process Monitoring S. Montani 1 , G. Leonardi 1 1 Dipartimento di Informatica, University of Piemonte Orientale, Alessandria, Italy

2. Business Process Management BP: collection of related activities which produce a service or a product BPM: defining, executing, monitoring and optimizing BPs BP Optimization: flexibly change and adapt BP: Expected situations (new laws, reengineering) Unanticipated exceptions (emergencies) BP not correctly applied

3. Business Process Monitoring Agile Workflow Technology: solution for process adaptation and overriding: Modifications of process schema (by process engineers) Ad-hoc changes of process instances (by end users) Business Process Monitoring: highlighting and organizing non-compliances Allows verification of process conformance Suggests long-term changes to deal with non-compliances

4. Case-based Business Process Monitoring Analysis based on retrieval of traces similar to the current one Trace: Sequence of actions occoured in a process instance Suggestions on how to modify the default process schema: Analysing most similar examples of changes in retrieved traces Inspecting clusters obtained on the base of trace similarity To perform analysis: Case definition Similarity measure

5. Case Structure Case: a trace of execution of a given process schema Process schema: general schema definition suggesting what has to be done Trace: sequence of actions actually performed by the working team Actions ordered by execution time, and stored with additional information: the timestamp of the event, the performer or originator of the event (i.e., the person/resource executing or initiating the action), data elements recorded with the event (e.g., the dosage of a drug) ID action start end actor 1. A1 1-1-08 10:30 1-1-08 10:35 Robert 2. A2 1-1-08 10:30 1-1-08 12:00 Frank 3. A1 2-1-08 12:00 3-1-08 00:30 Louise 4. A4 2-1-08 16:00 2-1-08 17:03 Carla 5. A5 2-1-08 23:45 3-1-08 04:55 Joseph 6.

6. Similarity Measure Edit Distance: given two traces T1, and T2, ED (T1, T2) = number of operation needed to obtain T1 = T2 Edit operations on traces: Substitute one action with a different one Insert a new action Delete an action Each edit operation has cost set to 1 Trace T1: A1, A2, A1, A4, A5 Trace T2: A1, A1, A3, A5 To obtain T1 = T2: - Delete A2 in trace T1 - Substitute A4 with A3 in trace T1 (or sub. A3 with A4 in T2) ED(T1, T2) = 2

7. Taxonomy of Actions In our case, the cost of a substitution is not always 1: The cost is a value [0, 1] The cost is domain dependent Actions are grouped in a taxonomy derived from domain knowledge Distance of actions (A1, A2) proportional to distance in the taxonomy All Actions Emergency Hospitalization Diagnostic Therapies Diagnostic Therapies C.A.T. Evaluation N.M.R. Rehabilitation

8. Taxonomical Distance Let and be two actions in taxonomy T; Let be the common ancestor of and ; Taxonomical distance between and : N1 = n. arcs in the path between and ; N2 = n. arcs in the path between and ; N3 = n. arcs between and the root of T All Actions Emergency Hospitalization Diagnostic Therapies Diagnostic Therapies C.A.T. Evaluation N.M.R. Rehabilitation dt(C.A.T., N.M.R) = (3 + 3) / (3 + 3 + 2*0) = 1 N1 = 3; N2 = 3; N3 = 0 All Actions Emergency Hospitalization Diagnostic Therapies Diagnostic Therapies C.A.T. Evaluation N.M.R. Rehabilitation dt(C.A.T., Evaluation) = (1 + 1) / (1 + 1+ 2*2) = 0.33 N1 = 1; N2 = 1; N3 = 2

9. Action Weight Also the cost of an insertion or deletion (indirection) is domain dependent: Some indirections can cause important effects (high weight value); other indirections cause minor impact (low weight value). Let be an action, its weight w is defined as: w [0, 1[ if generates an indirection w = 1 otherwhise

10. Trace Edit Distance Let P, Q be two traces of actions; let and be two actions taxonomy T. Trace Edit Distance d(P, Q) is defined as: Where: (e 1 , ... , e k ) transforms P into Q, and: C(e i ) = dt( , ) if e i is the substitution of with C(e i ) = 1*w if e i is the insertion (or deletion) of in P (from Q)

11. Trace Edit Distance: case 1 B and B are comparable: dt(B, B) : B and B considered aligned Next actions are checked (C and C)

12. Trace Edit Distance: case 2 B and B are not comparable: dt(B, B) , and A and A not comparable: B is substituted with B Dt(B, B) is added to d(P, Q)

13. Trace Edit Distance: case 3 B and B are not comparable: dt(B, B) , and A and A comparable, and B and C comparable: B aligned with C; B considered as indirection 1*w B is added to d(P, Q)

14. Trace Edit Distance: case 4 B and B are not comparable: dt(B, B) , and A and A comparable, and C and C comparable: B is substituted with B Dt(B, B) is added to d(P, Q)

15. Application Example: The Stroke Domain Stroke: loss of brain function due to ischemia or blood hemorrage Medical emergency, potentially leading to permanent damage or to death Number 2 cause of death worldwide (soon becoming number 1) Stroke patient management in 4 phases: Emergency management; Hospitalization; Discharge; Follow-up.

16. A correct Patient Management Emergency management: Symptoms recognition; C.A.T Hospitalization in neurological ward/stroke unit Diagnosis: Neurological evaluation + E.C.G. + other diagnostic investigations (i.e. chest X.R.) Therapy: antiaggregants; physical rehabilitation Discharge Follow-up: no data available

17. Retrieval Example: Query Case Description Atypical situation: Emergency: Stroke diagnosed directly after C.A.T., without need of further investigations antiaggregant therapy started immediately Hospitalization: Patient condition probably unstable, therefore further investigations: additional tests and a further anticoagulant therapy. Physical rehabilitation NOT started

18. Retrieval Example: Case Retrieved Query Case Most Similar Case Indirection: diagnosis and therapy started in Hospitalization Indirections: less investigations have been performed Anticoagulant and antiaggregants are comparable : very close in the taxonomy since very similar effect Query case: #103101 in the trace DB Trace DB composed by more than 300 traces K-Nearest Neighbor retrieval using: K = 20 Trace Similarity Distance Retrieval time: 1.2 seconds on an Intel Core2 Duo T9400 4Gb ram No physical rehabilitation even in the retrieved trace atypical situation captured

19. Conclusion and Future Work CBR-based BP Monitoring: Cases are traces of execution Trace Edit Distance Helps in adapting BP on the basis of deviations in past traces of the same BP Clustering to identify the most frequent changes to the BP schema Implemented (single linkage clustering), but not fully validated Integration in the Process Mining tool ProM (W. van der Aalst) Extension of Trace Edit Distance considering temporal delays