Reducing workflow model complexity for an automatic workflow discovery algorithm

With the increasing number of adverse patient outcomes and deaths due to medical errors, process mining techniques have been used to build automatic workflow discovery algorithms to assist teams during the treatment procedure, better understand the treatment practice and potentially help improve patient outcomes. These automatic workflow discovery algorithms use event logs that has all the information of the events that are executed in the treatment procedure, to extract a process model. Automatic workflow discovery algorithms are now used to analyze the trauma resuscitation process and improvise the patient outcomes. The widespread use of process mining techniques to discover and analyze workflows in healthcare, has motivated this study that has analyzed, how varying a parameter that contributes to the workflow discovery algorithm changes the workflow model, discusses the reasons behind those changes in the discovered workflow, presents the optimum value of the parameter that produces best results and backs up the theoretical discussion with numerical results. We use an expert-based model that is derived from hand drawn model by medical experts after multiple revisions, that serves as the ground truth and to compare the accuracy of our workflow model. Based on the comparison of the workflow model with the expert model, the variations observed could be classified into the categories : an incorrect addition of treatment activity, a redundant addition of treatment activity or a correct addition of treatment activity to the workflow. This work is an attempt in the direction to make the workflow model generated using the workflow algorithm, more accurate and exhibits least complexity, thereby making it easy to comprehend. The workflow discovery algorithm used for this work has two phases : (1) Construction of an event sequence of consensus activities that has the occurrence probability of more than the predefined threshold. This occurrence probability is determined using the event logs. This phase ensures that the workflow model comprises of the frequent activities that have appeared in the event logs. (2) Inclusion of non-consensus activities, that are common but dispersed between consensus activities in the workflow. This inclusion is done after multiple iterations through consensus activities to determine the activities that are interleaved between consensus activities and when considering a window of consecutive consensus activities, the combined probability of occurrence of the non-consensus is more than the predefined threshold. This study explores one of the contributing hyperparameters of the algorithm, to infer its effect on the workflow model. The analysis falls in agreement with the underlying notion behind this study, that varying the hyperparameter of workflow discovery algorithms results in varying complexity of the workflow model.M.S.Includes bibliographical reference