discovering unstructured processes
The aim of this section is to use the prepared event log of Diabetes type II patients for modelling the healthcare processes. Two techniques that are originally designed to cope with complex and unstructured processes are chosen. These techniques are Fuzzy miner and Local Process Models mining (LMs), that were discussed in Chapter 2. Both of these techniques support the concept of events abstraction. The results are explained below.
3.8.1
Fuzzy miner results
The fuzzy miner without abstraction has generated a complex model as presented in Figure 3.13. However, this technique supports an interactive abstraction where the nodes can be abstracted manually using a slider which provides the cut-off threshold of node significance. Figure 3.14 shows two different models (left), the top model with node significance cut-off ∼ 0.5 has four events which are input, discharge, chartevent and output and two clusters of random number 18 and 19 with significance 0.051 and 0.118 respectively. The significance of cluster is the sum of the significance of related events which is basically based on event frequency. On the other hand, the bottom abstracted model has one event, input (with the highest significance = 1) , and two clusters but with different related events and significance scores which are 0.261 and 0.235 for clusters 18 and 19 respectively.
Figure 3.13: The discovered model using fuzzy miner in ProM6.8
593.8. The practical limitations of some techniques used for discovering unstructured processes
preserved, the less frequent and high correlated events are clustered and the low frequent events are removed. A hierarchy modelling is supported as well when clicking on the clusters. Further analysis of the events that are included in cluster 19 and 18 is shown in Figure 3.14 as well. These clusters show that the gray edges represent the links from the previous abstracted model (with 0.5 cut-off). The actual number of elements inside cluster 19 is 4 elements (Prescription, CPTevent, Edreg and Edout) while cluster 18 has 5 elements (Datetimeevents, Admission, Transfer, Noteevent and Call). The Laboratory and Microbiology events are removed after abstraction since they were not significant based on this threshold.
Figure 3.14: Fuzzy model with abstraction
Although fuzzy miner has reduced model complexity, it has some practical limitations can be summarized in:
1. The fuzzy miner is an interactive technique and does not discover automated process model directly where it requires manual adaptation of discovery parameters and model tuning.
2. It models the sequence of events but cannot discover different process constructors such as parallel and choice.
for model evaluation and conformance checking with the related event log.
4. Events are assigned into clusters in one-to-one mapping which makes the model unable to discover same events with different context. This is a major limitation where no semantic clustering is supported. For instance, the examined healthcare process here is generated form a hospital with different Intensive Care Unit (ICUs). The charted events can be recorded in different ICUs, which represent different contexts of care. Therefore, clustering events based only on the frequency as provided by fuzzy miner is not efficient for distinguishing different care contexts/states.
5. The model may start with any random node where the start node might be changed after refreshing the model and no clear start and end node are presented at the process. 6. Low frequency events are removed from the abstract model such as Laboratory and Mi-
crobiology events. We believe that, low frequency events can be significant and may have an effect on the flow of healthcare process, therefore, such events are preferable to be presented in the abstract model in order to provide comprehensive process model.
3.8.2
Local process mining results
Based on the best of our knowledge, the latest method of unsupervised pattern extraction in process mining is mining local process models that is discussed in Chapter 2. The implemen- tation of this approach is supported by the ProM process mining tool. The aim is to discover abstract model and reduce model complexity with local process models detection. For this experiment we use the plug-in ‘Mine Local Process Models’ in ProM6.8.
In order to discover local models (LM), a number of settings should be set first such as; the number of local process models that will be discovered and what kind of process constructors that should be used for modelling. In this experiment, the parameters that are used before model discovery are 50 local process models and the process constructors of sequence, choice and loop. This has resulted in 50 local process model for 43 groups of events, where one group of event might be expressed using multiple local models. This will be explained in the results of LM below. The results of mining local models are ranked based on a score that is built on 4 metrics which are; confidence, determinism, language and coverage. These metrics are explained in Chapter 2 and for more information refer to [70].
Some samples of top ranked local models are presented below where, each event has two num- bers in the form of (event frequency in this pattern/total number of this event):
613.8. The practical limitations of some techniques used for discovering unstructured processes
Figure 3.15: Local process models of group 1
Figure 3.16: Local process models of group 2
(a) LM1 of group 3
(b) LM2 of group 3
(c) LM3 of group 3
Figure 3.17: Local process models of group 3
Figure 3.15 shows the top local process models for group 1 of events which includes admission and discharge where this local model represents the long dependency between admission and discharge. Admission event is observed 296 times in total but 295 times has occurred in this pattern. Also, discharge event is always observed in this pattern with 295 times in total. This local model has a score of 0.870 based on the different criteria such as frequency, confidence and coverage that are used in the discovery process. Figure 3.16 shows the pattern of emergency department register(edreg) then transfer to a hospital ward where a loop of transfer event is possible. The event (edreg) is observed 227 out of 245 in this pattern and transfer is observed 569 out of 697 in this patter as well.
Figure 3.17 presents the three top local models of group 3 which includes emergency department register(edreg), emergency department out(edout) and discharge events. These three local
models are ordered based on their scores which are 0.839,0.831 and 0.826. It should be noted that, similar models are discovered for the same group of events but not presented here to prevent redundancy.
It is important to note that, there are 50 local process models that are generated hence, the above models are picked manually to illustrate a sample of the results of mining local models. Although mining local model has extracted local pattern in unsupervised way and has captured the long dependency between events, it has a number of limitations can be outlined as follows: 1. Mining local process models aims to discover internal patterns which are restricted to three to five event types and cannot discover the whole process represented in a start-to-end model.
2. Mining local model requires time and careful selection of several parameters that should be set before doing process discovery. Four main parameters which are the number of local process models to be discovered, operators type (whether sequence or loop and so on), maximum and minim number of pattern occurrence and some temporal constraints such as time gap between subsequent events. These parameters have a heavy impact on the resulted models.
3. Although mining local process provides unsupervised pattern extraction, it generates a large number of local process models with the similar score and may have overlapping events. This requires a domain preference to select the most representative local model for each group of events.
4. The local process model cannot be used for hierarchy modelling such as the one that is supported by fuzzy miner.
5. Local process mining can be inefficient method and cannot scale with large event log. We have tried using this technique with the colorectal cancer patients logs however, the local models could not be discovered due to multiple crashes of ProM tool that have happened during local models discovery.
Exploring these two current techniques of process mining has emphasized the need for developing more robust abstraction method that can do the following; supports an automatic abstraction for start to end process model and discovers the general care of pattern for a complex large event log with the ability of handling process variations. Also, the required method should be able to generate a process model that can be evaluated and assessed within the available process mining frameworks in addition to distinguishing care events that may occur in different contexts of the process.