In last week’s Process Mining Café, we talked about process mining in healthcare with Luise Pufahl, a postdoctoral researcher at TU Berlin in Germany, and Fran Batchelor, a nursing informatics specialist at UW Health in the United States. You can now watch the recording here.
When you apply process mining to a healthcare process then at first everything seems to be very clear: The patient ID should be the case ID, the steps are the diagnosis, treatment, or scheduling activities that took place, and the timestamps are the date and time when the activity happened. However, there are many challenges that make things more difficult in practice.
We discussed the specific challenges of process mining in healthcare along the phases of a typical process mining project.
1. Scoping
First, you need to answer the question where does the process start and where does it end? Simply taking the patient ID as the case ID means that the scope spans the lifetime of the patient. Usually, this is too big and you want to limit the analysis to a smaller scope like a surgery. Another way to focus the analysis is to select a subset of activities, for example, based on the medical guidelines for a specific diagnosis and treatment pathway.
2. Extracting and preparing the data
During the data preparation, often different data sources need to be merged to get all the information that is needed. In this phase, understanding the data and dealing with data quality issues are the biggest problems. For example, there can often be data quality problems if the data is manually recorded. As more data is collected automatically (also by medical devices), the availability and quality of the data improves while data privacy concerns become more important as well.
3. Dealing with complexity
Once you import your data set into Disco, you need to deal with the complexity of the process even more than you would for most other processes. For example, it can easily happen that a data set with 1000 cases has 1000 variants, because every patient follows a unique path. The grouping of cases, leaving out details, activity aggregation, but also unfolding can help to get the data set to the right level for the analysis.
4. Analyzing and communicating the results
To interpret the analysis results correctly, domain knowledge is very important. The process visualizations that can be produced with process mining are more complicated than the manual Visio models that are often traditionally created, because they show all the unexpected flows, exceptions, and inefficiencies. However, in contrast to the manual models they show the actual flow of the process and help a lot in the communication with the medical staff.
Thanks again to Luise and Fran, and to all of you, for joining us!
Links
Here are the links that we mentioned during the session:
-
BPI Challenge Dataset from 2011 (open project file directly in Disco)
-
Bart van Acker explains during his presentation at Process Mining Camp 2015 how selecting the right activities was important for their analysis of the Head and Neck Care chain at Radboudumc hospital
-
Fran explains during her presentation at Process Mining Camp 2018 how surgical services can be analyzed with process mining and showed how she had to unfold data for one of her analyses to get a detailed enough view
-
Simplification strategies for complex process maps
-
Luise’s paper on event log generation for a hospital process related to the treatment of low back pain
-
HICSS Workshop on Process Mining in Healthcare (Paper submission deadline is June 15, 2021)