The jupyter notebook "Event_Log_Generation.ipynb" in this repository creates the event log for the evaluation in the paper "Change Detection in Dynamic Event Attributes". It uses the MIMIC-IV (https://physionet.org/content/mimiciv/1.0/) database, stored in a Postgres database. First, all cases related to acute kidney failure are retrieved. Then, the hospital data for their respective hospital admission is fetched, including laboratory values and demographic data. The database can be accessed via physionet, which requires CITI training for access (https://mimic.mit.edu/iv/). The change detection is implemented as a python script in "Change_Detection.ipynb", which starts with the classification of event attributes according to their type (continuous vs. categorical) and their process characteristic (static, semi-dynamic, and dynamic), where this contribution focusses on dynamic event attributes. Then, it performs the statistical tests to generate the respective change analysis cells in the change detection cube. The visualization of the change detection cube is implemented in the "UI.ipynb" notebook, allowing to perform the OLAP operations (Slice, Dice, Pivot, and Drill up/down). The UI can show views of slices of the cube by illustrating an event attribute change matrix, including the results of the change detection analysis for each change analyses cell. So far, the UI focusses on the analysis of continuous event attributes and allows the event attribute connection with categorical event attributes. However, it can be adjusted to display the results of the categorical statistical tests as well.
Therefore, consider the following steps to make it work:
- Generate the event log by executing the jupyter notebook "Event_Log_Generation.ipynb".
- Execute the jupyter notebook "Change_Detection.ipynb" taking the generated event log as its input.
- Use the modified pm4py package provided by this repository. Modified files within the dfg folder: visualizer.py, frequency.py
- Execute the jupyter notebook "UI.ipynb" taking the generated change detection files as input. The last cell in the notebook shows the UI.
Below are additional views of the change detection cube:
We further investigated, if we can detect changes in different process variants. For that, we looked, together with the medical expert, in the differences between patients visiting different ICU's (Intensive Care Units). In the picture below, one can observe the value changes of Cceatinine, which is a major marker for kidney disease. Whereas the RBC values are high for non-cardiac related ICU's, the RBC values are lower for patients visiting the cardiac related ICU's (variants 4 and 7). The variants can be seen in the variants selection box, which are enumerated according to their order (0-10). We discussed, that patients in the cardiac ICU tend to be treated regarding their issues with the heart, instead of their kidney. Hence, the creatinine is not changing as much as for the other patients.
To further investiagte that, we looked into another laboratory value hematocrit, where low hematocrit is prominent in patients having heart failure. In the matirx below, patients going through the cardiac icu tend to have a decreased hematocrit value in comparison to the visited ward before. As cardiac problems can occur in patients with kidney failure frequently, the other icu's experience a changing behaviour as well. However, the highest RBC value is related to the pre-icu cardiology:cardiac icu relation within the variant, where patients go through cardiology related wards only (4).
The value changes of creatinine and hematocrit can be visualized in the process model, which is shown below. It can be seen, that the cardiac related patients have a higher RBC value for changes before the icu in hematocrit, whereas the other patients have a higher RBC value for changes in creatinine. A possible explanation for that could be that there are different aims for treatment, where patients in the cardiac icu tend to be primarily treated according to their heart problems and the others according to their kidney related problems.
