Context-Aware Change Pattern Detection in Event Attributes of Recurring Activities: Implementation and Evaluation
This repository provides the implementation and further evaluation details of the paper entitled Context-Aware Change Pattern Detection in Event Attributes of Recurring Activities. The implementation includes four jupyter notebooks and a modified pm4py implementation. The jupyter notebooks enumerated from 1 to 3 represent the approach of the paper, starting with the detection and transformation of recurring events. Then, the change pattern detection is applied in the second jupyter notebook. The third jupyter notebook includes the UI to interact with the identified change patterns and to illustrate them in a process model. The modified pm4py package is only required, if one wants to visualize change patterns in the process model.
The Sepsis dataset provides an event log representing the hospital treatment process of patients having sepsis. It includes laboratory measurements as recurring activities, where the result of the measurement is stored as an event attribute. The event log includes 1,050 cases with 15,214 events representing 16 activities.
MIMIC-IV is a relational database including hospital processes of different patients, with procedures performed, medications given, laboratory values taken, image analysis conducted, and more. Its purpose is to support research in healthcare and is therefore publicly available (https://mimic-iv.mit.edu/).
The event log extracted from MIMIC-IV incorporates common activities in an intensive care unit for acute kidney failure patients, including different forms of ventilation and dialysis. Additionally, laboratory measurements are recorded, including nine measurements conducted at one activity ``Measurement'', which are important for monitoring the disease progression, such as creatinine and urea nitrogen. It contains 3,528 hospital process instances with 26,370 events representing nine different treatment activities.
The event log from MIMIC-IV cannot be shared directly due to a required data use agreement. Thus, we provide an event log generation script in the already mentioned GitHub repository.
To reproduce the results of the Sepsis event log, the scripts are ready to be executed, as the Sepsis event log is already available in the Logs folder. The outputs of the preprocessing steps, including the transformed event log and the detected change patterns, are also already available for Sepsis, so one can immediately execute the UI jupyter notebook if desired.
To reproduce the MIMIC results, one needs access to the MIMIC-IV database, which requires CITI training. Usually, that does not take much more than a day and access is granted within a week. If access is granted, the event log can be retrieved. We implemented an event log generation tool for MIMIC-IV, which allows to provide a config file as an input, which results in an ready-to-use event log. Use the config file in this repository to retrieve an event log by executing the following command: python extract_log.py --config MIMIC_Config.yml. Some post-processing is required, which is conducted in this jupyter notebook. After that, the other jupyter notebooks can be executed with the MIMIC event log.
The implementation is not limited to the above mentioned event logs and can be used with all event logs. It should be noted, that the event logs require recurring activities to make sense for further analysis. It is only required, that the event logs are provided as a .csv file and that the mandatory attributes case id, activity, and timestamp are renamed in the first jupyter notebook accordingly.
As mentioned in the paper, the following presents detailed results regarding the detection of recurring activities and change pattern results for the MIMIC event log. First, the results of the vectors dfr/dpr representing the directly follows and directly preceding ratios for all activities in MIMIC and Sepsis are shown. The repetition score is then the sum of one row, which are also illustrated below. The matrices are supposed to be read row-wise. For example, the "Measurement" activity in Tab. 1 is followed by "START Invasive Ventilation" in 84.5% of the occurences of "START Invasive Ventilation". Tab. 1 and Tab. 2 show, that the "Measurement" is mostly conducted before the START and after the END of a treatment activity. Tab. 3 and Tab. 4 show the respective results for Sepsis. In Sepsis, "CRP" and "Leucocytes" are almost never following "Release A", but precede it relatively often.
 |
|---|
| Tab. 1 Directly-Follows ratios for activities in MIMIC |
 |
|---|
| Tab. 2 Directly-Precedes ratios for activities in MIMIC |
 |
|---|
| Tab. 3 Directly-Follows ratios for activities in Sepsis |
 |
|---|
| Tab. 4 Directly-Precedes ratios for activities in Sepsis |
Tab. 5 and Tab. 6 show the repetition scores, which are the average scores of the rows in the matrices.
 |
 |
|---|---|
| Tab. 5 Repetition scores for all activities in MIMIC | Tab. 6 Repetition scores for all activities in Sepsis |
We now compare the results of the transformed event logs to the results of the original event logs. With the identified recurring activities and contexts of interest, the methods of change pattern detection can be applied. Without considering their context, the resulting change patterns can be analysed only in relations from the activities to themselves. This is illustrated in Fig. 1 for the Sepsis event log, where the discovered process model is enhanced with change patterns of event attribute values CRP and Leucocytes, respectively.
 |
|---|
| Fig. 1 Enhanced process model with change patterns detected on the raw Sepsis event log |
The dotted lines illustrate change patterns when the activities are performed multiple times, where the blue colour illustrates a decrease of the values of CRP and Leucocytes. Please note, that we speak of activities, e.g. "CRP", when quoted and of event attributes, such as CRP, when written without quotes. As the event attributes are not related to other activities, only the relation from the recurring activities to themselves can be considered. The number of statistically significant p < 0.05 results for the original Sepsis event log is only one, which is the relation from "Leucocytes" to "Leucocytes". We also observe small effect sizes (0.13 for Leucocytes and 0.03 for CRP (not statistically significant!)), expressing that there is only a small indication of value decreases.
 |
|---|
| Fig. 2 Enhanced process model with change patterns detected on the transformed Sepsis event log |
Fig.2 presents the discovered process model for the transformed Sepsis event log. It includes the transformed measurement activities describing their context, such as "CRP AFTER ER Sepsis Triage" and "Leucocytes BEFORE Release A". With that, multiple activities include the event attribute representing the respective measurement (Leucocytes/CRP) now. It can be observed, that, dependent on where the measurements are conducted, the change patterns differ. For example, looking at the CRP measurements, CRP has a tendency to increase from "CRP AFTER ER Sepsis Triage" to "CRP Admission NC" and then drastically decreases from "CRP AFTER Admission NC" to "CRP BEFORE Release A" for almost all patients with an effect size of
The following figures show detailed results of the change pattern detection in MIMIC. Fig. 1 and Fig. 2 show a process model enhanced with change patterns. Fig. 1 illustrates a process model without the detection and transformation of recurring activities and Fig. 2 shows then the process model based on the transformed event log. It can be seen, that the "Measurement" activity is highly recurring and acts as a central point in the process, which is also represented as high values in dfr/dpr. Thus, the insights of process discovery are very limited. This is the case for the change pattern detection as well, where a decrease of Anion Gap was idnetified with a very low test statistic of -0.13, meaning that there is a very small indication of a value change only.
 |
|---|
| Fig. 1 Enhanced process model with change patterns detected on the raw MIMIC event log |
Fig. 2 illustrates an enhanced process model with a few change patterns identified based on the transformed event log. As for Sepsis, more change patterns with increased effect sizes could be detected. Especially the measurements associated to the "Dialysis" activities have high effect sizes (0.78-0.88), where Blood Urea Nitrogene (BUN) decreases after dialysis and increases again, before "Dialysis" is conducted again [1]. This looping pattern could not be identified before. Additionally, a significant increase of systolic blood pressure could be identified after the beginning of dialysis until end of invasive ventilation, which is a sign of intradialytic hypertension [2]. Furthermore, the process mostly starts with measurements before Invasive Ventilation is conducted. It can also be seen, that "Invasive Ventilation" is conducted until the end of treatment and "Dialysis" is repeated multiple times during treatment. Thus, the structure of the process model improved as well and more insights regarding changing patterns could be derived.
 |
|---|
| Fig. 2 Enhanced process model with change patterns detected on the transformed MIMIC event log |
Fig. 3 shows another view of the change detection cube, which allows to see two-dimensional slices of the it. On the X-axes, three event attributes, and on the Y-axes, seven selected eventually follows relations are shown. There are some more, as seen in the list above, which can be explored with the UI jupyter notebook. Some of the illustrated change patterns in the process model can be seen here again and some more. The values in the cells represent the respective test statistic RBC.
 |
|---|
| Fig. 3 Change Pattern Matrix of selected event attributes and relations independent of a trace variant |
[1] Schiffl, H.: Discontinuation of renal replacement therapy in critically ill patients with severe acute kidney injury: predictive factors of renal function recovery. Int Urol Nephrol 50(10), 1845–1851 (Oct 2018)
[2] Inrig JK. Intradialytic hypertension: a less-recognized cardiovascular complication of hemodialysis. Am J Kidney Dis. 2010 Mar;55(3):580-9. doi: 10.1053/j.ajkd.2009.08.013. Epub 2009 Oct 22. PMID: 19853337; PMCID: PMC2830363.