CAiSE´2024, Limmasol

EXPLORING OBJECT
CEN TRIC P ROCESS MININ G
WITH MIMIC IV:
UNLOCKING INSIGHTS IN
H E A LT H C A R E

Authors : Aamod Vyas*, Anukriti Tripathi, Aneesh, Yuvraj, Shivam,

Swetank Pandey, O.P.Vyas

*[email protected]
 • Introduction

• Background (Process Mining &

OCPM)
• Datasets

• Event Log Generation

CONTENTS • Relationship between objects

• Analysis Results

• Conclusion

• References
P R O C E SS M I N I N G

• Process mining is a method of applying

specialized algorithms to event log data to
identify trends, patterns and details of how a
process unfolds. Process mining applies data
science to discover, validate and
improve workflows. (IBM)
O B J E C T- C E N T R I C P R O C E S S M I N I N G

• Object-Centric Process Mining is like having a super detailed

map that shows how different objects (like toy parts,
machines, and orders) interact and move through the factory.
• It doesn't just focus on one process but looks at everything
happening around each object. This gives you a much clearer
picture of how different parts of the factory work together.

van der Aalst, W.M.: Object-centric process mining: unraveling the fabric of real processes. Mathematics 11(12),
2691 (2023)
T R A D I T I O N A L P R O C E SS M I N I N G &
O B J E C T- C E N T R I C P R O C E S S M I N I N G

Traditional Process Mining: Tracks the Object-Centric Process Mining: Tracks

steps to assemble a toy. toy parts, machines, workers, and
orders, showing how they all interact to
create the toy.
M E D I C A L I N F O R M AT I O N M A RT
FOR INTENSIVE CARE (MIMIC)

MIMIC is a database that provides de-identified medical data from

critically ill patients. It includes a wealth of information such as:
• Patient details

• Vital signs

• Lab results

• Medications

• Clinical notes

Cremerius, J., Pufahl, L., Klessascheck, F., Weske, M.: Event log generation in MIMIC-iv : International Conference on
Process Mining. pp. 302–314. Springer (2022)
M E D I C A L I N F O R M AT I O N M A RT F O R
INTENSIVE CARE (MIMIC)

Researchers use MIMIC for various purposes,

including:
•Training machine learning algorithms

•Predicting patient risks

•Studying treatment effectiveness

•Developing tools to improve healthcare delivery

O B J E C T- C E N T R I C P R O C E S S
M I N I N G ( O C P M ) I N H E A LT H
CARE

OCPM provides a unique perspective by focusing on objects

rather than individual cases. These objects can include:
• Patients

• Medications

• Lab results

• Referrals
OBJECTIVE OF OUR WORK

The goal is to transform event logs from the MIMIC IV healthcare database
into Object-Centric Event Logs. This involves:
• Selecting relevant objects and their relationships
• Extracting and analyzing interactions between these objects
Purpose:
• To gain deeper insights into patient journeys within the MIMIC IV data
• To achieve a comprehensive understanding of healthcare workflows and
patient care pathways by focusing on objects and their connections
M I M I C - I V H E A LT H C A R E D A T A S E T
MIMIC-IV is a comprehensive healthcare database containing data from over
60,000 patients from a Boston hospital, covering patient stays until 2018. It
includes:

• Detailed Information: Procedures, tests, and activities

• Privacy Measures: Timestamps are shuffled, but event order is preserved

for analysis integrity

The database is organized into modules, each providing a valuable resource

for healthcare research:

1. Core: Basic patient information like demographics

2. Hosp: Hospital-wide data on a patient’s journey, focusing on continuity of

care

3. ED:
Edin, Information
J., Junge, from
A., Havtorn, J.D., the
Borgholt, Emergency
L., Maistro, Department,
M., Ruotsalo, including
T., Maaløe,L.: Automated initial
medical coding diagnosis
on MIMIC-iii
and MIMIC-iv: A critical review and replicability study. In: Proceedings of the 46th International ACM SIGIR Conference on
andand
Research triage
Development in Information Retrieval. pp. 2572–2582 (2023)
M I M I C - I V H E A LT H C A R E D A T A S E T

Building on Existing Work:

• Developed a Python script to extract and analyze healthcare data from the MIMIC-IV
database.

Features:

• Cohort Definition: Users can define a patient cohort using criteria like:
• Admission and subject IDs
• Age ranges

• Refinement: Further refine the cohort using ICD (International Classification of

Diseases) and DRG (Diagnosis-Related Group) codes.
M I M I C - I V H E A LT H C A R E D A T A S E T

Data Extraction:

• Extracts case attributes such as:

• Gender
• Age
• Admission type
• Diagnostic/procedural codes

• Data Levels: Users can choose between subject-level or admission-level data.

Flexibility:

• Optionally save cohort and case attribute data as CSV files for further analysis.
E V E N T L O G G E N E R AT I O N
 PROCESS DISCOVERY & CONFORMANCE CHECKING
After generating the event log, our next objective was to understand the process flow it
depicted through process discovery. We used the ‘Interactive Data-Aware Heuristic Miner’
algorithm within the ProM tool, customizing threshold values (Frequency: 0.054, Dependency:
0.097, Binding: 0.643, Condition: 0.11) to fine-tune the mining process.

Outcome:

• This analysis resulted in a comprehensive process model, illustrated in Figure 1.

Alignment and Assessment:

• To ensure the accuracy of the discovered process model, we employed an optimal

alignment conformance technique. This technique compares the event log with the
discovered model and identifies three types of deviations:

1. Log Deviation (Yellow): Indicates where the log contains events not captured by the
model.

2. Model Deviation (Magenta): Shows where the model predicts events that are not found
in the log.
Visualization:
In Figure 1, the deviations are color-coded to help assess and understand the discrepancies and accuracies in the
process model. Yellow highlights log deviations, magenta shows model deviations, and green areas depict where the
log and model align perfectly. These visual markers provide a clearer picture of the actual workflow.
U N L O C K I N G PA T I E N T J O U R N E Y S
W I T H O B J E C T- C E N T R I C P R O C E S S
MINING IN MIMIC IV

Traditional process OCPM in MIMIC IV OCPM reveals MIMIC IV Petri net discovery
mining is linear and links events to dependencies, exploration uncovers patterns
misses healthcare multiple objects like bottlenecks, and identifies objects for detailed patient
complexities. patients and treatment and their journey analysis.
medications. variations. interactions,
creating Object-
Centric Event Logs.
SELECTION OF OBJECTS FOR MIMIC-IV

• Understanding patient flows in a hospital is key to optimizing healthcare delivery. This study examined
patient trajectories using an event log, focusing on three attributes: SubjectId, hadmId, and transferId.

• Subject ID (SubjectId): Uniquely identifies a patient across admissions.

• Hospital Admission ID (hadmId): Uniquely identifies a specific hospital stay.

• Transfer ID (transferId): Tracks patient transfers between different units or facilities, detailing the
movement of patients within the hospital. This includes transitions between departments, wards, or
specialized care units, providing insights into patient flow and resource utilization.

• The process started with extracting these attributes from an extensive XES event log, which records a
sequence of hospital events and activities.
F I N D I N G R E L AT I O N S B E T W E E N
OBJECTS

To determine relationships between objects in an

event log, the values for corresponding attributes are
examined to see how they relate to each other within
the log.
• Algorithm 1 is employed to check these relationships.
The algorithm assesses the relationship by comparing
two attributes at a time.
• The results reveal that a hadmId (hospital admission
ID) is associated with a single subjectId (patient ID)
but can have multiple transferIds, indicating various
transfers within the hospital for a single admission.
 F I N D I N G R E L AT I O N S B E T W E E N
OBJECTS

• Figure 2 illustrates the patient flow in the hospital system.

Each subjectId (S1-S5) is the source node at the top.
Directly below, hadmId (H1-H5) nodes show the patient’s
journey through the hospital. Branching from each
hadmId, multiple transferId nodes (T1.1, T1.2, etc.) depict
patient movement across departments. This structure
highlights patterns of divergence and convergence.

• For example, subjectId S2 is linked to hadmId H2, which

branches into five transfers, indicating extensive
departmental interactions. In contrast, other subjects
show fewer interactions, suggesting a more streamlined
treatment path. Figure 2: Object relations to one
another
V I S U A L I Z AT I O N S O F R E L AT I O N S
BETWEEN OBJECTS & ACTIVITIES
We find the objects and activities
connection in figure 3.

It is noted that `hadmID' and

`transferId' are closely
associated with nearly all
activities out of 40, whereas
`subjectID' is linked to only a
limited number of activities.

`subjectID' represents the

patient ID, directly connected to
object such as `hadmID', but not
directly to activities.

Figure 3: Convergence & Divergence

between the objects & activities.
D I S C OV E RY I N O B J E C T C E N T R I C
PETRI NET MODEL
We discover the Object centric petri Net model by using the PM4PY tool. Our focus to find the
objects(SubjectID, HadmID, TransferID) relations within the whole process execution. Although the process
model is extensive in scale, it is more clearly visualized than the traditional process discovery approach.
When all activities are covered, our model may resemble spaghetti-like complexity, but with the application
of OCPM, our model encompasses all paths and activities while remaining clear and comprehensible.
 A N A LY S I S & F I N D I N G S

1. We transformed the dataset from a relational database into an executable object-centric event log,
focusing on tracking specific objects in the patient journey.

2. Using OCPM, we investigated the transformed log and identified various relationships between
objects, including one-to-many and one-to-one connections. Figure 5 illustrates these connections,
showing that patients are assigned a diagnosis-based hospital upon each visit and subsequently
interact with multiple departments. Moreover, patients may also visit other hospitals as needed.

3. Our analysis using OCPM techniques provided a more comprehensive understanding of the patient
journey. These tools facilitated exploration of object involvement, revealing insights into how
patients interact with various entities throughout their healthcare experience.
A N A LY S I S & F I N D I N G S

Figure 5: Object visits through the

other objects
CONCLUSION

Successful Demonstration of OCPM:

● Applied Object-Centric Process Mining (OCPM) to the MIMIC IV dataset.

● Transitioned from relational data to event logs for OCPM analysis.

Key Insights:

● Identified and analyzed relationships between objects in the emergency department.

● Uncovered complex one-to-many and many-to-many interactions among 40 unique activities.

Event Log Transformation:

● Converted event logs into OCEL format, essential for OCPM exploration with PM4PY tools.

Discovery and Visualization:

● Applied discovery algorithms to generate an OCPM Petri net.

● Visual representation highlights process dynamics and object interactions in the emergency department.
 REFERENCES
[1] van der Aalst, W.M.: Object-centric process mining: Dealing with divergence and convergence in event data. In: Software Engineering and
Formal Methods: 17th International Conference, SEFM 2019, Oslo, Norway, September 18–20, 2019, Proceedings 17. pp. 3–25. Springer
(2019)
[2] van der Aalst, W.M.: Object-centric process mining: unraveling the fabric of real processes. Mathematics 11(12), 2691 (2023)
[3] van der Aalst, W.M., Berti, A.: Discovering object-centric petri nets. Fundamenta informaticae 175 (1-4), 1–40 (2020)
[4] Cremerius, J., Pufahl, L., Klessascheck, F., Weske, M.: Event log generation in mimic-iv research paper. In: International Conference on
Process Mining. pp. 302–314. Springer (2022)
[5] Edin, J., Junge, A., Havtorn, J.D., Borgholt, L., Maistro, M., Ruotsalo, T., Maaløe,L.: Automated medical coding on mimic-iii and mimic-iv: A
critical review and replicability study. In: Proceedings of the 46th International ACM SIGIR Conference on Research and Development in
Information Retrieval. pp. 2572–2582 (2023).
[6] Ghahfarokhi, A.F., Park, G., Berti, A., van der Aalst, W.M.: OCEL: A standard for object-centric event logs. In: European Conference on
Advances in Databases and Information Systems. pp. 169–175. Springer (2021).
[7].Van Der Aalst, W.: Process mining: Overview and opportunities. ACM Transactions on Management Information Systems (TMIS) 3(2), 1–
17 (2012)
[8] Ohnson, A., Bulgarelli, L., Pollard, T.,Horng, S., Celi, L.A., Mark, R.: Mimic-iv. PhysioNet. Available online at:
https://physionet.org/content/mimiciv/1.0/(accessed August 23, 2021) pp. 49–55 (2020).
[9] Mannhardt, F., De Leoni, M., Reijers, H.A.: Heuristic mining revamped: an interactive, data-aware, and conformance-aware miner. In: 15th
International Conference on Business Process Management (BPM 2017). pp. 1–5. CEUR-WS. org (2017)
10.github
 T H A N K YO U
FOR YO U R
PAT I E N C E