Pith. sign in

REVIEW 3 major objections 5 minor 26 references

Curation and Analysis of MIMICEL -- An Event Log for MIMIC-IV Emergency Department

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A public event log now maps 425,000 emergency-department stays, extracted from MIMIC-IV-ED data and ready for process mining.

desk verdict MIMICEL is a genuinely useful public event log with code and validation, but the synthetic triage timestamp makes the 'no errors introduced' claim untenable and turns the triage-to-discharge statistics into arithmetic on length of stay. read the letter →

arxiv 2505.19389 v1 pith:35ZO7BGH submitted 2025-05-26 cs.DB

Jia Wei , Chun Ouyang , Bemali Wickramanayake , Zhipeng He , Keshara Perera , Catarina Moreira This is my paper · ORCID
classification cs.DB
keywords MIMIC-IV-EDeventlogprocessminingemergencydepartmentpatientflowdataqualityXESlengthofstay
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper's central claim is that MIMICEL, an event log curated from the public MIMIC-IV-ED emergency-department database, captures the end-to-end patient journey through the ED well enough for process-mining analysis. It reports that extraction followed a systematic nine-step guideline, yielding 425,028 ED stays with 7,568,824 events across six activity types, and that technical validation found no extraction-introduced errors: the quality problems detected were inherited from the source data. If this is right, process-mining researchers get a ready-made, reproducible dataset for studying patient flow, acuity-driven care intensity, length of stay, and crowding, without rebuilding extraction pipelines. The paper's demonstration analyses show that higher-acuity patients receive more frequent and tightly timed vital-sign and medication cycles, while lower-acuity patients are more often discharged directly after triage.

What carries the argument

The machinery that carries the argument is the mapping from the throughput phase of the ED-crowding conceptual model to the MIMIC-IV-ED tables, which fixes the process boundaries and six activity types, together with the nine-step extraction guideline that determines case IDs, timestamps, attributes, and invalid-case filtering. The pivotal mechanism is the temporal ordering of each stay's events into a sequence, with the invented triage event inserted one second after entry. That ordering is what turns static relational snapshots into an event log suitable for process discovery and performance analysis. A secondary mechanism is the quality-assessment framework used to certify that the extraction did not introduce errors.

What would settle it

Obtain real triage times for a sample of visits, either from a future version of the source database that records triage timestamps or from a chart audit, and compare them with the synthetic entry-time-plus-one-second values; if the gap is more than a few minutes, the paper's triage-to-discharge intervals and any process map containing triage as a distinct step would need revision.

Watch

Extended reading notes

Core claim

The core discovery is a publicly reusable event log derived from the MIMIC-IV-ED relational tables. The authors map six activities—Enter the ED, Triage in the ED, Vital sign check, Medicine reconciliation, Medicine dispensation, and Discharge from the ED—onto the source tables, using each stay's intime and outtime as the entry and discharge boundaries, chart times from the vital-sign and medication tables, and one synthetic timestamp: triage is set to entry time plus one second because the triage table has no timestamps. Each stay_id is a case and all other source fields become case or event attributes. Validation under a systematic quality framework reports missing acuity values in 1.64% of cases, 15.1% of home-discharge cases retaining a hospital admission ID, 29% of pain values outside the 0-10 range, and mixed timestamp granularities and temperature units; the paper attributes all of these to the source dataset. Demonstration analyses then reveal monotonic acuity trends, such as consecutive vital-sign-check intervals rising from 30 minutes at acuity 1 to 120 minutes at acuity 5, and a triage-to-discharge path that becomes more common and faster as acuity drops.

Load-bearing premise

Everything that involves triage timing rests on the assumption that the triage event happened exactly one second after the patient entered the ED, because the source data records no triage timestamp.

Editorial extensions

If this is right

  • Researchers can apply process-mining discovery, conformance, and performance-analysis techniques directly to MIMICEL without re-extracting from the relational source.
  • The acuity analyses imply a measurable gradient in care intensity: from acuity 1 to acuity 5, the median interval between consecutive vital-sign checks rises from 30 to 120 minutes.
  • The triage-to-discharge direct path rises from 1.23% of acuity-1 cases to 18% of acuity-5 cases, with median duration falling from 126 to 60 minutes.
  • For high-acuity patients with prolonged length of stay, vital-sign self-loops occur in 88% of cases and transitions between medication and monitoring take about twice as long as in fast-tracked urgent cases.
  • Crowded ED conditions are associated with a higher proportion of admitted patients: about one-third of admitted patients were treated in crowded conditions versus one-fourth of patients discharged home.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The synthetic triage timestamp makes the triage activity a proxy; researchers should test whether process maps change when triage is removed or given a realistic delay.
  • The 15.1% of home discharges that retain a hospital admission ID could corrupt any downstream linkage between ED events and inpatient data, so users should clean disposition-dependent attributes before joining.
  • Because medication and vital-sign timestamps are minute-granular while entry and discharge are second-granular, sub-minute durations are unobservable and overlap-based crowding estimates are approximate.
  • MIMICEL could serve as a benchmark corpus for process-mining algorithm evaluation, since its ground truth is the public source data and its quality issues are documented and inherited.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper describes the curation of MIMICEL, an event log extracted from the MIMIC-IV-ED dataset, following the nine-step guideline of Jans et al. The log records six activity types for 425,028 ED stays, is provided in CSV and XES formats on PhysioNet, and is accompanied by publicly available extraction and validation code. The authors validate the log with the DaQAPO framework and present three demonstration analyses (acuity-based, length-of-stay, and crowdedness) to illustrate the log's utility for ED process mining. The central claim is that MIMICEL captures an end-to-end patient journey with high fidelity and that no errors were introduced during extraction.

Significance. If the concerns below are addressed, MIMICEL would fill a genuine gap: it provides a large, publicly available, reproducible event log for ED process mining and makes MIMIC-IV-ED accessible to the process-mining community. The authors follow a systematic extraction guideline, publish their SQL and Python scripts, and ship the log in two standard formats, which are concrete strengths. However, the paper's central validation claim is currently overstated because one activity is synthetic and because duplicate discharge events are treated as faithful records. The dataset remains valuable, but its description and the demonstrated analyses need to be made more cautious and precise.

major comments (3)
  1. [Step 7, Table 2, Figure 4, Tables 9-10] The 'Triage in the ED' activity is assigned the synthetic timestamp intime + 1 second because the MIMIC-IV-ED triage table has no timestamp. This makes triage a deterministic relabeling of 'Enter the ED' rather than an observed event. Consequently, the 'Triage to Discharge' path in Figure 4 and Tables 9-10 can occur only for cases with no other recorded activity between entry and discharge, and the reported interval is exactly outtime - intime - 1 second (i.e., the length of stay minus one second). The paper discloses this construction, but it does not assess how the acuity-based conclusions are affected; for example, the statement that 'lower-acuity patients are more likely to be discharged directly after triage' is an interpretation of an artifact. Please either remove triage-based analyses, explicitly relabel them as 'entry-to-discharge with no intermediate recorded activity,' or provide evidence that intime+1 is a clinically valid proxy for triage time.
  2. [Technical Validation, 'Remark' and 'Repeated activities with identical timestamps'] The statement that 'no errors were introduced during the log extraction process' is contradicted by two extraction choices. First, the synthetic triage event is introduced by the authors, not inherited from MIMIC-IV-ED. Second, converting each diagnosis row into a separate 'Discharge from the ED' event causes 60.13% of cases to have multiple discharge events at identical timestamps; this is an extraction-induced representation, not a source-data issue. These choices inflate the total event count (7,568,824) and can distort process-discovery results, including self-loops and path frequencies. Please revise the validation claim to distinguish source-data issues from extraction-induced representations, and either aggregate discharge events per stay or publish a variant of the log with one discharge event per case.
  3. [Usage Notes, 'Acuity-based analysis' and 'LoS-driven analysis'] The two main demonstration analyses rely on threshold-like choices that are derived from the same dataset: the 500-minute LoS threshold is the 75th percentile of the data, and the crowdedness threshold of 12 simultaneous patients is the 75th percentile of the simultaneous-patient distribution. The paper presents these as descriptive cohort comparisons, which is acceptable, but it should explicitly state that the thresholds are data-derived and that the resulting 'normal' versus 'prolonged' or 'crowded' versus 'non-crowded' categories are relative to this cohort rather than externally validated. Otherwise readers may over-interpret the quantitative claims in Tables 9-12.
minor comments (5)
  1. [Table 6] The years in the timestamp column (2165, 2169, 2170) appear to be typos; please replace them with the actual years present in MIMIC-IV-ED, and correct 'W ALK IN' to 'WALK IN'.
  2. [Usage Notes, 'Acuity-based analysis'] In the bullet list, 'In contrast.' should be 'In contrast,' and in the fourth bullet the phrase '50 and 36 minutes (Table 10' is missing a closing parenthesis.
  3. [Technical Validation] The text says the incomplete_cases function was used to detect incomplete ED visits, but no result for this check is reported in Table 7 or in the narrative; please add the result or remove the mention.
  4. [Usage Notes, 'LoS-driven analysis'] The claim that 500 minutes 'aligns with the internationally recommended acceptable ED LoS of ≤8 hours' is not quite right: 8 hours is 480 minutes. Please correct the comparison or use a threshold of 480 minutes if the intent is to align with the cited recommendation.
  5. [Figure 4] The process map is dense and the thick red arrow may be difficult to read in a black-and-white print version; please increase font size or add a zoomed inset for the triage-to-discharge path.

Circularity Check

1 steps flagged · score 6.0 of 10

The synthetic triage timestamp makes 'Triage to Discharge' intervals equal length-of-stay minus one second by construction.

  1. self definitional [Methods, Step 7 (Table 2); Usage Notes, Acuity-based analysis (Tables 9-10)]
    "we assign an artificial timestamp to the triage activity by adding "one second" to the time when the patient enters the ED (i.e., intime of the edstays table) ... The path from triage to discharge is relatively uncommon, appearing in only 2.65% of cases (with acuity level 3). The median time interval between these two activities is 2.1 hours."

    By construction, timestamp(Triage in the ED) = intime + 1 second, so for any case in which Triage is immediately followed by Discharge, the reported 'Triage to Discharge' interval equals outtime - (intime + 1s) = LoS - 1s. Thus the Table 10 column 'Triage to Discharge' is not an independent measure of triage-to-discharge behavior; it is the case length of stay shifted by one second. The associated insight that lower-acuity patients have a shorter triage-to-discharge interval is therefore a restatement of the LoS distribution for that subset, not evidence about a real triage step. The construction is disclosed in Step 7, but Tables 9-10 and Figure 4 still present the synthetic event as an observed activity.

full rationale

MIMICEL is largely a faithful extraction of observed MIMIC-IV-ED records with transparent table-to-activity mapping, and the acuity, LoS, and crowdedness analyses are descriptive rather than predictive; the 500-minute and 12-patient thresholds are explicitly data-derived splits, not fitted predictions. The one place where a derived result is forced by definition is the synthetic 'Triage in the ED' event: because its timestamp is defined as intime + 1 second, the triage-to-discharge interval in Tables 9-10 is exactly LoS - 1s for the cases that exhibit that path, making that particular finding tautological. The Technical Validation remark that 'no errors were introduced during the log extraction process' overstates fidelity given that a synthetic event was inserted, but that is a validity concern rather than an additional circular step. No load-bearing self-citation chain exists. Score 6 reflects partial circularity: one analysis result reduces by construction, while the rest of the log and its analyses remain independent.

Assumptions & free parameters 3 free parameters · 3 assumptions · 1 invented entities

The central artifact, MIMICEL, depends on three free parameters (the one-second triage offset and two 75th percentile thresholds), three domain assumptions about table-to-activity mapping and timestamp granularity, and one invented entity (the synthetic triage activity). These are all documented, but they mean the event log is not a purely raw extraction; it embeds analyst choices that directly shape downstream analyses.

free parameters (3)
  • Triage timestamp offset = 1 second
    Chosen by hand to create an artificial timestamp for the 'Triage in the ED' activity, because the triage table in MIMIC-IV-ED has no timestamp. This offset directly defines the triage event time and therefore affects all triage-related analyses.
  • Length of stay threshold = 500 minutes
    Set as the 75th percentile of the LoS distribution in the extracted log. Used to classify cases as normal versus prolonged LoS in the LoS-driven analysis.
  • Crowdedness threshold = 12 simultaneously treated patients
    Set as the 75th percentile of the number of simultaneously treated patients in the ED. Used as the cutoff for labeling a visit as occurring in a crowded ED.
assumptions (3)
  • domain assumption The four cornerstone ED activities (arrival, triage/room placement, diagnostic evaluation/treatment, disposition) can be fully mapped to the MIMIC-IV-ED tables as listed in Table 1.
    Invoked in Methods, Step 3, where all tables in MIMIC-IV-ED are mapped to the cornerstone activities. If any table maps to more than one activity or if a key activity is absent from these tables, the event log would miss process steps.
  • ad hoc to paper The triage time is approximated by the ED admission time (intime) plus one second.
    Stated in Methods, Step 7, citing dataset documentation. This is a convenience assumption because no real triage timestamp exists, and it directly creates the synthetic triage event.
  • domain assumption Repeated activities with identical timestamps represent distinct medications, diagnoses, or measurements and should be kept as separate events.
    Discussed in Technical Validation for medicine reconciliation, medicine dispensation, and discharge events. This assumption determines event counts and affects process maps and path frequency analyses.
invented entities (1)
  • Triage in the ED activity
    purpose: Represent the triage step as an event in the event log, since the MIMIC-IV-ED triage table lacks a timestamp.
    The activity is not present in the source data with its own time; it is created by taking the patient's ED entry time and adding one second. The paper discloses this, but the activity has no real timestamp and could mislead users who treat it as a measured event.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Curation and Analysis of MIMICEL -- An Event Log for MIMIC-IV Emergency Department." pith.science (2026). https://pith.science/paper/35ZO7BGH

@misc{pith2026250519389,
  author       = {Pith},
  title        = {Pith review of: Curation and Analysis of MIMICEL -- An Event Log for MIMIC-IV Emergency Department},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/35ZO7BGH}},
  note         = {Machine review of arXiv:2505.19389}
}
read the original abstract

The global issue of overcrowding in emergency departments (ED) necessitates the analysis of patient flow through ED to enhance efficiency and alleviate overcrowding. However, traditional analytical methods are time-consuming and costly. The healthcare industry is embracing process mining tools to analyse healthcare processes and patient flows. Process mining aims to discover, monitor, and enhance processes by obtaining knowledge from event log data. However, the availability of event logs is a prerequisite for applying process mining techniques. Hence, this paper aims to generate an event log for analysing processes in ED. In this study, we extract an event log from the MIMIC-IV-ED dataset and name it MIMICEL. MIMICEL captures the process of patient journey in ED, allowing for analysis of patient flows and improving ED efficiency. We present analyses conducted using MIMICEL to demonstrate the utility of the dataset. The curation of MIMICEL facilitates extensive use of MIMIC-IV-ED data for ED analysis using process mining techniques, while also providing the process mining research communities with a valuable dataset for study.

Figures

Figures reproduced from arXiv: 2505.19389 by the authors.

Figure 1
Figure 1. An overview of the method for generating MIMICEL 2/17 [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Relationships between key tables in the MIMIC-IV-ED dataset 3/17 [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. A snippet of extracted MIMICEL in XES format example, the attribute hadm_id (hospital admission ID) will be null if a patient is not admitted to the hospital (e.g. when the disposition is “ADMITTED”). In addition, the number of incomplete patient ED visit records serves as an essential quality metric21. Each ED visit record in the MIMIC-IV-ED dataset is expected to include mandatory activities such as entering, tria… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: The process map of ED visits with the acuity level of 3 [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Distribution of ED visits in accordance with LoS and acuity levels [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Comparison of the process between Q1 Urgent Quick zone and Q4 Urgent Slow zone Path Home Cohort (min) Admitted Cohort (min) Time duration of consecutive “Vital sign check” activities execution 120 64 Time duration taken by the path from “Vital sign check” to “Medicine …
Figure 7
Figure 7. Figure 7: Comparison of the process between non-crowded and crowded ED Path Crowded ED - Admitted (min) Crowded ED - Home (min) Time interval of consecutive “Vital sign check” activities 64 120 Time duration taken by the path from “Vital sign check” to “Discharge from the ED” 58…
Figure 8
Figure 8. Figure 8: The annotated process map of ED visits with the acuity level of 3 17/17 [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 25 canonical work pages

  1. [1]

    A., Mandingorra, J

    Ibanez-Sanchez, G., Celda, M. A., Mandingorra, J. & Fernandez-Llatas, C. Interactive process mining in emergencies. In Interactive process mining in healthcare, 165–180 (2021)

    work page 2021

  2. [2]

    & Aringhieri, R

    Duma, D. & Aringhieri, R. An ad hoc process mining approach to discover patient paths of an emergency department. Flex. Serv. Manuf. J.32, 6–34 (2018)

    work page 2018

  3. [3]

    Savioli, G.et al.Emergency department overcrowding: Understanding the factors to find corresponding solutions.J. Pers. Medicine12(2022)

    work page 2022

  4. [4]

    emergency nursing36, 303–310 (2010)

    Brenner, S.et al.Modeling and analysis of the emergency department at university of kentucky chandler hospital using simulations.J. emergency nursing36, 303–310 (2010)

    work page 2010

  5. [5]

    & Ferreira, D

    Rebuge, Á. & Ferreira, D. R. Business process analysis in healthcare environments: A methodology based on process mining.Inf. systems37, 99–116 (2012)

    work page 2012

  6. [6]

    & Matsatsinis, N

    Delias, P., Manolitzas, P., Grigoroudis, E. & Matsatsinis, N. Applying process mining to the emergency department. In Encyclopedia of business analytics and optimization, 168–178 (2014)

    work page 2014

  7. [7]

    Munoz-Gama, J.et al.Process mining for healthcare: Characteristics and challenges.J. Biomed. Informatics127, 103994 (2022)

    work page 2022

  8. [8]

    Martin, N.et al.Recommendations for enhancing the usability and understandability of process mining in healthcare.Artif. Intell. Medicine109, 101962 (2020)

    work page 2020

Show all 26 references
  1. [9]

    journal environmental research public health17, 6290 (2020)

    Cho, M.et al.Process mining-supported emergency room process performance indicators.Int. journal environmental research public health17, 6290 (2020). 10.van der Aalst, W. M.Process Mining: Data Science in Action(2016)

    2020

  2. [11]

    A., Mark, R., & Horng, S

    Johnson, A., Bulgarelli, L., Pollard, T., Celi, L. A., Mark, R., & Horng, S. Mimic-iv-ed (version 2.2).PhysioNet https://doi.org/10.13026/5ntk-km72 (2022)

    2022 doi

  3. [12]

    & Jouck, T

    Jans, M., Soffer, P. & Jouck, T. Building a valuable event log for process mining: an experimental exploration of a guided process.Enterp. Inf. Syst.13, 601 – 630 (2019). 15/17

    2019

  4. [13]

    P., Böttinger, E

    Remy, S., Pufahl, L., Sachs, J. P., Böttinger, E. & Weske, M. Event log generation in a health system: a case study. In International Conference on Business Process Management, 505–522 (Springer, 2020)

    2020

  5. [14]

    Sci.7, 302 (2017)

    Rojas, E.et al.Question-driven methodology for analyzing emergency room processes using process mining.Appl. Sci.7, 302 (2017)

    2017

  6. [15]

    Andrews, R.et al.Quality-informed semi-automated event log generation for process mining.Decis. Support. Syst.132, 113265 (2020)

    2020

  7. [16]

    T., ter Hofstede, A

    Andrews, R., Suriadi, S., Wynn, M. T., ter Hofstede, A. H. M. & Rothwell, S. Improving patient flows at st. andrew’s war memorial hospital’s emergency department through process mining. InBusiness Process Management Cases(2018)

    2018

  8. [17]

    R.et al.A conceptual model of emergency department crowding.Annals emergency medicine42, 173–180 (2003)

    Asplin, B. R.et al.A conceptual model of emergency department crowding.Annals emergency medicine42, 173–180 (2003)

    2003

  9. [18]

    IEEE Standard for eXtensible Event Stream (XES) for Achieving Interoperability in Event Logs and Event Streams

    IEEE. IEEE Standard for eXtensible Event Stream (XES) for Achieving Interoperability in Event Logs and Event Streams. IEEE Std 1849-20161–50 (2016)

    2016

  10. [19]

    L.et al.Physiobank, physiotoolkit, and physionet: components of a new research resource for complex physiologic signals.Circulation[online]101, e–215–e–220 (2000)

    Goldberger, A. L.et al.Physiobank, physiotoolkit, and physionet: components of a new research resource for complex physiologic signals.Circulation[online]101, e–215–e–220 (2000)

    2000

  11. [20]

    Berti, A., Van Zelst, S. J. & van der Aalst, W. Process mining for python (PM4Py): Bridging the gap between process- and data science. InProceedings of the ICPM Demo Track 2019, co-located with 1st International Conference on Process Mining (ICPM 2019), 13–16 (2019)

    2019

  12. [21]

    & Braekers, K

    Vanbrabant, L., Martin, N., Ramaekers, K. & Braekers, K. Quality of input data in emergency department simulations: framework and assessment techniques.Simul. Model. Pract. Theory91, 83–101 (2019)

    2019

  13. [22]

    Bose, J. C. J. C., Mans, R. S. & van der Aalst, W. M. Wanna improve process mining results?2013 IEEE Symp. on Comput. Intell. Data Min. (CIDM)127–134 (2013)

    2013

  14. [23]

    & Janssenswillen, G

    Martin, N., Van Houdt, G. & Janssenswillen, G. Daqapo: Supporting flexible and fine-grained event log quality assessment. Expert. Syst. with Appl.191, 116274 (2022)

    2022

  15. [24]

    & Van Nieuwenhuyse, I

    Vanbrabant, L., Braekers, K., Ramaekers, K. & Van Nieuwenhuyse, I. Simulation of emergency department operations: A comprehensive review of kpis and operational improvements.Comput. & Ind. Eng.131, 356–381 (2019)

    2019

  16. [25]

    Di Somma, S.et al.Overcrowding in emergency department: an international issue.Intern. Emerg. Medicin10, 171–175 (2015)

    2015

  17. [26]

    journal trauma, resuscitation emergency medicine20, 1–7 (2012)

    Rose, L.et al.Emergency department length of stay for patients requiring mechanical ventilation: a prospective observational study.Scand. journal trauma, resuscitation emergency medicine20, 1–7 (2012)

    2012

  18. [27]

    Healthcare10, 1625 (2022)

    Sartini, M.et al.Overcrowding in emergency department: causes, consequences, and solutions—a narrative review. Healthcare10, 1625 (2022). Appendix Interpretation of process maps The process maps in this study were created using Disco, a process mining tool 2 designed for visua...

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.