---
title: 'EHRWorld-110K: Patient-Centric Clinical Trajectories'
url: https://www.emergentmind.com/topics/ehrworld-110k
type: topic
---

# EHRWorld-110K: Patient-Centric Clinical Trajectories

EHRWorld-110K is a large-scale, patient-centric longitudinal clinical dataset derived from the MIMIC-IV electronic health record (EHR) database, designed explicitly to facilitate research on temporal modeling, simulation, and counterfactual reasoning in the context of hospitalization episodes. It comprises over 110,000 de-identified hospitalization trajectories with richly structured, time-stamped clinical events and static annotations, tailored for the development and evaluation of medical world models under long-horizon, intervention-driven dynamics [2602.03569].

## 1. Cohort Definition and Dataset Construction

EHRWorld-110K construction proceeds in three explicit stages:

- **Patient cohort selection** targets all adult (age ≥ 18) inpatient stays within MIMIC-IV (2008–2019) containing at least one medication administration and at least one laboratory or vital-sign event. Exclusions are applied to:
    - Admissions with zero therapy events,
    - Episodes with total event count below the 10th percentile or above the 90th percentile (to suppress uninformative and outlier stays),
    - Administrative duplicates and incomplete records.
- **Time span and observation window** represent each episode as a trajectory from admission ($\tau_1$) to discharge ($\tau_T$), omitting episodes under one day or above the 90th-percentile length of stay (approximately 100 days). There is no imposed maximal episode length beyond data-driven filtering.
- **Scale**: The resultant dataset comprises $N_{visits} = 110,513$ hospitalization episodes and approximately $N_{events} = 17.5$ million individual, timestamped events. Episodes correspond to de-identified patients (grouped by subject\_id), each potentially contributing multiple visits.

Expressed mathematically, with $T_i$ as the number of visits for patient $i$:
$$
N_{visits} = \sum_{i=1}^{N_{pat}} T_i = 110,513
$$
and with $E_j$ the number of events in episode $j$:
$$
N_{events} = \sum_{j=1}^{110,513} E_j \approx 17.5 \times 10^6
$$

## 2. Structure and Content of Each Trajectory

Each of the 110,513 episodes is encoded as a JSON object containing both static and longitudinal information:

- **Static attributes** (episode-level context) include:
    - Demographics: age (in years), gender, allergy history.
    - Hierarchical diagnoses: one "Primary Diagnosis" and an ordered set of "Secondary Diagnoses," each annotated with free-text "Content" and an associated "Reason" (both extracted from discharge summaries).
- **Time-indexed event sequence:** for $t = 1 \ldots T$ (admission to discharge), a chronologically ordered list captures:
    - **Inquiry events** ($A_{inq}$): laboratory tests, vital signs, physical examination findings. Each is recorded as 
      ```
      {"code": <test_code>, "timestamp": τ_t, "value": <numeric_or_categorical>, "units": <units>}
      ```
    - **Intervention events** ($A_{int}$): medication administrations or procedures, formatted as 
      ```
      {"code": <med_code_or_proc_code>, "timestamp": τ_t, "dose": <dosage>, "route": <route>}
      ```
    - All events possess explicit raw timestamps ($\tau_t$), absolute or relative to admission, and are grouped as $A_t = A_{inq} \cup A_{int}$.

The schematic JSON structure for an episode is:
```json
{
  "static": {
    "age": ...,
    "gender": ...,
    "allergies": ...,
    "diagnoses": {
      "primary": {"content": ..., "reason": ...},
      "secondary": [{ "content": ..., "reason": ... }, ...]
    }
  },
  "events": [
    { "timestamp": τ_1, "inquiries": [ ... ], "interventions": [ ... ] },
    ...
  ]
}
```

## 3. Annotation Workflow and Quality Control

Key steps in processing and annotation for EHRWorld-110K are as follows:

- **Static information extraction:** Large language model (Qwen3-235B-Instruct) parses discharge summaries via prompts enforcing a rigid JSON schema. Only diagnoses actively treated during the indexed hospitalization are extracted; coincidental or historical comorbidities are omitted.
- **Event cleaning:** Administrative duplicates and entries lacking essential fields are removed. Only log entries with valid execution timestamps and values (for inquiries or interventions) are retained.
- **Normalization and missingness:** Laboratory and vital sign measurements are preserved in native clinical units; no global normalization (e.g., z-scoring) is performed during data release. Episodes with missing allergies or demographic attributes are excluded.
- **Statistical filtering:** To promote stability in downstream modeling, stays with length of stay (LOS) or event count outside the empirical 10th–90th percentile range are removed.
- **Patient-centric data split:** Stratified at subject\_id, ensuring all visits for a given patient fall exclusively in either training or test partitions. The held-out test set contains 579 episodes ($\approx$ 0.5% of all data), spanning 1,043 distinct primary/secondary diagnostic profiles.

## 4. Statistical Properties and Distributions

Key empirical statistics (expressed in LaTeX) characterize the EHRWorld-110K corpus:

- **Age:**
    - $\mu_{age} = \frac{1}{N_{pat}} \sum_{i=1}^{N_{pat}} age_i \approx 58.0~\text{years}$
    - $\sigma_{age} \approx 19.2~\text{years}$
    - Age distribution ranges from 18 to 90+, peaking at 50–60 years.
- **Length of Stay (LOS):**
    - For episode $j$, LOS in days is $\ell_j$.
    - $\mu_{LOS} = \frac{1}{N_{visits}} \sum_{j=1}^{N_{visits}} \ell_j \approx 5.2~\text{days}$
    - $\sigma_{LOS} \approx 4.8~\text{days}$
    - Max LOS $> 100$ days.
- **Event Counts per Episode:**
    - For episode $j$, $e_j = |\mathbf{E}_j|$.
    - $\mu_e = \frac{1}{N_{visits}} \sum_{j=1}^{N_{visits}} e_j \approx 158.5$
    - $\sigma_e \approx 115.3$
    - Distribution is right-skewed: some episodes exceed 1,000 events.
- **Event Intensity (per-day event rate):**
    - Defined as $\frac{e_j}{\ell_j}$, averaging $\approx 30$ events/day, with spectrum from $<10$ to $>200$ (lower in routine, higher in intensive care episodes).

## 5. Intended Use, Scope, and Limitations

EHRWorld-110K is released solely for research into longitudinal patient-level modeling and simulation. The following guidelines and caveats apply:

- **Privacy and compliance:** All patient information is de-identified under the MIMIC-IV IRB protocol. The risk of re-identification is nominal; however, use requires compliance with the MIMIC data use agreement.
- **Limitations:**
    - Autoregressive modeling approaches may accumulate errors over extended trajectory rollouts.
    - The dataset omits unstructured clinical narrative text apart from discharge summaries (i.e., progress notes and clinician free-text are excluded).
    - No explicit outcome labels (e.g., mortality or readmission) are provided; EHRWorld-110K represents "pure" trajectory data, not a supervised prediction corpus.
- **Biases:** The underlying population reflects the demographic and care patterns of a US academic hospital system, potentially encoding and perpetuating those patterns in simulated outputs.
- **Recommended applications:** The dataset is suitable for:
    - Training and evaluating sequential models mapping (time, action) pairs to resulting observations,
    - Supporting counterfactual policy-learning research if users define their own reward functions,
    - Model development, benchmarking, and ablation studies in world modeling and simulation.  
It is not intended for prospective deployment in live clinical decision support without further validation.

## 6. Context and Implications for World Modeling

EHRWorld-110K was developed to address shortcomings in applying traditional large language models to dynamic, temporally evolving clinical simulation. Evaluations demonstrate that sequence models trained directly on causally grounded, longitudinal observation/intervention data—such as that provided by EHRWorld-110K—exhibit greater stability, more consistent modeling of rare or clinically significant events, and superior simulation fidelity over long horizons compared to LLMs that rely solely on static medical knowledge [2602.03569]. A plausible implication is that high-fidelity, patient-centric trajectory datasets are indispensable for robust, trustworthy clinical world modeling in sequential settings.

Source: https://www.emergentmind.com/topics/ehrworld-110k