# AI in Emergency Medicine - Turkiye — Data

来源：https://www.kaggle.com/competitions/ai-emergency-medicine-turkiye/data

采集时间：2026-10-19T08:13:13Z

## Dataset Overview

This competition provides a large-scale tabular clinical dataset comprising **1,000,000 emergency 
department encounters** from 25 cities across all NUTS-2 statistical regions of Turkiye, spanning 
January 2019 through December 2024. The dataset is designed to support research-grade predictive 
modeling for two binary clinical outcome targets.

| Property | Value |
|----------|-------|
| Total Encounters | 1,000,000 |
| Training Set | 700,000 encounters |
| Test Set | 200,000 encounters |
| Supplemental Set | 100,000 encounters |
| Feature Columns | 135 (train) / 132 (test) |
| Time Period | January 2019 through December 2024 |
| Geographic Coverage | 25 cities, all NUTS-2 regions of Turkiye |
| Primary Target Prevalence | ~41% |
| Secondary Target Prevalence | ~14% |

---

## Files

### train.csv
The primary training dataset. Contains **700,000 rows** and **135 columns** including both 
target variables (`adverse_outcome` and `readmission_30d`). Use this file to build and 
validate your models.

### test.csv
The competition test dataset. Contains **200,000 rows** and **132 columns**. The two target 
columns have been removed. You must predict both targets for all 200,000 encounters and 
submit them in the format described on the Evaluation page.

### sample_submission.csv
A correctly formatted submission file containing all test `encounter_id` values with 
placeholder predictions of 0 for both targets. Use this as the structural template for 
your submission file. Do not alter the `encounter_id` column.

### supplemental_data.csv
An additional **100,000 labeled encounters** using the same schema as `train.csv`, 
including both target variables. These records are not part of the train/test split. 
Participants may use this file freely for any purpose, including additional model training, 
pseudo-labeling, calibration, or external validation.

### hospital_reference.csv
Infrastructure metadata for the 25 cities represented in the dataset. Join to the main 
data on the `city` column.

| Column | Description |
|--------|-------------|
| city | City name |
| nuts2_region | NUTS-2 region name |
| latitude | City latitude (decimal degrees) |
| longitude | City longitude (decimal degrees) |
| population_estimate | City population estimate |
| num_state_hospitals | Number of state hospitals |
| num_university_hospitals | Number of university hospitals |
| num_private_hospitals | Number of private hospitals |
| num_city_hospitals | Number of city hospitals |
| num_training_research_hospitals | Number of training and research hospitals |
| total_icu_beds | Total ICU beds across the city |
| total_ed_beds | Total ED beds across the city |
| helicopter_ems_available | 1 if helicopter EMS is available |
| stroke_center | 1 if a certified stroke center is present |
| trauma_center_level | Level I, Level II, or Level III |

### icd10_reference.csv
Reference table for the primary ICD-10 diagnosis codes used in the dataset. Join to the 
main data on the `primary_icd10_code` column.

| Column | Description |
|--------|-------------|
| icd10_code | ICD-10 code |
| diagnosis_description | Full text diagnosis name |
| icd10_chapter | Chapter letter |
| chapter_name | Chapter description |
| typical_los_days | Typical inpatient length of stay (days) |
| typical_mortality_pct | Typical in-hospital mortality rate (%) |

---

## Feature Groups

### Temporal Features
Covers arrival year, month, day, hour, day of week, season, night shift flag, weekend flag, 
pandemic period label, and COVID-19 period binary flag.

### Geographic and Institutional Features
City, NUTS-2 region, latitude, longitude, urban/rural classification, hospital type, 
annual ED volume, and bed count category.

### Patient Demographics
Age, age group, gender, pediatric flag, geriatric flag, height, weight, and BMI.

### Socioeconomic Features
Insurance type (SGK, Private, Green Card, Uninsured, Emekli Sandigi) and a composite 
socioeconomic status index scored 1 to 10.

### Triage and Presentation
Arrival mode, Manchester Triage System level (1 to 5), triage label, and chief complaint 
(30 categories).

### Vital Signs
Heart rate, systolic and diastolic blood pressure, mean arterial pressure, pulse pressure, 
respiratory rate, temperature, SpO2, Glasgow Coma Scale total, and NRS pain score.

### Laboratory Values
29 laboratory columns covering complete metabolic panel, complete blood count, coagulation 
studies, cardiac biomarkers, inflammatory markers, and organ function tests. See the 
missingness section below for columns with elevated missing rates.

### Comorbidities
18 binary comorbidity columns (prefix: `comorbidity_`) plus the computed Charlson 
Comorbidity Index score.

### Medications
10 binary medication class columns (prefix: `med_`) plus a polypharmacy indicator for 
patients on 5 or more drug classes.

### ED Procedures and Interventions
14 binary procedure columns (prefix: `proc_`), IV fluid volume in mL, specialist 
consultation flag, and consulting specialty.

### Operational Times
Triage-to-physician time (minutes) and total ED length of stay (minutes).

### COVID-19 Features
COVID-19 tested flag, positive result flag, and vaccination status flag.

### Clinical Severity Scores
Computed approximations of MEWS (Modified Early Warning Score), NEWS2 (National Early 
Warning Score 2), and SOFA (Sequential Organ Failure Assessment).

---

## Missing Data

Seven laboratory columns contain **informative missingness (MNAR: Missing Not At Random)**. 
These tests are selectively ordered for higher-acuity patients, meaning that the fact a 
test was not ordered is itself clinically meaningful. Treating these columns as Missing 
Completely At Random (MCAR) will underperform relative to approaches that model missingness 
explicitly.

**Recommended approach:** Before imputation, create a binary indicator feature 
`{column}_missing` for each of the following columns.

| Column | Approximate Missing Rate | Clinical Reason for Selective Ordering |
|--------|--------------------------|----------------------------------------|
| troponin_i_ngml | ~35% | Ordered for suspected cardiac events |
| bnp_pgl | ~40% | Ordered for suspected heart failure |
| d_dimer_mgfl | ~42% | Ordered for suspected thromboembolism |
| lactate_mmoll | ~25% | Ordered for suspected shock or sepsis |
| lipase_ul | ~38% | Ordered for suspected pancreatitis |
| procalcitonin_ngl | ~30% | Ordered for suspected bacterial infection |
| albumin_gdl | ~28% | Ordered for nutritional or hepatic assessment |

---

## Important Notes for Participants

**On the `disposition` column:** This column is present in `train.csv` and 
`supplemental_data.csv`. It reflects the ED discharge decision (Discharged Home, 
Admitted to Ward, Admitted to ICU, Transferred Out, Left AMA, Died in ED) and is a 
direct post-decision proxy for both target variables. **Using this column as a predictor 
constitutes data leakage and will result in disqualification.** It is provided solely 
for exploratory data analysis.

**On derived columns:** The columns `map_mmhg`, `pulse_pressure_mmhg`, `hematocrit_pct`, 
and `anion_gap_meql` are mathematically derivable from other features in the dataset. 
Participants may use them as-is or re-engineer them.

**On the COVID-19 period:** Encounters from 2020 and 2021 exhibit structured distribution 
shift in testing patterns, presentation volumes, and disposition rates relative to 
non-pandemic years. The `covid_period_flag` column marks these encounters explicitly. 
Participants may find it valuable to model this period separately or include it as an 
interaction term.

**On temporal validation:** A time-based train/validation split (e.g., training on 
2019-2022 and validating on 2023-2024) is recommended as a more realistic evaluation 
strategy than random k-fold cross-validation, as it better simulates prospective 
deployment conditions.