Mimic Table Relationships

MIMIC-IV Table Relationships

Understanding the identifier hierarchy and table relationships is essential for correct query construction. Incorrect joins can cause data duplication or missing records.

When to Use This Skill

• Writing complex queries joining multiple tables
• Understanding why queries return unexpected row counts
• Debugging duplicate or missing data issues
• Learning MIMIC-IV data structure

Identifier Hierarchy

subject_id (patient)
    └── hadm_id (hospital admission)
            └── stay_id (ICU stay)
                    └── Events (chartevents, labevents, etc.)

subject_id

• Unique per patient
• Persists across all hospitalizations and ICU stays
• Links to: patients table

hadm_id

• Unique per hospital admission
• One patient can have multiple hadm_ids (readmissions)
• Links to: admissions, diagnoses_icd, prescriptions, most lab/hospital tables

stay_id

• Unique per ICU stay
• One hospital admission can have multiple stay_ids (ICU readmission)
• Links to: icustays, chartevents, ICU-specific tables

Core Table Relationships

Hospital Module (mimiciv_hosp)

patients             -- 1 row per subject_id
    |
    +-- admissions   -- 1 row per hadm_id
    |       |
    |       +-- diagnoses_icd
    |       +-- procedures_icd
    |       +-- prescriptions
    |       +-- labevents
    |       +-- microbiologyevents
    |
    +-- transfers    -- Multiple per hadm_id (ward movements)

ICU Module (mimiciv_icu)

icustays            -- 1 row per stay_id
    |
    +-- chartevents  -- Vitals, assessments
    +-- inputevents  -- Medications, fluids
    +-- outputevents -- Urine, drains
    +-- procedureevents
    +-- datetimeevents

Common Join Patterns

Patient -> Hospital -> ICU

SELECT p.subject_id, a.hadm_id, ie.stay_id
FROM mimiciv_hosp.patients p
INNER JOIN mimiciv_hosp.admissions a
    ON p.subject_id = a.subject_id
INNER JOIN mimiciv_icu.icustays ie
    ON a.hadm_id = ie.hadm_id;

Labs to ICU Stay (Time-Bounded)

-- Labs drawn during ICU stay only
SELECT ie.stay_id, le.charttime, le.valuenum
FROM mimiciv_icu.icustays ie
INNER JOIN mimiciv_hosp.labevents le
    ON ie.hadm_id = le.hadm_id
    AND le.charttime >= ie.intime
    AND le.charttime <= ie.outtime;

Labs Within N Hours of ICU Admission

-- First 24 hours
SELECT ie.stay_id, le.charttime, le.valuenum
FROM mimiciv_icu.icustays ie
INNER JOIN mimiciv_hosp.labevents le
    ON ie.hadm_id = le.hadm_id
    AND le.charttime >= ie.intime
    AND le.charttime <= DATETIME_ADD(ie.intime, INTERVAL 24 HOUR);

Critical Join Warnings

1. Hospital Labs Duplicate Across ICU Stays

If a patient has multiple ICU stays in one hospitalization, joining labs by hadm_id only will duplicate lab values:

-- WRONG: Duplicates labs for patients with multiple ICU stays
SELECT ie.stay_id, le.*
FROM mimiciv_icu.icustays ie
INNER JOIN mimiciv_hosp.labevents le
    ON ie.hadm_id = le.hadm_id;  -- No time filter!

-- CORRECT: Add time bounds
SELECT ie.stay_id, le.*
FROM mimiciv_icu.icustays ie
INNER JOIN mimiciv_hosp.labevents le
    ON ie.hadm_id = le.hadm_id
    AND le.charttime BETWEEN ie.intime AND ie.outtime;

2. Derived Tables Already Filtered

Many mimiciv_derived tables are pre-joined to ICU stays:

-- These already have stay_id and time-bounded data
SELECT * FROM mimiciv_derived.vitalsign;  -- Already per stay_id
SELECT * FROM mimiciv_derived.chemistry;  -- Has subject_id and hadm_id

3. Multiple Measurements Per Time Point

Aggregate or select appropriately:

-- Get worst GCS per hour
SELECT stay_id,
       DATETIME_TRUNC(charttime, HOUR) AS hour,
       MIN(gcs) AS worst_gcs
FROM mimiciv_derived.gcs
GROUP BY stay_id, DATETIME_TRUNC(charttime, HOUR);

Cardinality Reference

Example: Verify Join Correctness

-- Check for unexpected duplicates
WITH joined AS (
    SELECT ie.stay_id, COUNT(*) AS n_labs
    FROM mimiciv_icu.icustays ie
    INNER JOIN mimiciv_hosp.labevents le
        ON ie.hadm_id = le.hadm_id
    GROUP BY ie.stay_id
)
SELECT
    COUNT(*) AS n_stays,
    AVG(n_labs) AS avg_labs_per_stay,
    MAX(n_labs) AS max_labs  -- Very high = possible duplication
FROM joined;

BigQuery vs DuckDB Syntax

MIMIC concepts are written for BigQuery. Key syntax differences (table names use the same canonical schema.table format on both backends):

Table names like mimiciv_hosp.patients and mimiciv_icu.icustays work on both backends.

References

• MIMIC-IV Documentation: https://mimic.mit.edu/docs/iv/
• Johnson AEW et al. "MIMIC-IV, a freely accessible electronic health record dataset." Scientific Data. 2023.