[MIMIC-IV v3.1] Cross-table inconsistencies in transfers and death-related fields

[chenhaoran2068]

Hello MIMIC team,

We performed read-only aggregate quality checks on MIMIC-IV v3.1 (hosp and ICU modules) after loading the official release files into PostgreSQL.

We reviewed the official table documentation and the related discussions/issues: #1406, #966, #1865, and #1945.

Aggregate observations

1. Transfer/admission subject mismatch

   • One transfers row has a non-null hadm_id that matches an admission, but transfers.subject_id differs from admissions.subject_id.
   • The row has eventtype='ED'.

2. Hospital death flag without an admission death timestamp

   • Eleven admissions have hospital_expire_flag=1 and deathtime IS NULL.
   • All eleven corresponding patients have a non-null patients.dod.

3. Admission death timestamp after discharge timestamp

   • There are 745 admissions where deathtime > dischtime.
   • All have hospital_expire_flag=1.
   • 1 difference is <= 1 minute; 17 are > 1 minute and <= 1 hour; 727 are > 1 hour and <= 24 hours; none are > 24 hours.

4. Patient death date differs from the admission death date

   • One admission has a calendar-date mismatch between patients.dod and admissions.deathtime.
   • The admission deathtime date is 2-7 days later.

Aggregate reproduction queries

SELECT count(*)
FROM mimiciv_hosp.transfers t
JOIN mimiciv_hosp.admissions a USING (hadm_id)
WHERE t.subject_id IS DISTINCT FROM a.subject_id;

SELECT count(*)
FROM mimiciv_hosp.admissions
WHERE hospital_expire_flag = 1 AND deathtime IS NULL;

SELECT count(*)
FROM mimiciv_hosp.admissions
WHERE deathtime > dischtime;

SELECT count(*)
FROM mimiciv_hosp.admissions a
JOIN mimiciv_hosp.patients p USING (subject_id)
WHERE a.deathtime IS NOT NULL
  AND p.dod IS NOT NULL
  AND a.deathtime::date <> p.dod;

Current conservative handling

We currently preserve all source values without imputation, timestamp clipping, ID replacement, or source-row deletion. We plan to expose explicit quality-status fields so downstream analyses can identify the affected records.

Questions

1. Is the transfers subject/admission mismatch a known MIMIC-IV v3.1 issue?
2. For this mismatch, should users treat hadm_id as unreliable, subject_id as unreliable, or retain and flag the relationship as unresolved?
3. Is preserving hospital_expire_flag, deathtime, and dod as independent signals the recommended approach?
4. Should deathtime values after dischtime be retained without correction?
5. Are there release-specific validation statistics or planned corrections for these categories?

No patient identifiers, row-level values, raw timestamps, or source records are included in this post.

Thank you.