Suspicion Of Infection

• The culture time if culture was obtained BEFORE antibiotic
• The antibiotic time if antibiotic was given BEFORE culture

This represents when infection was first clinically suspected.

Culture Matching Logic

Each antibiotic is matched to cultures in two directions:

Culture Before Antibiotic (Primary)

• Culture obtained within 72 hours before antibiotic start
• If multiple cultures, uses the EARLIEST culture before the antibiotic

Culture After Antibiotic (Secondary)

• Culture obtained within 24 hours after antibiotic start
• If multiple cultures, uses the EARLIEST culture after the antibiotic

Priority: Culture-before-antibiotic takes precedence when both exist.

Critical Implementation Notes

1. One Row Per Antibiotic: Each antibiotic prescription gets its own row, potentially matched to one culture. A single culture may be matched to multiple antibiotics.

2. Culture Positivity Not Required: Negative cultures still count as suspected infection — the flag captures clinical suspicion, not confirmed infection.

3. All Culture Types Included: Blood, urine, sputum, wound, CSF, etc. The specimen column identifies the type.

4. Systemic Antibiotics Only: Topical formulations (eye drops, ear drops, creams, ointments) must be excluded. The specific route codes vary by dataset.

General Limitations

1. Proxy for Clinical Suspicion: Not all antibiotic + culture pairs represent true infection suspicion. Routine screening cultures (e.g., weekly surveillance) paired with prophylactic antibiotics may be misclassified as suspected infection.

2. Time Window Is an Operationalization: The 72h/24h windows are from the Seymour et al. operationalization. There is no biological basis for these specific cutoffs — they are pragmatic choices that balance sensitivity and specificity.

3. Misses Antibiotic-Only or Culture-Only Events: Patients treated empirically without cultures sent, or cultures obtained without subsequent antibiotics, will not be flagged.

4. Does Not Distinguish Empiric vs Targeted Therapy: The concept captures the initial antibiotic-culture pairing regardless of whether the antibiotic was empiric (before results) or targeted (after susceptibility).

Dataset Availability

MIMIC-IV

Suspicion of infection is available as a pre-computed derived table. Materialize with:

m4 init-derived mimic-iv          # All derived tables including suspicion_of_infection

SELECT
    subject_id,
    stay_id,
    hadm_id,
    ab_id,                     -- Unique antibiotic ID per patient
    antibiotic,                -- Antibiotic name
    antibiotic_time,           -- When antibiotic started
    suspected_infection,       -- 1 if meets criteria, 0 otherwise
    suspected_infection_time,  -- Onset time of suspected infection
    culture_time,              -- When culture was obtained
    specimen,                  -- Culture specimen type
    positive_culture           -- 1 if culture positive, 0 if negative
FROM mimiciv_derived.suspicion_of_infection;

BigQuery users already have this table via physionet-data.mimiciv_derived.suspicion_of_infection without running init-derived.

MIMIC-IV implementation details:

• The derived tables originate from the MIT-LCP mimic-code repository. The full SQL query is in scripts/mimic-iv.sql.
• Antibiotics sourced from mimiciv_derived.antibiotic (which filters the prescriptions table for systemic routes, excluding topical routes: OU, OS, OD, AU, AS, AD, TP, and topical formulations like creams, gels, ophthalmic ointments).
• Cultures from mimiciv_hosp.microbiologyevents. Positive culture identified by non-null org_name excluding itemid 90856 ("NEGATIVE").
• stay_id is populated when antibiotic timing overlaps with an ICU stay. May be NULL for floor patients.
• Chart dates vs times: Microbiology cultures sometimes only have dates (not times). When charttime is null, the query falls back to chartdate with day-level matching (72h becomes 3 days, 24h becomes 1 day).

MIMIC-IV limitations:

• Prescription duplication: each ICU stay in a hospitalization gets a copy of all prescriptions for that admission, which the query handles via unique ab_id per patient.
• Antibiotic route filtering relies on MIMIC's route coding. Miscoded routes could include topical antibiotics or exclude systemic ones.

eICU

Suspicion of infection is not pre-computed in eICU. Both components must be derived from raw tables:

eICU limitations:

• Center variability in missingness: Medication documentation and culture practices vary substantially across the 208 hospitals. Some sites have near-complete medication records; others have significant gaps. This non-random missingness affects infection identification rates.
• Medication naming: drugname is free-text and varies across sites. The same antibiotic may appear as "Vancomycin", "VANCOMYCIN", "vancomycin 1g IV", etc. Building a reliable antibiotic identification filter requires extensive text matching and validation.
• Route coding: routeadmin also varies by site. Excluding topical routes requires site-aware filtering.
• Offset-based timing: Both drugstartoffset and culturetakenoffset are in minutes from unit admission (not absolute timestamps). The antibiotic-culture pairing logic must use offset arithmetic rather than datetime comparisons.
• No upstream antibiotic table: Unlike MIMIC (which has a derived antibiotic table that pre-filters systemic antibiotics), eICU requires building the antibiotic filtering from scratch.

An eICU script is not yet available.

Example: Infection Events Per Patient

SELECT
    subject_id,
    COUNT(*) AS n_suspected_infections,
    SUM(positive_culture) AS n_positive_cultures
FROM mimiciv_derived.suspicion_of_infection
WHERE suspected_infection = 1
GROUP BY subject_id
ORDER BY n_suspected_infections DESC;

Example: Most Common Antibiotics in Suspected Infection

SELECT
    antibiotic,
    COUNT(*) AS n_prescriptions,
    SUM(positive_culture) AS n_positive,
    ROUND(AVG(positive_culture), 2) AS positive_rate
FROM mimiciv_derived.suspicion_of_infection
WHERE suspected_infection = 1
GROUP BY antibiotic
ORDER BY n_prescriptions DESC
LIMIT 20;

Related Skills

• sepsis-3-cohort — Combines this concept with SOFA >= 2 for Sepsis-3 identification
• sofa-score — Organ dysfunction component of Sepsis-3
• sirs-criteria — Historical inflammatory response criteria (pre-Sepsis-3)

References

• Singer M et al. "The Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis-3)." JAMA. 2016;315(8):801-810.
• Seymour CW et al. "Assessment of Clinical Criteria for Sepsis." JAMA. 2016;315(8):762-774.