ADS599-Capstone/interim_data

Cohort Data

Processed ED patient cohort derived from MIMIC-IV. One row per ED visit (ed_stay_id). Built from cohort_base (BigQuery) with the following post-processing steps:

• AGAINST ADVICE discharge_location records removed — patient-driven departures are out of scope.
• Consecutive ED visits sharing a hadm_id collapsed into single rows; second ed_stay_id stored in ed_stay_id_2.
• stay_window_start/stay_window_end columns added covering ED arrival to final discharge.
• ed_boarding_time_min added: minutes in ED after admission decision (null for ED-only patients).

Intended as the primary cohort table for feature engineering and RL state construction.

Dispensed Meds

Dispensed medication records for all cohort patients across two phases and two populations.

ED phase (in_er=True): Pyxis dispenses for all patients (admitted and ED-only) while physically in the ED, derived from ed.pyxis.
Ward phase (in_er=False): eMAR administration records for admitted patients from ED departure up to ICU transfer, derived from hosp.emar.

event_txt filtered to administration and rate-change events only.

Medication names mapped to a 22-class drug vocabulary via regex. Each record includes minutes_into_stay (time since ED arrival) and time_step bucket (floor(minutes / time_block)).

ECG Data

ECG acuity labels derived from mimiciv_ecg.machine_measurements for cohort patients. Each of the 18 report columns is classified independently via regex with priority order: abnormal (2) > neutral (1) > normal (0) > empty (-1).

Row-level max across all 18 columns gives ecg_acuity: 0=normal, 1=neutral/unknown, 2=abnormal.

Patients with all 18 columns empty are assigned 1 (missing ECG ≠ normal ECG).

Multiple ECGs per ED stay resolved to one row: highest acuity kept, ties broken by earliest ecg_time.

Window covers ED arrival through hospital ward stay (capped at ICU transfer if applicable), with ±1 hour buffer for ECG machine clock drift.

Labs Data

Laboratory results from hosp.labevents for cohort patients during their stay window.

Grouped by category x fluid (19 unique combinations) rather than individual test label to reduce action space.

Each result includes the abnormal flag for worst-case aggregation. Intended as the primary lab state feature source.

Medrecon Data

Medication reconciliation records from ed.medrecon — medications the patient was taking prior to ED arrival. One row per medication per ED visit. Represents pre-arrival medication state, not actions taken during the visit.

Microbiology Data

Microbiology culture events from hosp.microbiologyevents for cohort patients during their stay window. Includes culture order time, specimen type, organism name, and antibiotic sensitivity results.

culture_ordered is the real-time state signal; culture_positive is a retrospective label only (~2% of results available before ED discharge).

OMR Data

Height and weight measurement records from hosp.omr for cohort patients — separated to take single measurements of height as an average of all available height measurements for a patient and retaining all separate weight measurements. Duplicates were dropped. Used to supplement ED and inpatient state features with recent baseline measurements.

Radiology Data

Radiology report text from mimiciv_note.radiology for cohort patients. Covers all imaging modalities (CXR, CT, MRI, ultrasound, etc.). Primary reports only (note_type=RR). Window covers ED arrival through hospital ward stay (capped at ICU transfer if applicable).
hadm_id is NULL for ED-only patients and populated for admitted patients. exam_name and cpt_code included from radiology_detail to identify imaging modality.

Triage_Vitals

Combined triage and vital signs time series — one row per (ed_stay_id, charttime) timestep. Triage records (ed.triage) provide the baseline reading at ed_intime; subsequent vitals (ed.vitalsign) provide follow-up readings. stay_id remapping applied so that vitals from a patient's second consecutive ED visit are attributed to the canonical (first) stay_id. Rows recorded before ed_intime are dropped; records from vitals with charttime == ed_intime are dropped as duplicates of the triage records.

Preprocessing: pain column normalized and coerced to numeric, physiologically implausible vital values nulled. Missing indicators created before forward-fill. Vitals forward-filled within stay, then mean-imputed for any remaining NaN.

Feature engineering: time_since_last_min, per-column deltas and rates of change (units/min), 1-hour rolling averages, mean arterial pressure (dbp + (sbp - dbp) / 3). Triage rows get 0 for all delta/rate columns. Final vital columns renamed to current_{col}.