Data Dictionary v1

Table of contents

1. File Tree
2. Entity Relationship Diagram
3. Patient
4. ECG Waveforms
   1. Long Leads - 100 Hz
   2. All Leads - 500 Hz
5. ECG Metadata
6. Emergency Department Encounters

Dataset Changes

v1.0.1 Changes (Mar 2024)

• The 12 “short” leads were added to the dataset with a sample rate fo 500 Hz.
• The 3 “long” leads with a sample rate of 500 Hz were added.

File Tree

.
└── ed-bwh-ecg
    └── v1
        ├── ecg-ed-enc.csv
        ├── ecg-metadata.csv
        ├── ecg-npy-index.csv
        ├── ed-encounter.csv
        ├── ecg-waveform.h5
        ├── ecg-waveform.npy
        └── patient.csv

Entity Relationship Diagram

erDiagram
    PATIENT {
        string patient_ngsci_id PK
        string demographics
    }
    ECG-METADATA {
        string ecg_id PK
        string patient_ngsci_id FK
        string data
    }
    ECG-NPY-INDEX {
        string ecg_id PK,FK
        int npy_index
    }
    ED-ENCOUNTER {
        string enc_id PK
        string patient_ngsci_id FK
        string data
    }
    ECG-ED-ENC {
        string ecg_id PK,FK
        string enc_id FK
    }
    ECG-WAVEFORM {}

    PATIENT ||--|{ ECG-METADATA : patient_ngsci_id
    PATIENT ||--|{ ED-ENCOUNTER : patient_ngsci_id
    ECG-METADATA ||--|| ECG-NPY-INDEX : ecg_id
    ECG-METADATA ||--o| ECG-ED-ENC : ecg_id
    ED-ENCOUNTER ||--|{ ECG-ED-ENC : enc_id
    ECG-NPY-INDEX ||--|| ECG-WAVEFORM : npy_index

Entity Relationship Diagram

Patient

patient.csv contains patient

Column Name	Description	Data Type	Example
patient_ngsci_id	Unique patient identifier Pattern: pat{8 digit hex}	string	pat089c033f
sex	Patient Sex	string	Female
black	Patient Race/Ethnicity - The patient will only have 1 in one of the categories.	int	1
hispanic		int	0
white		int	0
other		int	0
agi_under_25k	Adjusted Gross Income (AGI) distributions from block-level census data based on patient address	float	0.36921
agi_25k_to_50k		float	0.28902
agi_50k_to_75k		float	0.15070
agi_75k_to_100k		float	0.07804
agi_100k_to_200k		float	0.09849
agi_above_200k		float	0.01453

• Rows: 44,713
• Columns: 12

ECG Waveforms

The ECGs for this dataset were collected as PDF files. The waveforms were visibly rendered on the PDF files and needed to be processed in order to be converted into numeric form. The images appeared as in Figure 2.

Sample 12 Lead ECG

Long Leads - 100 Hz

The waveforms for leads V1, II, and V5 were smoothed with a rolling average and the sample rate was reduced to 100 Hz. The time duration for these leads is 10 seconds.

ecg-npy-index.csv

Column Name	Description	Data Type	Example
ecg_id	Unique ECG identifier Pattern: ecg{10 digit hex}	string	ecg3df45120a4
npy_index	Index of the NumPy array	int	523

• Rows: 112,900
• Columns: 2

ecg-waveform.npy and ecg-waveform.h5

Shape: (112900, 3, 1000)

All Leads - 500 Hz

datasets/ed-bwh-ecg/v1/ecg-waveforms-npz/{first two digits of ecg_id}/{ecd_id}.npz

We have also included the numeric waveforms for all leads. These waveforms have a sample rate of 500 Hz.

The waveforms are stored as compressed NumPy files. There is a separate file for each ECG, with the ECG ID in the file name. The array contained in each file has the dimensions of 15 x 5000. The first 12 rows are the waveforms for the 12 leads. However, each of these leads only have a duration of about 2.5 seconds. As seen in Figure 2, leads I, II, and III are measured for the first 2.5 seconds. Leads aVR, aVL, and aVF are measured for the second 2.5 seconds. Leads V1, V2, and V3 are measured for the third 2.5 seconds. And leads V4, V5, and V6 are measured for the last 2.5 seconds. Not seen in Figure 2, leads V1, II, and V5 are measured for the full 10 seconds. Those represent the last 3 rows of the array.

A notebook interacting with these files can be seen on the platform here datasets/ed-bwh-ecg/supplementary/notebooks/accessing-waveforms.ipynb

ECG Metadata

Date Shift

Dates in this dataset have been shifted by a random amount for each patient. This is done to create anonymity while preserving the temporal relationship between events for patients.

ecg-metadata.csv

Column Name	Description	Data Type	Example
patient_ngsci_id	Unique patient identifier Pattern: pat{8 digit hex}	string	pat089c033f
ecg_id	Unique ECG identifier Pattern: ecg{10 digit hex}	string	ecg3df45120a4
date	Shifted date and time of the ECG	string	2110-07-29T11:27:56Z
p-r-t_axes	P-R-T axes	string	52 9 27
p_axes	P axes	int	52
r_axes	R axes	int	9
t_axes	T axes	int	27
pr_interval	PR interval	int	176
pr_interval_units	PR interval units	string	ms
qrs_duration	QRS duration	int	74
qrs_duration_units	QRS duration units	string	ms
qtqtc	QTQTc	string	432/413 ms
qt_interval	QT interval	int	432
qt_interval_units	QT interval units	string	ms
qtc_interval	QTc interval	int	413
qtc_interval_units	QTc interval units	string	ms
vent_rate	Vent rate	int	55
vent_rate_units	Vent rate units	string	BPM
has_bbb	Flags for whether search terms were present in the cardiology remarks	int	0
has_afib		int	0
has_st		int	0
has_pacemaker		int	0
has_lvh		int	0
has_normal		int	1
has_normal_ecg		int	1
has_normal_sinus		int	0
has_depress		int	0
has_st_eleva		int	0
has_twave		int	0
has_aberran_bbb		int	0
has_jpoint_repol		int	0
has_jpoint_eleva		int	0
has_twave_inver		int	0
has_twave_abnormal		int	0
has_nonspecific		int	0
has_rhythm_disturbance		int	0
has_prolonged_qt		int	0
has_lead_reversal		int	0
has_poor_or_quality		int	0

• Rows: 112,900
• Columns: 39

Columns with names that start with has_ indicate whether certain search terms were present in the cardiology remarks. Below are the search terms each flag label.

Column Name	Regex Search Terms
has_bbb	bbb or bundle\s+branch\s+block
has_afib	atrial\s+flutter or atrial\s+fibrillation
has_st	st\s+
has_pacemaker	pacemaker or paced
has_lvh	lvh or ventricular\s+hypertrophy
has_normal	(normal\s+sinus\s+rhythm and not abnormal\s+sinus\s+rhythm) or (normal\s+ecg and not abnormal+ecg)
has_normal_ecg	normal\s+ecg and not abnormal\s+ecg
has_normal_sinus	normal\s+sinus\s+rhythm and not abnormal\s+sinus\s+rhythm
has_depress	st\s*\w*\s*depress
has_st_eleva	st\s*\w*\s*eleva
has_twave	t.wave
has_aberran_bbb	bbb or bundle\s+branch\s+block or aberran
has_jpoint_repol	j\s+point or early repol
has_jpoint_eleva	st\s*\w*\s*eleva or j\s+point or early repol
has_twave_inver	t.wave and inter
has_twave_abnormal	t.wave.abnormal
has_nonspecific	nonspecific
has_rhythm_disturbance	premature (atrial|ventricular)|PAC|PVC or aberran or intraventricular conduction or ectop or arrythmia or junctional or fusion complex or a-v|atrioventricular
has_prolonged_qt	prolonged qt
has_lead_reversal	lead reversal
has_poor_or_quality	poor or quality

Emergency Department Encounters

Date Shift

Dates in this dataset have been shifted by a random amount for each patient. This is done to create anonymity while preserving the temporal relationship between events for patients.

ed-encounter.csv

Column Name	Description	Data Type	Example
patient_ngsci_id	Unique patient identifier Pattern: pat{8 digit hex}	string	pat089c033f
ed_enc_id	Unique ED encounter identifier Pattern: enc{8 digit hex}	string	enc5ba023af
start_datetime	Shifted start of the date and time of the ED encounter	string	2110-07-29T11:06:00Z
end_datetime	Shifted end of the date and time of the ED encounter	string	2110-07-29T12:31:00Z
age_at_admit	Patient age	int	75
macetrop_030_pos	Major adverse cardiovascular events (MACE) & pos troponin in 30 days after visit	bool	FALSE
death_030_day	Death in 30 days after visit - This variable comes from Social Security Death Index data, so it captures both death in and out of the hospital.	bool	FALSE
macetrop_pos_or_death_030	Adverse Events (30days)	bool	FALSE
stent_010_day	Stent within 10 days after visit	bool	FALSE
cabg_010_day	Coronary artery bypass graft surgery (CABG) within 10 days after visit	bool	FALSE
stent_or_cabg_010_day	Stent or CABG within 10 days	bool	FALSE
ami_day_of	Acute myocardial infarction “heart attack” day of visit days_to_ami == 0	bool	FALSE
days_to_ami	Number or days to soonest AMI, missing if no AMI	int	5
maxtrop_sameday	Max troponin lab results on day of visit	float	0.25
tn_group_sameday	Categorized maxtrop_sameday into following bins - missing - 0 - (0-0.05] - (0.05-0.1] - (0.1-05] - >0.5	string	(0.1,0.5]
disch_disp	Discharge code	string	a
disch_obs	Flag for whether patient is dispatched to observation (disch_disp == e | disch_disp == edobs)	bool	FALSE
test_010_day	Stress test or cath test within 10 days after visit	bool	FALSE
stress_010_day	Stress testing (10days)	bool	FALSE
cath_010_day	Catheterization (10days)	bool	FALSE
days_to_stress	Days to earliest stress test	int	1
days_to_cath	Days to earliest catheterization test	int	2
first_test	Whether earliest test is stress or cath; if they have both we generally assume first test is stress even if this doesn’t seem true by timestamps	string	cath
excl_flag_c_int	Flag for cardiac intervention in previous 30d	bool	FALSE
excl_flag_chronic	Flag for chronic illness	bool	FALSE
excl_flag_death	Flag for discharge = death	bool	FALSE
exclude_modeling	Exclusion flag for training models = (excl_flag_c_int | excl_flag_chronic | excl_flag_death | (ami_day_of & !test_010_day))	bool	FALSE
exclude	Exclusion flag for analysis = (exclude_modeling | age_at_admit >= 80 | (!test_010_day & maxtrop_sameday > 0))	bool	FALSE

• Rows: 71,460
• Columns: 29

ecg-ed-enc.csv

Column Name	Description	Data Type	Example
ecg_id	Unique ECG identifier Pattern: ecg{10 digit hex}	string	ecg3df45120a4
ed_enc_id	Unique ED encounter identifier Pattern: enc{8 digit hex}	string	enc5ba023af

• Rows: 103,952
• Columns: 2