Link Search Menu Expand Document

Diagnosing ‘silent’ heart attack using ECG waveforms: A Nightingale Open Science dataset

Authors: Rajiv Pramanik1, Bhumil Shah1, Anna Roth1, Honga Wei1, Ted Castillo1, Katie Lin2, Sachin Shah2, Stelios Serghiou2, Nick Foster2, Josh Risley2, Katy Haynes2, Ziad Obermeyer2,3

1 Contra Costa Health Services
2 Nightingale Open Science
3 University of California, Berkeley

Lead Nightingale analyst: Nick Foster

When using this resource, please cite: more options
Pramanik, R., Shah, B., Roth, A., Wei, H., Castillo, T., Lin, K., Shah, S., Serghiou, S., Foster, N., Risley, J., Haynes, K., & Obermeyer, Z. (2021). Diagnosing ’silent’ heart attack using ECG waveforms [Data set]. Nightingale Open Science.

Additionally, please cite: more options
Mullainathan, S., & Obermeyer, Z. (2022). Solving medicine’s data bottleneck: Nightingale Open Science. Nature Medicine, 28(5), 897–899.

The problem

Every year, millions of heart attacks happen around the world. But up to 78% of them are undiagnosed or “silent”. This means a large fraction of people with heart attack never get the cocktail of drugs known to save lives, by preventing future heart attacks and sudden death.

Today, doctors can order tests (like MRIs or ultrasounds) to diagnose patients when they suspect a prior heart attack. But the reason so many heart attacks remain silent is precisely because doctors and patients don’t even suspect a heart attack has happened.

Finding new ways to diagnose these undiagnosed heart attacks at scale could dramatically expand access to life-saving medications. And because of our close partnership with the county health system that sourced these data, algorithms developed on the platform, once validated, have a clear pathway for making it into clinical use and helping real patients.

Electrocardiograms (ECGs) are a cheap, widespread test done everywhere in the health care system: during annual checkups, ER visits, before surgical procedures, etc. Doctors have learned to diagnose some limited signs of prior heart attack on ECGs (like ‘Q waves’), but these coarse findings still miss about 80% of prior heart attacks. We know that algorithms can match human performance on ECG interpretation—but could they do better, by systematically mining ECG waveforms for signals that might identify prior heart attacks? We don’t know, because there have not historically been datasets linking ECGs to high-quality labels on prior heart attack.

Dataset overview

This dataset will link 48,788 ECG waveforms from 13,438 patients to data on prior heart attack: the results of detailed cardiac ultrasound tests (echocardiograms), done in a one-year window around the ECGs, which can visualize scars in the wall of the heart formed by prior heart attack. Linking ECGs to cardiac ultrasound labels will allow researchers to train algorithms to identify patients with prior or future heart attack.

Each observation in the dataset corresponds to a single 12-lead ECG. We identified all ECGs done as an inpatient or outpatient by the Contra Costa Health Services (CCHS) county health system between January 1st, 2013 and December 31st, 2020, using the Philips TraceMasterVue ECG Management System (now known as IntelliSpace ECG), which stores ECG data from all Philips cardiographs and bedside monitors (more details). ECG waveforms were shared with us as an XML file, which we parsed into an array of 5,500 points for each one of twelve leads. Note that, in the eventual v1 dataset to be released imminently, there may be multiple ECGs per patient.

Our partners

CCHS is a public, county health system that serves 190,000 people in Contra Costa County Contra Costa County, California. It comprises a federally-qualified Health Maintenance Organization (HMO) health insurer, one regional medical center, and eight health centers and clinics.

This dataset was conceived of and created by Rajiv Pramanik, CCHS Chief Medical Informatics Officer and Bhumil Shah, CCHS Chief Analytics Officer, and thanks to the leadership of Anna Roth, the Chief Executive Officer of CCHS. We think this dataset is unique because it comes from a kind of health system typically under-represented in machine learning: CCHS is not a well-resourced academic health center, or a private health systems, but rather a public, county system that cares for a variety of under-served patient populations. For this reason, this dataset—and the many like it we plan to release over the coming months—holds the promise of expanding access to high-quality medical diagnostics for traditionally under-served patients.

We are deeply grateful to the Gordon and Betty Moore Foundation, who supported this work with a grant from their Diagnostic Excellence Initiative.

Dataset details


This dataset v0: The v0 dataset is a subset of 5,000 ECGs from 5,000 unique patients who had a cardiac ultrasound in the year before the ECG. These patients are randomly chosen from the v1 dataset (coming soon, pending formal certification of deidentification). Each row contains the waveforms for the 12 leads of the ECG, and a label that identifies whether a regional wall motion abnormality (RWMA) was identified in the cardiac ultrasound (more detail in the Key Variables section below). In this dataset 9.6% of the 5,000 ECGs have a positive label for RWMA.

What’s next for v1 (target release date: Summer 2022): We’ll add all remaining ECGs—the full set of patients who had a cardiac ultrasound in the year before their ECG, as well as all patients who had a cardiac ultrasound in the year after their ECG.

We’ll add meta-data for each ECG (e.g. heart rate, QRS interval, QT interval, interpretation, etc.), as automatically identified by the ECG machine and provided on TraceMasterVue. We’ll also add Information on patients’ diagnoses and medications, both before and after the ECG.

Finally, we’ll add ECGs from a completely new set of patients: those who had ECGs done during primary care or ER visits, but did not have ultrasounds. These are exactly the kinds of patients in whom we’d like to identify undiagnosed prior heart attacks—they are the ultimate hold-out set, because we don’t (yet) have labels for them. On the other hand, we do have some other ways to understand whether physicians have already diagnosed and treated the heart attacks: these patients’ diagnoses and medications. We can use this to start to understand whether algorithmic predictions are providing genuinely new information.

Dataset schema

Dataset construction and key outcome variables are shown in the schematic below. A note on color choices: the burnt siena (orange) indicates the node that corresponds to the observations (rows) in the dataset, and the grape (purple) indicates key patient outcomes.

Note that this section describes the v1 dataset (as opposed to the random sample of 5000 in v0).

Key variables

Cardiac ultrasound:

We identified all cardiac ultrasounds done as an inpatient or outpatient between January 1st, 2013 and December 31st, 2020. We then linked these to ECGs for the same patient done within 1 year of the cardiac ultrasound (backwards and forwards in time; note the v0 dataset contains only cardiac ultrasound results from the year before the ECG).

v1 ECGs identified in the CCHS system between 7/1/2012 and 3/1/2020

ECGs 147,453
Patients 64,128
ECGs linked to cardiac ultrasounds 48,788
Patients 13,438
Cardiac Ultrasound Reports 20,159

Regional wall motion abnormalities:

We extracted text-based cardiac ultrasound reports in order to identify Regional Wall Motion Abnormalities (RWMA): abnormalities in the contractile function of the left cardiac ventricle, which suggest prior injury due to myocardial infarction (i.e. a heart attack). The presence or absence of RWMA is ascertained by the cardiologist who interprets the images.

v1 Cardiac ultrasound report feature frequency

Echocardiogram Reports (N) 24,211  
Unique Patients (N) 15,183  
Features Count % of total echo reports
NormalWM 12,168 50
GlobalWM 2,569 11
RWMA 2,193 9.1
Prior 4,531 19
TechnicalLimited 6,828 28

We determined the presence or absence of RWMA in cardiac ultrasound reports by using regular expression matching. The process of selecting the appropriate regex terms for each feature was an iterative process with the oversight of a clinician. Reports were parsed in the following way.

Sample of the free text in a report:

Technical Comments:
The study quality is good.  
Left Ventricle:
The left ventricular chamber size is normal. There is no left
ventricular hypertrophy.   Global left ventricular wall motion and
contractility are within normal limits. There is normal left ventricular
systolic function. The ejection fraction is calculated to be 63% using
the Method of Disks.  Age appropriate diastolic function. 
Left Atrium:
The left atrial chamber size is normal.

Additional details on these labels:

Normal Wall Motion NormalWM:
The report includes an observation of normal wall motion. The word “normal” for wall motion must be mentioned.

Global Wall Motion Abnormality GlobalWM:
A global wall motion is an observed impairment of multiple segments of heart muscle suggesting an underlying process that affects the entire heart. Note that this is not typically the consequence of heart attack, which affects a specific section of the wall.

Regional Wall Motion Abnormality RWMA:
Regional wall motion abnormality is an observed impairment of a particular segment(s) of the heart wall, suggesting heart attack. This typically results from a blocked vessel (but may also occur in the absence of coronary artery disease: myocarditis, sarcoidosis and takotsubo cardiomyopathy).

Prior Observation Prior:
The finding is a prior finding.

Technically Limited TechnicalLimited:
A technically limited cardiac ultrasound may not have enough information captured to make an accurate assessment.

The scripts for this regular expression analysis can be found at this repository.

Table of contents

Copyright © 2021 Nightingale Open Science. All rights reserved.