ECG data and endpoint

Predictor variables are ECG, age, and sex. Digitally stored 12-lead ECG data, amounting to 5000 numbers for each lead was recorded over 10 s (500 Hz). We removed 1 s each at the beginning and end of ECG because these areas have more artifacts than other parts. Because of this, the length of each was 8 s (4000 numbers). We made a dataset using the entire 12-lead ECG data. We also used partial datasets from 12-lead ECG data, limb 6-lead (aVL, I, − aVR, II, aVF, and III). We selected these sets of leads because they can easily be recorded with wearable devices, pads, and other daily living devices while in contact with the extremities¹⁰. Consequently, when we developed and validated an algorithm using 12-lead ECGs, we used a dataset of 2-dimensional (2D) data of 12 × 4000 numbers. To make the input 2D ECG data, we rearranged the data in the order of V1, V2, V3, V4, V6, aVL, I, -aVR, II, aVF, and III. The convolutional neural network (CNN), a method of deep learning, is a well-known architecture for learning 2D image data¹⁶. In the same manner, when we developed and validated an algorithm using a 6-lead; we used datasets that were 6 × 4000.

The primary endpoints were type 1 and 2 MIs defined in the fourth universal definition of MI¹⁷. Four cardiologists manually reviewed the patients’ medical records and results of cardiac enzymes, coronary angiography (CAG), and echocardiography to label the occurrence of an MI at the time of CAG. A cardiologist labeled the development dataset, and three cardiologists not involved in the development data only labeled the internal and external validation data. After labeling the validation data, we used a voting system to render a decision when an agreement from the cardiologists is not reached. Cardiologists labeled MI using the gold standard of CAG findings such as a primary coronary event such as plaque erosion and/or rupture, fissuring, or dissection, and supply/demand imbalance such as coronary spasm and elevation of cardiac troponins, which were above the 99th percentile upper reference limit at the time of CAG. With regard to development, internal validation, and external validation, the ECGs that were acquired within 24 h before MI CAG were labeled as MI. The ECGs that were acquired within 24 h before non-MI CAG were labeled as non-MI. We only used ECGs acquired less than 24 h before the initiation of CAG.