Endpoint and predictive variables

The primary endpoint was the presence of HFpEF, which was defined as left ventricular diastolic dysfunction (LVDD) in the presence of normal or near-normal LV ejection fraction and had symptoms and signs of HFpEF. Normal or near-normal LV ejection fraction defined as an ejection fraction of 50% or more. LVDD was defined in accordance with the most recent guidelines with the following cut-off values suggesting abnormal diastolic function: (i) septal e′ < 7 cm/s and/or lateral e′ <10 cm/s; (ii) averaged E/e′ > 14; (iii) tricuspid regurgitation velocity > 2.8 m/s; and (iv) left atrial volume index > 34 mL/m². Patients who satisfied ≥ more than one-half of these criteria were defined as having an abnormal diastolic dysfunction.¹⁷ We defined the symptoms and signs of HFpEF as chest discomfort, palpitation, exercise intolerance, fatigue, oedema, dyspnoea, syncope, and general weakness, and confirmed the information from the echocardiography report, in which the cardiologist has mentioned the reason for conducting echocardiography. The echocardiographic findings were obtained from comprehensive two-dimensional (2D) Doppler echocardiography. Acquisitions and measurements were performed by licensed sonographers and cardiologists who were blinded to any other study data. We used demographic information and ECG as predictive variables. We used demographic information and ECG as predictive variables. We used four variables (age, sex, weight, and height) as demographic information. These four variables are simple and can be objectively collected consistently during the screening evaluation. We did not use past medical history, because the information relied on the patient’s memory, and there was a possibility of error and undetected disease. Only definite epidemiologic information was used for the evaluation.

We used the raw data from each 12-lead ECG, amounting to 5000 data points for each lead, recorded over 10 s (500 Hz), and 60 000 data points from each ECG. We used 8 s of ECG data by excluding the first and last 1-s periods because more artefacts were contained within this range. We created a dataset using the entire 12-lead ECG data. Furthermore, we used partial datasets from the 12-lead ECG data, such as the limb six-lead (I, II, III, aVL, aVR, and aVF) and single lead (I or II). We selected those leads as they can be easily recorded by wearable and lifestyle devices in contact with the patient’s limbs.¹⁸ Consequently, when we developed and validated an algorithm using 12-lead ECG, we used a dataset comprising 12 × 4000 2D data points. Similarly, for the six-lead and single-lead ECG signals, we used datasets comprising 6 × 4000 and 1 × 4000 data points, respectively.