Shortening heart rate variability measurement time to 1 minute using deep learning: Implication of real-time measurement of heart rate variability

Abstract

Heart rate variability (HRV) is an effective predictor of cardiovascular diseases. The current standard 5-min recording time is lengthy compared with routine clinical examinations such as blood pressure measurement. Previous studies have observed that the indices of 3-min HRV data are as clinically meaningful as those of 5-min HRV data; however, shorter durations are considered unreliable, and there have been no attempts to challenge this notion. This study aimed to validate the outcomes of 1-min HRV recordings reconstructed using deep learning algorithms. Three-minute HRV recordings from 34,885 participants were included in the analysis. Of the recordings, 60% (20,931), 30% (10,465), and 10% (3,489) were allocated to the training, validation, and test sets, respectively. Data from 1-min excerpts of the 3-min recordings were used as the input for the deep learning models to predict the data of the 3-min recordings. Various deep learning models were applied to each indicator, and the model that produced the lowest mean absolute error was selected as that particular indicator’s learning model. There was no statistical difference between the values of the 1-min recordings reconstructed by deep learning and those of the 3-min recordings. The 1-min recordings reconstructed by deep learning demonstrated a higher correlation with the 3-min recordings when compared with the 1-min recordings that were not processed by deep learning. They also strongly agreed with the 3-min recordings in the Bland–Altman analysis. The 1-min HRV recordings reconstructed by deep learning were as reliable as the 3-min HRV recordings, suggesting that a 1-min recording could serve as a proxy for real-time HRV monitoring in the future. Keywords: heart rate variability; deep learning; Bland-Altman analysis; ultrashort-term; one minute HRV