Comparative Analysis of Machine Learning Techniques for Heart Rate Prediction Employing Wearable Sensor Data.

Abstract

Monitoring heart rate (HR) is vital for health management and athletic performance, and wearable technology enables scientists to obtain real-time cardiovascular insights. This study compares Machine Learning (ML) techniques, including Long Short-Term Memory (LSTM) networks, Physics-Informed Neural Networks (PINNs), and 1D Convolutional Neural Networks (1D CNNs). Then, we develop a hybrid Singular Spectrum Analysis (SSA)-Augmented ML technique to predict HR using wearable sensor data. Additionally, we investigate the impact of incorporating auxiliary physiological inputs, such as breathing rate (BR) and RR intervals, on predictive accuracy. The study utilizes the cardiorespiratory data acquired through wearable sensors while practising sports, including 126 recordings from 81 participants (53 males, 28 females) engaged in 10 different sports. Physiological signals were collected at 1 Hz using the BioHarness 3.0 (Zephyr Technology, Mangaluru, India). The dataset includes individuals with varied levels of sports experience (beginner, intermediate, and advanced), allowing for a more comprehensive evaluation of HR variability across different expertise levels. Our results demonstrate that the hybrid SSA-LSTM model reaches the lowest prediction error by effectively capturing HR dynamics. Furthermore, integrating HR, BR, and RR data significantly enhances accuracy over single or dual parameter inputs. These findings support adopting multivariate machine learning models for health monitoring, improving HR prediction accuracy for fitness and preventive healthcare.