Heart Rate and HRV Estimation Using PPG Based on Superlet Transform and LSTM Network

Abstract

Photoplethysmography (PPG) is a widely used, noninvasive method for monitoring cardiovascular parameters in wearable devices. However, wrist-based PPG signals are often affected by motion artifacts and poor sensor contact, which can compromise heart rate (HR) estimation accuracy. This article presents a novel algorithm that combines deep learning and spectro-temporal analysis to enhance HR estimation and HR variability (HRV) assessment from PPG signals. A long short-term memory (LSTM) network is employed to model temporal patterns in preprocessed signals, followed by spectral analysis to extract HR-relevant features. The method is evaluated on a custom dataset collected from 15 subjects under six motion conditions, including walking, climbing stairs, and hand movements. Experimental results show that the proposed approach achieves a mean absolute error (MAE) of 0.93 beats per minute (bpm), outperforming existing state-of-the-art methods with improvements ranging from 7.92% to 66.06% in the MAE across all subjects. The method demonstrates consistently low absolute errors (AEs) in diverse motion scenarios, with a minimum AE of 0.15 bpm, indicating high precision in HR estimation. Additionally, the proposed method aligns closely with ground truth in all HRV metrics, with an IBI mean difference of 0.051 s, SDNN difference of 0.063 s, and RMSSD difference of 0.127 s. In the frequency domain, low-frequency (LF) and high-frequency (HF) power differ by 0.01 normalized units (n.u.) each, while the LF/HF ratio differs by 0.13. Nonlinear measures also show close alignment, with approximate entropy (ApEn) and detrended fluctuation analysis (DFA) differing by just 0.031 and 0.07, respectively. These findings highlight the method’s robustness in capturing both linear and nonlinear HRV characteristics and its effectiveness in improving the reliability of wearable PPG monitoring in real-world scenarios.