A Deep Learning Approach for Heart Rate Variability Error Correction

This study evaluates the effectiveness of Convolutional Neural Network, Long Short-Term Memory (CNN-LSTM) networks in correcting errors in the RR-interval series that is necessary for assessing cardiovascular health and autonomic nervous system function through Heart Rate Variability (HRV) parameters, i.e., time-domain, frequency-domain, and nonlinear parameters. These parameters are used in clinical studies and medical diagnosis; however, errors in the RR-interval series, often arising from measurements in ambulatory conditions, use of low-cost wearable devices, and signal processing limitations, necessitate preprocessing and error correction. By employing Monte Carlo simulations, this study compares CNN-LSTM series reconstruction with error-containing series characterized by Gaussian additive noise and the traditional cubic spline interpolation method. Specifically, the RR-interval series were segmented into piece-wise stationary sections according to physical activity, and errors were uniformly introduced in 3300 samples in each scenario, with varying error presence (10% to 30%) and duration (Is to 7s) to assess the correction efficacy for different HRV parameters. The CNN-LSTM model significantly outperformed cubic spline interpolation (CSI) in reconstructing RR-series, with differences in key metrics such as RMSSD, MaxRR, HF, and LF reaching 98.12%, 98.80%, 100.00%, and 99.99%, indicating less variability and distortion. The CSI showed less deviation for alpha2, SampEn, and ApEn. While both methods show considerable differences compared to the erroneous series, CNN-LSTM more effectively mitigates errors between the erroneous series and CNN-LSTM reconstruction as the duration and presence of noise increase. This finding is valuable for enhancing the accuracy of HRV parameters, thereby supporting and improving medical diagnosis.