Unveiling Chaotic Dynamics for HRV Signals by Machine Learning Methods: A Comparative Study

Abstract

Heart Rate Variability (HRV), derived from RR intervals in ECG signals, reflects autonomic nervous system activity but is challenging to predict due to its nonlinear, chaotic nature. This study explores machine learning for HRV time-series forecasting, by comparing Support Vector Regression (SVR) with a linear kernel against deep learning models, Artificial Neural Networks (ANN), simple Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) networks, and a hybrid LSTM+CNN model. Using normalized RR intervals as the sole feature, we applied a sliding window approach for phase space reconstruction (input: 50 consecutive RR values; output: next value). Deep learning models used the ReLU activation function, while SVR used none. Models were trained in MATLAB R2024b on 18 RR sequences (14 training, 4 testing) from the MIT-BIH Normal Sinus Rhythm Database. Evaluated via Root Mean Square Error (RMSE), the hybrid LSTM+CNN model outperformed others, achieving the lowest RMSE across test signals. These findings highlight the effectiveness of hybrid deep learning in capturing HRV’s chaotic dynamics, with implications for clinical monitoring and early diagnosis.