Detection and Comparative Evaluation of Noise Perturbations in Dynamical Systems and ECG Signals Using Complexity-Based Features

Noise can substantially distort both chaotic and physiological dynamics, obscuring deterministic patterns and altering the apparent complexity of signals. Accurately identifying and characterizing such perturbations is essential for reliable analysis of dynamical and biomedical systems. This study combines complexity-based features with supervised learning to characterize and predict noise perturbations in time series data. Using two chaotic systems (Rössler and Lorenz) and synthetic electrocardiogram (ECG) signals, we generated controlled Gaussian, pink, and low-frequency noise of varying intensities and extracted a diverse set of 18 complexity metrics derived from both raw signals and phase-space embeddings. The analysis systematically evaluates how these metrics behave under different noise regimes and intensities and identifies the most discriminative features for noise classification tasks. Approximate Entropy, Mean Absolute Deviation, and Condition Number emerged as the strongest predictors for noise intensity, while Condition Number, Sample Entropy, and Permutation Entropy most effectively differentiated noise categories. Across all systems, the proposed framework reached an average accuracy of 99.9% for noise presence and type classification and 96.2% for noise intensity, significantly surpassing previously reported benchmarks for noise characterization in chaotic and physiological time series. These results demonstrate that complexity metrics encode both structural and statistical signatures of stochastic contamination.