Enhancing biomedical signals through genetic algorithm optimized Exponentiated transmuted weibull denoising techniques

Biomedical signals are frequently corrupted by physiological and environmental noise, which obscures diagnostically relevant features and complicates clinical interpretation. This study introduces a denoising framework that integrates the Exponentiated Transmuted Weibull Distribution (ETWD) with Independent Component Analysis (ICA) to model complex, non-Gaussian noise patterns. The ETWD’s tri-parametric structure $$\:\left(\varvec{\alpha\:},\varvec{\beta\:},\varvec{\lambda\:}\right)$$ generalizes eleven classical distributions, enabling data-driven adaptability beyond conventional heavy-tailed models. A novel score function derived from ETWD is embedded within the FastICA algorithm, with parameters optimized via a Genetic Algorithm (GA). Sparsity constraints in the wavelet domain are applied to preserve transient signal features while suppressing noise. The framework is evaluated on electroencephalogram (EEG), electrocardiogram (ECG), and medical imaging datasets using standardized protocols with 70/30 development/test splits and 10 independent runs. Results demonstrate statistically significant improvements over conventional methods (Gauss, Pow3, Skew, Tanh). For EEG, Sparse ETWD achieved SNR of 7.77 ± 0.07 dB ( $$\:\varvec{p}<0.01$$ ) and improved epileptic spike detection accuracy from 71% to 94%. For ECG, ETWD achieved SNR of 21.34 ± 0.12 dB ( $$\:\varvec{p}<0.01$$ ) and improved R-peak detection F1-score from 0.89 to 0.97. For medical images, after correcting the evaluation protocol for normalized data (PSNR = $$\:-10{\mathbf{l}\mathbf{o}\mathbf{g}}_{10}\left(\text{MSE}\right)$$ ), ETWD achieved 32.01 dB under Gaussian noise, outperforming baselines across speckle (29.84 dB) and Rician noise (30.92 dB). Cross-dataset validation confirmed robustness within each modality, and comparison with a convolutional autoencoder under identical conditions showed competitive or superior performance without requiring training data. The framework offers a training-free, computationally efficient alternative to deep learning methods, and its application led to improved performance in downstream tasks such as R-peak and spike detection.