An Efficient Multi-Channel EEG Compression Using Intelligent Tunable <i>Q</i> -Wavelet Transform: A Machine Learning-Driven Approach

Processing of multichannel EEG (electroencephalography) or MEEG signals in real-time is challenging due to the high data volume, which affects storage, transmission and computational efficiency. This study presents a 2-stage approach leveraging the correlations between EEG channels using a machine learning-based intelligent tunable Q-wavelet transform (TQWT) and multiscale principal component analysis (MSPCA) for efficient compression of MEEG. TQWT first decomposes each channel into subbands to capture intra-channel correlations. Then, MSPCA is applied to the subbands to identify inter-channel correlations, enabling enhanced data compression. Principal components (PCs) are encoded via delta coding and run-length encoding to improve compression efficiency further. The signal reconstruction quality is ensured by optimizing the Q-factor and TQWT decomposition levels for each 3s EEG snippet through a multilayer feedforward neural network (MLPFFNN) trained offline with particle swarm optimization (PSO). A deep autoencoder (DAE) extracts relevant features as inputs for the MLPFFNN. Testing on four benchmark datasets, including PhysioNet and brain-computer interface (BCI) competition data, results in high compression rates (CR of 14.74) and low error (PRD of 2.72), with a mean squared coherence (MSC) of 0.97. The method shows competitive or superior compression performance, with low memory and latency demands, suitable for real-time EEG applications.