APPLICATION OF LARGE LANGUAGE MODELS FOR ELECTROMYOGRAPHIC NOISE REMOVAL IN ECG RECORDINGS

This study addresses the relevant scientific and practical problem of adaptive suppression of electromyographic (EMG) interference in electrocardiograms (ECG) using hybrid artificial intelligence architectures. The unsuitability of directly using large language models (LLMs) as numerical filters is substantiated due to tokenization-induced noise and the lack of improvement in the signal-to-noise ratio. A hybrid LLM-assisted DSP architecture is proposed, in which the language model acts as an intelligent planner that analyzes the time-frequency characteristics of the interference and generates a configuration JSON plan, while numerical filtering is performed by a deterministic wavelet module. Experimental results demonstrate that segmentation with a window length of N = 500 samples (1.0 s), 50% overlap, and edge padding provides an improvement in the signal-to-noise ratio of nearly 3 dB. It is demonstrated that, according to a set of clinical and morphological metrics, the developed hybrid architecture outperforms the classical wavelet filter with fixed parameters, providing more accurate R-peak localization and less distortion of QRS-complex width, which is critical for automated computer-aided diagnosis.