Transformer-Based Framework for Multi-Dimensional Biomedical Signal Interpretation and Diagnosis

Abstract

Clinical diagnosis requires the correct interpretation of biomedical signals, including electrocardiograms (ECG), electroencephalograms (EEG), electromyograms (EMG). Their high dimensionality, non-stationary nature, and noise, however, make them less practical with such conventional machine learning and deep learning methods. This work suggests a Transformer-based architecture of multi-dimensional biomedical signal interpretation and diagnosis, which is designed to focus on long-range temporal dependencies as well as inter-signal relationships, achieved via self-attention. The suggested architecture unites time, frequency, and statistical characterizations within a single architecture, allowing improved learning of representations with diverse signal modalities. Benchmark datasets, such as PhysioNet ECG and EEG repositories, are evaluated with a common training testing protocol with experimental evaluation. The suggested framework is contrasted with the standard models e.g., CNN, LSTM and CNN-LSTM frameworks. Findings indicate better performance, with an accuracy of 91.2, F1-score of 0.90 and AUC of 0.93 with less variance of performance between the 2 runs. The results point to the fact that transformer-based models are a solid and scalable answer to the multi-dimensional analysis of biomedical signals. The suggested framework assists in the creation of intelligent healthcare systems as it allows more accurate and effective automated diagnoses, and may be utilized in real-time clinical decision support settings.