Machine learning methods for classification problems in biomedical signal processing

Abstract

Routine physiological data contains rich diagnostic cues, yet clinicians still rely chiefly on manual visual inspection of raw waveforms. This subjective approach does not consider informative patterns hidden in the time-frequency domain of the signal. Here, we ask whether modern machine-learning algorithms, applied to spectral representations of biomedical signals, can uncover latent biomarkers and turn them into actionable clinical insights. The project addresses this overarching question by focusing on two distinct signal types: (i) electrical by classifying intramuscular electromyography (iEMG) to distinguish spontaneous skeletal muscle activity, and (ii) mechanical by predicting treatment outcomes in obstructive sleep apnea from nasal airflow recordings obtained during polysomnography. First part of this work investigates skeletal muscle channelopathies - a group of neuromuscular disorders that disturb the cell membrane excitability, which results clinically in myotonia. The genetic aetiology of some of these disorders can be traced to mutations in the SCN4A or CLCN1 genes, which encode sodium and chloride channels, respectively. These channels are specialized proteins in the cell membrane that play a crucial role in generation and propagation of action potentials. Commonly, sodium or chloride channel defects lead to pathophysiological hyperexcitability of muscles, which is observable as myotonic discharges on iEMG recordings. However, there is an ongoing debate about whether the properties of the myotonic discharge can differentiate the type of channel defect, such as sodium versus chloride. This discussion leads to the broader question of whether a stable genotype-to-EMG-phenotype relationship exists in muscle channelopathies. Accurate identification of the underlying ion-channel defect is also essential for tailored treatment and informed prognosis. At present, clinicians depend exclusively on genetic testing to distinguish sodium- from chloride-channel defects. Although detailed, manual inspection of iEMG recordings can potentially reveal defect-specific patterns in research settings, the procedure is too time-consuming and complex for routine clinical practice. Our study demonstrated the existence of distinct spectral features in myotonic discharges of patients with sodium and chloride channel defects and addressed the need for their automated classification. We developed and validated a method for this purpose, transforming iEMG recordings into their spectral representations (scalograms) via wavelet transform. These scalograms were subsequently classified using an ensemble of pre-trained deep neural networks. The resulting ensemble achieved a balanced accuracy of approximately 81% and a Brier score of 0.14 on unseen test data. A selective-prediction analysis further indicated that at high-confidence thresholds (greater than 0.85), the model's accuracy exceeded 90%. These results show the potential clinical utility of this approach for enhancing diagnostic efficiency, specifically by helping to prioritize genetic testing for a specific mutation. One of our objectives was to identify physiological signal characteristics specific to each subtype of ion-channel defect. Electrophysiological analysis can reveal how distinct channel defects produce characteristic discharge patterns, clarifying the underlying mechanisms of myotonia. Gradient-weighted saliency mapping identified discriminative spectral features, including broadband, early-burst energy characteristic of chloride-channel defects, while sustained high-frequency spectral components were observed in sodium-channel defect class samples. To support the physiological interpretation of our findings, we generated synthetic myotonic discharges using the biophysical model developed by Klotz et al. 2019. The convolutional neural network detected class-specific spectral differences within these simulated signals. This reinforces the physiological relevance of the spectral motifs identified as markers for distinct channelopathies, thereby enhancing the interpretability of our deep learning model. To further test the robustness and general applicability of the framework, we used the same classification pipeline to distinguish voluntary motor unit activity from spontaneous fibrillation potentials - an electrophysiological hallmark of degenerative and neurogenic muscle disease. Without altering any network architecture or hyperparameters, the ensemble maintained high performance, achieving a balanced accuracy of 87% and a Brier score of 0.11. These findings confirm that the spectral feature learning approach transfers across various pathologies and could therefore serve as a versatile tool for automated EMG diagnostics. The second part of this study investigates obstructive sleep apnea, a common chronic disorder often treated with mandibular advancement splints. Roughly one-third of patients, however, do not respond adequately to this treatment. Because it is still difficult to predict patient responsiveness from baseline polysomnography, we examined whether the routinely recorded baseline nasal-airflow signal contains features that can predict treatment outcome. The signal's spectral characteristics were quantified by extracting dominant frequencies within low, mid, and high-frequency bands. Using these features, classical machine learning algorithms, particularly Random Forest classifiers, demonstrated promising results. Specifically, when trained on signals sampled at sufficiently high frequencies, these classifiers accurately identified responders with a recall approaching 90% and a Cohen's kappa of approximately 0.48. Conversely, downsampling the data to lower frequencies significantly impaired predictive performance, highlighting the critical role of mid- and high-frequency spectral content in this classification task. This study introduces the first automated, uncertainty-aware diagnostic pipeline capable of distinguishing skeletal muscle channelopathies directly from routine EMG signals. By employing spectral signal characteristics within a machine learning framework, it suggests broad applicability across various biomedical contexts. We developed a purely signal-based predictive tool with good clinical diagnostic potential.