Key result
Hybrid CNN-ELM models classify cardiac BCG segments from motion artifacts with ~95% accuracy.
Why the study?
Ballistocardiogram recordings are highly susceptible to motion artifacts, creating a need for a robust classification framework to distinguish informative cardiac segments from artifact-affected segments across heterogeneous multi-pathology recordings.
Does a hybrid 1D-CNN and Extreme Learning Machine classification framework accurately distinguish informative cardiac activity segments from motion artifacts in a multi-pathology BCG dataset?
Does a hybrid 1D-CNN and Extreme Learning Machine classification framework accurately distinguish informative cardiac activity segments from motion artifacts in a multi-pathology BCG dataset?
Hybrid 1D-CNN and ELM classifiers effectively distinguish reliable cardiac activity segments from motion artifacts in heterogeneous BCG recordings, offering a robust preprocessing step for physiological cardiac monitoring.
Objective Accurate identification of informative Ballistocardiogram (BCG) segments is essential for reliable cardiac monitoring because BCG recordings are highly susceptible to motion artifacts. This study aims to develop a robust classification framework for distinguishing informative BCG segments containing cardiac activity from non-informative segments affected by motion artifacts in heterogeneous multi-pathology recordings. Methods: Two hybrid architectures, CNN–ELM and CNN–OS-ELM, were developed by combining a one-dimensional Convolutional Neural Network (1D-CNN) for deep feature extraction with Extreme Learning Machine (ELM) and Online Sequential Extreme Learning Machine (OS-ELM) classifiers. A multi-pathology BCG dataset comprising healthy subjects, patients with heart failure, and individuals with various arrhythmias was used. The signals were preprocessed using a 0.5–15 Hz band-pass filter, standardized, segmented into overlapping 10-s windows, and automatically labeled using a recording-specific standard-deviation threshold. Model performance was evaluated using subject-wise grouped stratified 10-fold cross-validation to prevent subject-level overlap between training and test partitions. An alternative 85th-percentile labeling threshold was additionally evaluated as a sensitivity analysis. Results: CNN–ELM achieved a mean accuracy of 0.9514, recall of 0.9844, F1-score of 0.9733, and AUC of 0.9361, whereas CNN–OS-ELM achieved a mean accuracy of 0.9514, recall of 0.9847, F1-score of 0.9733, and AUC of 0.9364. The sensitivity analysis using the 85th-percentile threshold showed that both models maintained high classification performance, although performance decreased compared with the primary 90th-percentile criterion. Conclusions: The results demonstrate that CNN-based deep feature extraction combined with ELM-based classifiers provides an effective framework for BCG/non-BCG segment classification under leakage-resistant subject-wise evaluation. The proposed approach can serve as a preprocessing stage for identifying reliable cardiac segments before subsequent cardiac-cycle and physiological-parameter analysis.
No takes yet. Share an insight, caveat, or question.
Karahan et al. (2026) studied Heart failure and various arrhythmias. Hybrid CNN-ELM and CNN-OS-ELM classification frameworks was evaluated on Classification of informative cardiac activity segments from motion artifacts in BCG signals (AUC 0.9361 (CNN-ELM); AUC 0.9364 (CNN-OS-ELM)). Hybrid CNN-ELM and CNN-OS-ELM models classified cardiac BCG segments from motion artifacts with 95.14% accuracy and AUCs of 0.9361 and 0.9364, respectively.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: