Singular spectrum analysis for detection of abnormalities in periodic biosignals

High level of false positive alarms is one of the main issues in ambulatory monitoring in Intensive Care Units. The solution to it is the development of new methods that will be both reliable in detection of anomalies in the patient's state and robust to noise and artifacts. The current study is focused on the development of unsupervised automated approach to analysis of periodic biosignals. The proposed classification method for distinguishing anomalies from normal patterns is based on the combination of time series domain pattern recognition method Singular Spectrum Analysis and clustering techniques. The model itself includes preprocessing, analysis, classification and validation stages and one of its main benefits consists in automated approach to regular features (e.g., heartbeats) extraction without the need of analysing its morphologies, and further unsupervised classification of the obtained patterns. Still, this method has its limitations, as all unsupervised learning-based techniques, and the validation stage requires additional work. The results of testing on the series of biomedical signals (ECG, O₂, arterial pressure) from Physionet Database showed that this method is effective in anomalies detection tasks, highly independent of the periodic signal specificity and resistant to the average level of noise.