Automated artifact elimination of physiological signals using a deep belief network: An application for continuously measured arterial blood pressure waveforms

• Published in: Information sciences 2018-08, Vol.456, p.145-158
• Main Authors: Son, Yunsik, Lee, Seung-Bo, Kim, Hakseung, Song, Eun-Suk, Huh, Hyub, Czosnyka, Marek, Kim, Dong-Joo
• Format: Article
• Language: English
• Subjects: Arterial pressure; Computer-assisted; Monitoring; Neural networks (computer); Physiologic; Signal processing
• ISSN: 0020-0255; 1872-6291
• DOI: 10.1016/j.ins.2018.05.018

•The deep belief network (DBN) model excels at artifact classification in pulse waveforms.•The DBN allows rapid detection and elimination of artifacts in physiological signals.•Various types of artifacts in arterial blood pressure can be detected by the DBN. Artifacts in physiological signals acquired during intensive care have the potential to be recognized as critical pathological events and lead to misdiagnosis or mismanagement. Manual artifact removal necessitates significant labor-time intensity and is subject to inter- and intra-observer variability. Various methods have been proposed to automate the task; however, the methods are yet to be validated, possibly due to the diversity of artifact types. Deep belief networks (DBNs) have been shown to be capable of learning generative and discriminative feature extraction models, hence suitable for classifying signals with multiple features. This study proposed a DBN-based model for artifact elimination in pulse waveform signals, which incorporates pulse segmentation, pressure normalization and decision models using DBN, and applied the model to artifact removal in monitoring arterial blood pressure (ABP). When compared with a widely used ABP artifact removal algorithm (signal abnormality index; SAI), the DBN model exhibited significantly higher classification performance (net prediction of optimal DBN = 95.9%, SAI = 84.7%). In particular, DBN exhibited greater sensitivity than SAI for identifying various types of artifacts (motion = 93.6%, biological = 95.4%, cuff inflation = 89.1%, transducer flushing = 97%). The proposed model could significantly enhance the quality of signal analysis, hence may be beneficial for use in continuous patient monitoring in clinical practice.