J. Ranta – Sleep classification in infants using a bed mattress sensor and analysis of ECG

Jukka Ranta

Thursday 2018-05-31 14.00 – 14.30

Seminar room B337, T-building

Sleep classification in infants using a bed mattress sensor and analysis of ECG

Infants spend the majority of their time asleep. Although extensive studies have been carried out, the role of sleep for infant cognitive, psychomotor, temperament and developmental outcomes is not clear. The current contradictory results may be due to the limited precision when monitoring infant sleep for prolonged periods of time, from weeks to even months. Sleep-wake cycle can be assessed with sleep questionnaires and actigraphy, but they cannot separate sleep stages. The gold standard for sleep state annotation is polysomnography (PSG), which consist of several signal modalities such as electroencephalogram, electrooculogram, electrocardiogram (ECG), electromyogram, respiration sensor and pulse oximetry. A sleep clinician manually assigns sleep stages for 30 sec epochs based on the visual observation of these signals. Because method is obtrusive and laborious it is not suitable for monitoring long periods. There is, therefore, a need for an automatic and unobtrusive sleep staging approach.

In this work, a set of classifiers for infant sleep staging was created and evaluated. The cardiorespiratory and gross body movement signals were used as an input. The different classifiers aim to distinguish between two or more different sleep states. The classifiers were built on a clinical sleep polysomnography data set of 48 infants with ages ranging from 1 week to 18 weeks old (a median of 5 weeks). Respiration and gross body movements were observed using an electromechanical film bed mattress sensor manufactured by Emfit Ltd. ECG of the PSG setup was used for extracting cardiac activity. Signals were preprocessed to remove artefacts and an extensive set of features (N=81) were extracted on which the classifiers were trained.

The NREM3 vs other states classifier provided the most accurate results. The median accuracy was 0.822 (IQR:0.724-0.914). This is comparable to previously published studies on other sleep classifiers, as well as to the level of clinical interrater agreement. Classification methods were confounded by the lack of muscle atonia and amount of gross body movements in REM sleep.

The proposed method could be readily applied for home monitoring, as well as for monitoring in neonatal intensive care units.