Apnoea detection: human performance and reliability of a computer algorithm

Abstract

We examined the consistency of apnoea recognition between three human experts. The hypothesis was that computer detection of apnoea could emulate human expert apnoea recognition. The aim was to detect apnoeas with the highest possible accuracy from a single breathing signal, by both human experts and computer. Three human experts independently examined recordings of breathing wave-form from overnight sleep studies from 10 infants aged 3-17 weeks. All apnoeas of 5 s or more were identified and reviewed. However, there still remained 10% disagreement. A computer apnoea detector was implemented. An algorithm analysed statistical properties of the signal to find breathing pauses. Optimal performance was 1% missed apnoeas (compared with the agreed apnoeas identified by the three experts) and 29% false detections. This computer algorithm reliably identified most apnoeas but did not replace the human expert.