APNEA DETECTION - HUMAN-PERFORMANCE AND RELIABILITY OF A COMPUTER ALGORITHM

We examined the consistency of apnoea recognition between three human experts. The hypothesis was that computer detection of apnoea could em ulate human expert apnoea recognition. The aim was to detect apnoeas w ith the highest possible accuracy from a single breathing signal, by b oth human experts and computer. Three human experts independently exam ined recordings of breathing waveform from overnight sleep studies fro m 10 infants aged 3-17 weeks. All apnoeas of 5 s or more were identifi ed and reviewed. However, there still remained 10% disagreement. A com puter apnoea detector was implemented. An algorithm analysed statistic al properties of the signal to find breathing pauses. Optimal performa nce was 1% missed apnoeas (compared with the agreed apnoeas identified by the three experts) and 29% false detections. This computer algorit hm reliably identified most apnoeas but did not replace the human expe rt.