AUTOMATED DETECTION AND CLASSIFICATION OF SLEEP-DISORDERED BREATHING FROM CONVENTIONAL POLYSOMNOGRAPHY DATA

Efficient automated detection of sleep disordered breathing (SDB) rom routine polysomnography (PSG) data is made difficult by the availabili ty of only indirect measurements of breathing. The approach we used to overcome this limitation was to incorporate pulse oximetry into the d efinitions of apnea and hypopnea. In our algorithm, 1) we begin with t he detection of desaturation as a fall in oxyhemoglobin saturation lev el of 2% or greater once a rate of descent greater than 0.1% per secon d (but less than 4% per second) has been achieved and then ask if an a pnea or hypopnea nias responsible; 2) an apnea is detected if there is a period of no breathing, as indicated by sum respiratory inductive p lethysmography (RIP), lasting at least 10 seconds and coincident with the desaturation event; and 3) if there is breathing, a hypopnea is de fined as a minimum of three breaths showing at least 20% reduction in sum RIP magnitude from the immediately preceding breath followed by a return to at least 90% of that ''baseline'' breath. Our evaluation of this algorithm using 10 PSG records containing 1,938 SDB events showed strong event-by-event agreement with manual scoring by an experienced polysomnographer. On the basis of manually verified computer desatura tions, detection sensitivity and specificity percentages were, respect ively, 73.6 and 90.8% for apneas and 84.1 and 86.1% for hypopneas. Ove rall, 93.1% of the manually detected events were detected by the algor ithm. We have designed an efficient algorithm for detecting and classi fying SDB events that emulates manual scoring with high accuracy.