A statistical pattern recognition model for sleep apnea syndrome

Intervals of reduced or absent respiration (nasal/oral events) during sleep causing a drop in arterial blood oxygen saturation (hypoxemia) can result in life threatening cardiac and hemodynamic changes and sleep fragmentation. Due to the volume of data involved, comprehensive manual analysis of hypoxemic and nasal/oral events for an all-night recording is prohibitively time consuming. An automatic technique is required to quantify these events.;Four things contribute to the difficulty of recognizing nasal/oral events: (1) the output from nasal/oral airflow transducers (thermistors) is often artifactual, (2) structures of interest in the %SaO2 and nasal/oral waveforms have a wide dynamic range, (3) event duration is variable and (4) the system is nonstationary. A structural pattern recognition method is used to segment the continuous nasal/oral flow and %SaO2 waveforms into sequences of discrete structures. Primitive pattern elements are identified using a nonparametric statistical trend technique which is robust in the presence of data irregularity and capable of responding to subtle as well as steep trends. Statistical relationships between structures segmented from the %SaO2 waveform and structures segmented from the nasal/oral waveform provide an a-priori estimate of the location and duration of a hypothesized nasal/oral event. A hypothesize and test paradigm using a nonparametric change in distribution location test is used to detect the presence of an event within the estimated range.;Validation of the model is based on clinically accepted criteria. 179 events were scored both manually and automatically. Event duration estimates between the two techniques were within +/...