Do not go with just the flow: machine learning in oximetric versus flow-based sleep apnoea scoring

Introduction: Traditionally apnoeas and hypopneas have been defined using respiratory flow, but relying on flow alone cannot identify false-positive events that would over-diagnose sleep apnoea. Aim: To confirm our clinical impression that false-positive flow-determined events can be confirmed and more precisely described by using cluster analysis, an unsupervised form of machine learning. Methods: Traditional flow-based apnoea hypopnoea indices (AHI) and validated oximetry-based AHI estimates (ODI) were obtained from the automated scoring of 1000 sleep polygraphs submitted for interpretation. K means clustering was performed on the paired standardized indices. The within-cluster sum of squares (wss) was plotted against the number of potential clusters to select an optimal number of clusters. The optimal number of clusters was then plotted. Results: A bend in the plot of the wss versus the number of potential clusters suggested the optimal number of clusters was 4. Two clusters of lower and midrange AHI clusters were not segregated by ODI. Higher AHI subjects were divided into two clusters roughly midway through the range of ODI values. Clinically the higher AHI with lower ODI cluster subjects often have low amplitude baseline nasal pressure, intermittent mouth breathing, or other technical errors. Conclusion: Clinical observation and unsupervised machine learning both confirm that separate oximetry-based scoring can identify false-positive flow-based scoring of respiratory events to prevent the over-diagnosis of sleep apnoea.