Summary Between January 1994 and July 1997, 793 patients suspected of having sleep-disordered breathing had unattended overnight oximetry in their homes followed by laboratory polysomnography. From the oximetry data we extracted cumulative percentage time at SaO2 < 90% (CT90) and a saturation variability index (ΔIndex, the sum of the differences between successive readings divided by the number of readings – 1). CT90 was weakly correlated with polysomnographic apnea/hypopnea index (AHI), (Spearman ö=0.36, P< 0.0001) and with ΔIndex (ö=0.71, P< 0.0001). ΔIndex was more closely correlated with AHI (ö=0.59, P< 0.0001). In a multivariate model, only ΔIndex was significantly related to AHI, the relationship being AHI=18.8 ΔIndex +7.7. The 95% CI for the coefficient were 16.2, 21.4, and for the constant were 5.8, 9.7. The sensitivity of a ΔIndex cut-off of 0.4 for the detection of AHI≥15 was 88%, for detection of AHI≥20 was 90% and for the detection of AHI≥25 was 91%. The specificity of ΔIndex ≥0.4 for AHI≥15 was 40%. In 113 further patients, oximetry was performed simultaneously with laboratory polysomnography. Under these circumstances ΔIndex was more closely correlated with AHI (ö=0.74, P< 0.0001), as was CT90 (ö=0.58, P< 0.0001). Sensitivity of ΔIndex ≥0.4 for detection of AHI≥15 was not improved at 88%, but specificity was better at 70%. We concluded that oximetry using a saturation variability index is sensitive but nonspecific for the detection of obstructive sleep apnea, and that few false negative but a significant proportion of false positive results arise from night-to-night variability.