Skip to main content
. Author manuscript; available in PMC: 2022 Aug 5.
Published in final edited form as: IEEE J Biomed Health Inform. 2021 Aug 5;25(8):2906–2916. doi: 10.1109/JBHI.2020.3048901

TABLE IV.

Diagnostic performance of AHICNN vs. ODI3 and AHIMLP in the CHAT, UofC and BUH test databases

Test set ICC RMSE 4-class kappa 4-class Accuracy (%)
CHAT AHICNN 0.960 2.89 0.515 72.8
ODI3 0.871 4.63 0.417 65.1
AHIMLP 0.832 5.51 0.377 63.3
UofC AHICNN 0.917 5.45 0.422 60.2
ODI3 0.861 6.21 0.372 56.6
AHIMLP 0.890 6.02 0.381 56.9
BUH AHICNN 0.583 10.44 0.423 61.0
ODI3 0.520 10.64 0.369 57.6
AHIMLP 0.500 11.05 0.306 52.4

AHICNN = apnea-hypopnea index (AHI) estimated by our convolutional neural network architecture, ODI3 = 3% oxygen desaturation index, AHIMLP = AHI estimated by the multi-layer perceptron neural network trained with features from the blood oxygen saturation (SpO2) signal, ICC = intra-class correlation coefficient, RMSE = root mean squared error, kappa = Cohen’s kappa index, CHAT = Childhood Adenotonsillectomy Trial, UofC = University of Chicago, BUH = Burgos University Hospital.