Predicting diabetic retinopathy and identifying interpretable biomedical features using machine learning algorithms

Table 7 Performance comparison with previous studies

Approaches	Data Sets	Patients	Features	AUC	Acc.	Sens.	Spec.
Comparison 1
Hosseini et al.	Taiwan	212	10	0.796	0.717	0.745	0.689
Hosseini et al.	Iran	3734	11	0.704	NA^b	0.603	0.694
Comparison 2
Oh et al.	Taiwan	212	10	0.823	0.771	0.784	0.757
Oh et al.	South Korea	490	37	0.820	0.752	0.721	0.760
Comparison 3
Ogunyemi et al.	Taiwan	212	10	0.744	0.667	0.682	0.650
Ogunyemi et al.	United States	513	24	0.720	0.735	0.692	0.559

AUC, accuracy, sensitivity, and specificity of our Taiwan data set are compared with the Iran data set in Comparison 1 (i.e., using Hosseini et al.’s approach), with the South Korea data set in Comparison 2 (i.e., using Oh et al.’s approach), with the United States data set in Comparison 3 (i.e., using Ogunyemi et al.’s approach)
^aBest evaluation measures in each comparison are underlined
^bNA stands for “Not Available” because this evaluation measure was not reported in the study

ISSN: 1471-2105