Fig. 2: Mitigating ageism of AI in ophthalmology.
From: Fairer AI in ophthalmology via implicit fairness learning for mitigating sexism and ageism
a Sample distribution of patient age (in 10-year intervals) for the OculoScope test set. The unbalanced sample distribution of patient age may cause AI to favor the patients in the age group that contains more samples in the screening dataset, which will lead to the problem of AI ageism. In this study, we show that implicit fairness learning based on the relationship between ophthalmic diseases and fundus features can not only mitigate the unfairness that arises due to different sensitive attributes but also improve the screening accuracy for multiple diseases. b FairerOPTH is developed and validated with the OculoScope dataset, which contains data from 8405 patients with ages ranging from 0 to 90. Within each age division, FairerOPTH has obvious advantages in four fairness metrics (ΔA (average screening disparity), ΔM (max screening disparity), DPM (demographic disparity metric), and EOM (equality of opportunity metric)2) compared with the baseline model. The smaller ΔA and ΔM are, the better. c Details of the fairness metric ΔD (screening quality disparity) for 38 diseases. d Details of PQD (predictive quality disparity)2 for 38 diseases. Source data are provided as a Source Data file.
