Fig. 5: Uncertainty analysis across demographic groups for models trained with 25% and 100% of labeled data. Violin plots show entropy distributions for four models across. | npj Digital Medicine

Fig. 5: Uncertainty analysis across demographic groups for models trained with 25% and 100% of labeled data. Violin plots show entropy distributions for four models across.

From: Clinically informed semi-supervised learning improves disease annotation and equity from electronic health records: a glaucoma case study

Fig. 5: Uncertainty analysis across demographic groups for models trained with 25% and 100% of labeled data. Violin plots show entropy distributions for four models across.The alternative text for this image may have been generated using AI.

a race, b gender, and c age groups. Light colors indicate 25% training data; dark colors indicate 100% data. Ci-SSGAN (gray) maintains lowest uncertainty ( < 0.2 entropy) regardless of data size, while BERT models (red/green) show high uncertainty with 25% data (0.6–1.4 entropy) that reduces substantially with 100% data (0.2–0.4 entropy). SSGAN (blue) exhibits intermediate performance. The horizontal line indicates the median and the error bars represent the minimum and maximum values observed for each group.

Back to article page