Abstract
Purpose: Artificial intelligence (AI) solutions for medical diagnosis require thorough evaluation to demonstrate that performance is maintained for all patient sub-groups and to ensure that proposed improvements in care will be delivered equitably. This study evaluates the robustness of an AI solution for the diagnosis of normal chest X-rays (CXRs) by comparing performance across multiple patient and environmental subgroups, as well as comparing AI errors with those made by human experts. Methods: A total of 4,060 CXRs were sampled to represent a diverse dataset of NHS patients and care settings. Ground-truth labels were assigned by a 3-radiologist panel. AI performance was evaluated against assigned labels and sub-groups analysis was conducted against patient age and sex, as well as CXR view, modality, device manufacturer and hospital site. Results: The AI solution was able to remove 18.5% of the dataset by classification as High Confidence Normal (HCN). This was associated with a negative predictive value (NPV) of 96.0%, compared to 89.1% for diagnosis of normal scans by radiologists. In all AI false negative (FN) cases, a radiologist was found to have also made the same error when compared to final ground-truth labels. Subgroup analysis showed no statistically significant variations in AI performance, whilst reduced normal classification was observed in data from some hospital sites. Conclusion: We show the AI solution could provide meaningful workload savings by diagnosis of 18.5% of scans as HCN with a superior NPV to human readers. The AI solution is shown to perform well across patient subgroups and error cases were shown to be subjective or subtle in nature.
Abstract (translated)
URL
https://arxiv.org/abs/2209.09204
