Interaction between clinicians and artificial intelligence to detect fetal atrioventricular septal defects on ultrasound: how can we optimize collaborative performance?

OBJECTIVES: Artificial intelligence (AI) has shown promise in improving the performance of fetal ultrasound screening in detecting congenital heart disease (CHD). The effect of giving AI advice to human operators has not been studied in this context. Giving additional information about AI model workings, such as confidence scores for AI predictions, may be a way of improving performance further. Our aims were to investigate whether AI advice improved overall diagnostic accuracy (using a single CHD lesion as an exemplar), and to see what, if any, additional information given to clinicians optimized the overall performance of the clinician-AI team.

METHODS: An AI model was trained to classify a single fetal CHD lesion (atrioventricular septal defect, AVSD), using a retrospective cohort of 121,130 cardiac four chamber images extracted from 173 ultrasound scan videos (98 with normal hearts, 75 with AVSD). A ResNet50 model architecture was used. Temperature scaling of model prediction probability was performed on a validation set, and gradient-weighted class activation maps (grad-CAMs) produced. Ten clinicians (two consultant fetal cardiologists, three trainees in pediatric cardiology, and five fetal cardiac sonographers) were recruited from a center of fetal cardiology to participate. Each participant was shown 2000 fetal four chamber images in a random order (1,000 normal and 1,000 AVSD). The dataset was comprised of 500 images, each shown in four conditions: 1) image alone without AI output; 2) image with binary AI classification; 3) image with AI model confidence; 4) image with gradient-weighted class activation map image overlays. The clinicians were asked to classify each image as normal or AVSD.

RESULTS: 20,000 image classifications were recorded from 10 clinicians. The AI model alone achieved an accuracy of 0.798 (95% CI 0.760 - 0.832), sensitivity of 0.868 (95% CI 0.834 - 0.902) and specificity of 0.728 (95% CI 0.702 - 0.754, and the clinicians without AI achieved an accuracy of 0.844 (95% CI 0.834 - 0.854), sensitivity of 0.827 (95% CI 0.795 - 0.858) and specificity of 0.861 (95% CI 0.828 - 0.895). Showing a binary (normal or AVSD) AI model output resulted in significant improvement in accuracy to 0.865 (p <0.001). This effect was seen in both experienced and less experienced participants. Giving incorrect AI advice resulted in significant deterioration in overall accuracy from 0.761 to 0.693 (p <0.001), which was driven by an increase in both type I and type II error by the clinicians. This effect was worsened by showing model confidence (accuracy 0.649, p <0.001) or grad-CAM (accuracy 0.644, p <0.001).

CONCLUSIONS: AI has the potential to improve performance when used in collaboration with clinicians, even if the model performance does not reach expert level. Giving additional information about model workings such as model confidence and class activation map image overlays did not improve overall performance, and actually worsened performance for images where the AI model was incorrect. This article is protected by copyright. All rights reserved.