Can incorrect artificial intelligence (AI) results impact radiologists, and if so, what can we do about it? A multi-reader pilot study of lung cancer detection with chest radiography

Abstract

Objective: To examine whether incorrect AI results impact radiologist performance, and if so, whether human factors can be optimized to reduce error.

Methods: Multi-reader design, 6 radiologists interpreted 90 identical chest radiographs (follow-up CT needed: yes/no) on four occasions (09/20-01/22). No AI result was provided for session 1. Sham AI results were provided for sessions 2-4, and AI for 12 cases were manipulated to be incorrect (8 false positives (FP), 4 false negatives (FN)) (0.87 ROC-AUC). In the Delete AI (No Box) condition, radiologists were told AI results would not be saved for the evaluation. In Keep AI (No Box) and Keep AI (Box), radiologists were told results would be saved. In Keep AI (Box), the ostensible AI program visually outlined the region of suspicion. AI results were constant between conditions.

Results: Relative to the No AI condition (FN = 2.7%, FP = 51.4%), FN and FPs were higher in the Keep AI (No Box) (FN = 33.0%, FP = 86.0%), Delete AI (No Box) (FN = 26.7%, FP = 80.5%), and Keep AI (Box) (FN = to 20.7%, FP = 80.5%) conditions (all ps < 0.05). FNs were higher in the Keep AI (No Box) condition (33.0%) than in the Keep AI (Box) condition (20.7%) (p = 0.04). FPs were higher in the Keep AI (No Box) (86.0%) condition than in the Delete AI (No Box) condition (80.5%) (p = 0.03).

Conclusion: Incorrect AI causes radiologists to make incorrect follow-up decisions when they were correct without AI. This effect is mitigated when radiologists believe AI will be deleted from the patient's file or a box is provided around the region of interest.

Clinical relevance statement: When AI is wrong, radiologists make more errors than they would have without AI. Based on human factors psychology, our manuscript provides evidence for two AI implementation strategies that reduce the deleterious effects of incorrect AI.

Key points: • When AI provided incorrect results, false negative and false positive rates among the radiologists increased. • False positives decreased when AI results were deleted, versus kept, in the patient's record. • False negatives and false positives decreased when AI visually outlined the region of suspicion.

Keywords: Artificial intelligence; Cognitive science; Psychology.

Fig. 1

Study flow chart. Overall experimental procedure is displayed. The order of Keep AI (No Box) and Delete AI (No Box) was counterbalanced such that n = 3 participated in Keep AI (No Box) first, and n = 2 participated in Delete AI (No Box) first

Fig. 2

Procedural overview. A brief description of the major procedural elements within each session is shown

Fig. 3

False negatives (incorrect AI feedback) by experimental condition. False negative percent (y-axis) is shown for the No AI (red), Keep AI (No Box) (brown), Delete AI (No Box) (green), and Keep AI (Box) (blue) conditions (x-axis). Mean (circle) and 95% confidence intervals are displayed. Results display the four conditions for the 4 cases where AI provided false negative feedback

Fig. 4

False positives (incorrect AI feedback) by experimental condition. False positive percent (y-axis) is shown for the No AI (red), Keep AI (No Box) (brown), Delete AI (No Box) (green), and Keep AI (Box) (blue) conditions (x-axis). Mean (circle) and 95% confidence intervals are displayed. Results display the four conditions for the 8 cases where AI provided false positive feedback