Frequency and characteristics of errors by artificial intelligence (AI) in reading screening mammography: a systematic review.

purposeArtificial intelligence (AI) for reading breast screening mammograms could potentially replace (some) human-reading and improve screening effectiveness. This systematic review aims to identify and quantify the types of AI errors to better understand the consequences of implementing this technology.

methodsElectronic databases were searched for external validation studies of the accuracy of AI algorithms in real-world screening mammograms. Descriptive synthesis was performed on error types and frequency. False negative proportions (FNP) and false positive proportions (FPP) were pooled within AI positivity thresholds using random-effects meta-analysis.

resultsSeven retrospective studies (447,676 examinations; published 2019-2022) met inclusion criteria. Five studies reported AI error as false negatives or false positives. Pooled FPP decreased incrementally with increasing positivity threshold (71.83% [95% CI 69.67, 73.90] at Transpara 3 to 10.77% [95% CI 8.34, 13.79] at Transpara 9). Pooled FNP increased incrementally from 0.02% [95% CI 0.01, 0.03] (Transpara 3) to 0.12% [95% CI 0.06, 0.26] (Transpara 9), consistent with a trade-off with FPP. Heterogeneity within thresholds reflected algorithm version and completeness of the reference standard. Other forms of AI error were reported rarely (location error and technical error in one study each).

conclusionAI errors are largely interpreted in the framework of test accuracy. FP and FN errors show expected variability not only by positivity threshold, but also by algorithm version and study quality. Reporting of other forms of AI errors is sparse, despite their potential implications for adoption of the technology. Considering broader types of AI error would add nuance to reporting that can inform inferences about AI's utility.