Real-world performance of a Food and Drug Administration-approved artificial intelligence tool for mammography in patients with prior breast surgery or indwelling biopsy clips.
Authors
Affiliations (1)
Affiliations (1)
- Russell H. Morgan Department of Radiology and Radiological Science, Johns Hopkins University School of Medicine, Baltimore, MD 21287, United States.
Abstract
Artificial intelligence (AI) tools are increasingly used in clinical screening mammography, but their performance in patients with prior breast surgery or indwelling biopsy clips, who represent a substantial proportion of screening populations, has not been well characterized. To evaluate the real-world performance of an Food and Drug Administration-approved AI mammography tool in patients with prior breast surgery or biopsy clips. We retrospectively reviewed all screening mammograms performed at our institution between August 1, 2024, and February 28, 2025, after implementation of a commercially available AI tool (Transpara 1.7.4-A), merging clinical outcomes with the AI risk score (low, intermediate, or elevated). Patients were stratified by procedural history into 3 groups: prior breast surgery, biopsy clip, and reference (ie, neither intervention). Sensitivity, positive predictive value (PPV), and false positive rate (FPR) at the elevated and intermediate + elevated thresholds were compared using Fisher exact tests presented as percentage (95% confidence interval [CI]). A total of 24 427 screening mammograms were included (mean age 61 ± 13 years): 19 287 in the reference group, 2647 in the surgery group, and 2493 in the biopsy clip group. AI sensitivity at the intermediate + elevated threshold was preserved (reference 92.1% [95% CI, 86.9%-95.3%], surgery 90.5% [95% CI, 71.1%-97.3%], biopsy clip 90.0% [95% CI, 69.9%-97.2%]; both <i>P</i> > .6). The AI-PPV of an elevated exam was correspondingly reduced in altered-breast patients (reference 13.6% [95% CI, 11.1%-16.6%], surgery 5.5% [95% CI, 2.9%-10.0%], biopsy clip 4.9% [95% CI, 2.5%-9.4%]; both <i>P</i> < .005 vs reference). The AI-FPR was higher in both subgroups at the intermediate + elevated threshold (reference 35.6% [95% CI, 34.9%-36.3%] vs surgery 47.9% [95% CI, 46.0%-49.9%] and biopsy clip 47.9% [95% CI, 45.9%-49.8%]; both <i>P</i> < .001) and more than doubled at the elevated threshold (reference 2.7% [95% CI, 2.5%-2.9%] vs surgery 5.9% [95% CI, 5.1%-6.9%] and biopsy clip 6.2% [95% CI, 5.3%-7.2%]; both <i>P</i> < .001). AI examination-level sensitivity for breast cancer was preserved in patients with prior breast surgery or indwelling biopsy clips. However, an elevated AI risk score was less likely to reflect a cancer in these patients than in the reference group, with a lower AI-PPV and higher AI-FPR.