AI may predict interval breast cancers years before diagnosis
New research suggests AI can identify high-risk patterns in normal mammograms, though experts warn that prospective clinical data is still lacking.
Artificial intelligence may be able to predict the development of interval breast cancers years before they become visible to human radiologists. By analyzing patterns invisible to the human eye, AI systems can assign risk scores to mammograms that appear technically normal, offering a potential window for earlier intervention.
According to a review of AI systems for risk prediction, the technology has demonstrated a significant ability to flag future cases in retrospective data. In one study, AI assigned high-risk scores to 23% of women who eventually developed interval cancer three screening rounds before their actual diagnosis. This predictive accuracy increased to 39% on the mammogram immediately preceding the diagnosis.
The Challenge of Interval Cancers
Interval cancers are those that emerge between scheduled screening rounds, often appearing unexpectedly. While traditional radiology relies on detecting existing masses or calcifications, AI analyzes subtle textural and structural patterns to estimate future risk. However, the current landscape of research is fragmented. Experts note that the field lacks standardization, as different studies use varying definitions of interval cancer, different screening intervals, and diverse AI algorithms.
Clinical Tension and Overdiagnosis
The ability to flag a "high-risk" patient based on a negative mammogram creates a complex clinical dilemma. There is a significant tension between the goal of earlier detection and the risk of overdiagnosis. If clinicians act on AI risk scores without a visible tumor, it could lead to a surge in false positives, unnecessary patient recalls, and invasive biopsies that may not have been required.
The Path to Prospective Proof
Despite the promising retrospective results, medical experts caution that the technology is not yet a proven clinical solution. The primary concern is that finding a cancer in a historical dataset is fundamentally different from implementing the tool in a live environment.
"Finding a cancer retrospectively is very different from demonstrating that using AI during routine screening would have led to an earlier diagnosis," said Dr. Hannah Milch, an associate professor of radiology at the David Geffen School of Medicine at UCLA. Milch emphasized that the industry requires prospective evidence to prove that AI actually improves patient outcomes rather than simply increasing the burden on healthcare systems.
What remains to be seen is whether standardized, real-world trials can validate these risk scores. Until prospective studies can confirm that AI-driven screening schedules reduce mortality without causing widespread over-treatment, the technology remains a promising research tool rather than a standard of care.