Objectives To evaluate the association between the cancer detectability by artificial intelligence (AI) and long-term survival outcomes in invasive breast cancer. Materials and methods This retrospective study analyzed consecutive women diagnosed with invasive breast cancer who underwent preoperative mammography between January and December 2013. Mammograms were analyzed using FDA-cleared AI software (Lunit INSIGHT MMG v1.1.8.2). Cancers were classified as AI-detected if correctly localized by AI, and AI-undetected if AI missed or mislocalized. Propensity score matching was performed using 29 clinical, pathological, and treatment-related covariates. Recurrence-free survival (RFS) and overall survival (OS) were compared using Kaplan–Meier estimates and Cox proportional hazards models. Results Among 879 women (mean age ± standard deviation, 50.3 ± 10.2 years), AI correctly identified cancers in 83%. Before matching, the AI-detected group had higher recurrence (11% vs 5%; p = 0.02) and mortality rates (7% vs 1%; p = 0.003). Distant recurrence was also more prevalent in AI-detected cases (p = 0.04). After matching, no differences were observed in RFS (HR, 1.7 [95% CI: 0.8, 3.9]; p = 0.20) or OS (HR, 4.1 [95% CI: 0.5, 38.1]; p = 0.21). AI detectability was not associated with RFS (HR, 1.9 [95% CI: 0.9, 3.8]; p = 0.07) or OS (HR, 5.5 [95% CI: 0.8, 40.7]; p = 0.09) in multivariable analysis. Conclusion AI-detected breast cancers showed higher recurrence and mortality rates in the unadjusted analysis. However, after adjusting for confounders, AI detectability was not associated with RFS or OS, suggesting that AI may preferentially detect tumors with aggressive characteristics. Key Points <Emphasis Type="BoldItalic">Question</Emphasis> AI-based software for mammography interpretation is increasingly being integrated into practice, but the long-term prognostic implications of breast cancers detected or undetected by AI remain unclear. <Emphasis Type="BoldItalic">Findings</Emphasis> In this retrospective study, AI detectability was not associated with recurrence-free (HR, 1.7; p = 0.20) or overall survival (HR, 4.1; p = 0.21) after propensity score matching. <Emphasis Type="BoldItalic">Clinical relevance</Emphasis> AI may be more likely to detect biologically aggressive tumors, rather than directly influencing survival. Graphical Abstract

Long-term prognostic implications of AI-detected versus AI-undetected breast cancers on mammography: a propensity score-matched analysis

期刊论文
发表日期：

2026年4月2日

查看全文

Hee Jeong Kim,
Eun Young Chae,
Hye Joung Eom,
Woo Jung Choi,
Hee Jung Shin,
Joo Hee Cha,
Hak Hee Kim

摘要

Objectives

To evaluate the association between the cancer detectability by artificial intelligence (AI) and long-term survival outcomes in invasive breast cancer.

Materials and methods

This retrospective study analyzed consecutive women diagnosed with invasive breast cancer who underwent preoperative mammography between January and December 2013. Mammograms were analyzed using FDA-cleared AI software (Lunit INSIGHT MMG v1.1.8.2). Cancers were classified as AI-detected if correctly localized by AI, and AI-undetected if AI missed or mislocalized. Propensity score matching was performed using 29 clinical, pathological, and treatment-related covariates. Recurrence-free survival (RFS) and overall survival (OS) were compared using Kaplan–Meier estimates and Cox proportional hazards models.

Results

Among 879 women (mean age ± standard deviation, 50.3 ± 10.2 years), AI correctly identified cancers in 83%. Before matching, the AI-detected group had higher recurrence (11% vs 5%; p = 0.02) and mortality rates (7% vs 1%; p = 0.003). Distant recurrence was also more prevalent in AI-detected cases (p = 0.04). After matching, no differences were observed in RFS (HR, 1.7 [95% CI: 0.8, 3.9]; p = 0.20) or OS (HR, 4.1 [95% CI: 0.5, 38.1]; p = 0.21). AI detectability was not associated with RFS (HR, 1.9 [95% CI: 0.9, 3.8]; p = 0.07) or OS (HR, 5.5 [95% CI: 0.8, 40.7]; p = 0.09) in multivariable analysis.

Conclusion

AI-detected breast cancers showed higher recurrence and mortality rates in the unadjusted analysis. However, after adjusting for confounders, AI detectability was not associated with RFS or OS, suggesting that AI may preferentially detect tumors with aggressive characteristics.

Key Points

Question AI-based software for mammography interpretation is increasingly being integrated into practice, but the long-term prognostic implications of breast cancers detected or undetected by AI remain unclear.

Findings In this retrospective study, AI detectability was not associated with recurrence-free (HR, 1.7; p = 0.20) or overall survival (HR, 4.1; p = 0.21) after propensity score matching.

Clinical relevance AI may be more likely to detect biologically aggressive tumors, rather than directly influencing survival.

Graphical Abstract