Commercially Available Artificial Intelligence Score on Preoperative Mammography for Prediction of Future Breast Cancer After DCIS Treatment.

Mammographic artificial intelligence (AI) systems have been explored for future breast cancer risk prediction. To investigate associations of scores from a commercial AI system for mammographic breast cancer detection and diagnosis with development of second breast cancers after DCIS treatment and to compare AI predictive performance with existing clinical risk models. This retrospective five-center study included 1740 patients (median age, 55.0 years) who underwent surgery for DCIS between January 2012 and December 2017 and had ≥1 year of postoperative follow-up. Medical records were reviewed to identify second breast cancers (ipsilateral recurrences after breast-conserving surgery [BCS] or mastectomy or contralateral breast cancers). A commercial AI system for breast cancer detection and diagnosis processed preoperative mammograms. AI scores were dichotomized using the Youden index for second breast cancer prediction. Univariable and multivariable cause-specific hazards models with competing-risk analysis assessed associations with second breast cancers. Cumulative incidence rates (CIRs) were compared between dichotomized AI scores using log-rank tests. Time-dependent AUCs were compared between AI scores and clinical risk models incorporating pathologic information (Van Nuys prognostic index [VNPI]; MSKCC nomograms) using bootstrapping. Twenty-eight patients developed post-BCS ipsilateral recurrence; seven developed post-mastectomy ipsilateral recurrence; 25 developed contralateral breast cancer. AI scores were dichotomized at a threshold of ≥73.5%. Post-BCS ipsilateral recurrence showed a significant independent association with AI score ≥73.5% (HR=2.88). CIR for post-BCS ipsilateral recurrence was higher for AI score ≥73.5% than for AI score <73.5% at 5 years (4.13% vs 0.86%, p<.001) and 10 years (7.26% vs 3.72%, p<.001). AUC for predicting post-BCS ipsilateral recurrence was not significantly different between AI scores and VNPI or MSKCC nomogram at 5 years (0.70 vs 0.73 [p>.99] and 0.63 [p=.82], respectively) and 10 years (0.66 vs 0.75 [p=.66] and 0.68 [p>.99], respectively). AI scores were not associated with other second breast cancer events in hazard models and CIR analyses (p>.05). AI scores showed independent associations with ipsilateral recurrence after BCS for DCIS and had predictive performance not significantly different from existing clinical models. AI scores, readily obtained noninvasively on preoperative mammography, may help inform DCIS treatment and surveillance strategies.