Image-only deep learning risk model performance in patients with known germline mutations.

536 Background: Mammograms contain highly predictive biomarkers of future breast cancer risk, which can be identified by deep learning (DL) risk stratification models. DL model performance in discerning risk in patients with known germline mutations has yet to be established. The aim of this study is to assess image-only DL model risk score performance in patients with and without a germline mutation undergoing screening breast MRI. Methods: This retrospective, multisite study included the first screening MRI of consecutive females ≥30 years from 3/1/2011 to 11/30/2021 at four facilities, with a prior bilateral screening mammogram containing a DL risk score and at least one year of follow-up. A mammography-based DL five-year model was used to assess risk, with an increased-risk threshold of ≥2.2 and high-risk of ≥5.8. Patient demographics, germline mutation status, and traditional risk factors were extracted from electronic medical records. Standard screening performance metrics including cancer detection rate (CDR) within one year of MRI were determined. Cancer outcomes were determined via linkage to a regional tumor registry. DL model performance was compared using Pearson’s chi-squared tests. Results: 4025 patients who underwent screening breast MRI met inclusion criteria. 381(9.5%) had a known germline mutation. There was no significant difference in age of those with or without a germline mutation (54.9y ± 10.5 vs 54.8y ± 8.8, p = 0.284) or family history of breast cancer (68.5% vs 65.3%, p = 0.205). Patients with a germline mutation were more likely to be White (95.0 vs 91.4%, p < 0.05), have non-dense breasts (52.8% vs 35.3%, p < 0.001) and less likely to have a personal history of breast cancer (25.7% vs 36.3%, p < 0.001). The CDR in those with a germline mutation was higher than those without a mutation (39.4 vs 20.3, p < 0.05). The median DL risk score in those with a germline mutation was lower than those without a mutation: 1.9 (IQR: 1.4-2.5) vs 2.3 (IQR: 1.7-3.8), p < 0.001. In patients with a germline mutation, the CDR in those with elevated DL scores was higher than those with low DL scores (47.6 vs 34.2, p = 0.512: increased-risk threshold; 60.6 vs 37.4, p = 0.512: high-risk threshold). A similar trend was seen in those without a germline mutation (26.4 vs 12.8, p < 0.01: increased-risk threshold; 27.1 vs 19.0, p = 0.203: high-risk threshold). Trends were demonstrated across all comparisons, but significance was only reached in the largest group. Conclusions: Compared to patients without germline mutations, those with mutations had higher CDRs on MRI but lower DL risk scores. Patients with higher DL risk scores had higher CDRs in those identified as increased or high-risk by DL model, regardless of germline mutation status. Further study to elucidate mechanisms to explain differential performance of DL models in patients with germline mutations is ongoing.