Table 1.

List of structured and extracted variables*

Structured	Variables extracted using NLP
• Age	• Calcification distribution
• Family history (of breast cancer)†	• Calcification morphology¥
• Personal history (of breast cancer)	• Mass margins
• Prior surgery‡	• Mass shape
• Palpable lump	• Architectural distortion
• Breast density	• Focal asymmetric density
• BI-RADS assessment
• Indication for exam if diagnostic
• Principal mammography findingΨ
• Mass size

*These variables were used as input to the stepwise regression to produce the models for older and younger women.

†Defined as family history of breast cancer (Minor = one or more relatives more distant than first-degree relatives, Strong = one first-degree relative with unilateral postmenopausal breast cancer, Very Strong = more than one first-degree relative with unilateral postmenopausal breast cancer, one first-degree relative with bilateral breast cancer, or one first-degree with premenopausal breast cancer).

‡Defined as prior breast surgery of any kind.

ΨPrincipal mammographic finding: architectural distortion, calcifications, asymmetry (one view), focal asymmetry (two views), developing asymmetry, mass, single dilated duct, both calcifications and something else.

¥To overcome low frequency categories, features are grouped into high probability malignancy, intermediate and typically benign categories, as described in the Breast Imaging and Reporting Data System (BI-RADS) lexicon [18].