Development and validation of an AI-enabled digital breast cancer assay to predict early-stage breast cancer recurrence within 6 years

Table 2 PDxBr training and validation: image and clinical features

Training dataset
Concordance index (C-index):	0.78
*Training sensitivity/specificity threshold:	57.77
Train sensitivity:	0.72
Train specificity:	0.77
Validation dataset
Concordance index (C-index):	0.75
Test sensitivity:	0.61
Test specificity:	0.77

Feature	Weight in final model
Proliferative activity	− 17.11
Nuclear pleomorphism	− 28.53
Age and size composite	− 11.07
Age at diagnosis	− 23.14
Stage	− 12.21
Tumor-infiltrating lymphocytes	23.48
Positive lymph nodes	− 29.34
Tumor sheets/architecture	8.36
Intact tubules	42.91

Image features: Proliferative activity: mitotic figure count; nuclear pleomorphism: nuclear shape, size, contour, chromatin content; tumor-infiltrating lymphocytes: number of intra-tumoral lymphocytes; tumor sheets/architecture: concentrated islands of tumor with and without intervening stroma; intact tubules: varying sized gland structures composed of epithelial cells with an intact lumen and adjacent stromal components. Clinical feature: age and size composite; novel feature that balances impact of tumor size as a function of age
*Model threshold of 57.77 was rounded to 58 for subsequent risk categorization and reporting

ISSN: 1465-542X