Background:  Active surveillance is currently considered only for low/intermediate-grade ductal carcinoma in situ (DCIS). The key vulnerability is baseline misclassification: a lesion labelled as G1–G2 on vacuum-assisted breast biopsy (VABB) may harbour occult aggressive disease (high-grade G3 DCIS and/or invasion) revealed only at surgery. Mammographic radiomics could provide additional, reproducible information to reduce this misclassification. Objectives: To evaluate whether mammographic radiomic features and clinical variables improve preoperative identification of occult aggressive disease in VABB-diagnosed DCIS, with implications for active surveillance selection. As a complementary analysis, we assessed prediction of invasive upstaging alone using radiomic and clinical features. Inter-reader segmentation agreement and robustness of radiomic features between two readers were assessed. Methods: We retrospectively included consecutive patients with DCIS diagnosed by VABB at the European Institute of Oncology (December 2009–November 2016) who underwent surgical excision. Clinical data including menopausal status, the presence of residual lesion after VABB procedure and the molecular receptor profile were collected. Two independent radiologists segmented each patient’s mammogram and radiomic features were extracted. The Dice Similarity Coefficient (DSC) and Interclass Correlation Coefficient (ICC) were calculated to compare the two segmentations. The radiomic features were selected by applying the minimum redundancy maximum relevance method (mRMR) on 1000 different versions of the original radiomic data. Logistic models with the most frequently selected features were estimated to generate the radiomic score (RS) for each patient. A 5?fold cross?validation (CV) scheme was implemented, computing AUCs on the corresponding test sets and comparing model performance using the DeLong test. Results: Among 394 patients, 96 (24%) were upstaged to invasive breast cancer at surgery. A total of 56 radiomic features were extracted. The median DSC was 0.82 (IQR 0.74-0.87) and only 13% of the patients had values lower than 0.7; the median ICC was 0.95 (IQR 0.92-1.00), signalling good overlapping between ROIs. A clinical-only multivariable model showed moderate discrimination for invasive upstaging (AUC 0.66, 95% CI 0.60-0.73): premenopausal status, residual lesion after VABB and Ki67 >14% were positively associated with upstaging, whereas PgR >50% and HER2 expression were inversely associated. Adding the RS did not significantly enhance model performance over clinical variables alone (radiologist 1: AUC=0.69, 95% CI 0.63-0.76, Delong p-value 0.09; radiologist 2: AUC=0.70, 95% CI 0.63-0.76, Delong p-value 0.46). In the main analysis focused on occult aggressive disease (invasive or G3 DCIS versus G1-G2 lesions), the association between clinical variables and outcome was consistent with what was obtained above, whereas RS improved discrimination compared to the clinical model (AUC without RS 0.72, 95% CI 0.67-0.77; AUC with RS 0.76, 95% CI 0.71-0.81; DeLong p=0.02). In contrast, no improvement in discrimination was observed for features derived from the second segmentation, despite RS remaining statistically significant. Conclusions: Mammographic radiomic features were reproducible across readers. While incremental value for predicting invasive upstaging alone was limited, radiomics improved discrimination of an aggressive phenotype (G3 and/or invasive disease), the misclassification most consequential for safe selection to active surveillance. External validation, ideally within biopsy G1–G2 candidates, is warranted.