The present study is the first to develop a scoring system to predict the probability of upgrade in patients diagnosed with ADH at stereotactic/DBT-guided VAB. ADH is still a challenge for breast specialists: although most ADH is benign, surgical excision is still recommended because in the impossibility of identifying a subgroup of these lesions with a sufficiently low upgrade rate to obviate surgery [11, 15]. This leads to a huge number of unnecessary surgical excisions that represent an overtreatment, especially in the era of de-escalation where even for DCIS there is growing evidence that monitoring select cases is a safe alternative to standard surgical therapy [18, 19].

We selected only ADH diagnosed with stereotactic/DBT-guided VAB to reduce heterogeneity of data. Moreover, previous studies and a recent meta-analysis demonstrated that upgrade rate is significantly lower (about half) in this type of biopsy if compared to US-guided or MRI-guided biopsies [12,13,14,15], probably related to lesion type (prevalence of microcalcifications), larger core specimens and using the vacuum-assisted device, so a part of these lesions could be potentially suitable for follow-up instead of surgical excision, if correctly identified. We also decided to exclude mutations carrier patients, due to the high malignancy association of B3 lesions in this category of patients, which suggests that these lesions must always be surgically excised in high-risk women [20].

In our study, the underestimation rate was 18.7%, in line with previously published results [9, 10, 15, 21] and, even when only upgrade rate to invasive cancers is considered, a 5.3% of underestimation was found, indicating a too high percentage to safely avoid surgery. The present study identified higher age, concurrent breast cancer, 4b/4c/5 BIRADS classification and multiple foci of ADH as independent predictors of upgrade, while post-menopausal status was a negative predictor. We found that younger and pre-menopausal women are at higher risk of upgrade, as previously found by several studies [22,23,24]. Also, BIRADS classification of the lesion was confirmed to be a factor associated with upgrade [11]. The presence of an ipsilateral or contralateral BC resulted significantly associated with the risk of upgrade, justifying the simultaneous excision of both the breast cancer and the site of ADH (with one large lumpectomy or two separate lumpectomies) to accomplish the goal of removing the known cancer and ruling out additional malignancy at the site of ADH. Our results also confirmed that multiple foci of ADH on biopsy are associated with more frequent upgrade, as previously reported in several previous studies, using a cut-off of either < 2 foci [25, 26] or < 3 foci [27, 28]. We did not found a statistically significant difference in the upgrade rate between stereotactic and DBT guidance: most of ADH lesions (87.5% in our cohort) present as microcalcifications that can be correctly identified and biopsied under stereotactic guidance, without a significant improvement when DBT guidance is used (as we would expect for architectural distortions, better depicted with DBT). The complete removal of the target lesion did not enter our model. This factor was first considered as a safe condition to justify follow-up, but was subsequently demonstrated that the upgrade rate (also in cases with complete removal of target lesion) was still too high and the recommendation for follow-up in these cases was removed from the second edition of the consensus conference on B3 lesions [5]. The dimension of the lesion was not identified as a predictor of upgrade, too, while most of the previously published studies found a significant association between residual lesion and upgrade [11, 25, 26] and also the univariate analysis of our data demonstrated an association between diameter of the lesion and upgrade (p = 0.012, data not shown). However, since the purpose of our study is to identify possible predictors of the outcome (ADH upgrade) to build a scoring system, we used a backward stepwise linear regression, which builds a regression model from a set of candidate predictor variables by removing predictors based on p values, in a stepwise manner, using an automated method. Stepwise regression did not confirm the results of univariate analysis, probably because the diameter of the lesion has a less significant impact on upgrade if compared to the other predictors that entered the model. About ADH pattern, contrasting results have been published [25, 27] and our results did not find that micropapillary pattern significantly increased the risk of malignancy. Previous studies reported that the upgrade rate is significantly reduced when ADH is found only in specimens showing microcalcifications [11], while our data did not confirm this association. Finally, our data confirmed previously published results which showed that there is no association between ADH upgrade and the presence of other B3 lesions [12, 29,30,31,32]; it seems that, when ADH is present, its own risk of upgrade overwhelms the risk associated with other B3 lesions such as FEA, LN, papilloma or RS.

Our model showed an AUC = 0.85 (95% CI 0.76–0.94), indicating a good discriminatory power. A scoring systems was then developed as a statistical tool to predict the probability of upgrade and assist clinicians in decision-making. Ko et al. [33] previously proposed a scoring system for ADH diagnosed at ultrasound-guided CNB based on clinical, imaging and pathologic features, but they tested this score in only 34 patients and a subsequent validation study [34] demonstrated the low reproducibility of this score. The present scoring system demonstrated a high diagnostic performance to identify women at low risk (< 2%) of malignant upgrade when the total score is ≤ 1, thus this cut-off value can be used to define a subset of “probably benign” lesions, corresponding to a BIRADS category 3. These lesions could be safely sent to follow-up, reducing over-treatment and consequently morbidity and economic burden. Even if novel approaches, such as artificial intelligence [35, 36] or molecular markers [37, 38], seem to be promising ways forward, our tool, based on clinical, radiological and histopathological data, easy to retrieve in any reality, can be a useful tool in daily practice.

Limitations of the present study include its retrospective design and lack of a dataset to validate the performance of our model so we are working to validate our work prospectively on a larger, independent cohort. Moreover, our model is built only on ADH diagnosed with VAB under stereotactic/DBT guidance so it can be useful for decision-making only for a subset of ADH diagnosis. Lastly, we included in the study also women who did not undergo surgery (21/112, 18.7%) so, even if the median follow-up time is relatively long (48 months), it could be possible that indolent low-grade in situ cancer in patients who were managed conservatively may have not become evident in the imaging follow-up period.

In conclusion, our scoring system, based on clinical, radiologic and histopathologic parameters, seemed a promising easy-to-use decision support tool for management of ADH, decreasing unnecessary surgeries, reducing patients’ overtreatment and healthcare costs. Further work is needed to validate our model on independent datasets.