Dual-Center Ultrasound Radiomics Model Predicts Complete Response to Neoadjuvant Chemotherapy in 659 HER2-Positive Breast Cancers

One scan, taken before the first cycle

The input is the pre-treatment ultrasound study, the same examination that produced the biopsy target. From it the group derived a radiomics signature, pruned the feature set for redundancy and dimensionality "to mitigate overfitting" in their phrasing, ran LASSO regression to select what survived, then folded in the clinical variables that came out as independent risk factors. Against the surgical specimen, the clinical model alone reached an AUC of 0.699 in the training cohort and the radiomics signature alone 0.817. The integrated model came in at 0.861, and the authors report the DeLong test confirming that gap.

The numbers the abstract withholds

Europe PMC lists the full-text availability as subscription required, and Elsevier's paywall bears that out. The peer-reviewed abstract, the PubMed record and the Crossref metadata were the whole of what could be read. Several things a reader would want are not in that material. No confidence intervals accompany any of the three AUC figures. The 0.699 and 0.817 values are tagged to the training cohort while the 0.861 is described as holding "across all cohorts," which leaves the external-validation figure unstated as a separate number. Nothing appears on calibration, decision-curve analysis, or how the model compares with a radiologist reading the same images. The abstract also does not say which pathological complete response definition was applied, whether residual ductal carcinoma in situ was permitted, or whether the neoadjuvant regimens included HER2-directed antibodies, a detail that governs the response rate in this population and therefore the class balance the model was trained against.

Center 2 supplied 255 cases the model never saw

Center 1 contributed 404 of the 659 patients, split 283 for training and 121 for internal validation, leaving 255 external cases from Center 2. Holding back an entire hospital, rather than a random slice of one hospital's patients, is the harder test, and a holdout of that size is substantial by radiomics standards. The two accrual windows overlap only partially, with Center 1 reaching back to 2015 and Center 2 starting in 2018, which means the two sets were imaged on different equipment across different years of HER2 treatment practice. The authors' stated conclusion is that "the clinical-radiomics combined model demonstrates improved and stable predictive performance for early response to NACT compared with radiomics-only and clinical-only models in both internal and external validation cohorts." Stability across two centers is the claim. The paper makes no claim about transferability to a scanner fleet outside Jiangsu and Shanghai.

Why This Matters to the APO|APE Reader

The label this model is trying to anticipate comes out of the surgical pathology report months later, which makes the histology bench the arbiter of every AUC quoted in the paper, including the 0.861. A test that guesses residual disease status off a pre-treatment scan changes nothing about how that specimen is grossed or graded, but it does change which patients arrive at surgery having had their regimen altered midstream, and any laboratory asked to supply ground truth for an external validation set will be asked for its pCR criteria in writing. Until the full text is readable, the criteria this group used remain unpublished outside the paywall.