AI and Cancer Research ── Breast ── 2026-10-31
Breast cancer: treatment, trials and AI
- Abstract
Background/Objectives: Peritumoral ultrasound may capture tissue characteristics relevant to treatment sensitivity. In this study, we examined whether malignancy-related prediction scores (PS) were associated with pathological response to neoadjuvant therapy in breast cancer.
Methods: Support vector machine classifiers were developed for 1, 3, 5, and 7 mm peritumoral regions using 1716 pathologically confirmed lesions. We assessed PS against postoperative Miller–Payne (MP) grade in 132 patients who completed neoadjuvant therapy and surgery. Ordinal logistic regression examined overall associations, interactions with hormone receptor (HR) status and Ki67, and treatment-stratified associations adjusted for both. Holm correction addressed multiple testing. The 7 mm classifier was refitted using 1584 patients outside the response cohort while retaining the originally selected features.
Results: Higher 7 mm PS was associated with lower MP grade overall (odds ratio [OR] per 0.1-unit increase, 0.79; 95% confidence interval [CI], 0.66–0.95; Holm-adjusted p = 0.048). HR and Ki67 interactions remained significant at 1 mm (adjusted p = 0.049 and 0.036, respectively). In 60 complete cases receiving chemotherapy plus anti-HER2 therapy, the original 7 mm PS was associated with MP grade after adjustment (OR, 0.60; 95% CI, 0.42–0.85; adjusted p = 0.037). The refitted PS showed an attenuated association in these patients (OR, 0.709; 95% CI, 0.526–0.955; raw p = 0.0239; Holm-adjusted p = 0.0478). Associations in the original classifier test subsets (n = 22–26) were not significant after correction.
Conclusions: The 7 mm score may complement HR and Ki67 for response assessment during chemotherapy plus anti-HER2 therapy. Interactions at 1 mm suggest variation with molecular characteristics. These findings warrant independent validation.
Journal IF-equivalent: 4.0 (OpenAlex 2-year mean citedness, value as of 2026-10-29, retrieved 2026-10-31; not the official Clarivate IF)Reference: Wang Y, Shi A, Lian J, Liang R, Wang J, Jiang J. Multiscale Peritumoral Ultrasound Machine Learning Scores and Pathological Response to Neoadjuvant Therapy in Breast Cancer. Diagnostics. 2026 Oct 29;16(21):3580. doi:10.3390/diagnostics16213580.Checked: Abstract only - Abstract
Grade III/IV chemotherapy-induced neutropenia (CIN) is a serious complication of breast cancer chemotherapy. Existing prediction models often rely on post-chemotherapy laboratory values, limiting their utility at the point of prescribing. We aimed to develop machine learning (ML) models predicting Grade III/IV CIN using exclusively pre-chemotherapy data, and to evaluate cross-cycle stability within the same cohort. We retrospectively analysed breast cancer patients who received chemotherapy from September 2011 to January 2025. Only pre-chemotherapy variables were used as inputs. Cycle 1 records were split into training (80%) and internal validation (20%) sets. Cycles 2-8 records formed the cross-cycle test set. Nine candidate ML classifiers were evaluated. Hyperparameter tuning, Youden-index threshold optimisation, calibration, decision curve analysis, SHAP-based feature attribution, and per-cycle AUC were assessed. On the Cycle 1 validation set, the tuned Logistic L1 and Extra Trees models achieved AUCs of 0.760 (95% CI 0.703-0.812) and 0.763 (95% CI 0.707-0.815), respectively. On the pooled Cycles 2-8 cross-cycle test set, the corresponding AUCs were 0.746 (95% CI 0.727-0.764) and 0.729 (95% CI 0.709-0.748). Logistic L1 maintained an AUC above 0.70 across all cycles. SHAP attribution identified docetaxel and epirubicin as dominant predictors. Baseline neutrophil count, total bilirubin, and albumin were the leading non-drug clinical features, particularly in Logistic L1. Models trained exclusively on pre-chemotherapy data retained broadly similar discrimination when applied to subsequent cycles within the same cohort. Logistic L1 showed more consistent cross-cycle discrimination, which supports prospective, cycle-by-cycle risk stratification supporting G-CSF prophylaxis and dose-modification decisions.
Journal IF-equivalent: 2.0 (OpenAlex 2-year mean citedness, value as of 2026-10-29, retrieved 2026-10-31; not the official Clarivate IF)Reference: Lin M, Weng Y, Chen W, Cai C, Wu Y, Guo J, et al. Pre-chemotherapy machine learning models for predicting grade III/IV chemotherapy-induced neutropenia in breast cancer: training on cycle 1 and cross-cycle evaluation. Support Care Cancer. 2026 Oct 29;34(11):1173. doi:10.1007/s00520-026-11402-6. PMID: 42955369.Checked: Abstract only