Machine learning models predict severe chemotherapy-induced neutropenia in breast cancer using only pre-treatment data

Researchers have developed machine learning models that predict severe chemotherapy-induced neutropenia in breast cancer patients using exclusively pre-chemotherapy data, according to a new study published in Supportive Care in Cancer. The work addresses a common limitation of existing prediction tools, which often depend on laboratory values measured after chemotherapy has been given, restricting their usefulness at the moment of prescribing.

Grade III/IV chemotherapy-induced neutropenia is a serious complication of breast cancer chemotherapy. The research team, led by Min Lin and Quanyao Chen with colleagues at multiple institutions, aimed to build models that could flag patients at high risk before treatment begins, and to test how well those models hold up across successive treatment cycles within the same cohort.

The study was retrospective, analysing breast cancer patients who received chemotherapy between September 2011 and January 2025. Only pre-chemotherapy variables were used as model inputs. Records from the first treatment cycle were split into a training set containing 80 percent of cases and an internal validation set containing the remaining 20 percent, while records from cycles 2 through 8 formed a separate cross-cycle test set.

Nine candidate machine learning classifiers were evaluated. The researchers applied hyperparameter tuning, optimised classification thresholds using the Youden index, and assessed model calibration, decision curve analysis, SHAP-based feature attribution, and per-cycle area under the receiver operating characteristic curve.

On the cycle 1 validation set, the tuned Logistic L1 and Extra Trees models achieved AUCs of 0.760 (95 percent CI 0.703 to 0.812) and 0.763 (95 percent CI 0.707 to 0.815), respectively. On the pooled cycles 2 to 8 cross-cycle test set, the corresponding AUCs were 0.746 (95 percent CI 0.727 to 0.764) for Logistic L1 and 0.729 (95 percent CI 0.709 to 0.748) for Extra Trees.

Logistic L1 maintained an AUC above 0.70 across all cycles, indicating more consistent discrimination when applied to later treatment cycles than the other models tested. The authors report that models trained exclusively on pre-chemotherapy data retained broadly similar discrimination when applied to subsequent cycles within the same cohort.

SHAP-based attribution identified the chemotherapy agents docetaxel and epirubicin as the dominant predictors of severe neutropenia. Among non-drug clinical features, baseline neutrophil count, total bilirubin, and albumin were the leading contributors, particularly in the Logistic L1 model.

The authors conclude that the approach supports prospective, cycle-by-cycle risk stratification to inform decisions about G-CSF prophylaxis and dose modification. Because the study is retrospective and evaluated within a single cohort, the models require prospective validation before clinical implementation, and the reported performance reflects discrimination within this dataset rather than demonstrated effects on patient outcomes.

Subject of Research: Machine learning prediction of severe chemotherapy-induced neutropenia in breast cancer using pre-chemotherapy data

Article Title: Pre-chemotherapy machine learning models for predicting grade III/IV chemotherapy-induced neutropenia in breast cancer: training on cycle 1 and cross-cycle evaluation

Article References: Lin, M., Weng, Y., Chen, W., Cai, C., Wu, Y., Guo, J., Yu, X., Xi, W., Wu, W., Chen, Y., & Chen, Q. (2026). Pre-chemotherapy machine learning models for predicting grade III/IV chemotherapy-induced neutropenia in breast cancer: training on cycle 1 and cross-cycle evaluation. Supportive Care in Cancer, 34(11), 1173. https://doi.org/10.1007/s00520-026-11402-6

Image Credits: AI Generated

DOI: 10.1007/s00520-026-11402-6

Keywords: breast cancer, chemotherapy-induced neutropenia, machine learning, pre-chemotherapy prediction, G-CSF prophylaxis, docetaxel, epirubicin, neutrophil count, risk stratification, Supportive Care in Cancer, Pre-chemotherapy, machine