Incremental value of pretreatment MRI features for predicting Miller–Payne grade 5 after neoadjuvant therapy in breast cancer: model development and exploratory temporal evaluation

Abstract Background To quantify the incremental value of pretreatment magnetic resonance imaging (MRI) measurements beyond clinicopathological variables for predicting Miller–Payne grade 5 (MP5) after neoadjuvant therapy (NAT), and to describe performance in a same-centre temporal cohort. Methods Patients receiving NAT were divided chronologically into development (n = 371)\(\:n=371\) and same-centre temporal (n = 93)\(\:n=93\) cohorts. Core clinicopathological, MRI-integrated, and Expanded predictor sets were evaluated with six algorithms using nested five-fold cross-validation repeated five times; hyperparameters were tuned within each outer training fold. After the unified development comparison, the MRI-integrated eXtreme Gradient Boosting (XGBoost) model was selected post hoc as a candidate for further study. Results Development MRI-integrated receiver operating characteristic areas under the curve (ROC-AUCs) ranged from 0.8447 to 0.8644 across the six algorithms. XGBoost had the highest five-repeat mean (0.8644), only 0.0002 above Random Forest (0.8642). Its Core-to-MRI ΔAUC was + 0.0124 (Bonferroni simultaneous bootstrap 95% CI for the six Core-to-MRI comparisons, − 0.0036 to + 0.0289; separately, Holm-adjusted P = 0.221); all six adjusted intervals crossed zero. In the temporal cohort (43/93 MP5), XGBoost achieved ROC-AUC 0.8507 (95% CI, 0.7618–0.9222), compared with 0.8535 (0.7665–0.9239) for Random Forest. The post hoc paired AUC difference was − 0.0028 (XGBoost minus Random Forest; DeLong 95% CI, − 0.0194 to 0.0139; two-sided P = 0.743). XGBoost had a Brier score of 0.1632 (95% CI, 0.1314–0.1994), joint-fit calibration intercept of 0.093 (95% CI, − 0.427 to 0.634), and slope of 1.373 (95% CI, 0.888–2.395). Conclusions Clinicopathological variables provided good MP5 discrimination. Routine MRI measurements yielded small, uncertain incremental gains. XGBoost was selected after the unified development comparison. Its temporal discrimination and calibration estimates are exploratory results from already available data; independent validation of discrimination, calibration, and clinical utility is needed before clinical use.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-17
DOI
https://doi.org/10.21203/rs.3.rs-11266553/v1
Primary Topic
MRI in cancer diagnosis
Type
preprint

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Incremental value of pretreatment MRI features for predicting Miller–Payne grade 5 after neoadjuvant therapy in breast cancer: model development and exploratory temporal evaluation

Hu Meng, Renyi Zhu, Yongqiang Xing, Xiang Zhuang et al.
2 links
Research Square
MRI in cancer diagnosis
preprint

Incremental value of pretreatment MRI features for predicting Miller–Payne grade 5 after neoadjuvant therapy in breast cancer: model development and exploratory temporal evaluation

Hu Meng, Renyi Zhu, Yongqiang Xing, Xiang Zhuang, Xinying Zhang, Qing Yang, Zhihui Zhang, Hui Zhang, Jia Hao, Kaiqi Tan, Chenwei Hou
preprint en

Abstract

Abstract Background To quantify the incremental value of pretreatment magnetic resonance imaging (MRI) measurements beyond clinicopathological variables for predicting Miller–Payne grade 5 (MP5) after neoadjuvant therapy (NAT), and to describe performance in a same-centre temporal cohort. Methods Patients receiving NAT were divided chronologically into development (n = 371)\(\:n=371\) and same-centre temporal (n = 93)\(\:n=93\) cohorts. Core clinicopathological, MRI-integrated, and Expanded predictor sets were evaluated with six algorithms using nested five-fold cross-validation repeated five times; hyperparameters were tuned within each outer training fold. After the unified development comparison, the MRI-integrated eXtreme Gradient Boosting (XGBoost) model was selected post hoc as a candidate for further study. Results Development MRI-integrated receiver operating characteristic areas under the curve (ROC-AUCs) ranged from 0.8447 to 0.8644 across the six algorithms. XGBoost had the highest five-repeat mean (0.8644), only 0.0002 above Random Forest (0.8642). Its Core-to-MRI ΔAUC was + 0.0124 (Bonferroni simultaneous bootstrap 95% CI for the six Core-to-MRI comparisons, − 0.0036 to + 0.0289; separately, Holm-adjusted P = 0.221); all six adjusted intervals crossed zero. In the temporal cohort (43/93 MP5), XGBoost achieved ROC-AUC 0.8507 (95% CI, 0.7618–0.9222), compared with 0.8535 (0.7665–0.9239) for Random Forest. The post hoc paired AUC difference was − 0.0028 (XGBoost minus Random Forest; DeLong 95% CI, − 0.0194 to 0.0139; two-sided P = 0.743). XGBoost had a Brier score of 0.1632 (95% CI, 0.1314–0.1994), joint-fit calibration intercept of 0.093 (95% CI, − 0.427 to 0.634), and slope of 1.373 (95% CI, 0.888–2.395). Conclusions Clinicopathological variables provided good MP5 discrimination. Routine MRI measurements yielded small, uncertain incremental gains. XGBoost was selected after the unified development comparison. Its temporal discrimination and calibration estimates are exploratory results from already available data; independent validation of discrimination, calibration, and clinical utility is needed before clinical use.

Research Square
Inner Mongolia University of Science and Technology (CN), Chengdu University of Traditional Chinese Medicine (CN)
National Natural Science Foundation of China, Natural Science Foundation of Inner Mongolia
Good health and well-being
MRI in cancer diagnosis
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon