Predicting Four-Class Residual Cancer Burden (RCB) Before Neoadjuvant Chemotherapy in Breast Cancer: Block-Structured Model Comparison, Selection Criterion and Decision Threshold
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background and Objective
Residual cancer burden (RCB) after neoadjuvant chemotherapy (NAC) stratifies prognosis more finely than binary pathologic complete response, yet prediction models routinely collapse this four-tier scale. We predicted all four classes from pre-treatment data and examined how the selection criterion and decision threshold determine the outcome in an imbalanced multiclass problem.
Methods
In a single-centre retrospective cohort of 328 patients treated with NAC followed by surgery, 64 candidate variables were organised into six thematic blocks; four were excluded, two for missingness and two because they are not known before NAC, leaving 60 features. A 17-variable radiology model (Model R) was compared with a 60-variable multimodal model (Model ALL) across 100 stratified splits. Within each split, selection was repeated by five-fold inner cross-validation on the training data using the mean of macro AUC and macro-F1; the held-out test set was used once, after selection. Reporting follows TRIPOD+AI.
Results
RCB classes were distributed as RCB-0 27.4% (n=90), RCB-I 10.4% (n=34), RCB-II 43.6% (n=143) and RCB-III 18.6% (n=61). Across 100 splits the median test AUC was 0.823 (IQR 0.803–0.838) for Model ALL and 0.773 (0.746–0.803) for Model R, Model ALL being superior in 88% of splits ( d z =1.18), in 84% for macro-F1 and in 98% for ordinal agreement (median quadratic weighted kappa 0.504 versus 0.260; d z =1.69). For the most frequently selected configuration, class-wise precision ranged from 0.300 to 1.000, sensitivity from 0.250 to 0.722 and specificity from 0.703 to 1.000. On one fixed split Model R fell at the 98th percentile of its own distribution and Model ALL at the 31st, reversing the ranking. The default 0.50 threshold drove sensitivity to zero at the intensification boundary (RCB-III versus the rest) in 48 of 100 splits for Model ALL and 51 for Model R; a threshold selected on training data raised median sensitivity to 0.833 and lowered specificity to 0.537.
Conclusions
The multimodal model is consistently superior in discrimination and markedly so in ordinal agreement. The selection criterion and the decision threshold are separate design decisions, and evaluation resting on a single split can reverse the ranking of models.
