KEY POINTS
- The study retrospectively analyzed 2,067 breast cancer patients from the prospective multicenter REQUITE cohort, comparing 12 machine-learning algorithms using clinical data, SNP data or both. Eight late toxicity endpoints were evaluated at 24 months using five-times repeated 10-fold cross-validation.
- The eight endpoints were arm lymphedema, breast edema, nipple retraction, breast atrophy, hyperpigmentation, telangiectasia and induration inside or outside the tumor bed. Breast atrophy was most common (47.3%), followed by tumor-bed induration (39.5%); arm lymphedema occurred in only 3.7%. Most toxicity was grade 1, and outcomes were dichotomized as grade 0 versus grade ≥1.
- For arm lymphedema, clinical variables alone produced an AUC around 0.84 with Random Forest. SNP-only logistic regression reached approximately 0.87, while combined clinical-genomic models exceeded 0.90; in the patient-aligned statistical comparison, combined modeling achieved AUC 0.93 versus 0.78 with clinical data alone and 0.89 with SNPs alone.
- Because lymphedema prevalence was low, precision-recall performance was also examined. Logistic regression improved from AUC-PR 0.22 with clinical variables to 0.47 with SNPs and 0.52 with combined data, compared with a prevalence-based chance level of approximately 0.04.
- Genomic information was not equally useful for every endpoint. Breast-edema discrimination increased from AUC 0.67 to 0.82 with combined data, with a large effect size (Cohen’s d 1.78); nipple-retraction prediction also improved strongly (d 2.03). By contrast, telangiectasia showed essentially no improvement (AUC 0.620 clinical vs 0.616 combined).
- Frequently selected clinical predictors included age, smoking, mean heart dose, number of examined nodes and tumor size. Genetic pathway analysis linked lymphedema-associated SNPs to regulation of cadherin-19/type-II classical cadherins (FDR 1.3×10⁻³) and nipple retraction to cornified-envelope and keratinization pathways (FDR 1.9×10⁻³).
- Important limitations temper the high AUCs: the SNPs were derived from prior GWAS work within REQUITE, genomic analyses were restricted to European ancestry, detailed dose-volume data were not consistently available, outcomes collapsed mild and severe toxicity together, and no independent external validation cohort was used. The authors explicitly position the work as a modeling benchmark rather than a deployable clinical tool.
CLINICAL TAKEAWAY
Genomic information appears to add substantial predictive value for specific breast RT toxicities—most convincingly arm lymphedema—rather than improving toxicity prediction uniformly. The cohort is large and multicenter, but independent validation and a demonstration that prediction changes prevention or treatment decisions are required before genotyping becomes part of routine radiotherapy planning.
SOURCE
International Journal of Radiation Oncology, Biology, Physics