An interpretable breast cancer risk stratification model via multi-omics integration: multi-method development and cross-cohort validation.

[BACKGROUND] Breast cancer remains a formidable global health challenge, and the absence of suitable survival guidance models persists, with most existing models suffering from methodological limitations, inadequate comparisons, and lack of external validation.

[METHODS] The data used in this study were obtained from the latest version of The Cancer Genome Atlas (TCGA) database. A total of 1063 samples were randomly divided into a training cohort (n = 745) and a testing cohort (n = 318) in a 7:3 ratio. The training cohort was then randomly divided into a training set (n = 558) and a validation set (n = 187) in a 3:1 ratio. The training set was used for the model to learn data features, while the validation set was used to evaluate and select optimal hyperparameter combinations. The testing cohort was used to assess the performance of all models and determine the final selection. The ultimately established model was subsequently evaluated in an independent validation cohort and compared with outstanding models from prior studies.

[RESULTS] The final conclusion demonstrates that multi-omics models are superior to single-omics models. Among multi-omics models, multi-kernel learning approaches outperform other single-kernel methods. An independent external validation set demonstrates robust generalization performance when confronted with unseen data.

[CONCLUSION] The study developed a new risk stratification model for cancer patients based on a multi-omics integration approach. Our work holds promise for supporting stratification, precision therapy, and prognostic prediction in cancer patients.