Work

Feature tokenizer Transformer based multi task learning for integrating clinical pathological and genomic features in breast cancer prognosis

Year: 2026
Type: article
Abstract: Predicting breast cancer outcomes remains challenging because of the heterogeneity of clinical and genomic information and the complex interactions among these variables. This study presents a multi-task breast cancer prediction framework based on the Feature Tokenizer Transformer (FT-Transformer) for jointly modelling heterogeneous patient-level information. Feature-specific tokens represent clinical characteristics and genomic alterations, while multi-head self-attention captures nonlinear interactions among the resulting feature representations. A shared Transformer encoder is coupled with task-specific prediction heads for PAM50 subtype classification, ER and HER2 status prediction, relapse prediction, and survival-risk estimation. The model is optimized using a composite objective comprising cross-entropy, binary cross-entropy, and Cox proportional hazards losses. On the held-out METABRIC test set, the ER and HER2 prediction heads achieve accuracies of 93.56% and 89.62%, respectively, while relapse prediction achieves an AUROC of 0.9124 and an AUPRC of 0.8934. Survival-risk estimation achieves a concordance index (C-index) of 0.7963. These task-specific results demonstrate the capability of the proposed framework to learn heterogeneous clinical–genomic relationships across classification, relapse, and time-to-event prediction tasks. Furthermore, attention-based feature attribution provides interpretability by identifying clinical and genomic variables that contribute to individual model predictions. The proposed framework provides a promising computational approach for integrated breast cancer risk assessment and outcome prediction.
Language: English
Subfield: Cancer Research
Open Access status: gold
Estimated APC paid: $1,520
OpenAlex ID: W7235970447
FWCI:
Cites: 32
Cited by:
Related to: 10