Benefit-Hematotoxicity Stratification in Patients with Triple-Negative Breast Cancer Receiving Platinum-Based Neoadjuvant Therapy: A Multitask Deep Learning Study.

Objectives: Platinum-based neoadjuvant therapy can improve pathological response in triple-negative breast cancer (TNBC), but severe hematotoxicity may compromise treatment delivery. This study developed and validated a multitask learning framework to jointly predict pathological complete response (pCR) and severe hematotoxicity and to support benefit-hematotoxicity stratification. Methods: This multicenter retrospective study included 2060 consecutive patients with TNBC receiving platinum-based neoadjuvant therapy at three institutions. Patients were assigned to a training cohort (n = 1406), an internal validation cohort (n = 351), or an external validation cohort (n = 303). A multitask TabNet (MT-TabNet) model was developed using pretreatment clinical, pathological, imaging, laboratory, electrocardiographic, and planned treatment exposure variables to jointly estimate pCR and severe hematotoxicity. Model performance was evaluated using discrimination, calibration, decision curve analysis, and interpretability analyses. The predicted probabilities were further integrated into a utility-based framework for benefit-hematotoxicity stratification. Results: Overall, 639 patients (31.0%) achieved pCR, and 651 (31.6%) developed severe hematotoxicity. MT-TabNet achieved AUCs of 0.874, 0.842, and 0.819 for pCR prediction and 0.859, 0.846, and 0.818 for severe hematotoxicity prediction in the training, internal validation, and external validation cohorts, respectively. The model showed generally acceptable calibration and clinical net benefit. pCR prediction was predominantly associated with tumor-related characteristics, whereas severe hematotoxicity prediction was more strongly associated with planned treatment exposure and host-related laboratory indicators. Utility-based stratification showed progressively higher pCR rates and lower severe hematotoxicity rates from the low- to high-benefit groups across all three cohorts. Survival differed significantly among benefit groups for both progression-free survival (PFS) and overall survival (OS) in the training and internal validation cohorts; in the external validation cohort, OS differed significantly, whereas PFS showed a similar but nonsignificant trend. Conclusions: MT-TabNet enabled joint estimation of pCR and severe hematotoxicity in patients with TNBC receiving platinum-based neoadjuvant therapy. The utility-based framework integrated efficacy and toxicity predictions into clinically interpretable benefit-hematotoxicity stratification and warrants further prospective evaluation for individualized risk assessment and toxicity monitoring.