Multi-task Learning
Across six reported tasks, the multi-task LoRA LLM achieved average accuracy 0.981, macro F1 0.976, and AUROC 0.996. The largest multi-task benefit appeared for HER2, where the joint model outperformed the single-task adapter under sample…
1 sources - 6 claims
Across six reported tasks, the multi-task LoRA LLM achieved average accuracy 0.981, macro F1 0.976, and AUROC 0.996. The largest multi-task benefit appeared for HER2, where the joint model outperformed the single-task adapter under sample imbalance. The multi-task architecture reduced deployment from six models to one and cut trainable parameters relative to single-task LoRA adapters. Multi-task learning improved aggregate accuracy and macro F1 compared with independent single-task LoRA adapters. The article interprets multi-task learning as statistically beneficial because related clinical variables share narrative context in pathology reports. The model jointly predicted T stage, N stage, M stage, histologic grade, ER, PR, and HER2.