The research team employed a twelve-metric matrix on the EBM-NLP word-level PICO classification task, including imbalance-sensitive metrics such as balanced accuracy, macro-averaged F1, and ROC-AUC