Predicting phase III clinical trial outcomes with an evidence-driven large language model

Phase III clinical trials are pivotal for regulatory approval and clinical guidelines, yet are costly, time-consuming, and failure prone. Existing clinical trial outcome prediction models often rely on registry-level summaries and text-to-label classification, limiting temporal validity, endpoint fidelity, and mechanistic interpretability. To address this challenge, we develop and evaluate Xuanwu-TrialMind, an evidence- and mechanism-conditioned in silico evaluation framework. We first curated Xuanwu-Corpus, a dataset comprising 500 phase III clinical trials from four leading general medical journals with 15,867 background references. 350 trials published before March 1, 2024, were used for training the model and 150 later trials for temporally external validation. Then, given the full protocol, and a strictly pre-publication evidence pack as inputs, Xuanwu-TrialMind outputs a success-likelihood prediction accompanied by traceable, evidence-linked rationale. On temporal validation, Xuanwu-TrialMind achieved 93.33% accuracy, 96.64% recall, 80.65% specificity, F1=0.96, and Matthews correlation coefficient=0.79. Bootstrap resampling (5,000 iterations) showed stable accuracy (median 93.33%; 95% CI 89.33%-96.67%) and Matthews correlation coefficient (median 0.79; 95% CI 0.66-0.91). Additionally, Xuanwu-TrialMind outperforms established baselines on the TOP benchmark, achieving PR-AUC 0.920, F1 0.818, and ROC-AUC 0.803, with the largest relative gains in recognizing negative trials.