https://jerry-terrasse.github.io/lm_center_homepage/tags/dataset/DatasetHugo Blox Builder (https://hugoblox.com)zhFri, 17 Oct 2025 00:00:00 +0000https://jerry-terrasse.github.io/lm_center_homepage/media/icon_hu_6873430e16214d30.pnghttps://jerry-terrasse.github.io/lm_center_homepage/tags/dataset/https://jerry-terrasse.github.io/lm_center_homepage/post/2025-10-17-trialbench-scientific-data-2025/Fri, 17 Oct 2025 00:00:00 +0000https://jerry-terrasse.github.io/lm_center_homepage/post/2025-10-17-trialbench-scientific-data-2025/<p>临床试验是新药从实验室走向患者的关键桥梁，但其过程充满挑战：平均成功率不足15%，耗时超过十年，成本高达数十亿美元。</p> <p>2025年9月，由香港科技大学（广州）陈晋泰、南京大学符天凡、哈佛、斯坦福及临床试验公司IQVIA等团队联合开发的<strong>TrialBench平台</strong>在Nature子刊Scientific Data正式发表，成为全球首个<strong>面向AI的多模态临床试验预测数据集</strong>。</p> <p> <figure > <div class="flex justify-center "> <div class="w-100" ><img alt="图片" srcset=" /lm_center_homepage/post/2025-10-17-trialbench-scientific-data-2025/paper01_hu_7620d51be21c6fe7.jpg 400w, /lm_center_homepage/post/2025-10-17-trialbench-scientific-data-2025/paper01_hu_4aaa86739dc9d05f.jpg 760w, /lm_center_homepage/post/2025-10-17-trialbench-scientific-data-2025/paper01_hu_7c2d2d6f553bf2e0.jpg 1200w" src="https://jerry-terrasse.github.io/lm_center_homepage/post/2025-10-17-trialbench-scientific-data-2025/paper01_hu_7620d51be21c6fe7.jpg" width="760" height="289" loading="lazy" data-zoomable /></div> </div></figure> </p> <h3 id="平台核心价值">平台核心价值</h3> <p>TrialBench系统整合了23个子数据集，涵盖8大核心预测任务：</p> <ul> <li> <p>预测试验时长</p> </li> <li> <p>预测患者退出率</p> </li> <li> <p>预测严重不良事件</p> </li> <li> <p>预测死亡事件</p> </li> <li> <p>预测试验是否获批</p> </li> <li> <p>识别失败原因</p> </li> <li> <p>自动生成入选标准</p> </li> <li> <p>推荐合理给药剂量</p> </li> </ul> <p> <figure > <div class="flex justify-center "> <div class="w-100" ><img alt="八大临床试验预测问题总结" srcset=" /lm_center_homepage/post/2025-10-17-trialbench-scientific-data-2025/paper02_hu_4f45e0d23d338d45.jpg 400w, /lm_center_homepage/post/2025-10-17-trialbench-scientific-data-2025/paper02_hu_db842f9f42ddfee4.jpg 760w, /lm_center_homepage/post/2025-10-17-trialbench-scientific-data-2025/paper02_hu_d70b7ed3870d1439.jpg 1200w" src="https://jerry-terrasse.github.io/lm_center_homepage/post/2025-10-17-trialbench-scientific-data-2025/paper02_hu_4f45e0d23d338d45.jpg" width="760" height="154" loading="lazy" data-zoomable /></div> </div></figure> </p> <p>八大临床试验预测问题总结</p> <h3 id="技术特色">技术特色</h3> <p> <figure > <div class="flex justify-center "> <div class="w-100" ><img alt="TrialBench平台技术框架" srcset=" /lm_center_homepage/post/2025-10-17-trialbench-scientific-data-2025/paper03_hu_39504e77b3020e21.jpg 400w, /lm_center_homepage/post/2025-10-17-trialbench-scientific-data-2025/paper03_hu_9e863135d9fea542.jpg 760w, /lm_center_homepage/post/2025-10-17-trialbench-scientific-data-2025/paper03_hu_bc88d2be2313038e.jpg 1200w" src="https://jerry-terrasse.github.io/lm_center_homepage/post/2025-10-17-trialbench-scientific-data-2025/paper03_hu_39504e77b3020e21.jpg" width="760" height="590" loading="lazy" data-zoomable /></div> </div></figure> </p> <p>平台集成了多源数据，采用先进AI技术：</p> <ul> <li> <p>图神经网络处理药物分子结构</p> </li> <li> <p>Bio-BERT解析临床文本</p> </li> <li> <p>层级注意力模型理解疾病编码</p> </li> </ul> <p>同时提供完整的基线模型、评估指标和多模态融合方法，支持Python与R语言工具包，实现“开箱即用”。</p> <h3 id="应用成果">应用成果</h3> <p> <figure > <div class="flex justify-center "> <div class="w-100" ><img alt="TrialBench实验结果" srcset=" /lm_center_homepage/post/2025-10-17-trialbench-scientific-data-2025/paper04_hu_930050982b7f4b23.jpg 400w, /lm_center_homepage/post/2025-10-17-trialbench-scientific-data-2025/paper04_hu_b4c2c69324fa7292.jpg 760w, /lm_center_homepage/post/2025-10-17-trialbench-scientific-data-2025/paper04_hu_6fcc0073106c07db.jpg 1200w" src="https://jerry-terrasse.github.io/lm_center_homepage/post/2025-10-17-trialbench-scientific-data-2025/paper04_hu_930050982b7f4b23.jpg" width="760" height="528" loading="lazy" data-zoomable /></div> </div></figure> </p> <p>实验结果显示，在14个二分类任务中，多模态模型在11个任务中F1分数超过0.7，展现出强大的预测能力。目前，<strong>Google DeepMind</strong>已在TxGemma模型中应用TrialBench进行不良事件预测，<strong>AUTOCT</strong>项目也将其作为基准评估平台。</p> <h3 id="开放获取">开放获取</h3> <p>TrialBench已向全球研究者开放，旨在推动AI与医疗研究的深度融合，助力优化临床试验设计、加速新药研发进程。</p> <p>平台链接：https://huyjj.github.io/Trialbench/</p> <p><a href="https://mp.weixin.qq.com/s/W67qpWpovmYYvTSF1hc6tA" target="_blank">查看原文</a></p>