Identifying the association rules between adverse events and concomitant medicines in clinical trial data management using random forest

Adverse events (AEs) and concomitant medications (CMs) underreporting remains a recurrent issue in clinical trials. This study aimed to build a mapping relationship between the AE and CM using a random forest (RF) model that can be embedded in the current Electronic Data Capture (EDC) system, to enable a reliable detection of underreporting AE or CM. Current data including 524 AEs and 684 CMs for 13,905 individuals, were taken from three cancer clinical trials and six other studies that are non-public. An additional validation dataset was consisted of 448 independent patients with 269 unique AEs and 407 CMs from a cancer clinical trial on Paclitaxel. We found that the machine learning method exhibited in a learning task between AE and CM well for a common AE like hypertension that was caused by a single cause, but it was ineffective when the AE, such as the increased blood alkaline phosphatase, was caused by complex reasons or just an associated symptom of some diseases. This study suggests the potential of automatically detecting the underreported AE and CM in detail, and it will improve further safety and validity inspections from clinical trials. <b>ABBREVIATIONS:</b> AE, adverse event; ATC, anatomical therapeutic chemical; CM, concomitant medication; EDC, electronic data capture; FDA, Food and Drug Administration; GCP, good clinical practice; MedDRA, medical dictionary for regulatory activities; NMPA, National Medical Products Administration; PT, preferred term; RCT, randomized controlled trial; RF, random forests; SDV, source data verification; VIM, variable importance measure

不良事件（Adverse Event, AE）与合并用药（Concomitant Medication, CM）的漏报始终是临床试验中反复出现的共性问题。本研究旨在借助随机森林（Random Forest, RF）模型构建AE与CM之间的映射关系，该模型可嵌入现有电子数据采集（Electronic Data Capture, EDC）系统，从而实现对漏报AE或CM的可靠检测。

本次研究所用的现有数据来自3项癌症临床试验与6项未公开的其他研究，涉及13905名受试者、524例不良事件与684种合并用药。额外的验证数据集则取自一项针对紫杉醇（Paclitaxel）的癌症临床试验，包含448名独立患者、269种独特不良事件及407种合并用药。

研究发现，针对高血压这类由单一诱因引发的常见不良事件，本研究所用的机器学习方法在AE与CM的关联学习任务中表现良好；但对于血清碱性磷酸酶升高这类由复杂原因引发的不良事件，或是仅作为某些疾病伴随症状的不良事件，该方法则效果不佳。

本研究表明，自动检测漏报AE与CM的详细方案具备应用潜力，将有助于进一步优化临床试验的安全性与有效性核查工作。<b>ABBREVIATIONS:</b> AE：不良事件（Adverse Event）；ATC：解剖治疗化学分类（Anatomical Therapeutic Chemical）；CM：合并用药（Concomitant Medication）；EDC：电子数据采集系统（Electronic Data Capture）；FDA：美国食品药品监督管理局（Food and Drug Administration）；GCP：药物临床试验质量管理规范（Good Clinical Practice）；MedDRA：监管活动医学词典（Medical Dictionary for Regulatory Activities）；NMPA：国家药品监督管理局（National Medical Products Administration）；PT：首选术语（Preferred Term）；RCT：随机对照试验（Randomized Controlled Trial）；RF：随机森林（Random Forest）；SDV：源数据验证（Source Data Verification）；VIM：变量重要性度量（Variable Importance Measure）