Automating risk of bias assessment for clinical trials

摘要: In medicine, the publication of clinical trials now far out-paces clinicians' ability to read them. Systematic reviews, which aim summarize entirety available evidence on a specific question, have therefore become linchpin evidence-based decision making. A key task in systematic reviews is determining whether results included studies may be affected by biases, e.g., poor randomization or blinding. This called risk bias assessment and standard practice. Standardized tools are used perform these assessments; notable example being Cochrane tool, covers seven different types potential biases involves researchers extracting sentences from articles support their assessments. These assessments crucial interpreting published evidence, but due exponential growth biomedical literature base, manually assessing has grown burdensome for researchers. Aiming mitigate this workload, we explore automating assessment. We demonstrate that distantly supervise text mining models, obviating need annotated trial reports. Specifically, leverage data Database Reviews (a large repository reviews), link reports structured same found CDSR produce pseudo-annotated labeled corpus. then develop joint model which, using (the PDF of) report as input, predicts risks each aforementioned areas while simultaneously fragments supporting