A comparative analysis of semi-supervised learning: The case of article selection for medical systematic reviews

Abstract

While systematic reviews are positioned as an essential element of modern evidence-based medical practice, the creation of these reviews is resource intensive. To mitigate this problem, there have been some attempts to leverage supervised machine learning to automate the article triage procedure. This approach has been proved to be helpful for updating existing systematic reviews. However, this technique holds very little promise for creating new reviews because training data is rarely available when it comes to systematic creation. In this research we assess and compare the applicability of semi-supervised learning to overcome this labeling bottleneck and support the creation of systematic reviews. The results indicated that semi-supervised learning could significantly reduce the human effort and is a viable technique for automating medical systematic review creation with a small-sized training dataset.