A randomised trial of integrated machine-learning for systematic review 'Risk of bias' assessments

Article type
Authors
Arno A1, Wallace B2, McKenzie J3, Marshall I4, Thomas J1, Elliott J3
1Institute of Education, University College London
2College of Computer and Information Science, Northeastern University
3School of Public Health and Preventive Medicine, Monash University
4Department of Primary Care and Public Health Sciences, King's College London
Abstract
Background:
Assessment of study risk of bias (RoB) is a key step in a systematic review but is very time-consuming. RobotReviewer is an open-access platform which partially automates RoB assessment using machine-learning (ML) and natural language processing. Covidence is a cloud-based systematic review production tool, recommended by Cochrane for the production of Cochrane Reviews.

Objectives:
The purpose of this trial is to evaluate the use of ML in RoB assessments in systematic reviews. Specifically, we sought to:
1) evaluate the accuracy of RoB assessments produced using a combination of human effort and ML compared to those produced by humans alone; and
2) determine whether the person-time required to complete RoB using a combination of human effort and ML is less than those completed by humans alone.

Methods:
A two-arm, parallel-group randomised controlled trial is ongoing. Each included study has initial RoB assessment performed by two independent reviewers (Human 1 and Human 2). RobotReviewer assistance will be provided to one of the two reviewers, with this assistance assigned randomly (1:1) for each study. RoB consensus will be performed by a third reviewer (Human 3) blinded to the allocation of RobotReviewer assistance.

Results:
The trial commenced recruitment in January, 2018 and is ongoing. At current rate of recruitment, final results will be available by mid-2018 and the co-primary outcomes will be presented.

Conclusions:
Following completion of the trial, results will be disseminated through journal publications and conference presentations. Study data will be used to inform next steps in machine-learning design and promotion.

Patient and healthcare consumer involvement:
While patients and consumers are unlikely to interact directly with the results of this study, increased automation holds exciting potential for increased availability of high quality, up-to-date health information. This trial is being conducted under the larger goal of building living evidence processes to provide reliable, up to date evidence for health decision-making.