Machine Learning Methods in Systematic Reviews
Identifying Quality Improvement Intervention Evaluations
Published in: Machine Learning Methods in Systematic Reviews: Identifying Quality Improvement Intervention Evaluations / Susanne Hempel et al. Methods Research Report (Prepared by the Southern California Evidence-based Practice Center under Contract No. 290-2007-10062-I). AHRQ Publication No. 12-EHC125-EF. (Rockville, MD : Agency for Healthcare Research and Quality, Sep. 2012), 55 p
Posted on RAND.org on September 01, 2012
BACKGROUND: Electronic searches typically yield far more citations than are relevant, and reviewers spend a substantial amount of time screening titles and abstracts to identify potential studies eligible for inclusion in a review. This is of particular relevance in complex research fields such as quality improvement. We tested a semiautomated literature screening process applied to the title and abstract screening stage of systematic reviews. A machine learning approach may allow literature reviewers to screen only a fraction of a search output and to use a predictive model to learn and then emulate the reviewers' decisions. Once learned, the model can apply the selection process to an essentially unlimited number of citations. METHOD: Two independent literature reviewers screened 1,591 quasi-randomly selected citations in a training dataset used to predict decisions on the remaining citations in a MEDLINE search output of 9,395 citations. We explored different prediction algorithms and tested results against reference samples screened by experts in quality improvement. Qualitative (relevance cutoff determined in ROC curve) and quantitative predictions (probability rank order of citations) were determined. RESULTS: The agreement between independent literature reviewers ranged from κ= 0.55 to 0.57. Across two reference samples, the predictive performance of the machine learning approach demonstrated 90.1 percent sensitivity, 43.9 percent specificity, and 32.1 percent PPV. This translates to a reduction of 36.1 percent in citation screening if applied. The predictive performance was affected by reviewer disagreements: a subgroup analysis restricted to citations both reviewers agreed on showed a sensitivity of 98.8 percent (specificity 43.9 percent). CONCLUSION: Machine learning approaches may assist in the title and abstract inclusion screening process in systematic reviews of complex, steadily expanding research fields such as quality improvement. Increased reviewer agreement appeared to be associated with improved predictive performance.