Improving Consensus Scoring of Crowdsourced Data Using the Rasch Model: Development and Refinement of a Diagnostic Instrument

Background: Diabetic retinopathy (DR) is a leading cause of vision loss in working age individuals worldwide. While screening is effective and cost effective, it remains underutilized, and novel methods are needed to increase detection of DR. This clinical validation study compared diagnostic gradings of retinal fundus photographs provided by volunteers on the Amazon Mechanical Turk (AMT) crowdsourcing marketplace with expert-provided gold-standard grading and explored whether determination of the consensus of crowdsourced classifications could be improved beyond a simple majority vote (MV) using regression methods. Objective: The aim of our study was to determine whether regression methods could be used to improve the consensus grading of data collected by crowdsourcing. Methods: A total of 1200 retinal images of individuals with diabetes mellitus from the Messidor public dataset were posted to AMT. Eligible crowdsourcing workers had at least 500 previously approved tasks with an approval rating of 99% across their prior submitted work. A total of 10 workers were recruited to classify each image as normal or abnormal. If half or more workers judged the image to be abnormal, the MV consensus grade was recorded as abnormal. Rasch analysis was then used to calculate worker ability scores in a random 50% training set, which were then used as weights in a regression model in the remaining 50% test set to determine if a more accurate consensus could be devised. Outcomes of interest were the percent correctly classified images, sensitivity, specificity, and area under the receiver operating characteristic (AUROC) for the consensus grade as compared with the expert grading provided with the dataset. Results: Using MV grading, the consensus was correct in 75.5% of images (906/1200), with 75.5% sensitivity, 75.5% specificity, and an AUROC of 0.75 (95% CI 0.73-0.78). A logistic regression model using Rasch-weighted individual scores generated an AUROC of 0.91 (95% CI 0.88-0.93) compared with 0.89 (95% CI 0.86-92) for a model using unweighted scores (chi-square P value