Predicting item response theory estimates and response correctness with cyclic dual latent discovery in the Korean Medical Licensing Examination: a methodological study

PURPOSE: We evaluated whether student and item representations learned from response correctness provide information about examinee and item characteristics, including item response theory (IRT) estimates, and whether this information translates into better response prediction. METHODS: We analyzed Korean Medical Licensing Examination (KMLE) responses from 3,259 examinees to 360 items. We froze 64-dimensional representations learned by cyclic dual latent discovery (CDLD) and assessed IRT prediction using nested cross-validation. Exploratory analyses compared the addition of CDLD, 64-dimensional representations based on singular value decomposition (SVD64), or random inputs to augmented response summaries. Across 5 splits, response prediction was compared with Rasch and 2-parameter logistic (2PL) models; 0-40 responses from new examinees were used for adaptation. RESULTS: In exploratory comparisons, adding CDLD to augmented response summaries increased R2 from 0.9289 to 0.9843 for 2PL ability and from 0.8759 to 0.9216 for log discrimination. The increment for log discrimination was 0.0458 (95% confidence interval, 0.0163 to 0.0842), exceeding the increment with SVD64 by 0.0267 (0.0091 to 0.0457). Prediction of Rasch difficulty decreased negligibly, whereas improvement in 2PL difficulty was uncertain. Among existing KMLE examinees, CDLD and 2PL achieved areas under the curve of 0.8446 and 0.8451, respectively, for response prediction. With 40 adaptation responses from new examinees, the difference in score correlation was -0.0179, and noninferiority was not established. CONCLUSION: Exploratory comparisons showed that CDLD provided additional information about KMLE 2PL ability and log discrimination, but did not support its adoption as a superior predictor of responses or student scores.