Hye Jin Cohng
Sunchon National University, South Korea
Abstract Title:
Machine Learning Prediction Model of Post-Transplant Diabetes Mellitus in Kidney Transplant Recipients in South Korea
Biography:
Dr. Hye Jin Chong is a Associate Professor in Department of Nursing, Sunchon National University, Republic of Korea. Her research interests include kidney transplantation, post-transplant metabolic complications, and clinical prediction modeling using multicenter cohort data and machine-learning methods. She is actively involved in clinical research aimed at improving long-term outcomes in kidney transplant recipients.
Research Interests:
Post-transplant diabetes mellitus (PTDM) is a clinically important complication after kidney transplantation. We aimed to develop a baseline-only machine-learning model for predicting 1-year PTDM. We retrospectively analyzed the KoreaN Cohort Study for Outcome in Patients With Kidney Transplantation. Recipients with pretransplant diabetes, fasting glucose ?126 mg/dL, or glycated hemoglobin ?6.5% at baseline were excluded. PTDM was defined as recorded PTDM or fasting plasma glucose ?126 mg/dL at 1 year. Thirty-eight pretransplant and transplant-time predictors were evaluated using logistic regression, random forest, gradient boosting, and extreme gradient boosting with stratified 5-fold cross-validation. Median/mode imputation was used primarily; K-nearest-neighbor and MICE-like iterative imputation were assessed in sensitivity analyses. Of 1,080 recipients, 646 met eligibility criteria. Among 584 recipients with evaluable 1-year outcomes, 90 developed PTDM (15.4%). With median/mode imputation, random forest performed best, with an area under the receiver operating characteristic curve (AUROC) of 0.663, precision-recall AUC of 0.241, and F1 score of 0.370. Results were similar with alternative imputation methods. Random forest with K-nearest-neighbor imputation yielded an AUROC of 0.663 and precision-recall AUC of 0.266, whereas logistic regression with MICE-like imputation yielded an AUROC of 0.658 and precision-recall AUC of 0.265. A machine-learning model showed moderate performance for predicting 1-year PTDM. More complex imputation methods did not materially improve discrimination. Further refinement and external validation are warranted.
