A Framework for Locally Imputing and Predicting Biomarker Trajectories Under Irregular Monitoring: Application to Chronic Myeloid Leukemia

Abstract Irregular monitoring and missing data limit the utility of longitudinal biomarkers in real-world practice. We developed a generalizable framework that combines interval-aligned preprocessing, localized multiple imputation, and machine-learning forecasting to generate complete trajectories and predict future biomarker values under routine clinical conditions. Using BCR::ABL1 monitoring in chronic myeloid leukemia as a case study, we aligned measurements to 90-day intervals, applied a windowed, uncertainty-propagating imputation strategy, and trained recurrent neural network (RNN) and XGBoost models to forecast values three and six months ahead. Full Information models achieved RMSEs of 1.22–1.24 for 3-month predictions—well below the biomarker’s observed variability—and maintained accuracy even when the most recent visit was intentionally omitted, simulating extended follow-up. This framework preserves local temporal structure, supports individualized monitoring decisions, and is directly adaptable to other continuous biomarkers measured under irregular real-world schedules.