To anticipate critical events, clinicians intuitively rely on multidimensional time-series data. It is, however, difficult to model such decision process using machine learning (ML), since real-world medical records often have irregular missing and data sparsity in both feature and longitudinal dimensions. Here we propose a nonparametric approach that updates risk score in real time and can accommodate sampling heterogeneity, using forecasting of severe acute graft-versus-host disease (aGVHD) as the study case. The area under the receiver operator characteristic curve (AUC) rose steadily after transplantation and peaked at >0.7 in both adult and pediatric cohorts. Various numerical experiments provided guidelines for future applications.