In treating malignant cerebral edema after a large middle cerebral artery stroke, clinicians need quantitative tools for real-time risk assessment. Existing predictive models typically estimate risk at one, early time point, failing to account for dynamic variables. To address this, we developed Hybrid Ensemble Learning Models for Edema Trajectory (HELMET) to predict midline shift severity, an established indicator of malignant edema, over 8-h and 24-h windows. The HELMET models were trained on retrospective data from 623 patients and validated on 63 patients from a different hospital system, achieving mean areas under the receiver operating characteristic curve of 96.6% and 92.5%, respectively. By integrating transformer-based large language models with supervised ensemble learning, HELMET demonstrates the value of combining clinician expertise with multimodal health records in assessing patient risk. Our approach provides a framework for accurate, real-time estimation of dynamic clinical targets using human-curated and algorithm-derived inputs.