Hospital readmission within 30 days of Intensive Care Unit (ICU) discharge remains a significant challenge, linked to increased mortality and healthcare costs. We present BERTBoost, a novel hybrid machine learning framework that integrates structured electronic health record (EHR) variables (e.g., demographics, diagnoses, lab values) with semantic embeddings derived from discharge summaries using ClinicalBERT. The model was trained on the MIMIC-III dataset, with class imbalance addressed using Synthetic Minority Over-sampling Technique (SMOTE), and evaluated on a held-out test set. The final cohort included 7,917 ICU stays, with approximately 30% experiencing a 30-day readmission. BERTBoost achieved 89.7% accuracy, 95.4% precision, 72.6% recall, an F1-score of 0.83, and a ROC–AUC of 0.922, outperforming structured-only base-lines (∼70% accuracy, AUC 0.75). The inclusion of unstructured text improved recall from 50% to 72.6%, representing a 46% relative gain over the structured-only XGBoost baseline, by capturing clinical context often missed by coded data. Key predictors included comorbidity indices and principal components derived from discharge summaries. These findings highlight the advantages of combining large language models with traditional EHR features to improve ICU readmission prediction, enable earlier interventions, and enhance overall healthcare quality.