This study introduces a heart failure (HF) prediction framework built on electronic health records (EHR) from the MIMIC-IV and MIMIC-IV-ED databases. The proposed system integrates structured clinical variables (e.g., demographics, vitals, laboratory results, comorbidities) with unstructured admission notes encoded using PubMedBERT embeddings. After preprocessing with Winsorization, k-nearest neighbor imputation, and Z-score normalization, multiple machine learning (ML) and deep learning (DL) algorithms were trained and compared, including Random Forest (RF), Logistic Regression, Decision Tree, Naïve Bayes, AdaBoost, Dense Neural Network (DNN), Long Short-Term Memory (LSTM), and Convolutional Neural Network (CNN).