Heart failure (HF) is a crucial health concern globally that requires prompt diagnosis in order to be effectively treated clinically. In order to overcome the difficulties presented by unstructured data formats as free-text notes and scanned documents, this study suggests a machine learning (ML)-based method to predict HF using unstructured electronic health records (EHRs). We use a hybrid ensemble model that combines logistic regression and support vector machines (SVM) after integrating OCR and NLP for feature extraction. Monte Carlo dropout and probability calibration were used to create a confidence measure that ensures accurate and comprehensible forecasts, hence improving therapeutic trust. The findings show that by bridging the gap between unstructured EHRs and ML-based predictive analytics, the model has better applicability in actual clinical situations.