Abstract
The ability to forecast ICU mortality represents a fundamental clinical duty because it enables health professionals to begin lifesaving interventions early on. The current medical practice predominantly uses structured measurements of heart rate as well as test results while ignoring important diagnostic findings from medical notes. Medical Information Model for Critical Care version 3 includes complete information about structured and unstructured data types. The project uses initial models including Logistic Regression and Random Forest before moving onto deep learning models consisting of LSTM–Transformer and BioClinicalBERT. The prediction system uses MEWS scores and discharge summaries to enhance its forecasting capabilities. The text processing component of BioClinicalBERT produces embeddings from textual input and the prediction certainty determination rests with Monte Carlo Dropout.