A Bio BERT-Based Multimodal AI for Comprehensive ICU Outcome Prediction

Abstract

Efficient management of the Intensive Care Unit (ICU) is crucial to improving survival rates, enhancing hospital resource allocation, and improving clinical decision-making. The current project describes a sophisticated predictive model that utilizes Large Language Models (LLMs) (specifically fine-tuned BioBERT) to continuously process both structured data (e.g., vitals and labs) and unstructured data (e.g., clinical notes and radiology reports). The goal is to predict outcomes in the ICU, specifically mortality risk, length of stay, sepsis, acute kidney injury (AKI), cardiac arrest, pneumonia, and acute respiratory distress syndrome (ARDS). Using real-world ICU patient medical records from the MIMIC-IV public dataset, the proposed approach to construct the predictive model uses extensive data preprocessing, innovative feature fusion, and strong model training, validation, and performance evaluations. Unlike traditional machine learning models such as logistic regression and random forests which heavily rely on structured datasets, the proposed model enhances prediction performance by leveraging the ability to extract complex patterns in unstructured data. The preliminary evidence by accuracy, loss, mean squared error, and various Area Under the Receiver Operating Characteristic (AUC-ROC) metrics shows that can produce a better model for predicting restoration in ICU patients. Future research will look at scaling the model for wider applicability, deploying the model in real hospital settings including challenges of explainability, continuous learning and ethical challenges, and comparing the model against other LLMs in the domain.

Ähnliche Arbeiten