Development of a multimodal AI framework for ICU outcome mortality prediction using clinical notes and laboratory data

Abstract

Objectives: This study presents a multimodal artificial intelligence framework aimed at improving mortality prediction in intensive care units (ICUs) by integrating structured clinical data with unstructured clinical notes. The framework combines traditional machine learning algorithms with transformer-based language models to capture both numerical patterns and nuanced textual information recorded during patient care. For structured data analysis, machine learning models including Random Forest (RF), Support Vector Machine (SVM), Logistic Regression (LR), Decision Tree (DT), and Naive Bayes (NB) were employed using laboratory measurements, vital signs, and demographic features. 

Methodology: To process unstructured textual data, the study fine-tuned BioBERT and ClinicalBERT models, which are specifically designed to interpret medical language and are pre-trained on large-scale clinical corpora. These models transform clinical narratives into meaningful contextual representations. In parallel, structured variables, including laboratory results and vital signs, were processed independently to generate complementary feature representations. 

Results: The findings demonstrate that integrating clinical text with structured data significantly enhances predictive performance compared to using either data source independently. The combination of transformer-based language models with machine learning techniques and self-attention mechanisms contributes to more robust and reliable mortality prediction. 

Conclusion: The proposed framework provides a practical and scalable foundation for clinical decision support systems and shows strong potential for improving risk assessment and patient management in intensive care settings.