Towards Accurate and Reliable ICU Outcome Prediction: A Multimodal Learning Framework Based on Belief Function Theory...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Clinical Data Science

Background:

• Accurate Intensive Care Unit (ICU) outcome prediction is crucial for patient care and resource management.
• Current methods primarily use structured electronic health record (EHR) data, neglecting valuable information in clinical notes.
• A need exists for frameworks that can effectively integrate heterogeneous EHR data, including free-text notes.

Purpose of the Study:

• To develop and evaluate a multimodal framework for accurate and reliable ICU outcome prediction.
• To fuse heterogeneous structured EHR data and free-text clinical notes using belief function theory.
• To address prediction uncertainty and data conflicts inherent in multimodal EHR data.

Main Methods:

• A novel multimodal framework based on belief function theory was developed.
• The framework fuses structured EHR data (demographics, vital signs) with unstructured clinical notes.
• Fusion strategy accounts for intra-modality uncertainty and inter-modality conflicts.

Main Results:

• The proposed framework significantly outperformed existing methods on two large ICU datasets (MIMIC-III and ZICIP).
• Mortality prediction F1 score and AUPRC improved by 6.51% and 3.72% on MIMIC-III.
• Predictive reliability increased, evidenced by an 18.08% decrease in Brier score for mortality prediction.

Conclusions:

• The multimodal framework offers superior accuracy and reliability for ICU outcome prediction.
• Improved prediction supports more precise triage and efficient allocation of critical care resources.
• The framework is a versatile tool for multimodal EHR analysis with broad clinical applications.