Improving clinical outcome predictions using convolution over medical entities with multimodal learning

Area of Science:

• Biomedical Informatics
• Artificial Intelligence in Healthcare
• Clinical Data Science

Background:

• Accurate prediction of patient mortality and length of stay (LOS) is crucial for clinical decision-making and resource management.
• Electronic Health Records (EHR) offer valuable data, but challenges exist in utilizing unstructured clinical notes due to their sparse and high-dimensional nature.
• Existing models often fail to fully leverage the rich information contained within clinical narratives.

Purpose of the Study:

• To develop a deep multimodal architecture for improved prediction of patient mortality and LOS.
• To investigate the impact of incorporating extracted medical entities from clinical notes as supplementary features.
• To compare the effectiveness of different embedding techniques (Word2vec, FastText) for medical entities.

Main Methods:

• A novel convolution-based multimodal architecture was proposed, integrating time-series Intensive Care Unit (ICU) signals with medical entities extracted from clinical notes.
• Medical entities were identified and embedded using techniques like Word2vec and FastText.
• The model was evaluated on its ability to predict patient mortality and LOS, comparing its performance against baseline and other multimodal approaches.

Main Results:

• The proposed deep multimodal method demonstrated superior performance compared to all baseline models.
• Mortality prediction accuracy improved by approximately 3% in terms of Area Under the Receiver Operating Characteristics (AUROC) and Area Under Precision-Recall Curve (AUPRC).
• Length of stay (LOS) prediction showed an improvement of around 2.5% over the time-series baseline model.

Conclusions:

• Integrating medical entities extracted from clinical notes significantly enhances the predictive power of patient outcome models.
• The developed deep multimodal architecture offers a robust approach for leveraging diverse data sources within EHRs.
• This method provides a promising advancement for early risk stratification and resource optimization in critical care settings.