DeepMPM: a mortality risk prediction model using longitudinal EHR data.

Area of Science:

• Artificial Intelligence in Healthcare
• Clinical Informatics
• Machine Learning for Medical Prediction

Background:

• Conventional mortality risk prediction methods struggle with longitudinal electronic health records (EHRs), failing to capture complex variable interactions and temporal dependencies.
• Existing deep learning models often overlook the interplay between multiple diseases and conditions when predicting patient outcomes.

Purpose of the Study:

• To develop an advanced deep learning model for accurate mortality risk prediction in intensive care unit (ICU) patients.
• To effectively utilize longitudinal EHR data, including disease and treatment information, for enhanced risk modeling.

Main Methods:

• Developed DeepMPM, a deep learning model incorporating a two-level attention mechanism (visit-level and variable-level).
• Leveraged patient EHRs encompassing multiple diseases and diverse conditions to train the model.
• Utilized the MIMIC III database for model evaluation and validation.

Main Results:

• DeepMPM achieved a high Area Under the ROC Curve (AUC) of 0.85 on the MIMIC III database.
• Demonstrated superior performance in mortality risk prediction compared to other deep learning approaches.
• Successfully modeled complex interactions between diseases and treatments for improved representation learning.

Conclusions:

• DeepMPM offers a powerful tool for accurate mortality risk prediction in ICU patients by integrating disease and treatment data.
• The model's ability to capture inter-disease relationships enhances the learning of patient risk status.
• Case studies indicate DeepMPM's potential for providing insights into feature correlations and model interpretability.