Integrating Structured and Unstructured EHR Data for Predicting Mortality by Machine Learning and Latent Dirichlet...

Area of Science:

• Critical Care Medicine
• Medical Informatics
• Machine Learning in Healthcare

Background:

• Intensive care units (ICUs) provide critical care for severely ill patients.
• Accurate mortality prediction in ICUs can enhance patient outcomes and resource allocation.
• Existing models often overlook valuable unstructured clinical data, such as physician notes.

Purpose of the Study:

• To develop an improved mortality risk prediction model for ICU patients.
• To investigate the impact of incorporating unstructured clinical data into prediction models.
• To leverage Latent Dirichlet Allocation (LDA) for analyzing unstructured diagnostic notes.

Main Methods:

• Utilized the MIMIC-III database for patient data.
• Developed a two-part model: one using eight structured variables (vitals, GCS, age), and another incorporating unstructured data from initial physician diagnoses analyzed by LDA.
• Combined structured and unstructured data using machine learning for mortality prediction.

Main Results:

• The combined model significantly improved the accuracy of predicting clinical outcomes.
• Achieved an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.88, indicating strong predictive performance.
• The model successfully identified key predictive variables and demonstrated temporal prediction capabilities.

Conclusions:

• Integrating easily collectible structured variables with unstructured clinical notes analyzed by LDA enhances ICU mortality prediction.
• Initial clinical observations and diagnoses contain vital information for improving patient care decisions.
• This approach offers a valuable tool for ICU medical and nursing staff to aid in clinical decision-making.