A Predictive Model for Medical Events Based on Contextual Embedding of Temporal Sequences

Area of Science:

• Computational medicine
• Artificial intelligence in healthcare
• Clinical informatics

Background:

• Medical concepts present ambiguity and errors due to human fallibility, limiting classical machine learning applications in early disease prediction.
• Existing methods struggle to fully leverage the semantic richness of medical data for predictive tasks.

Purpose of the Study:

• To develop a novel machine-friendly representation for medical concepts that captures their semantics.
• To create a sequential predictive model for medical events using this new representation.

Main Methods:

• Developed contextual embedding techniques to integrate diverse medical events (diagnoses, prescriptions, lab tests).
• Converted each medical event into a numerical vector representing its semantics for similarity measurement.
• Built predictive models using these vectors to forecast new diagnoses.

Main Results:

• Evaluated the sequential prediction model on the MIMIC-III dataset for disease risk estimation.
• Achieved an Area Under the Receiver Operating Characteristic curve (AUC) of 0.79 for chronic systolic heart failure.
• Obtained an average AUC of 0.67 across 80 common diagnoses.

Conclusions:

• Proposed a general early prognosis predictor for 80 diagnoses.
• The method generates numerical representations of medical events to reveal their underlying meaning.
• Demonstrated the method's efficiency in providing more accurate diagnoses for patients and physicians.