Personalized event prediction for Electronic Health Records.

Area of Science:

• Medical Informatics
• Computational Biology
• Health Data Science

Background:

• Clinical event sequences are crucial for patient care but exhibit significant patient-specific variability.
• Population-wide predictive models often fail to capture individual patient dynamics due to diverse clinical conditions.

Purpose of the Study:

• To develop and evaluate new predictive models for clinical event sequences that account for patient-specific variability.
• To improve the accuracy of patient condition interpretation, adverse event prediction, and overall patient care.

Main Methods:

• Proposed and investigated multiple novel event sequence prediction models.
• Developed methods for refining population models to subpopulations and enabling self-adaptation.
• Implemented meta-level model switching for adaptive prediction selection.
• Tested models on clinical event sequences from the MIMIC-III database.

Main Results:

• The developed models demonstrated improved ability to adjust predictions for individual patients.
• Methods for subpopulation refinement and adaptive model switching showed promise in handling patient-specific dynamics.
• Performance analysis on MIMIC-III data validated the effectiveness of the proposed approaches.

Conclusions:

• Addressing patient-specific variability is essential for accurate clinical event sequence prediction.
• Adaptive and refined modeling strategies offer a pathway to more personalized and effective patient care.
• These advancements have the potential to significantly enhance clinical decision support systems.