Dynamic Survival Analysis for EHR Data with Personalized Parametric Distributions.

Area of Science:

• Machine Learning
• Biostatistics
• Health Informatics

Background:

• High-dimensional electronic healthcare record (EHR) datasets enable clinical insights and risk predictions.
• Machine learning is increasingly applied to dynamic survival analysis for updated time-to-event risk predictions from EHR data.
• EHR data poses challenges in data representation, modeling, interpretability, and evaluation for dynamic survival analysis.

Purpose of the Study:

• Propose a new approach to dynamic survival analysis addressing EHR data challenges.
• Develop a modeling approach for population characteristics and individual risk prediction.
• Introduce a new dynamic C-Index for clinically meaningful evaluation.

Main Methods:

• Learning a global parametric distribution for population characteristics.
• Dynamically locating individuals on the time-axis based on their histories.
• Validating the approach on real-world datasets for COVID-19, cardiovascular disease (CVD), and primary biliary cirrhosis (PBC).

Main Results:

• The proposed modeling approach demonstrates competitive performance against established methods.
• The approach offers potential advantages in individual-level prediction interpretability.
• Successful dynamic risk prediction was achieved on diverse clinical datasets.

Conclusions:

• The novel dynamic survival analysis method effectively utilizes EHR data.
• The approach provides competitive and interpretable risk predictions.
• This work advances the application of machine learning in clinical risk prediction.