Learning and diSentangling patient static information from time-series Electronic hEalth Records (STEER)

Area of Science:

• Health Informatics
• Machine Learning in Healthcare
• Data Privacy

Background:

• Machine learning in healthcare raises patient privacy and algorithmic fairness concerns.
• Self-reported race can be predicted from medical data lacking explicit racial information, but the extent is unknown.
• Developing models minimally affected by sensitive attributes is challenging.

Purpose of the Study:

• To systematically investigate the predictive power of time-series electronic health record (EHR) data for patient static information.
• To assess the ability of raw and learned representations from ML models to encode sensitive patient attributes.
• To develop a privacy-preserving approach for machine learning models using EHR data.

Main Methods:

• Systematic investigation of time-series EHR data for predicting static patient information.
• Training machine learning models on raw and learned representations to predict biological sex, age, and self-reported race.
• Development of a variational autoencoder (VAE) approach for disentangling sensitive attributes.

Main Results:

• Machine learning models achieved high predictive performance for biological sex (AUC 0.851), binarized age (AUC 0.869), and self-reported race (AUC 0.810) from EHR data.
• High predictive performance was consistent across different tasks, cohorts, model architectures, and databases.
• The VAE-based approach successfully learned a latent space to disentangle sensitive patient attributes from time-series data.

Conclusions:

• Time-series EHR data and learned representations contain significant patient-sensitive information.
• Existing machine learning models can inadvertently encode and predict sensitive attributes.
• A novel VAE-based method offers a generalizable approach to protect patient privacy in healthcare AI.