The application of unsupervised deep learning in predictive models using electronic health records

Area of Science:

• Machine Learning
• Health Informatics
• Data Science

Background:

• Patient-level electronic health record (EHR) data presents a rich resource for predictive modeling.
• Unsupervised deep learning algorithms, such as autoencoders, can extract meaningful features from complex EHR data.
• The generalizability of autoencoder-derived features across various predictive tasks is an area of active research.

Purpose of the Study:

• To explore the utility of autoencoder-generated features from EHR data in diverse predictive modeling tasks.
• To evaluate the performance of autoencoder features against traditional machine learning models.
• To assess the effectiveness of combining autoencoder features with response-specific variables for enhanced prediction.

Main Methods:

• Utilized an autoencoder, an unsupervised deep learning algorithm, to generate lower-dimensional features from EHR data.
• Compared the predictive performance of autoencoder features against logistic regression with LASSO and Random Forest.
• Developed and evaluated predictive models using response-specific variables (Simple Reg) and a combination of these with autoencoder features (Enhanced Reg).
• Validated models on both simulated EHR data and real-world EHR data from multiple hospitals.

Main Results:

• Autoencoder features demonstrated competitive precision on simulated data, outperforming Random Forest and nearing LASSO.
• On real EHR data for predicting 30-day readmission rates, autoencoder features showed comparable performance to LASSO.
• The Enhanced Reg model, combining autoencoder features with response-specific variables, achieved competitive prediction performance and often required fewer features.
• Autoencoder features proved applicable to a wide array of predictive tasks, showcasing their general utility.

Conclusions:

• Autoencoders can generate effective, generalizable features from the entirety of EHR data for various predictive applications.
• Integrating autoencoder features with key response-specific predictors leads to efficient and robust predictive models.
• This approach reduces the labor involved in data extraction and model training for predictive analytics in healthcare.