MetaPred

MetaPred applies meta-learning to improve clinical risk prediction from limited longitudinal Electronic Health Records (EHR) data.

Key Features:

• Meta-learning framework: Trains a meta-learner across related risk prediction tasks to learn how to produce effective predictors from limited samples.
• Meta-learner training: Learns parameter initialization and training strategies by leveraging multiple source risk prediction tasks.
• Base predictors: Employs Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) as configurable base models.
• Fine-tuning: Applies the learned meta-learner model directly to target risks and further refines parameters using available target-domain samples.
• Longitudinal EHR handling: Targets temporal, sparse, and irregular characteristics of longitudinal EHR data in the predictive modeling process.
• Empirical validation: Demonstrated performance using real patient EHR data from Oregon Health & Science University.

Scientific Applications:

• In-hospital mortality prediction: Predicts risk of in-hospital mortality from limited longitudinal EHR records.
• Hospital readmission prediction: Predicts risk of hospital readmission using learned models adapted to target cohorts.
• Chronic disease onset prediction: Predicts onset risk of chronic diseases from sparse patient EHR histories.
• Condition exacerbation prediction: Predicts risk of condition exacerbation using temporally structured EHR data.
• General clinical risk prediction with limited samples: Improves predictive accuracy for various clinical risks when target-domain samples are scarce.

Methodology:

Train a meta-learner on a collection of related risk prediction tasks, use CNNs and RNNs as base predictors, and fine-tune the learned model on target-domain samples; validated on Oregon Health & Science University EHR data.