CATNet

CATNet predicts medical events from heterogeneous, irregular electronic health record (EHR) histories using a cross-event attention-based, time-aware neural network.

Key Features:

• Unified heterogeneous and temporal modeling: Integrates varied medical event types (medications, diagnosis codes, laboratory tests, procedures, outcomes) with temporal information and addresses irregular timing at both local and global levels.
• Cross-event attention mechanism: Exploits correlations among different medical event types via cross-event attention to capture relationships between heterogeneous historical events and target events.
• Time-awareness and task adaptivity: Incorporates time-aware components sensitive to temporal dynamics and can be adapted to multiple medical event prediction (MEP) tasks without major architectural changes.
• Attention-based neural architecture: Employs an attention-driven neural network architecture for sequence modeling of EHR data.

Scientific Applications:

• Medical event prediction: Forecasts medications, diagnosis codes, laboratory tests, procedures, and outcomes from patient EHR histories.
• EHR temporal modeling: Models heterogeneous and irregular temporal patterns in EHR data for predictive analytics.
• Benchmarking on MIMIC-III and eICU: Evaluated on the MIMIC-III and eICU public datasets and reported to outperform state-of-the-art methods across multiple MEP tasks.

Methodology:

Implements a neural network architecture that uses cross-event attention and time-aware components to model irregular temporal characteristics at local and global levels and to provide task adaptivity.