PAVE

PAVE embeds observed clinical events with their real-valued measurements and timestamps and applies a self-attention mechanism to predict patient risk and provide interpretable contributions for outcomes such as sepsis onset and mortality without imputing missing values.

Key Features:

• Handling Missing Values: Embeds only observed values directly into vectors rather than imputing missing values, avoiding imputation-induced bias.
• Incorporation of Real Value Medical Events: Integrates real-valued medical events such as lab tests and vital signs alongside Boolean events like diagnosis codes to capture informative signals for acute outcomes.
• Self-Attention Mechanism: Applies a self-attention mechanism to identify and weight significant patterns among medical events, improving predictive accuracy.
• Interpretable Outputs: Exposes self-attention weights to indicate the contributions of event patterns to risk predictions, enabling interpretation for outcomes such as mortality.

Scientific Applications:

• Clinical risk prediction: Represents temporal and value information of medical events to support clinical predictive tasks requiring accurate and interpretable risk assessments.
• Sepsis onset prediction: Leverages lab tests and vital signs alongside event timings to predict sepsis onset.
• Mortality prediction: Identifies event patterns associated with increased mortality risk and supports mortality prediction.

Methodology:

Medical events with associated real-valued measurements and timestamps are embedded into vectors (embedding only observed values), and a self-attention mechanism processes these embeddings to focus on relevant patterns for prediction and interpretation.