An interpretable risk prediction model for healthcare with pattern attention.

Area of Science:

• Clinical informatics
• Machine learning in healthcare
• Predictive modeling

Background:

• Massive data enables clinical predictive tasks, with deep learning showing promise.
• Existing methods struggle with missing data imputation bias, limited input features (Boolean only), and lack of pattern contribution analysis.
• Real-value medical events are crucial for acute disease and mortality prediction but often ignored.

Purpose of the Study:

• To propose a novel interpretable Pattern Attention model with Value Embedding (PAVE) for clinical risk prediction.
• To address limitations of existing methods by avoiding imputation bias and incorporating real-value medical events.
• To enhance model interpretability by identifying contributions of medical event patterns.

Main Methods:

• PAVE utilizes embeddings of medical events, their values, and timestamps as input.
• A self-attention mechanism is employed to identify meaningful patterns among medical events for risk prediction.
• The model avoids imputation by embedding only observed values, thus preventing imputation bias.

Main Results:

• PAVE demonstrated superior performance in sepsis onset and mortality prediction on MIMIC-III and proprietary EHR datasets.
• The model successfully identified meaningful medical event patterns associated with mortality through self-attention weight analysis.
• Experimental results confirm PAVE's effectiveness in outperforming existing predictive models.

Conclusions:

• PAVE enhances clinical risk prediction by learning effective medical event representations incorporating values and timestamps.
• The self-attention mechanism in PAVE captures patient health states and provides interpretable contributions of medical event patterns.
• PAVE advances interpretable clinical risk prediction by offering insights into the drivers of patient risk.