Interpretable machine learning prediction of all-cause mortality.

Area of Science:

• Epidemiology
• Computational Biology
• Biostatistics

Background:

• Traditional linear models for all-cause mortality analysis have limitations in capturing complex, non-linear relationships.
• Complex machine learning models offer enhanced capabilities for identifying novel risk factors and improving prediction accuracy.
• Explainable Artificial Intelligence (XAI) is crucial for understanding the insights derived from these advanced models.

Purpose of the Study:

• To introduce the IMPACT framework, utilizing XAI to explain a tree ensemble model for mortality prediction.
• To analyze all-cause mortality using the IMPACT framework on a large-scale dataset (NHANES).
• To identify previously unrecognized risk factors and interactions influencing mortality over various follow-up periods.

Main Methods:

• Development and application of the IMPACT framework, an XAI technique.
• Utilizing a state-of-the-art tree ensemble model for mortality prediction.
• Analysis of the NHANES dataset (47,261 samples, 151 features) for 1-, 3-, 5-, and 10-year mortality.

Main Results:

• IMPACT models demonstrated superior accuracy compared to linear models and neural networks.
• Identification of overlooked risk factors and significant interaction effects on mortality.
• Development of accurate, efficient, and interpretable mortality risk scores, validated temporally and externally (UK Biobank).

Conclusions:

• IMPACT provides explainable predictions, revealing complex non-linear relationships in mortality data while maintaining high accuracy.
• Explainable risk scores enhance health self-awareness for individuals and aid clinicians in high-risk patient identification.
• The IMPACT framework represents a significant advancement in applying XAI to epidemiological research.