Interpretable instance disease prediction based on causal feature selection and effect analysis.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Causal Inference

Background:

• Machine learning in healthcare often relies on correlation, not causation.
• This limitation can lead to ineffective or harmful interventions.
• Healthcare research requires methods adhering to causality norms.

Purpose of the Study:

• To develop a two-stage prediction method for instance disease prediction.
• To incorporate causal inference into machine learning models for healthcare.
• To provide both qualitative and quantitative causal explanations for predictions.

Main Methods:

• A two-stage prediction method involving instance feature selection and causal effect analysis.
• Feature selection based on counterfactuals using a reinforcement learning framework.
• A model comprising three neural networks trained with the actor-critical method.
• Improving neural network attribution algorithms for quantitative causal effect calculation.

Main Results:

• The proposed method provides qualitative and quantitative causal explanations alongside predictions.
• Experiments on synthetic, open-source, and real medical data validate the approach.
• The method successfully integrates causal reasoning into disease prediction models.

Conclusions:

• Causality enables deeper exploration of essential variable relationships.
• Causal feature selection and effect analysis enhance disease prediction model reliability.
• This approach moves beyond correlation to establish more trustworthy medical AI.