Robust and interpretable unit level causal inference in neural networks for pediatric myopia

Area of Science:

• Artificial Intelligence
• Clinical Medicine
• Causal Inference

Background:

• Deep learning models in medicine often function as 'black boxes', hindering clinical trust and adoption.
• Interpretability and reliability are paramount for integrating AI into healthcare settings.

Purpose of the Study:

• To develop and validate a causal inference framework integrated into neural networks for enhanced AI transparency in medicine.
• To assess the direct and indirect causal effects of individual features on predictions using a pediatric ophthalmology cohort.

Main Methods:

• Integration of a causal inference framework with neural networks.
• Application to a prospective pediatric ophthalmology cohort (>3000 children) with longitudinal data.
• Estimation of direct/indirect causal effects and validation through refutation experiments.

Main Results:

• The proposed method achieved good performance in predicting myopia progression.
• Identified clinically plausible causal pathways influencing myopia progression in children.
• Refutation experiments confirmed the robustness and reliability of the estimated causal effects.

Conclusions:

• The model-agnostic, causal inference framework enhances transparency and reliability of deep learning in healthcare.
• This approach supports explainable AI for digital health interventions and advances precision medicine.
• Facilitates equitable healthcare by providing interpretable AI insights.