Unit-Level Causal Analysis in Neural Networks for Childhood Myopia

Overview

This study introduces a causal inference framework integrated into neural networks to elucidate feature-level causal effects on childhood myopia progression. Using a large prospective pediatric cohort, the method demonstrated robust performance and identified clinically plausible causal pathways, enhancing interpretability and reliability in AI-driven myopia prediction.

Background

Childhood myopia is a growing public health concern with complex multifactorial etiology involving genetic and environmental factors. Traditional AI models for myopia prediction often lack transparency, limiting clinical trust and adoption. Causal inference methods can improve interpretability by distinguishing direct and indirect effects of risk factors. Integrating such causal reasoning into neural networks aligns with precision medicine goals and supports equitable healthcare interventions.

Data Highlights

The study analyzed a prospective cohort of over 3000 children with longitudinal follow-up data. The causal inference framework estimated direct and indirect effects of individual features on myopia progression. Multiple falsification strategies confirmed the robustness of causal effect estimates, supporting the reliability of the model's explanations.

Key Findings

• The proposed causal inference framework is model-agnostic and can be integrated with neural networks for unit-level causal effect estimation.
• Applied to a large pediatric ophthalmology cohort, the method identified clinically plausible causal pathways influencing myopia progression.
• Refutation experiments using multiple falsification strategies validated the robustness and reliability of the causal effects derived.
• The approach enhances transparency and interpretability of AI models, addressing a key barrier to clinical adoption in medicine.
• This method supports digital health interventions by providing explainable predictions aligned with precision medicine principles.

Clinical Implications

Incorporating causal inference into neural networks enables clinicians to understand the direct and indirect influences of risk factors on myopia progression, facilitating personalized intervention strategies. This transparency fosters greater trust in AI tools and supports their integration into pediatric ophthalmology practice. The model-agnostic nature allows broad applicability across digital health platforms targeting childhood myopia.

Conclusion

Integrating unit-level causal analysis within neural networks advances explainable AI for childhood myopia, offering robust, clinically meaningful insights that promote precision medicine and equitable care. This framework represents a significant step toward transparent, reliable AI adoption in ophthalmology.

References

1. Liang, J. 2025 -- Trend and projection of myopia in children and adolescents from 1990 to 2050: a comprehensive systematic review and meta-analysis
2. Lee, Y., Keel, S. & Yoon, S. 2024 -- Evaluating the effectiveness and scalability of the World Health Organization myopiaed digital intervention: mixed methods study
3. Yang, B.-Y. 2024 -- Significance of school greenspaces in preventing childhood myopia
4. Morgan, I. G., Ohno-Matsui, K. & Saw, S.-M. 2012 -- Myopia
5. Huang, T. et al. 2023 -- Artificial intelligence for medicine: progress, challenges, and perspectives