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

Understanding causal mechanisms in deep learning is essential for clinical adoption, where interpretability and reliability are critical. Most existing AI systems act as black boxes, limiting transparency in medicine. We propose a causal inference framework integrated into neural networks to assess the influence of individual features on predictions. Using a prospective pediatric ophthalmology cohort of over 3000 children with longitudinal follow-up, our method estimates direct and indirect causal effects through intervention. Applied to myopia progression in children, our approach not only achieved good performance but also identified clinically plausible causal pathways. Refutation experiments with multiple falsification strategies confirm the robustness and reliability of causal effects. The approach is model-agnostic and suitable for digital health interventions requiring explainability. By incorporating unit-level causal reasoning into deep learning, this work advances transparent and reliable AI systems aligned with the goals of precision medicine and equitable healthcare.