A Meta-Learning and Physics-Informed Neural Network Framework for Early Detection of Pregnancy-Associated Anaemia

Abstract

Physio-Nutri Meta-PINNs is introduced as a novel framework for the early detection of pregnancy-associated anaemia, integrating meta-learning with physics-informed neural networks (PINNs) to address challenges arising from sparse and noisy clinical data. The framework combines a Transformer-based meta-encoder for robust feature extraction, a PINN for modeling physiological–nutritional dynamics governed by domain-specific laws, and a cross-domain adapter to align pre-trained representations with real-world clinical measurements. The meta-encoder extracts latent features from heterogeneous early-trimester data while incorporating medical priors, ensuring physiologically plausible representations even under limited sample availability. The PINN enforces biophysical constraints through partial differential equations, effectively bridging data-driven inference with established physiological mechanisms. In addition, the cross-domain adapter mitigates distributional shifts between meta-features and clinical observations, thereby improving generalization. Conventional feature extraction layers are replaced by meta-encoder outputs, enabling seamless integration with existing nutritional models while supporting probabilistic prediction and uncertainty quantification. Implemented using state-of-the-art components, including a 12-layer Transformer and a conditional GAN-based adapter, the proposed framework demonstrates strong capability in handling sparse, noisy, and multi-modal data. This approach advances early anaemia detection by unifying meta-learning, physics-informed modeling, and transfer learning into a clinically interpretable and robust solution for maternal health monitoring.