This paper introduces an innovative method for precisely predicting haemoglobin levels non-invasively using video data of nail pallor through pre-trained models. Accurate haemoglobin level prediction is essential for diagnosing and managing a variety of medical conditions. Non-invasive methods offer significant advantages over traditional invasive procedures. Leveraging the power of pre-trained models, we extract high-level features from image data, capturing rich representations that encode relevant information for haemoglobin level estimation. To improve prediction accuracy and interpretability, we employ feature selection techniques including Mutual Information, Recursive Feature Elimination (RFE), Principal Component Analysis (PCA), and Uniform Manifold Approximation and Projection (UMAP). Mutual Information helps to identify informative features by quantifying the dependency between variables, while RFE iteratively selects features based on their contribution to model performance. PCA reduces the dimensionality of the feature space by capturing the most significant variations in the data and UMAP, a nonlinear dimensionality reduction technique, further refines the feature space, preserving both local and global structure. To perform the final prediction, we utilize Graph Neural Networks (GNNs) as a regression model. By integrating these techniques, we achieve more accurate and interpretable predictions of haemoglobin levels, enabling early detection and better management of medical conditions. Experimental results demonstrate the effectiveness of the proposed approach, showcasing its potential for non-invasive haemoglobin level estimation from video data.