The task of accurately classifying complex patterns in data remains a fundamental challenge in domains such as medical diagnostics and image recognition. The paper addresses the problem of fingerprint-based blood group classification, aimed at predicting blood groups from fingerprint images. Challenges include variations in image quality, noise, and the inherent complexity of fingerprint patterns. To overcome these, an ensemble stacking approach combining ResNet50, VGG16, and EfficientNetB0 as base models, with logistic regression is employed as the meta-model. The method achieves a classification accuracy of 93%, outperforming individual base models. To address class imbalance, the Synthetic Minority Oversampling Technique (SMOTE) is applied, resulting in improved performance across minority classes. This research has significant implications for medical diagnostics, offering a non-invasive and efficient method for blood group prediction.