Design and Experimental Validation of an Integrated Biomedical Instrumentation Framework for Non-Invasive ABO/Rh Phenotype Prediction Using Hybrid Deep Learning

Blood group determination is essential for transfusion safety and surgical planning; however, conventional serological methods are invasive and resource-dependent. This study presents an integrated bioengineering framework for non-invasive ABO/Rh phenotype prediction using dermatoglyphic fingerprint analysis and hybrid deep learning. The proposed system combines standardized optical fingerprint acquisition with computational feature extraction and classification. A dataset of 6000 publicly available fingerprint images was used for model development (85/15 training/testing split), followed by independent external validation on 300 real-world samples. Four network topologies were evaluated: ResNet-50, ResNet-101, an unweighted hybrid model, and a class-weighted hybrid model. The weighted hybrid configuration achieved the highest training accuracy (96.65%). During external validation, it demonstrated superior generalization performance with 96.73% accuracy and 10 misclassifications, whereas the unweighted hybrid achieved 94.33% accuracy. Statistical assessment using 95% Wilson confidence intervals and two-proportion z-tests confirmed significant improvements over single-backbone architectures (p < 0.001). These findings support the feasibility of a fingerprint-based predictive framework as a rapid, non-invasive screening and decision-support tool for preliminary assessment in emergency and resource-limited healthcare settings.