Lightweight Multimodal Biometric Authentication Via Adaptive Fingerprint–Iris Feature Fusion for Secure Patient Identification in Healthcare Systems

Abstract

Patient misidentification remains a critical source of medical errors in hospitals, contributing to adverse events such as incorrect medication administration and incompatible blood transfusions. Although biometric authentication offers a promising solution, unimodal systems are susceptible to spoofing, sensor noise, and environmental variability, undermining their reliability in high stakes clinical environments. This study proposes a multimodal biometric identification system that fuses fingerprint and iris traits using deep feature extraction. A pretrained EfficientNetB0 backbone independently derives 1,280-dimensional embeddings from each modality, which are concatenated into a 2,560-dimensional fusion vector and classified by a regularized neural network with dropout and L2 penalties. The system was trained on a 45-subject Kaggle dataset with stratified splitting (70/15/15) and augmentation (rotation, translation, zoom, flip). On the held-out test set, the model achieved 98.52% accuracy, a False Acceptance Rate of 1.48%, a False Rejection Rate of 1.48% (Equal Error Rate = 1.48%), a micro-averaged ROC-AUC of 0.999, and a Precision–Recall AUC of 1.00. Grad-CAM visualizations confirmed attention to discriminative regions (ridge minutiae in fingerprints, crypt/collarette texture in irises), enhancing interpretability. A Flask-based web prototype demonstrated real-time deployment feasibility. These findings indicate that feature-level fusion of pretrained CNNs offers an accurate, explainable, and lightweight solution for biometric patient identification, though larger, diverse datasets and anti-spoofing measures are needed prior to clinical adoption.