Abstract
Transthyretin amyloidosis (ATTR) is a rare, progressive, and fre- quently under-recognized systemic disease whose early clinical manifes- tations may overlap with common cardiac and neurological conditions. This work presents a proof-of-concept multimodal deep-learning frame- work for the binary classification of ATTR and non-ATTR cases by jointly analysing structured clinical narratives and electrocardiogram (ECG) im- ages. A curated dataset of 100 cases was assembled, including 60 literature- derived cases and 40 synthetic cases generated through a multimodal Gen- erative Adversarial Network (GAN) to mitigate data scarcity and class imbalance. The proposed architecture combines a frozen ResNet-50 vi- sual encoder with a frozen Italian BERT textual encoder; the resulting modality-specific embeddings are projected into a shared latent space and fused through a multilayer perceptron classifier. In a held-out test subset, the final multimodal configuration achieved an overall accuracy of 73%, an AUC-ROC of 0.78, and an ATTR recall of 0.89, indicating promising sensitivity for a preliminary screening scenario. These results should be interpreted as evidence of methodological feasibility rather than clinical readiness, given the limited cohort size, the inclusion of synthetic data, and the absence of external multicentre validation. Overall, the study sup- ports the feasibility of multimodal AI for rare-disease triage and provides a foundation for larger, fully validated clinical investigations.