Explainable Amyloidosis Detection
In collaboration with the TeCIP Institute of Scuola Superiore Sant'Anna and the Cardiology Hospital of Pisa, we developed an explainable classifier for cardiac amyloidosis from 6-lead ECGs. Amyloidosis is a rare but life-threatening condition caused by abnormal protein deposits in the heart, and early detection from ECG signals remains a challenging open problem with very few existing automated approaches.
This work uses a private clinical dataset (Castiglione et al., 2025) and was carried out in strict collaboration with a medical team to ensure clinical validity of both the predictions and the model's explanations.
Architecture
The core model is a mixed ResNet–Spatial Transformer network with skip connections, designed to capture both local morphological features and global rhythm patterns across the 6-lead ECG. This architecture outperformed the few existing works in the field on our clinical dataset.
To further boost performance, we combined it with a pure CNN branch into an ensemble. A final SVM classifier, trained on the probability outputs of the two models (rebalanced using ADASYN), performs the definitive classification decision. Key metrics are F2 score and AUC, chosen to penalise false negatives in this clinical context.
Explainability
A core contribution of this work is its interpretability component. We applied two complementary techniques to understand which parts of the ECG the model attends to:
- Grad-CAM — highlights the most discriminative time-steps in the convolutional feature maps
- Attention Rollout — propagates attention weights across the Transformer layers to produce a global attention map over the input signal
Both methods are validated qualitatively with the clinical team, who assessed whether the model’s focus regions align with known electrophysiological markers of amyloidosis.
