Cardiac amyloidosis remains underdiagnosed, partly because its electrocardiographic signatures are not reliably detected during routine clinical interpretation. While deep learning methods have enabled automated detection from ECG signals, most high-performing models depend on large labeled datasets and long signal durations, limiting their use in data-constrained environments with short-duration recordings.
We evaluated pre-trained transformer models and convolutional architectures trained from scratch on a balanced dataset of 461patients with 2.5-second twelve lead ECG recordings, using a 5-fold cross-validation strategy. Pre-trained transformer representations from HuBERT-ECG and ECG-JEPA were assessed under a linear probing setup without fine-tuning, while two CNNs (small and medium) and a hybrid CNN-LSTM were trained from scratch using identical validation splits and consistent preprocessing conditions.
The CNN-LSTM achieved the highest performance, yielding a cross-validated AUC of 0.832 (± 0.019). ECG-JEPA reached a comparable AUC of 0.827 (± 0.060), whereas HuBERT-ECG underperformed with a maximum AUC of 0.735 (± 0.060).
These results showed that, in settings with limited data and short signal durations, a hybrid CNN-LSTM architecture that jointly captures morphological and temporal structure retains a measurable advantage over frozen foundation models. However, the competitive performance of ECG-JEPA suggests that fine-tuning or the use of longer recordings may reduce this gap, supporting further exploration of self-supervised transformer models for cardiac amyloidosis detection under limited data conditions.