LSTrans targets automated ECG classification on resource-constrained wearable devices. Its architecture combines an interleaved 1D convolutional backbone for multi-scale rhythm and morphology features with a Transformer encoder for long-range temporal dependencies. LoRA is applied to critical layers to reduce trainable parameters, while homogeneous and heterogeneous knowledge distillation transfer diagnostic capability from larger teacher models. The authors report competitive sensitivity-efficiency tradeoffs across multiple benchmark datasets, with substantially lower peak memory usage and downstream adaptation latency. The abstract does not provide the datasets, numerical results, parameter counts, or deployment benchmarks.
No heat snapshots are available in the last 24 hours.