Variance-Aware Quantized Multi-Scale Conformer for Efficient ECG Rhythm Classification on Resource-Constrained Devices

Electrocardiogram (ECG) rhythm classification is vital for the early detection and continuous monitoring of cardiac disorders. To address this, we propose a Conformer-based architecture that effectively models both local waveform morphology and long-range temporal dependencies in ECG signals. However, deploying deep learning models on resource-constrained devices is challenging due to their high computational and memory demands. To address this, we introduce an adaptive layer-wise quantization framework driven by layer importance estimation. The sensitivity of each layer is quantified using a combination of parameter density, weight variance, and kurtosis, capturing both scale and higher-order distributional characteristics of network weights. This formulation enables fine-grained bit-width allocation across layers, applying conservative quantization to diagnostically critical modules while allowing aggressive quantization for redundant ones. Extensive experiments on the Chapman dataset demonstrate that the proposed method achieves up to 9.52× compression, consistently outperforming fixed-precision baselines. A comprehensive ablation study further validates the individual and joint contributions of each sensitivity component to overall performance.