Toward Energy-Efficient and Low-Power Arrhythmia Detection for Wearable Devices

As cardiovascular diseases remain a leading cause of global mortality, long-term monitoring is essential for detecting intermittent conditions like arrhythmia. However, current wearable devices are often bulky and uncomfortable, and existing deep learning (DL) models for automated ECG analysis are too computationally intensive for low-power, battery-operated hardware. This study explores whether approximation techniques—specifically data precision reduction and approximate multiplication—can enable efficient on-device arrhythmia detection without sacrificing diagnostic reliability.

The authors utilized the MIT-BIH Arrhythmia Database to train and validate a deep neural network (DNN) optimized for an application-specific integrated circuit (ASIC). The team implemented two primary optimization strategies:

• Precision Reduction: The model was converted from 16-bit to 8-bit fixed-point precision.
• Approximate Arithmetic: The team integrated various approximate multipliers into the hardware architecture, selecting those that balance power efficiency with error resilience.

To ensure the model remained accurate despite these approximations, the authors employed "approximation-aware training," where the model was retrained using the chosen approximate multipliers during the forward pass. This allowed the network to adapt its weights to compensate for the inherent errors introduced by the hardware approximations.

The proposed architecture demonstrates significant improvements in energy efficiency. At a 12 kHz operating frequency, the design consumes only 3.07 μW, representing a 64.9% reduction in power compared to the 8.75 μW reference architecture. At higher frequencies (100 MHz), the design achieves a 61.5% reduction in energy consumption. Despite these aggressive approximations, the system maintains a 93.7% classification accuracy and 92.1% sensitivity, meeting the requirements for effective clinical monitoring while substantially extending battery life.