Low-energy Formulations of Support Vector Machine Kernel Functions for Biomedical Sensor Applications

Although physiologically-indicative signals can be acquired in low-power biomedical sensors, their accurate analysis imposes several challenges. Data-driven techniques, based on supervised machine-learning methods provide powerful capabilities for potentially overcoming these, but the computational energy is typically too severe for low-power devices. We present a formulation for the kernel function of a support-vector machine classifier that can substantially reduce the real-time computations involved. The formulation applies to kernel functions employing polynomial transformations. Using two representative biomedical applications (EEG-based seizure detection and ECG-based arrhythmia detection) employing clinical patient data, we show that the polynomial transformation yields accuracy performance comparable to the most powerful available transformation (i.e., the radial-basis function), and the proposed formulation reduces the energy by over 2500x in the arrhythmia detector and 9.3-198x in the seizure detector (depending on the patient).