Feature Engineering for Accurate Low-Power Electrical Impedance Tomography in Hand Sign Recognition

Achieving high classification accuracy from raw measurements in Electrical Impedance Tomography (EIT) for hand sign recognition is a significant challenge due to the inherent ill-posed nature of the corresponding inverse problem to solve. Energy consumption is an issue, especially for wearable devices, so there is a growing preference to measure EIT with fewer electrodes or fewer current injections. In a previous study, we demonstrated that an 8-electrode hand gesture recognition system could achieve high accuracy with only 5 current injections instead of 8, resulting in significant power savings. In this study, we investigate feature extraction and selection techniques to improve the classification accuracy in the low-power EIT system based on less measurements. We evaluate the performance of 12 statistical features by analyzing the Feature Importance (FI) and identify 8 features with non-zero values: WL, MAV, Kurtosis, VORDER, DASDV, RMS, P2P and VAR. Using K-Nearest Neighbour (KNN), a comparative classification is performed using subsets of the best ranked features ranging from 2 to 8. Compared to the 96% accuracy achieved with raw data, our solution achieves 99.8% accuracy with the 8 extracted features. The results show that feature engineering significantly improves classification accuracy compared to the use of raw data.