Real-Time Detection of Skin-Electrode Adhesion Based on Embedded Neural Networks for Bioimpedance Spectroscopy

A module based on embedded Multi-Layer Perceptrons (MLPs) was developed for real-time monitoring of skin-electrode adhesion quality. It was designed to integrate with Insulin-Meter, an established 4-wire bioimpedance spectroscopy system for measuring insulin absorption in diabetic patients, reported in previous studies. The MLPs address two classification tasks in cascade: (i) adhesion vs. partial detachment and (ii) identification of the partially detached electrode. The MLPs can be deployed on the same microcontroller used for insulin absorption assessment, leveraging the same impedance spectroscopy data. In literature, adhesion monitoring based on impedance measurement has been implemented in applications with unfavorable signal-to-noise ratio (SNR), such as electroencephalography (EEG), where contact quality is typically verified prior to signal acquisition using threshold-based approach. For other biosignal measurements, the higher signal-to-noise ratio and shorter acquisition durations have generally made real-time monitoring of electrode-skin adhesion unnecessary. However, Insulin-Meter requires extended acquisition sessions under unfavorable SNR conditions. MLPs were compared to other machine learning algorithms, including Logistic Regression, Support Vector Machines and Random Forest. Hyperparameter optimization was performed with consideration for the memory footprint of all classifiers. The MLPs outperformed the other algorithms and were deployed on a low-cost, general-purpose microcontroller, requiring significantly less than 50 % of its flash memory. The system achieved an accuracy of 98 % $\pm ~3$ % for discriminating between adhesion and partial detachment, and 97 % $\pm ~13$ % for identifying the partially detached electrode. The microcontroller requires an average inference time of 4.286 ms to implement the two-step classification task.