Prediction of Electrotactile Stimulus Threshold in Real Time Using Voltage Waveforms Between Electrodes

Electrotactile stimulation provides tactile feedback by directly stimulating tactile nerves with electrical currents. However, variations in individual skin impedance hinder the practical application of this technology, and developing an algorithm to accurately estimate the sensation threshold in real-time remains challenging. In this study, we explored four methods to predict the electrotactile sensation threshold across all five fingers. First, we measured the correlation between skin impedance (resistance <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">R</i> and capacitance <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">C</i>) and the sensation threshold. Second, we applied Random Forest regression using <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">R</i> and <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">C</i> related data as inputs. Third, we employed an LSTM (Long Short-Term Memory) neural network with an attention mechanism, trained on raw voltage waveforms during charging and discharging phases at the detectable sensation threshold (LSTM-1). Fourth, similar to the third method, we trained the model on waveforms recorded at both detectable and undetectable sensation thresholds (LSTM-2). The results indicate that the LSTM-2 model, enhanced with attention, outperformed the other methods, achieving an <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">R</i><sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> of 0.910 and a mean squared error (MSE) of 0.661 mA. By utilizing voltage data from both charging and discharging phases, the LSTM-2 model captured the dynamic properties of skin impedance more effectively than methods relying solely on resistance and capacitance measurements. Furthermore, the LSTM-2 model provided stable tactile feedback over a 20-minute continuous stimulation period, maintaining consistent perception intensity and adapting to changes in skin impedance. Our results demonstrate that LSTM-based models can enhance the accuracy of current threshold estimation and provide stability during long-term stimulation, reducing the need for manual recalibration. The results also suggest that voltage waveforms during the discharging phase and at undetectable thresholds play a critical role in improving stimulation accuracy and should not be overlooked.