زينب محمد كاظم

Signal forecasting in the medical field has many applications, such as signal correction and anomaly detection. According to this application, robust forecasting is required to obtain a signal identical to the original signal. This study proposes a forecasting technique that obtains a robust signal that can be used in different applications. A long short-term memory neural network (LSTM-NN) was used to predict future samples from present and past samples. An Electroencephalography (EEG) dataset was used to test this technique. Four channels were used as input examples, one of which was the predicted output. All four channel samples were fed into the four networks to predict the future samples. To decrease complexity, only one hidden layer is used for this purpose. The statistical results are promising for applications that require an almost perfectly predicted signal. The number of hidden cells is first very low (five cells only), which gives a Root Mean Square Error of less than 20, whereas when the number of hidden cells is increased to 100, the Root Mean Square Error (RMSE) is approximately 7.5 for all four channels.

The identification of human motion intention through electromyography (EMG) signals is an important area of development in human–robot interaction. This technology aids amputees in controlling their prosthetic limbs in a more intuitive manner, facilitating the execution of daily activities. However, hand amputees face challenges in using dexterous prostheses due to control difficulties and low robustness in real-life situations. This study aims to enhance the accuracy of EMG gesture recognition by extracting spatial characteristics via multiple high density (HD) maps. A total of five HD-maps are generated utilizing the root mean square value (RMS), mean absolute value (MAV), zero crossings (ZC), sign slope changes (SSC), and waveform length (WL) features. The influence of each distinct HD-map, along with the synergistic effect of numerous HD-maps in the extraction of intensity features, is assessed with regard to its impact on classification accuracy. Three machine learning classifiers are employed to categorize nine hand movements of the Ninapro (DB5) dataset. The results show that features extracted from the combination of multiple HD-maps (CMHD) achieved a high accuracy in comparison to those of individual HD-maps. Moreover, the proposed features are superior to those of conventional TD features. The error rate is reduced by approximately 7.76% relative to time domain (TD) features. The results obtained confirm the significance of spatial features extracted from multiple HD-maps that ensure consistent information in different EMG channels. © 2025 Published by IRCS-ITB.