A novel approach for Motor Unit Number Estimation: integrating artificial intelligence and high-density electromyography

Determining the size and number of motor units within a specific muscle volume is essential for understanding both the function and structure of a muscle. This information is particularly relevant in the study of neuromuscular disorders, as it enables accurate diagnosis and monitoring of disease progression. Traditional Motor Unit Number Estimation (MUNE) techniques are of significant clinical interest as they provide a means to quantify the number of active motor units. Most MUNE techniques share the common approach of estimating motor units based on the ratio of the average electrical activity of a single motor unit to the maximal activity within the detection volume of the electrodes. A critical aspect of this process is determining the minimum stimulation required to detect the electrical activity of a single motor unit. However, these techniques must be applied cautiously as they face several challenges that can affect the accuracy of their estimates. One major issue is the phenomenon of “alternation”, where successive stimulations with similar current intensity lead to variations in the motor unit responses. This occurs when two or more motor units have nearly identical excitability thresholds, resulting in inconsistent recruitment behavior in response to the same stimulus. To avoid this, incremental stimulation must be fine-tuned to recruit only one motor unit at a time. While conceptually straightforward, this is difficult to achieve in practice. In this work, we propose a novel method for identifying the number of recruited motor units by utilizing high-density electromyographic (EMG) signals in combination with a neural network. The proposed approach uses a neural network to minimize quantification errors caused by alternation and the underestimation that occurs when multiple motor units are recruited in the same stimulation step. High-density EMG signals were simulated by modeling the gradual recruitment of motor units through incremental electrical stimulation of the nerve. These simulations were conducted on a muscle volume of the medial gastrocnemius, allowing the generation of EMG signals representative of varying levels of muscle activation, from the recruitment of the first motor unit to the activation of all motor units. A convolutional neural network (CNN) was developed specifically to identify motor units from EMG signals, designed to recognize the spatiotemporal characteristics of the signals presented as images of electrical activity, captured using an electrode grid. The dataset used for training and validation consisted of images representing different levels of muscle activation from 30 simulated subjects, under two adipose tissue conditions (4 mm and 8 mm) and three different signal-to-noise ratios (SNRs) (5, 10, and 15 dB). Finally, the proposed method was compared to the traditional incremental MUNE technique across various stimulation steps, to evaluate its robustness in scenarios with non-optimal incremental stimulations.