HD-MUNet: integrating artificial intelligence and high-density electromyography for motor unit number estimation

Objective: In this study we developed and evaluated the performance of HD-MUNet: a novel method for motor unit number estimation (MUNE), integrating high-density surface electromyograms (HD-sEMG) with artificial intelligence (AI). Methods: We designed a dual-branch neural network to process raw M-waves elicited by electrical stimulation and their incremental changes through convolutional and recurrent layers, producing stepwise estimates of recruited motor units (MUs). Because no experimental gold standard exists for counting MUs, a physiologically realistic model of the medial gastrocnemius was used to simulate HD-sEMG. Simulations included variations in MUs population (20-50-100-150-200-250-300), subcutaneous tissue thickness (4-8 mm), signal-to-noise ratio (0-8-16 dB), and stimulation protocol (50-100-200 steps), yielding 5040 subjects. HD-MUNet performance was benchmarked against two state-of the-art methods: Incremental MUNE and StairFit. Results: HD MUNet achieved consistently low median relative errors (≤ 7%) across conditions with high test-retest reliability (Pearson r > 0.995), whereas StairFit and Incremental MUNE showed strong sensitivity to noise and stimulation scenarios, with median errors ranging from −2% to −195% and −20% to −66%, respectively. Of practical relevance, differently from the other two methods, HD MUNet performed equally well for the three stimulation steps considered. We confirmed this result on a set of experimental data. Conclusion: By combining HD-sEMG with AI, HD-MUNet achieved robust and reliable MUNE across a broad range of simulated conditions, independently of stimulation protocol. Significance: This method has immediate potential to simplify MUNE protocols in both healthy and pathological conditions, with direct applicability to clinical neuromuscular disease diagnosis and monitoring.