DeepIPA: Deep Learning Augmented Needle EMG for Quantifying Nerve Injury

Introduction: Accurately assessing nerve health is essential for informed decision-making in complex nerve surgeries and predicting the outcomes of nerve transfers. Traditional semi-objective and quantitative electromyography (EMG) techniques often lack the precision required for clinical utility. While motor unit number estimation (MUNE) offers a more precise measure of muscle innervation, its applicability is limited to a few surface muscles that are rarely of clinical significance. To overcome these challenges, we propose a deep learning approach capable of rapidly and reliably predicting the number of motor units from needle EMG signals at high-force levels. This method enables the quantification of innervation in any muscle accessible with standard EMG techniques and without an additional time burden.

Methods: EMG signals from muscles with varying degrees of denervation were simulated, incorporating interference and artifacts. Signals were then preprocessed via frequency filtering and segmentation. Processed signals were fed into models including Long Short-Term Memory (LSTM) network to capture temporal dynamics and Audio Spectrogram Transformer (AST) fine-tuned for raw time-series EMG analysis.

Results: The LSTM model demonstrated strong performance, achieving low error (RMSE=2.11, MAE=1.53), and high correlation to true values (R²=0.94). The AST model yielded RMSE=3.60, MAE=2.75, and R²=0.81. Both models performed particularly well in severe denervation cases(MU≤5): 83.31% of LSTM predictions and 94.12% of AST predictions fell within ±1 of the true value.

Conclusion : DeepIPA has the potential to reliably and accurately assess nerve health. Both LSTM and AST architectures demonstrated the ability to quantify muscle innervation, highlighting their potential to enhance preoperative planning and postoperative monitoring in peripheral nerve surgery. Validation using clinical EMG data from patients with peripheral nerve injuries is necessary to confirm its clinical utility. However, early findings suggest significant promise for improving patient outcomes and refining treatment strategies in reconstructive neurosurgery.