Explainable Spike Detection Algorithm on C-fiber Microneurography Data using Weighted Hermite Variable Projection Neural Networks.

Microneurography data provides valuable insights into peripheral nerve responses critical for understanding sensory functions. This paper presents a unique approach by utilizing a neural network solution to detect spike events in microneurography data, employing the weighted Hermite Variable Projection Neural Network model to capture distinct neurophysiological patterns. Addressing the inherent challenges of low signal-to-noise ratios and extreme class imbalance, the proposed pipeline integrates a knowledge-augmented model, data balancing techniques and explainable AI methods to ensure both accurate and interpretable predictions. A sliding window technique is applied to the one-dimensional time series signal, where window-wise binary classifications are transformed back into the time series, resulting in continuous signal detection. The model's decision-making process is analyzed post-hoc, uncovering the neural patterns and feature extraction mechanisms it has discovered. This detection-focused approach is the first neural-network-based solution for spike detection in microneurography data. Source code is available at https://github.com/DominikaDarabos/explainable_MNG_spike_detection.git.Clinical relevance- Microneurography is essential for studying peripheral neural activity, but detecting spikes elicited by additional stimuli remains a significant challenge. Inaccurate detection can impact data interpretation, affecting insights into pain and itch mechanisms. The proposed spike detector introduces a fully automated solution, enhancing analysis reliability and potentially improving treatment and therapy.