gms | German Medical Science

Introduction: Microneurography, a unique tool within pain and itch research, is an electrophysiological method to monitor peripheral nerve fiber activity [1]. Given that recordings typically contain multiple active fibers, the analysis of neuronal discharge patterns requires spike sorting. It is a challenging task for microneurography due to low signal-to-noise ratios and the tendency of fibers to exhibit similar spike shapes that may evolve over time. In previous research, we introduced an initial spike sorting methodology for microneurography, utilizing manual feature extraction and support vector machine (SVM) classification [2]. Here, we evaluate the model-based network VPNet [3], which combines advanced computational methods with relatively low demands for training data size.

Methods: We evaluate the spike sorting performance of VPNet using 6 microneurography recordings. Within these recordings, we categorize the spikes into 2, 3, and 6 classes, respectively, depending on the number of fibers. Spike labeling was conducted using the marking method [4], which ensures that action potentials elicited by low-frequency stimulation are on “tracks”, providing ground truth data. For each fiber cluster, we have computed a template by averaging all tracked spikes. The template allows us to visualize inter-fiber differences. VPNet, based on variable projections (VP), presents a novel neural network architecture. We partitioned the data into 80% for training and 20% for testing, implementing 5-fold cross-validation to assess generalizability. The analysis involved averaging accuracies across each fold.

Results: Datasets A1-A3 include two active fibers, representing two distinct classes. The average accuracy of 5-folds for A1 and A2 stands at 0.91, with A3 trailing closely at 0.90. In the case of A4, which involves three classes, we observed an average accuracy of 0.67. Similarly, A5 (two classes) yields a score of 0.68, while A6, with five fibers, exhibits a score of 0.50.

Discussion: The VPN-net model demonstrated outcomes comparable to our optimized per-recording results computed with multiple feature extraction pipelines. The essential advantage of the VPNet approach is that it can be used universally, without the need to choose an optimal feature extraction method for each data set. VPNet offers distinct advantages over alternative neural network architectures due to its reduced variable count, resulting in a more compact structure and parameter interpretability.

Conclusion: In conclusion, we see the potential of the VPNet approach as a universal method for microneurography spike sorting. Since each recording has a relatively low number of spikes, we are further working on solutions that allow integrating multiple datasets via, for example, transfer learning for better performance.

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.