gms | German Medical Science

Introduction: Microneurography allows to record in-vivo human electrophysiological activity from nerve-fibers encoding sensory information [1]. Spike sorting is one of the key problems in microneurography due to variable spike shapes, successfully used in other electrophysiological domains [2], [3]. Therefore, for examining the activity of the recorded nerve fibers, a special stimulation protocol with regular background (0.125-0.25 Hz) pulses is used [4], [5]. The action potentials resulting from those pulses can be tracked according to their stable response latency allowing to determine the number of fibers. Temporal changes in their latency to the background stimulation are associated with preceding fiber activity [5], [6]. In the presented project we attempt to back-cast [7] those dependencies with a regression model for quantifying nerve activity which is otherwise inaccessible.

Methods: We used two in-vitro mice nerve recordings (MES-I and MES-II) mimicking microneurography data in order to have an initial spike sorting. We extracted the timing of the electrical stimulation protocol and the resulting action potentials and derived the normalized response latencies to the background pulses. These latencies (1–3 subsequent values) serve as input to regression models with varying polynomial degrees (1 to 4). The best parameters were set using a grid search. The output (back-casted) variable is the fiber activity, quantified by the spike count in the preceding fixed time interval. Goodness-of-fit was determined by R²-score, averaged over 3 folds using 3-fold cross-validation. Generalizability of the models was tested by training the model on the complete data of a single fiber and testing it on the second fiber.

Results: For MES-I, the best number of action potentials back-casted from latency changes was observed for 3 subsequent latencies and a cubic polynomial with the R²-score of 0.78. In contrast, for MES-II, the best results were obtained with a quadratic polynomial and a single latency as input. The R²-score is 0.47.

For testing the generalization, the model fitted with MES-I and tested on MES-II resulted in the R²-score of 0.32. When we train on MES-II and test on MES-I the R²-score drops to -0.26.

Discussion: The results support our hypothesis, that some information about the preceding activity could be obtained by examining fiber latency changes. However, the generalizability of the models between data sets has potential for improvement, which will be investigated further using calibration, adapted normalization methods and stratification of the data.

Another major problem is the ground truth reliability. In this project we compared the theoretical spike count based on the protocol with the sorted spikes and the significant differences found are currently being investigated.

The preliminary results presented here led to the design of a novel experimental protocol with stimulation paradigm enabling identification and sorting of every electrically evoked spike in a dataset in human microneurography to provide ground truth required for further work on complete spike sorting.

Conclusion: This work shows the feasibility of using regression models for partial backcasting of fiber activity from latency changes to support automatic [8] spike sorting and thus use the full potential of large microneurography datasets.

The authors declare that they have no competing interests.

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