In summary, clustering did not show good potential for spike sorting in microneurography, but classification seems promising for many recordings. In our clustering approach, employing F1 (amplitude and width) had the best scores for 17 out of 26 datasets. The raw waveform (F7) yielded the best performance for 19 datasets using SVM classification. The features of the SS-SPDF methods could provide additional information beyond the use of shape alone, which is important for fiber responses with very similar shapes in the same recording. To conclude, the classification approach empowered by the marking method demonstrates the higher potential.
The results strongly suggest that the identification of the most suitable input feature set requires individualized exploration for each dataset, although algorithmic approaches can be proposed for decision optimization.
In addition to the optimized sorting process, we can establish criteria to determine the general sortability of a recording. For instance, if the accuracy computed on the tracked spikes is below a predefined threshold, then the file can be marked as non-sortable. This threshold should depend on the specific research question and experimental objective, which would require an evaluation of the acceptable level of misclassifications. The first electrical stimulation protocol that is used for characterizing nerve fiber types in a microneurography experiment uses monotonous electrical stimulation with low frequency and can be employed for classification purposes. Thus, it can serve as training data and for the sortability check.
This work marks a significant step by introducing the first systematic spike shape analysis and reporting the challenges of microneurography data. It stands out as the first study that considers this amount of data from two different locations and hardware setups. What sets our study apart from the previous studies is the ability to evaluate the data against ground truth, facilitated by the marking method applied in microneurography.
Incorporating 26 datasets from two different locations allowed us to capture the diverse variability and variance present in microneurography recordings. Our selection aimed to encompass a wide spectrum of recording durations, fiber counts, and track counts, excluding recordings where fibers overlapped, as these instances could cause even more challenges for sorting.
Our spike sorting pipeline presents an opportunity to conduct microneurography analyses on a broader scale, employing straightforward methods that enhance transparency regarding the sorting process and its limitations within the context of microneurography. Our clustering and classification methods are adjustable, allowing for the incorporation of more advanced approaches suited to our task.
Signal pre-processing was essential for all datasets and involved techniques such as up-sampling for the Aachen data and smoothing for the Bristol data due to the different sampling frequencies of 10 and 30 kHz, respectively. While these pre-processing steps may alter the signal and potentially the spike shape, they were necessary to facilitate the computation of action potential derivatives for feature extraction using the SS-SPDF method.
Given the considerable variability and the large number of action potentials, we implemented automated outlier removal utilizing the computed templates. However, this approach may be too restrictive, particularly over extended recording periods, where action potential shapes can evolve and potentially fall outside the template range yet still belong to the tracks. Alternatively, we propose considering multiple templates per track, subsequently merging them, or implementing a manual upper limit for outlier removal when excessively many action potentials are filtered out, as, for example, observed in A4 with 1966 action potentials before filtering and 1399 after (see Table 1).
Considering the features utilized in the SS-SPDF method, our data deviates from the recording conditions presented in the literature. Specifically, the sampling frequencies for the Aachen and Bristol datasets (10 kHz and 30 kHz, respectively) are notably lower than those reported in the referenced paper (44 kHz). Moreover, microneurography recordings are sensitive to disruptions such as electrode movement, sudden overlay with other spikes originating from nerve fibers from the autonomic nervous system which bring spikes from the spinal cord to the periphery, and environmental electrical noise, which can impact the recording quality.
We decided on k-means clustering due to its interpretability and ability to specify the number of clusters, which aligns with our knowledge of the number of tracks present. A common challenge with k-means clustering is its sensitivity to initialization. Different initializations yield different results. To address this, we utilized a seed value to ensure reproducibility. Under favorable conditions, as observed in dataset A5, clustering alone suffices, and in such rare instances, the training process of the supervised approach could be disregarded.
For classification, we employed SVM classifiers with RBF kernels due to their robust performance and ability to effectively handle high-dimensional data. The challenge with using accuracy for evaluation arises when dealing with imbalanced datasets. In such cases, it is essential to consider class-specific measurements to ensure accurate evaluation. In this work, we reported the average results obtained through 5-fold cross-validation for accuracy and added class-specific measurements to the Supplementary Materials. For future work, we plan to utilize the cross-validation approach to identify the most suitable features for a specific dataset. After the identification, we intend to use all tracked action potentials for training and classify action potentials that are not on the tracks and responses to additional stimulation, such as those illustrated by the orange spikes in Fig. 6. Overall, our classification approach achieved successful classification rates of 60–70% for most action potentials. This is because detailed pattern analyses in microneurography are not sufficient and certain recordings may not reach this score due to similarities in action potential shapes across fibers or excessive activity from multiple fibers in a single recording. However, specific feature sets in individual datasets produce an acceptable sorting rate of above 80%. Thus, our pipeline indicates whether the action potentials in a given file can be sorted with a high enough degree of precision for a certain research question and with which method and according to which feature sets the highest accuracy is achieved.
The variability in the number of fibers is a challenge when comparing clustering and classification performance across single recordings. Each fiber has a distinct random choice probability, which must be considered when evaluating different results. The more fibers are within a recording the success rate drops. This means during the experimental phase careful consideration is necessary for picking the right recording for specific scientific questions. However, as demonstrated for dataset A8, it might be sufficient to sort the action potentials correctly to one specific nerve fiber and disregard the other non-separable fiber information as shown in the confusion matrix.
Another observation was the lack of a direct correlation between increased data amount in terms of recording time and improved performance, which additionally supports the idea that certain recordings are inherently more sortable than others.
This points out that conventional spike sorting algorithms should be used carefully and whenever possible validated against ground truth at least for some of the data for estimation of the reliability of the sorting process for a given recording and a specific nerve fiber activity.
In our future work, we aim to integrate more advanced classification methods, such as deep learning techniques. This needs, however, a sufficient amount of data. Therefore, we have also developed open-source software [17] and a metadata infrastructure [13] to improve data retrieval processes. As previously mentioned, this study serves as a proof-of-concept for the effectiveness of feature extraction methods in spike sorting. While our current focus has been on sorting spikes on tracks, we aspire to extend our analysis to include spikes elicited by extra stimuli, such as responses to pain- and itch-inducing substances.
Further, we expect that integrating latency information, given the unique activity-dependent slowing property of C-fibers, could enhance spike sorting by providing probabilities for action potentials to be affiliated with specific fibers. Together with advanced machine learning models enabling automatic feature extraction, this direction of research is promising to further improve spike sorting processes in microneurography necessary to the advancements of our knowledge on pain and itch signaling.