Machine learning powered tools for automated analysis of muscle sympathetic nerve activity recordings

FIGURE 1. Architecture of neural networks. The first line in each box refers to the type of layer, the second specifies further properties and the activation settings of the individual layer (Tensorflow activations relu, sigmoid, and softmax were employed), and the last line specifies the dimensions of the layer if applicable

FIGURE 2. Visualization of the quanitfication process between ECG and BP offset. ECG R‐waves are marked with red crosses and interrupted lines in the upper panel and interrupted lines in the lower panel. The lowest point of the BP cycle is marked in the lower panel with an orange dot and line, the highest with a green dot. Temporal difference between red and orange line was assessed in over 100,000 instances

FIGURE 3. Two excerpts of the microneurography recording in one participant. The upper panels show the recording after finding a valid MSNA signal (a, only the first 30 s of the 60 s excerpt on which Fourier transformation was carried out is displayed) and the histogram of the one dimensional discrete Fourier transformation of a 60 s except with the x‐axis just showing the absolute frequency of individual values (b). Panels (c) and (d) show the same analysis for an excerpt at the very beginning of the recording when no valid signal had been acquired yet

FIGURE 4. Performance parameters for the training for DTCT 1. Panels (a) and (b) show the average accuracy and loss of 100 individual, completely separated training repetitions of the algorithm for training and testing dataset. Panels (c) and (d) show histograms of the performance parameters AUC ROC and accuracy for each individual time the algorithm was trained

FIGURE 5. Examples from the labeled dataset for the second neural Network DTCT 2. A shows signal peaks that were rated to likely represent valid MSNA signals. b, shows examples of signal peaks rated not to represent likely valid MSNA signals. The signal is orientated in a way that its highest point is the center of the image and additionally marked with an orange, interrupted line

FIGURE 6. Performance parameters for the training for DTCT 2. Panel (a) and (b) show the average accuracy and loss of 100 individual training iterations of the algorithm for training and testing dataset. Panels (c) and (d) show histograms of the performance parameters AUC ROC and accuracy for each individual time the algorithm was trained

FIGURE 7. Examples from the labeled dataset for the third neural Network DTCT 3. Panel (a) shows signal peaks that were rated to likely represent valid MSNA signals. (b) shows examples of signal peaks rated not to represent likely valid MSNA signals. The signal is orientated in a way that its highest point is the center of the image and additionally marked with an orange, interrupted line

FIGURE 8. Performance parameters for the training for DTCT 3. Panels (a) and (b) show the average accuracy and loss of 100 individual training iterations of the algorithm for training and testing dataset. Panels (c) and (d) show histograms of the performance parameters AUC ROC and accuracy for each individual time the algorithm was trained

FIGURE 9. Distribution of time difference between ECG‐peaks (R‐wave) and lowest blood pressure signal. Panel (a) shows the final, focused distributions, and panel (b) (log‐scale) shows the distribution with long left, negative tail after z‐score based data selection and before clipping of extreme values below −250 ms

FIGURE 10. (a) shows the quality index derived from the Fourier transformed of a 3‐h long microneurography recording. The quality index is calculated for each detected signal peak during this time for a 60‐s window around the peak. The median of the Fourier transformation for these excerpts is associated with the individual peaks and plotted over time (a). Low, relatively constant intervals of this variable represent stable and valid MSNA recordings, higher and varying levels periods of positioning the needle or losing the signal for example after the participant moved. The excerpts at points (b, c and d) have been chosen randomly based on this quality index. Point (b) is located at a time point with high variance and generally high levels of the quality marker–indicating bad signal quality. The sub‐panel (b) shows the original microneruography signal corresponding to this time point–showing a noisy, non‐valid signal. At time point (c) the quality index suggest a stable, valid signal–as seen in the corresponding original data window in sub‐panel (c). At time point (d) the quality index has suddenly sored upwards again, indicating a loss of valid signal, confirmed by the corresponding…

FIGURE 11. Boxplot showing the distribution of median values of the one dimensional Discrete Fourier transformation (DFT) for 990 excerpts of MSNA recordings. These recordings are rated into valid and invalid recordings–analysis was carried out according to this grouping of the data. The middle bar of the box represents the median, borders of the boxes represent the interquartile ranges, notches the bootstrapped 95% confidence intervals. Whiskers and circles are showing outliers of the data

FIGURE 12. Signal window of microneuroggraphy recording of excellent quality. Signal peaks selected by the software are marked with vertical orange lines. The most important variables assigned to each detected signal peak are printed on top of each individual peak (first line consecutive signal number, second line binarized output of the three neural networks with 1 indicating valid signal and 0 non‐valid signal, third line absolute probability density of the individual peaks based on the timing model, last line median of the Fourier transformed 60‐s window around the signal). (b–d) are circular representations of the most important determined variables for each individual peak, subpanel (c) corresponds to the central peak marked in red in subpanel (a), the circles right and left to the corresponding next signal peak right and left of the marked central peak. The circular graphs visualizes the binary decision of the individual DTCT neural network algorithms (which are marked as DTCT_1, DTCT_2, and DTCT_3) surrounded each left and right by the continuous likelihood for a valid signal given out by the respective algorithm. The overall quality index (FFT signal Quality) rescaled from 0 to 1 (1…