This study investigated the negative effect of electrical crosstalk during reflex detection and revealed that the use of a previously validated scoring criterion may result in poor specificity in the presence of crosstalk. Two different standardized methods for estimation of muscle fiber CV were employed to demonstrate that significantly different CVs may be estimated during genuine reflexes and crosstalk, respectively. This discriminative feature was used to develop and assess a novel methodology for reflex detection from sEMG that is robust with respect to crosstalk.
Identification of genuine reflexes
The sEMG and iEMG signals recorded together with the set of fixed criteria described in the Methods section, based on physiological knowledge of the human withdrawal reflex, clearly indicate that the signals regarded as crosstalk are indeed so. This conclusion can mainly be drawn due to temporal observations. Studies in both animals [17–20] and humans [21, 22] have demonstrated that nociceptive withdrawal reflexes are modularly organized, meaning that each muscle or group of synergistic muscles has a bounded well-defined and unique cutaneous reflex receptive field (RRF). Noxious stimulation of the skin within the RRF may cause a reflex response involving the related muscles, whereas stimulation outside the RRF has no effect or may even inhibit activity in the same muscles [18, 19, 23]. The RRF is hence defined as the skin area from which a reflex can be evoked, which generally adheres to biomechanical function of the related muscle ensuring efficient withdrawal [17, 19]. During voluntary contraction, the two antagonistic muscles may likely be activated simultaneously (co-contraction) to stabilize a joint. However, due to the functional modular organization of the NWR, this is highly unlikely to occur during a reflex response. A few times during the data acquisition, activity in both muscles were observed within the reflex window. However, these occurrences were never synchronized in the recordings from the two muscles, but could be synchronically identified in both the sEMG and iEMG signals recorded from each of the two muscles, respectively. These recordings were not included in the data analysis, in accordance with the exclusion criteria.
Validation of automated scoring criteria
Several automated scoring criteria, including the interval peak z-score, have been demonstrated to enable accurate and reliable reflex detection from sEMG signals [4, 5]. This comprehensive validation of scoring criteria contributed to the standardization of reflex detection methodology and promoted the evaluation of reflex thresholds as a valuable experimental and clinical tool. However, the validation performed did not consider EMG crosstalk, and the applied gold standard (visual examination of the sEMG signals), did not allow a proper evaluation of the authenticity of any apparent reflex observed as illustrated in Figure 2.
The present study utilized iEMG to obtain an improved gold standard that allows for a distinction between crosstalk and genuine reflexes. The results demonstrated that crosstalk may easily fulfill both the investigated scoring criteria (interval peak z-score > 12) and the criteria for subjective assessment presented by Rhudy and France  and France et al. . Consequently, the validation of the interval peak z-score seems limited to the ability to detect ongoing electrophysiological activity. The origin of the electrical signal (the effectuating muscles) is not considered, rendering muscle-specific reflex detection problematic.
Crosstalk implications on reflex detection
The NWR utilizes complex muscle synergies to effectively withdraw a limb from potential noxious stimuli and will often involve activation of more than one muscle . This may cause a sEMG recording of a reflex response to be a mixture of signals originating from various muscles. Consequently, valid muscle specific reflex detection requires a distinction between signal components originating from adjacent muscles (i.e. crosstalk) and genuine muscle activity in the investigated muscle. This is of particular importance whenever the muscles executing the reflex response are of interest e.g. during mapping of RRF or assessment of RRF modulation for specific muscles in relation to variation in nociceptive sensitivity .
This study clearly demonstrated the possible consequences of disregarding crosstalk during reflex detection. The evaluation of interval peak z-scores calculated for SD sEMG signals (the standard recording of reflexes in both experimental and clinical settings) resulted in reflex detection with an extremely poor specificity (0.20 and 0.19 for TA and SOL respectively). In this study, nearly half of all detected reflexes were not genuine reflexes but merely reflecting crosstalk. The amount of crosstalk depends on anatomical conditions, such as thickness of the subcutaneous layer [7, 8], and clearly on the magnitude of the originating muscle activity.
The amplitude of reflexes in the present study may be modulated in several ways, most effectively by manipulation of stimulation intensities. Consequently, the specificity for reflex detections based on evaluation of interval peak z-scores may be influenced by varying stimulation intensities. However, the stimulation intensities applied in this study (1.2-1.5 times the reflex threshold) were lower than those generally applied in previous studies using similar experimental setups [22, 24]. This indicates that the magnitude of crosstalk observed in this study may be representative for NWR assessment using a similar experimental setup and hence that, under certain conditions, crosstalk may pose a serious problem during NWR assessment.
Distinction between crosstalk and genuine muscle activity
This study has clearly demonstrated that significantly different muscle fiber CVs may be estimated for genuine reflexes and crosstalk, respectively. The underlying hypothesis has been tested using two standardized methods for CVs estimation based on different types of sEMG. The apparent CVs estimated for crosstalk are unreasonable high from a physiological perspective, reflecting that the main components of the signals are not propagating at all but are observed roughly simultaneously at the two adjacent recording sites. The apparent CVs estimated due to electrical crosstalk alone were more than one order of magnitude higher than CVs estimated for genuine reflexes, rendering the precision of the simple cross-correlation technique sufficient to allow distinction between crosstalk and genuine reflexes. However, the limited precision of the cross-correlation technique and the resulting relatively large variability in CVs estimations did entail detection thresholds for muscle fiber CVs well above the physiological range.
Restrictive inclusion criteria for the individual sweeps recorded has been employed to ensure the existence of two separate datasets consisting exclusively of crosstalk and genuine reflexes, respectively. While appropriate and necessary in order to properly demonstrate the differences in CV for crosstalk and genuine muscle activity, it leaves room for more complex scenarios, involving both crosstalk and genuine reflex activity within the reflex window for further research. When practical measures are considered, crosstalk may be due to a combination of travelling and non-travelling signal components depending on the distance between the active muscle fibers and the detection point . In the case of a mixture of crosstalk and genuine muscle activity an estimation of CV will involve a weighted average of the temporal delay of both propagating and non-propagating signal components originating from both crosstalk and genuine muscle activity. Whether estimation of CV of a recorded signal indeed reflects the amount of crosstalk in a reliable manner needs to be investigated and could potentially allow not only detection but also estimation of crosstalk during reflex detection.
Conduction velocity estimation
Various CV estimation techniques differ with respect to specific definitions of the delay between signals which in practice have unequal shapes . Hence, minor discrepancies between CVs estimated using different methods are inevitable. However, the most evident difference between the two sets of muscle fiber CV estimations is the width of their confidence intervals. CV estimation using a sophisticated algorithm performed on several sEMG channels recorded using an electrode array had a higher precision than the much simpler cross-correlation technique performed on two SD sEMG channels. This was expected and the precise maximum likelihood estimator was intended for validation of estimated CVs and to constitute reference values for future work.
Reliable evaluation of whether a signal is dominated by propagating or non-propagating signals components requires estimation of CV with a certain degree of accuracy and precision. The simple and convenient method for muscle fiber CV estimation, constituting the core of this new methodology for reflex detection, is best suited for long superficial muscles with parallel fibers like TA. The performance of the novel methodology may therefore vary when applied on different muscles, but even evaluation of a bi-pennate muscle with short, non-parallel fibers like SOL did allow reflex detection with excellent accuracy.
Whereas the CVs estimated for TA were slightly higher than previous findings , the CVs estimated for SOL in the present study were definitely higher than physiologically reasonable. However, a considerable overestimation had to be expected considering the rather wide pennation angle for this muscle. SOL has a pennation angle of approximately 25 degrees at rest , reducing the effective inter-electrode distance along the orientation of the fibers about 10%, causing an equivalent overestimation of the average CV.
Furthermore, Broman et al.  reported CVs up to 8 m/s for TA estimated using the applied cross-correlation technique on SD signals, whereas application of the same technique on DD recordings eliminated these supra-physiological CV estimations. These high CVs may be caused by the existence of both propagating and non-propagating signal components due to inhomogeneity and anisotropic properties of the volume conductor and may possibly be reduced by evaluation of cross-correlations of DD signals instead of SD recording.
Additionally, the high-pass filtering prior to CV estimation using the cross-correlation technique may result in CV overestimation. The cut-off frequencies constitute a compromise between maintenance of a sufficient high SNR and rejection of distorted waves, and will accordingly differ between the two muscles due to marked difference in muscle fiber length. As such, an optimal cut-off frequency cannot be selected, especially for SOL. The applied cut-off frequency at 100 Hz attenuates signal components with wavelengths exceeding 40 mm considering an average CV of 4 m/s, whereas the length of SOL muscle fibers at rest are approximately 35–38 mm . Even under the most unlikely assumption (that the motor end plates are located at the end of the fibers), distorted waves remain. However, increasing the cut-off frequency will result in a SNR that would be too low to allow meaningful CV estimation.
Applicability and necessity of improved reflex detection
The validation of the interval peak z-score, based on visual examination of SD sEMG recordings, carried out in this study supports previous findings [4, 5]; a threshold value around 12 will allow accurate and reliable detection of apparent reflexes. However, in the presence of crosstalk, not all electrophysiological activity observed represents a genuine reflex involving the muscle investigated. This potential issue was elucidated by the application of a refined gold standard (visual examination of iEMG ), allowing distinction between crosstalk and genuine reflexes and resulting in improved validation. It was hereby revealed that application of an interval peak z-score threshold of 12 to achieve muscle specific reflex detection may result in an extremely poor specificity, especially when performed on SD sEMG signals. As shown in the plots of sensitivity and specificity (Figure 4), a strikingly improved specificity combined with a reasonable sensitivity could be achieved by setting a much higher threshold for the interval peak z-score. Evaluation of interval peak z-scores calculated for DD sEMG signals may allow a joint value of sensitivity and specificity of 0.95 if the threshold was set at 48 instead of 12.
It is nevertheless stressed that one optimal, fixed threshold for the interval peak z-score cannot be established, and that custom thresholds should be chosen with great care. The application of a very high threshold in order to distinguish genuine reflexes from crosstalk based on the magnitude of the electrophysiological measurements would work well on this specific dataset. However, this will probably not be the general case. Both the optimal interval peak z-score threshold and the resulting sensitivity and specificity may vary strongly depending on the data in question. In any case, sufficiently small reflexes will be mistaken for crosstalk and erroneously undetected. Since experimental and clinical protocols often emphasize evaluation of reflex thresholds, this poses a serious problem. This problem does not arise when applying CVA for reflex detection, which constitutes a major advantage of this novel methodology.
The optimal method for reflex detection depends on specific challenges and requirements, including the presence of crosstalk and also weighting of sensitivity and specificity. Reflex detection based on evaluation of interval peak z-scores performed on both SD and DD sEMG entailed perfect sensitivity, indicating great performance in the absence of crosstalk. However, in the presence of crosstalk, the evaluation of DD sEMG instead of SD signals may entail a significant improvement in detection accuracy. This is clear from the plots of sensitivity and specificity (Figure 4) where both sensitivity and specificity, for all interval peak z-score thresholds, are superior for evaluation of DD sEMG compared to SD sEMG. The statistical analysis and the box-plots in Figure 5 suggest that CVA seems to entail slightly lowered sensitivity, especially for SOL. Thus, the relative value of sensitivity and specificity respectively ought to be weighted prior to deciding whether to apply the novel methodology or to purely evaluate interval peak z-scores calculated for DD sEMG. Also the risk and magnitude of crosstalk should be considered. In cases with seldom and weak crosstalk, the specificity achieved by reflex detection based on evaluation of interval peak z-scores calculated for DD sEMG may be sufficient, rendering superior sensitivity. However, whenever muscle specific reflex detection with a reliable high specificity is required, CVA should be seriously considered.
Beyond reflex detection
CVA may be viewed as an additional binary evaluation following another reflex detection methodology, in order to assess whether a detected reflex indeed is a genuine reflex or merely the result of crosstalk. There seems to be no reason why this approach should be less efficient detecting crosstalk during static or voluntary contractions. Hence, this paper presents a convenient generic method for qualitative assessment of crosstalk, applicable on signals recorded using standard sEMG equipment and procedures which may possibly be utilized to ensure a more specific and reliable detection of genuine muscle activation e.g. during gait analysis, biofeedback therapy, prosthetic control, or other applications.