2.1 Participants

Eighteen (18) young male subjects with a mean age of 25.8 years were recruited to participate in this study. Male subjects were recruited to avoid confounds due to previously reported sex differences in motor neuron discharge rate (Peng et al., 2018; Tenan et al., 2015). Participants had no history of pain, surgery, or injury to the dominant upper extremity. Participants were also free from metabolic and neurological disorders, cardiovascular dysfunction, and did not take blood thinning medications. All methods were reviewed and approved by the Institutional Review Board of the U.S. Army Research Laboratory. Participants provided verbal and written informed consent prior to participation.

2.2 Maximal Voluntary Contraction (MVC) Procedures

Participants were seated at an adjustable height desk equipped with a custom instrumented grip (Figure 1). The height of the desk was adjusted such that the participant could sit up straight with their supine forearm resting flat on the surface of the desk. A custom adjustable cradle was mounted to the edge of the table to secure the position of the upper arm and standardize arm position between participants and between trials for the same participant (Figure 1 A). The grip consisted of a plastic rod with a trigger-like mechanism instrumented with a compression load cell (LCM302-50, Omega Engineering, Inc., Stamford, CT). The angular orientation of the grip was adjustable to afford a comfortable neutral position for the wrist while the forearm and hand remained in supine position. The distance of the grip from the participant was adjusted such that the middle phalanx of the second finger engaged the trigger mechanism, and the remaining fingers curled comfortably around the plastic rod. All finger flexion tasks were completed with the hand and forearm in this standardized supine position.

Prior to completing other experimental tasks, participants completed a series of maximum exertions to determine their maximum finger flexion force. Participants were instructed to produce force only by squeezing their pointer finger against the trigger mechanism and were asked to refrain from flexing their wrist or elbow. A monitor positioned at the participant's eye level presented visual prompts and force feedback via custom data acquisition script (Spike2, Cambridge Electronic Design Ltd., Cambridge, England, UK). Verbal encouragement was also provided during maximum voluntary contractions (MVCs). MVCs were sustained for 5 seconds each with a rest period of 1 minute between successive attempts. Force data were amplified using a bridge amplifier (Digitimer NL109, Digitimer Ltd., Ft. Lauderdale, FL) with a gain of 100 and were recorded at a rate of 1000 Hz via a Micro 1401-3 data acquisition unit (Cambridge Electronic Design Limited, Cambridge, England, UK). After three exertions, the values for each MVC were analyzed. If the three consecutive MVC values were within 10% of each other, the highest MVC was recorded as the maximum force value. If there was more than a 10% difference between the highest and lowest recorded MVC attempt, the participant was given 5 minutes of rest and another round of three MVCs were completed. This procedure continued until a maximum force value was identified. For all participants, this was accomplished in three or fewer rounds of MVC attempts (n=16 for 1 round, n=1 for 2 rounds, n=1 for 3 rounds).

Following the MVC procedures, each participant's individual sub-sensory threshold was determined. A vibrotactile stimulator (BM1C, Tactile Labs Inc., Montreal, Canada) was secured to the wrist of the participant's dominant hand using a neoprene strap (Figure 1 B). All participants were right-hand dominant as determined by self-report. A white Gaussian noise signal (\(\le\) 500 Hz) was generated with a custom Labview^TM (v8.5, National Instruments Corp., Austin, TX) program. The signal was generated using Labview's Gaussian white noise VI with a standard deviation of the Gaussian probability density function equal to 2.0. The resulting signal was low-pass filtered at 500 Hz to include primarily the range of frequencies known to stimulate cutaneous mechanoreceptors (30-350 Hz; Purves et al. (2001)). The signal amplitude ranged from +/- 1.0 with a mean of 0.0. An input of 1 V to these stimulators would produce an acceleration of approximately 2.5 G. This control signal was scaled with two different gradations: 5-100% of full-scale in 5% increments and 0.5-5% full scale in 0.5% increments. The second scaling gradation was necessary to provide a finer resolution for the identification of sensory threshold for participants’ whose threshold was under 5% of full scale (n=4). Similar white noise has been shown previously to improve performance of isometric force production tasks and includes the frequencies to which biological proprioceptors in the skin are sensitive (Trenado et al., 2014). The signals were transmitted to the wrist through the Micro 1401-3 unit via a custom Spike2 interface and delivered by the stimulator. Stimulation signals were presented for a duration of 5 seconds, and the magnitude of the signals and the time between successive stimulation signals was unknown to the participant. With eyes closed and hearing occluded with ear plugs, participants were asked to provide a thumbs up with their non-dominant hand when they detected stimulation. Researchers began by applying noise levels likely to be supra-sensory to confirm the participant understood instructions and were able to discern the vibratory stimulation. After approximately 5 trials of supra-sensory stimulation, the researcher applied a stimulation level which was expected to be on the cusp of perception. If the participant provided a positive response to the stimulation, the researcher followed by presenting the next 3 lower levels of noise stimulation. If the participant failed to provide a positive response to a stimulation, the researcher would apply noise two levels higher than that of the failed stimulation. The researcher would then proceed through the next 3 lower levels of noise stimulation. This process was repeated and afforded confirmation of threshold levels by ensuring that the participant would consistently report positive responses to levels above threshold and fail to detect other stimulation levels below threshold. The highest stimulation that the participant failed to detect three consecutive times was recorded as their sub-sensory threshold.

2.3 Principal Noise Stimulation Protocol

As this protocol included only a finite set of stimulation levels and optimal stimulation was not confirmed, the level eliciting the best performance was referred to as an individual's principal stimulation level. The principal noise stimulation level was determined using a ramp-and-hold finger flexion task. Participants were presented with a visual display of the force applied to the trigger mechanism (Figure 1 C). A visual cue would instruct the participant to gradually increase the force applied to the trigger over a period of 2.5 seconds to a maximum force of 20% MVC and maintain this exertion for 10 seconds. Participants were instructed to maintain the 20% MVC as steadily as possible by tracing the visually presented 20% MVC target force with the force profile generated by finger flexion (Figure 1 D). Ten practice trials were completed prior to data collection. Errors from these practice trials were retained to be included in analysis. To determine the principal sub-sensory stimulation level, noise signals were generated corresponding to 10-100% of their sub-sensory threshold. Participants completed 13 data trials in which they completed the ramp-and-hold task with or without stimulation applied to the wrist. The first and last trials were always sham trials in which no stimulation was applied. The sequence in which each noise signal (10, 20, 30, 40, 50, 60, 70, 80, 90, and 100% sub-sensory threshold) was applied was randomized for each participant. Following the initial sham trial, a block of 5 ramp-and-hold trials were completed with different levels of stimulation. Participants then rested for five minutes. The first trial following the break was a sham trial, then 5 more ramp-and-hold trials were completed with the remaining stimulation levels followed by the final sham trial. In total, 13 trials were completed including the initial sham trial, the first block of stimulation trials (5), a second sham trial, the second block of stimulation trials (5), and a final sham trial.

2.4 Dependent Measures

Force stability, the primary measure of performance, was defined as the RMS error of the difference between actual force and the target force of 20% MVC. This value was calculated during the middle 5 seconds of the sustained MVC hold for each trial resulting in a total of 13 values per participant. The central 5 seconds of the force stability data were selected for analysis because this window of time was free from overshoot and undershoot artifacts that occurred near the ramp portions of the trial and were considered the period of best sustained performance. The stimulation level corresponding to the lowest force stabilization RMS error (best performance) was recorded as the participant's principal noise stimulation level.

2.5 Statistical analyses

RMS errors from the practice trials were analyzed using a one-way repeated measures ANOVA in which trial number served as the single factor. Post hoc Tukey analysis was conducted to compare performance during each pair of trials, and statistical significance was adjusted for multiple comparisons. This analysis of practice data was used to ensure that any learning effect of the trials were mitigated and provided the researchers with a stable set of practice trials for use when comparing errors obtained during noise stimulation trials. The practice vs. stimulation assessments used a mixed-effects repeated measures ANOVA analysis with trial type (practice or stim) as a fixed effect and subject as a random effect. This analysis was used to determine whether stimulation had a systematic effect on task performance.

RMS error data were analyzed at the group level using a multilevel model with stimulation level and subject as fixed and random effects, respectively. Post hoc statistical comparisons were made using Dunnett's test to correct for multiple comparisons (Dunnett, 1955). At the individual level, one-sided Dixon Q tests were used to examine whether performance during the identified principal stimulation level was significantly different than the performance at all other levels of stimulation. Finally, Spearman correlations were used to identify whether relationships exist between threshold, MVC, and principal stimulation values. All statistical analyses were conducted with R software (3.5.0, R Foundation for Statistical Computing). Effect sizes were calculated using Cohen's d with pooled standard deviation.

3.1 Analyses of Baseline vs. Stimulation Trials

All participants completed 10 practice trials, and 3 participants completed an eleventh practice trial per their request. Repeated measures analysis revealed a main effect of trial number (F(10,148) = 8.15, p < .0001). Post hoc analyses for multiple comparisons showed that the RMS errors for the first trial were significantly greater (p < .001, d = 0.87 to 1.21) than RMS errors during all other trials (Figure 2). Significant differences in performance between other pairs of trials were not found.

To further investigate practice performance, the best practice trial was identified for each subject, and the remaining trials were identified as best-i or best+i where i preserved the sequence in which they were collected relative to the best trial. For example, if a subject's best performance was recorded for their fourth practice trial, this trial was identified as best. Practice trials 1-3 were labeled best-3, best-2, best-1, and trials following the best performance were labeled as best+1, best+2, etc. Using a linear mixed model with trial as a within-subjects factor and subject as a random effect, RMS error was compared between different trial levels. Post-hoc planned comparisons revealed significant differences (ranging from t(145) = -3.14, p = 0.035 to t(146) = -6.56, p < 0.0001) between the best practice trial and trials identified as best-9, best-8, best-7, and best-5. These results indicate that the trials following the first 5 practice trials best identify a performance plateau. Per these results, a subset of practice trials 6-10 or 6-11 where applicable were identified as the trials describing plateaued task performance during practice. This plateau subset would be used in subsequent analyses to draw comparisons to the stimulation trials.

3.2 Analyses of Stimulation Trials

Group-level analyses of force stabilization RMS errors were first evaluated by comparing performance at each stimulation level. Performance at each stimulation level varied greatly between participants (Figure 3). Multilevel model analyses were conducted with stimulation level as a fixed effect and subject as random intercept effect. No differences were observed between the RMS errors for the stimulation levels or the pre- or post-sham trials (F(11,187) = 1.45, p = .15). The largest RMS error magnitude (0.54 N) and standard error (0.084 N), however, occurred for the final sham trial (post-sham) following the completion of noise stimulation trials. To examine whether this increase is due to a fatigue effect, a Pearson correlation was calculated between RMS error and the sequential trial number. No fatigue-related time-on-task effects were found (r(16) = .12, p > .6).

To determine whether performance during stim trials differed from performance during practice, the RMS errors for the practice plateau subset were compared to the RMS errors during all stimulation trials. A one-way ANOVA analysis was completed with trial type as a within-subjects effect (two levels = practice, stim) and subject as a random effect. A log transform was used to satisfy the normality assumption, and results indicated no statistically significant difference (F(1,17) = 1.189, p = 0.291) between the mean RMS error during practice (0.54 \(\pm\) 0.17 N) and during all stim trials (0.49 \(\pm\) 0.13 N).

While the mean performance for all stim trials did not differ from practice performance, the primary interest of this study was whether individualized principal noise stim would result in improved performance. An individual's principal noise stimulation level was assigned as the noise stimulation trial for which they exhibited the best force stabilization performance (lowest RMS error). Figure 4 presents the RMS errors recorded for a single participant exhibiting principal stimulation at 30% sub-sensory threshold.

One-way ANOVA analyses were repeated comparing RMS errors between the plateau subset and the performance during individual principal noise stimulation. A significant main effect of trial type was found (F(1,16) = 26.313, p = 0.0001, d = 0.41). RMS error during principal noise stimulation (0.35 \(\pm\) 0.11 N) was significantly less than the RMS error during practice trials (0.54 \(\pm\) 0.17 N).

Principal stimulation levels varied widely across the group such that each sub-sensory stimulation level was found to be the principal stimulation level for at least one participant. To further examine the relationship between RMS errors and stimulation level for the group, the data was restructured such that the RMS error corresponding to an individual's principal stimulation level was assigned the label prin noise, and their RMS errors associated with the remaining stimulation trials were labeled as a positive or negative percent difference (%diff) from the individual's principal stimulation level. For example, for the data presented in Figure 4, the error at 30% sub-sensory threshold was identified as prin noise, and the errors at 10, 20, 40, and 50% were identified as 20, 10, +10, and +20 %diff, respectively, from the principal stimulation level. Once RMS errors for all participants were relabeled using this convention, the errors were averaged for the principal stimulation level and for each %diff stimulation level (Figure 5). The result of this data presentation is that the number of values contributing to the mean and error calculations are different between stimulation levels.

Group-level analysis was then conducted to determine if performance at the principal stimulation level was significantly better than performance at other stimulation levels expressed as %diff from this principal level. Model estimated marginal means, standard error, and confidence intervals are provided for each stimulation level in Table 1. When examining the RMS errors as a function of %diff from principal stimulation, the best performance as a group was observed at the principal stimulation level. Multilevel analyses revealed that the model estimate RMS error recorded for the principal stimulation level was significantly different (F(18,145) = 3.06, p < .0001) than the RMS errors for a number of stimulation levels using Dunnett's test (Dunnett, 1955) to correct for multiple comparisons. For stimulation levels ranging from -60% to +60%, all but the -30% level had t scores ranging from 3.28 \(\le\) t(145) \(\le\) 5.46 and significance levels of p \(\le\) .019 (d = 0.71 to 1.04). The stimulation level of -30% fell just outside the cutoff for significance with t(145) = 2.92, p = .053. Estimated RMS errors at stimulation levels of -90% and +80% were also significantly different than mean RMS error at the principal stimulation level (t(145) = 3.08, p = .035, d = 1.85 and t(145) = 3.61, p = .007, d = 0.89, respectively). In addition to the differences in estimated RMS values, the variation in performance as indicated by the box plot sizes also suggests that RMS errors at the principal stimulation level had a narrower distribution than errors recorded for other stimulation levels. To further validate this principal noise stimulation level, an identical analysis was performed on the errors measured during the practice trials. When completing the same ANOVA and post-hoc analyses on the ordered practice trial errors, the performance during the best practice trial was only significantly better than the first five attempts (t(146) = -6.57 to t(145) = -3.14, p = <0.0001 to 0.035, d = 2.46 to 0.92). This lent further support to the identification of the principal noise level as a special case.

Individual one-sided Dixon Q tests were conducted to determine whether the minimum RMS error recorded for a given participant could be identified as significantly different from the group of RMS errors recorded for other stimulation levels. Two participants (participants 4 and 11) demonstrated RMS errors that were significantly lower (Q = .50, p = .039 and Q = .54, p = .024, respectively) than the remaining distribution of errors recorded for the other stimulation levels. To further inspect the distribution of errors for each participant, parametric 95% confidence intervals were calculated for each individual (Figure 6). The difference between a participant's lowest RMS error and their lower 95% confidence limit ranged from 0.005 N to 0.21 N with a group mean difference of 0.08 \(\pm\) 0.06 N. As stated previously, no consistency was observed in the sub-sensory stimulation level eliciting the best force stabilization (Figure 6). Among the 18 participants, each sub-sensory stimulation level was found to be the principal level for at least one individual (Figure 6).

Following the force stabilization trials, the input voltages identified as the mean principal noise stimulation level had a range of 0.005-0.16 V with a group mean of 0.06 \(\pm\) 0.05 V. This mean principal stimulation voltage corresponded to acceleration magnitudes of 0.035-0.15 G across the frequency spectrum of the stimulation signals. To examine whether significant relationships exist between threshold, principal stimulation level, and MVC values, Spearman correlations were calculated for each pair of measures. Spearman, rather than Pearson, correlations were used due to the lack of normality for the threshold and principal stimulation level measures as determined by the Shapiro-Wilk test (W = .80, p = .002 and W = .83, p = .005, respectively). Moderate but significant Spearman correlations were observed between threshold and MVC (r_s = .49, p = .041) and principal stimulation level and MVC (r_s = .56, p = .016). Increases in MVC are associated with increases in both threshold and principal stimulation level. A strong significant Spearman correlation was observed between threshold level and principal stimulation level (r_s = .65, p = .003).

7.1 Data Accessibility

All data and analysis code used for this manuscript are available at https://osf.io/cah3j.

7.2 Author Contributions

Courtney A. Haynes contributed to this work as a co-investigator assisting with the design the experimental paradigm, developing custom data collection programs, pilot testing procedures, and executing the data collection. Dr. Haynes completed the analyses described in this paper and led the writing of the various manuscript drafts incorporating revisions provided by her co-authors. Dr. Haynes consents to the publication of this manuscript in its present form.

Matthew S. Tenan contributed to this work as a principal investigator leading the design of the experimental paradigm and executing the full data collection. Dr. Tenan has contributed substantially to the content by providing thorough reviews of the manuscript, providing additional content, and assisting with the interpretation of the study's findings. Dr. Tenan consents to the publication of this manuscript in its present form.

Antony D. Passaro is a co-investigator for this work assisting with the design of the experimental paradigm and providing his expertise regarding data analysis techniques to examine the study results. Dr. Passaro also provided valuable suggestions for data presentation and provided critical revisions for the manuscript drafts. Dr. Passaro has provided his consent for publication of this manuscript in its present form.

Andrew J. Tweedell served as a co-investigator assisting with design of the experiment, hardware acquisition and establishing the experimental apparatus, pilot testing, and executing the data collection with human subjects. Mr. Tweedell has provided extensive manuscript revisions and provided insights regarding data analysis and interpretation. Mr. Tweedell has provided his consent for publication of this manuscript in its present form.

Contributed to conception and design: CAH, MST, ADP, AJT. Contributed to acquisition of data: CAH, MST, AJT. Contributed to analysis and interpretation of data: CAH, MST, ADP, AJT. Draft and/or revised the article: CAH, MST, ADP, AJT. Approved the submitted version for publication: CAH, MST, ADP, AJT.

7.3 Conflict of Interest

The authors declare that this research was conducted independently and free of any personal or financial relationships that could be construed as a conflict of interest.

7.4 Funding

This work was funded by the Department of Defense Human Systems Integration (Cybernetics) research line at the U.S. Army Research Laboratory (74-A-HRCYB).

7.5 Acknowledgments

Many thanks to Mr. Mike Kosinski and Mr. Dave Kuhn for their hard work designing and fabricating the instrumented grip apparatus and the arm cradle used for data collection. Thanks also to Dr. J. Cortney Bradford, Mr. Paul Riggs, and the facilities staff at the Army Research Laboratory Mission Impact through Neuro-inspired Design (MIND) lab for the use of their equipment and lab facilities.