Author Proof CTED UNCORRE - University of North Carolina

UNCORRECTEDPROOF

1 Prediction of Individual Finger Forces Based on Decoded Motoneuron

2 Activities3

4 CHENYUN DAI,1,2 YIZHOU CAO,1,2,3 and XIAOGANG HU1,2

5 1Joint Department of Biomedical Engineering, University of North Carolina - Chapel Hill, 144 MacNider Hall, Chapel Hill,6 NC 27599, USA; 2North Carolina State University, Raleigh, NC, USA; and 3Medical College of Soochow University, Suzhou,7 Jiangsu, China

8 (Received 2 October 2018; accepted 28 February 2019)

9 Associate Editor Jane Grande-Allen oversaw the review of this article.10

11 Abstract—Accurate prediction of motor output based on12 neural signals is critical in human–machine interactions. The13 objective was to evaluate the performance of predicting14 individual finger forces through an estimation of the15 descending neural drive to the spinal motoneuron pool.16 High-density surface electromyogram (EMG) signals of the17 extensor digitorum communis muscle were obtained, and18 were then decomposed into individual motor unit discharge19 events. The frequency of the composite discharge events at20 the population level was used to derive the descending neural21 drive, which was then used to predict the finger forces. The22 global EMG-based approach was used as a control condi-23 tion. Compared with the EMG-based approach, the exper-24 imental results show that the neural-drive-based approach25 can better predict the individual finger forces with higher R2

26 values across different force levels and across different force27 trajectories (steady and varying forces). These findings28 indicate that the neural drive estimation based on motoneu-29 ron firing activities can be used as a reliable neural-machine30 interface signal involving individual fingers. However, real-31 time implementation of this approach is needed for future32 clinical translation.

33 Keywords—EMG signal processing, Motor unit decomposi-

34 tion, Finger force, Hand function, Muscle activity.35

36 ABBREVIATIONS

37 MU Motor unit

38 MUAP Motor unit action potential

39 EMG Electromyogram

40 sEMG Surface electromyogram

41HD High density

42EDC Extensor digitorum communis

43MVC Maximal voluntary contraction

44SNR Signal to noise ratio

45ANOVA Analysis of variance

4647

4849INTRODUCTION

50Decoding the desired motor output is a key com-

51ponent in human–machine interactions. This decoded

52signal can allow individuals with neuromuscular dis-

53orders to interact with machines, such as exoskeletons,

54prostheses, or neural-stimulation systems, which can

55help restore lost or diminished motor functions. Re-

56cent advancement in neurally implantable micro-elec-

57trodes or thin-film electrode arrays have facilitated the

58ability to decode neural signals sent directly from the

59brain to the periphery.2 However, there is still con-

60siderable challenge in applying these techniques to

61humans, specifically in clinical populations, largely due

62to the invasive nature of the procedures and the lack of

63long-term stability of the system interface.3

64Alternatively, non-invasive muscle activity record-

65ings, such as surface electromyogram (sEMG) signals,

66have been widely used as the neural control signals of

67human wearable robots or neuroprostheses.1,21 One

68common strategy uses the amplitude of sEMG as a

69proportional control input of different degrees of

70freedom of the machine.9 Another typical strategy uses

71different features/patterns of EMG to identify different

72preset movements involving multiple degrees of free-

73dom.25

74Nevertheless, these global EMG-based approaches

75have several limitations. First, the global sEMG sig-

Address correspondence to Xiaogang Hu, Joint Department of

Biomedical Engineering, University of North Carolina - Chapel Hill,

144 MacNider Hall, Chapel Hill, NC 27599, USA. Electronic mail:

[email protected]

Annals of Biomedical Engineering (� 2019)

https://doi.org/10.1007/s10439-019-02240-1

BIOMEDICALENGINEERING SOCIETY

� 2019 Biomedical Engineering Society

Journal : ABME MS Code : 10439 PIPS No. : 2240 h TYPESET h DISK h LE h CP Dispatch : 2-3-2019 Pages : 124 4

Au

tho

r P

ro

of

UNCORRECTEDPROOF

76 nals consist of the summation of hundreds of motor

77 unit action potentials (MUAPs) generated from

78 motoneuron discharge events, it may not accurately

79 reflect the cortical control of muscle activation.

80 Namely, the measured EMG amplitude can be biased

81 by intrinsic physiological factors, such as the cancel-

82 lation of superimposed MUAPs20 and the natural

83 variations of action potential amplitude from the

84 complex conductive processes of different tissue lay-

85 ers.11 Second, EMG signals can be further deterio-

86 rated/corrupted by external interference during signal

87 acquisition, such as changes in electrode–skin contact,

88 the crosstalk of multi-channel EMG,10 motion arti-

89 facts, and the variation of baseline noise. All these

90 factors can ultimately limit robust communications

91 between humans and machines.

92 In contrast, recent studies have used motoneuron

93 discharge timings as a promising neural control input

94 for human–machine interactions.12,27 Instead of using

95 the global sEMG signals, a decomposition step was

96 performed on the multiple-channel sEMG signals to

97 extract spinal motoneuron discharge events. Although

98 individual motoneurons have distinct discharge prop-

99 erties that may not directly reflect the descending drive,

100 the frequency/probability of the motoneuron discharge

101 at the population level can directly reflect high level

102 neural drive from the brain to the muscles. Essentially,

103 the spinal cord output signal (motoneuron discharge

104 spikes as a whole) is used to estimate/decode the spinal

105 input signal (descending command from the brain).

106 This decoded neural command based on the binary

107 events can potentially overcome some of the disad-

108 vantages of the global EMG patterns. However, the

109 feasibility of using this approach for fine motor con-

110 trol, such as involving individual finger movement has

111 not been evaluated. In addition, the decoding accuracy

112 of the neural-drive-based approach compared with the

113 EMG-based method has not been fully investigated

114 under different conditions (i.e., force level, signal

115 quality, or signal stability). To overcome these limita-

116 tions, the objective of this study was to evaluate the

117 performance of the neural-drive-based approach in

118 estimating the individual finger forces using both

119 simulated and experimental data under different con-

120 ditions, with a wide range of signal quality and muscle

121 contraction levels.

122 MATERIALS AND METHODS

123 Both simulated and experimental data were

124 obtained to verify the parameter selection and the

125 accuracy of the methods. Figure 1 shows the process of

126 the EMG-based (Fig. 1a) and the neural-drive-based

127 (Fig. 1b) estimations.

128Simulated Data

129A classic EMG model22 was used to generate the

130EMG data, in order to evaluate the accuracy of the

131methods. The signal generation were described as fol-

132lows:

133First, the individual spike trains were generated from

134awidely accepted physiologically-basedMUmodel.13A

135progressive recruitment pattern with an exponential

136threshold function and a distributed discharge rate of

137MUs were modeled as described in the original study.13

138The main parameters used were summarized in Table 1.

139Firing variations with a 10%Coefficient of Variation in

140the firing ratewere added to the spike trains. In addition,

141a moderate level (10%) of MU discharge synchroniza-

142tion was added based on Yao et al.29

143Second, for each spike train, waveforms of HD

144MUAP was randomly selected from a HD MUAP

145pool. The MUAP pool was derived from earlier

146experimental data6 using OT Biolab (OT Bioelettron-

147ica, Torino, Italy). The EMG signals were first

148decomposed into constituent spike trains, and the

149corresponding waveforms of HD MUAPs were

150derived from a spike trigger average algorithm.16 In

151addition, a 10% amplitude variation was added to the

152MUAPs for each firing. The MUAP grid was also

153scaled by a coefficient from a uniformly distributed

154random number ranging from 0.1 to 2.

155Third, each HD MUAP array was convolved with

156the corresponding spike train to generate an HD

157MUAP train grid. Then, the 8 9 8 noise free EMG

158was obtained by superimposing all the MUAP train

159grids from different MUs. Additional Gaussian noise

160(band-pass filtered at 10–900 Hz) with different signal-

161to-noise ratio (SNR = 5, 10 and 20 dB) was added.

162Fourth, the twitch force of each MU was also sim-

163ulated. The MU force and the resultant muscle force

164was subsequently calculated based on the excitation

165drive level and the MU discharge frequency using the

166force model described in the same MU model.13 The

167main parameters used for the force simulation was

168summarized in Table 2.

169To match the conditions of experimental data, two

170excitatory drive levels (20% and 50%) were simulated.

171The variation of the drive level was similar to that of the

172experimental data, including both sine-wave and trape-

173zoid. A total of 600 trials (two force levels 9 two trajec-

174tories 9 three SNRs 9 fifty repetitions) were simulated.

175Experimental Data

176Participants

177Nine neurologically intact (with no known neuro-

178logical injuries or disorders) subjects (six males, two

179females; aged 26.3 ± 4.9 years) were recruited. One



DAI et al.

Au

tho

r P

ro

of

UNCORRECTEDPROOF

180 was excluded due to a short forearm length. Therefore,

181 the experimental data were acquired from eight sub-

182 jects. A power analysis was performed to ensure that

183 eight subjects were not under powered. All subjects

184 provided written informed consent approved by the

185 Institutional Review Board at the University of North

186 Carolina at Chapel Hill.

187Experimental Setup

188The subjects sat in a height-adjustable chair with

189their dominant forearm comfortably placed on the

190experimental table and the elbow supported on a foam

191pad. Given that a large electrode pad was placed on

192their forearm, also covering the wrist extensor, the

FIGURE 1. Diagram of the EMG-based (a) and neural-drive-based (b) approaches. The diagram of the motor unit decompositionand neural drive estimation (c).

TABLE 1. The main parameters used for MU timing eventtrain generation.

Parameter Value

Total number of neurons in the pool 120

Range of recruitment threshold 30

Coefficient of recruitment threshold ln 30/120

Minimum firing rate 5

Maximum firing rate of first MU 35

Maximum firing rate of last MU 20

TABLE 2. The main parameters used for the forcegeneration.

Parameter Value

Range of peak twitch force (RP) 100

Coefficient of peak twitch force (ln RP)/120

Range of contraction time (RT) 3

Coefficient of contraction time logRT RP

Longest duration rise time 90 ms



Finger Force Estimation from Motoneuron Activities

Au

tho

r P

ro

of

UNCORRECTEDPROOF

193 subjects were asked to minimize the wrist extensor

194 activation. To restrict the wrist movement, their wrist

195 was immobilized within two padded wooden boards in

196 a neutral pronation-supination position and zero de-

197 gree flexion–extension. The four fingers (index, middle,

198 ring and little) were comfortably abducted. Each finger

199 was secured, using two Velcro straps (each at the

200 proximal and middle phalanges) to one load cell (In-

201 terface, SM-200N, with a resolution of 0.01 N) which

202 measured the finger forces with a 1 kHz sampling

203 frequency. During the experiment, the subjects were

204 asked to isometrically extend one designated finger in

205 each trial. The maximum voluntary contraction

206 (MVC) of each finger was separately measured. The

207 average force of each finger during the contraction

208 plateau of 2 s was calculated as the MVC.

209 During the main testing session, the subjects were

210 provided with a target force trajectory shown on a

211 screen. The finger force output was instantaneously

212 displayed as a feedback to the subjects via a custom-

213 built MATLAB program (MathWorks Inc). Only the

214 designated finger force was displayed, but the forces

215 from all the four fingers were recorded. They were

216 asked to adjust the displayed muscle force of a desig-

217 nated finger to track the trajectory. Two force target

218 trajectories, sine-wave and trapezoid, were tested. The

219 sine-wave force trajectory allowed us to evaluate the

220 force estimation performance of the EMG- and neural-

221 drive- based approaches during varied level of con-

222 tractions. Two contraction force levels (20% and 50%

223 MVC) were used in each target. For the trapezoid

224 force, the subject took 2 or 4 s (2 s for 20% or 4 s for

225 50%) to ramp up to the designated maximum effort,

226 maintained the force for 10 s, and lastly used 2 or 4 s

227 (2 s for 20% or 4 s for 50%) to ramp down to zero

228 effort. For the sine-wave force, the subject took 2 or

229 4 s to ramp up to the designated effort, maintained for

230 5 s, then performed three repeated sine-wave oscilla-

231 tion forces either from 10 to 20% or from 25 to 50%,

232 and finally used 2 or 4 s ramp down to zero effort.

233 Three repeated trials were performed for each condi-

234 tion. The order of the different conditions were ran-

235 domized during the experiment. A minimum of 2 min

236 rest was provided between trials to reduce the influence

237 of muscle fatigue. A total of 48 trials (four fin-

238 gers 9 two force levels 9 two trajectories 9 three

239 repetitions) were recorded for each subject.

240 EMG Recordings

241 Surface EMG signals were recorded over the

242 extensor digitorum communis (EDC) muscle via an

243 8 9 16 channel high density (HD) EMG electrode ar-

244 ray (see Fig. 1c) with 10 mm inter-electrode distance

245 (OT Bioelettronica, Torino, Italy). Each electrode was

246filled with conductive gel to ensure high conductivity

247between the electrodes and the skin surface. Prior to

248the electrode placement, the skin surface was scrubbed

249with alcohol pad to improve the signal quality. To

250standardize the electrode placement location across

251subjects, the HD array was centered at the midway

252between the olecranon process and the styloid process.

253The EMG signals were acquired from EMG_USB2+

254(OT Bioelettronica, Torino, Italy), sampled at 2048 Hz

255with a gain of 1000 and a bandwidth of 10–900 Hz.

256EMG Analysis

257Channel selection was first performed based on

258several previous studies.7,18,28 EMG activities from

259only a localized region of the extensor muscle can be

260obtained during individual finger extension. Therefore,

261a large number of the EMG channels contained pri-

262marily baseline noise. To remove the channels with no

263EMG information, only the EMG channels (64 out of

264128 channels) in proximity to the regions with EMG

265activities for individual fingers were selected for the

266analyses. Specifically, based on the muscle activation

267region of individual fingers, channels from row 1–8, 5–

26812, and 9–16 were used for index, middle, ring, and

269little fingers, respectively.

270EMG Decomposition and Neural-Drive-Based Estima-

271tion

272The EMG signals were decomposed into individual

273MU discharge spike trains using the fast independent

274component analysis (FastICA) method.19 Briefly, the

275algorithm includes four steps: (1) signal extension by

276adding eight delayed replicas of each original channel,24

277(2) signal whitening, (3) FastICA-based deconvolution,

278and (4) action potential timing identification through

279peak detection and clustering. The details of the algo-

280rithmand the parameter selection are described byNegro

281et al.24 Lastly, the silhouette measure (SIL), a separation

282index inclusteranalysis,wasused tofilter thoseMUswith

283low accuracy. Previous literature has shown that the

284decomposed MUs with larger SILs tended to have a

285higher accuracy.24 In the current study, MUs with SIL

286larger than 0.8 were accepted for further analysis.

287During decomposition, the algorithm can converge

288to the same source signal multiple times. To remove

289the duplicated MUs, a post-processing step was per-

290formed. If a pair of MU firing trains had > 50% of

291synchronized firing events within ± 1 ms after adjust-

292ing the time delay, the MU firing train with a lower

293SIL was removed. All the processed firing events trains

294were then pooled into a composite train, and the dis-

295charge rate of the composite firing event train was

296calculated (Fig. 1c).



DAI et al.

Au

tho

r P

ro

of

UNCORRECTEDPROOF

297 EMG-Based Estimation

298 The EMG-amplitude-based feature is widely used

299 for EMG-to-force prediction.23 The root-mean-square

300 of EMG has also been verified as the best EMG-am-

301 plitude feature especially at high effort levels (> 25%

302 MVC).5 Therefore, the root-mean-square of EMG was

303 used to estimate the force for experimental data as a

304 control condition compared with the neural-drive-

305 based estimation. Raw EMG signals were band-pass

306 filtered to reduce the influence of motion artifacts and

307 high frequency background noise (4th Butterworth

308 with a cut-off frequency at 50 and 500 Hz),23 and

309 notch filtered to eliminate the power-line interference

310 (2nd IIR filter at 60 Hz with a bandwidth of 1 Hz).

311 The root-mean-square values of all the 8 9 8 pro-

312 cessed EMG signals used for decomposition were cal-

313 culated, and were then averaged across all the

314 channels.

315 A 350 ms moving window with a moving step of

316 50 ms was used in both the neural-drive-based and

317 EMG-based approaches. The force prediction was

318 performed using a linear regression method. Specifi-

319 cally, a single trial from trapezoid condition was used

320 to calculate the regression coefficients from the neural-

321 drive-based or EMG-based estimate for each force le-

322 vel. The remaining trials were then used to evaluate the

323 force prediction. A cross-validation were also per-

324 formed such that any single trial was used for the

325 regression coefficient calculation, and the remaining

326 trials were used for evaluation/testing. Only the eval-

327 uation results were reported. The performance of the

328 two estimation approaches was evaluated by calculat-

329 ing the R2 values between the estimated values (EMG-

330 based or neural-drive-based approaches) and the ac-

331 tual force output. Prior to the R2 calculation, the time

332 series of the neural-drive-based or EMG-based esti-

333 mations were adjusted with the force recordings to

334 account for the neural-mechanical delays, using a

335 cross-correlation analysis.8

336Statistical Analysis

337Standard errors were used for all the figures, and

338standard deviations of the data were used for all other

339data analyses. The performance differences were tested

340using a repeated measures analysis of variance (AN-

341OVA) in SPSS 24 (IBM). Since the R2 values were

342bounded at 1, an arcsine-square-root transformation

343was performed to satisfy the normal distribution

344assumption of the ANOVA and regression analysis.

345For the simulated data, the performance of the neural-

346drive-based approach was examined on three factors

347[force level 9 trajectory 9 SNR]. For the experimental

348data, the evaluation was examined on four factors

349[force level 9 trajectory 9 finger 9 estimation meth-

350od]. Post hoc pairwise comparisons with Bonferroni

351corrections were conducted when a significance was

352found. A significance level of p < 0.05 was used.

353RESULTS

354Simulation Results

355The overall decomposition results (both accuracy

356and number of MUs detected) are shown in Table 3.

357Overall, the decomposition accuracy and yield revealed

358similar pattern across force trajectory. As the force

359level increased and the SNR decreased, the decompo-

360sition accuracy decreased from 97.33 ± 0.27% (the

361least challenging condition: 20% trapezoid shape at 20

362SNR) to 68.90 ± 0.76% (the most challenging condi-

363tion: 50% trapezoid shape at 5 SNR). Similarly, the

364corresponding number of MUs detected also dropped

365from 19.00 ± 0.24 to 11.92 ± 0.21.

366After decomposition, discharge events from all the

367MUs with SIL > 0.8 were pooled to estimate the

368neural drive. The sample time-series plots in Figs. 2a

369and 2b show that the estimated neural drive can largely

370match the variation of the simulated muscle force. The

371overall estimation results are summarized in Fig. 3.

372The R2 values varied from 0.90 to 0.96 across different

TABLE 3. The overall results of decomposition.

5 SNR 10 SNR 20 SNR

Sine 20% 78.90 ± 0.82% 88.32 ± 0.60% 95.32 ± 0.37%

(15.14 ± 0.21) (15.60 ± 0.19) (18.90 ± 0.24)

Sine 50% 71.49 ± 0.83% 81.26 ± 0.64% 92.58 ± 0.43%

(11.92 ± 0.21) (16.02 ± 0.20) (18.86 ± 0.27)

Trapezoid 20% 79.06 ± 0.84% 89.25 ± 0.67% 97.27 ± 0.27%

(15.36 ± 0.23) (15.66 ± 0.19) (19.00 ± 0.24)

Trapezoid 50% 68.73 ± 0.78% 79.61 ± 0.63% 92.18 ± 0.40%

(12.40 ± 0.27) (16.06 ± 0.20) (18.80 ± 0.21)

The accuracy (mean ± standard error) are shown. The numbers of motor units (MUs) detected are shown in brackets.




Au

tho

r P

ro

of

UNCORRECTEDPROOF

373 conditions. A three-way repeated measures ANOVA

374 [force level 9 trajectory 9 SNR] showed that all the

375 three main factors (p < 0.05) had significant differ-

376 ences. Further post hoc comparisons revealed that the

377 estimate became less accurate as the SNR decreased or

378 the force level increased, largely due to a decline of the

379 decomposition accuracy. In addition, the sine-wave

380 trajectory and lower force level showed higher R2

381 values.

382 To quantify the sensitivity of the R2 to the decom-

383 position performance, Fig. 4 illustrates the relation

384 between the R2 of the drive estimation and the

385 decomposition performance (the accuracy and the

386 number of MUs detected) after pooling all 600 simu-

387lated trials. The results indicated that a higher per-

388formance in both the accuracy and yield led to a higher

389R2 value. The decomposition accuracy seemed to be

390more critical in that a small number of MUs can lead

391to an accurate neural drive estimation, as long as the

392discharge timings were accurate. In contrast, a large

393number of inaccurate discharge timings can still lead to

394relatively low R2 values.

395Experimental Results

396A linear regression method was used to predict the

397muscle force. Figure 5 shows two representative

398examples of the time-series of EMG and neural drive,

FIGURE 2. Example time-series plots of two different drive shapes for simulated data. (a) Illustration of the detaileddecomposition results. The red trace illustrates one channel of EMG with the largest root-mean-square value. Each blue barrepresents one discharge event. (b) Sine wave. (c) Trapezoid.



DAI et al.

Au

tho

r P

ro

of

UNCORRECTEDPROOF

399 in comparison with the force output. The neural-drive-

400 based approach for both examples revealed better

401 performance compared with the EMG-based estima-

402 tion. The relation between the SNR and the R2 was

403 examined on both estimations across all the trials. A

404 robust linear regression was performed to quantify the

405 strength of the correlation (Fig. 6). Based on the

406 regression results, the R2 of the EMG-based approach

407 tended to decline as SNR increased (R = 0.32,

408 p < 0.05). However, the neural-drive-based approach

409 showed a weaker association with SNR (R = 0.19,

410 p < 0.05).

411 Four separate two-way (finger 9 estimation method)

412 repeated measures ANOVAs were performed on each

413 condition (each subplot in Figs. 7a–7d). The detailed

414 statistical outcomes are summarized in Table 4. Most

415 of the results showed that the estimation method was

416 significant, except for trapezoid 50% condition

417 (p > 0.05). The values of power analysis for the sig-

418nificant factors of the four two-way ANOVAs ranged

419from 0.80 to 0.95. Post hoc comparisons found that the

420neural-drive-based approach always had higher R2

421values than the EMG-based method for all the fingers.

422Since no significant difference was observed in the

423finger factor and no interaction was found, the mar-

424ginal mean values were calculated by averaging the R2

425values of four fingers for other different conditions

426(trajectory, force level, and estimation method) as

427shown in Fig. 7e. The R2 values ranged from 0.83 to

4280.92 for the neural-drive-based estimation, and from

4290.73 to 0.88 for the EMG-based estimation.

430A three-way (trajectory 9 force level 9 estimation

431method) repeated measures ANOVA was performed on

432R2 values. The results showed a significant difference for

433trajectory [F(1,7) = 14.355, p < 0.05], and a significant

434interaction [F(1,7) = 19.287, p < 0.05] between esti-

435mation method and force level. The values of power

436analysis of the three factors (trajectory, force level, and

437estimation method) are 0.92, 0.88, and 1, respectively.

438The sine-wave trajectory showed higher R2 values than

439the trapezoid trajectory (p < 0.05). Further Post hoc

440analysis were tested between the two estimation methods

441and the twodifferent force levels. For each force level, the

442neural-drive-based approach always had higher R2 val-

443ues than the EMG estimate (p < 0.05). For the EMG-

444based estimation method, 50% force level had higher R2

445values than the 20% force level (p < 0.05).

446DISCUSSION

447In this study, the feasibility of using motoneuron

448discharge events to estimate the individual finger forces

449was investigated. The simulation results showed that

450the decomposition accuracy was sensitive to the signal

451quality (SNR) and the number of active MUs in the

452signals (associated with the level of excitation), and

453that the decomposition yield was only sensitive to the

454signal quality. The experimental results showed that

455the neural-drive-based approach was consistently bet-

456ter than the EMG-based approach in estimating mus-

457cle forces across different conditions (individual finger,

458force level, and force trajectory). Overall, the superior

459performance of the neural-drive-based estimation of

460individual finger forces offers a promising neural

461interface signal for intuitive and robust control of

462rehabilitative/assistive techniques that can help restore

463individual finger movement.

464Implications of Simulation Results

465Given that the ground-truth of the neural drive

466(input) and MU firings (output) are known in the

467simulation, a direct evaluation on the performance of

FIGURE 3. The R2 values of the simulation results indifferent conditions. The bars represent the standard errors.The asteroids represent significant differences (p < 0.05).

FIGURE 4. The influence of the decomposition performance(accuracy and the number of MUs detected) on the R2 of theneural-drive-based estimation. The color bar shows thevalues of R2s. Note: the map was interpolated linearly threetimes just for visual presentation.




Au

tho

r P

ro

of

UNCORRECTEDPROOF

468 the neural-drive-based estimation in relation to the

469 performance of the MU decomposition was per-

470 formed. The descending neural drive can be estimated

471 from the discharge frequency/probability of the MU

472 pool. In order to reliably estimate the neural drive,

473 accurate discharge timings from a certain number of

474 MUs sampled from the MU pool are required to reflect

475 the pool behavior. Based on the simulation results,

476 both the accuracy and yield of the decomposition in-

477 deed affect the neural-drive-based estimation, in that a

478 lower decomposition accuracy and yield can lead to

479 inaccurate force estimations. However, the degree of

480influence of the decomposition accuracy and yield on

481neural drive estimation differs, with decomposition

482accuracy being the more sensitive variable. Specifically,

483if the decomposition accuracy is above approximately

48485%, a large range of decomposition yield (as low as 8

485MUs) can lead to an accurate estimation of the neural

486drive. In contrast, if the decomposition accuracy is low

487(below 80%), a large number of MUs (~ 15) would be

488needed to derive an accurate neural drive.

489On the other hand, the performance showed

490improvement as the number of MUs increased. This

491effect can arise from several factors associated with the

FIGURE 5. Example time-series plots of two different drive trajectories for experimental data. (a) Illustration of the detaileddecomposition results. The red trace illustrates one channel of EMG with the largest root-mean-square value, and thecorresponding waveforms of motor unit action potentials (MUAPs). Each blue bar represents one discharge event. SIL is thesilhouette measure. (b) Sine wave. (c) Trapezoid.



DAI et al.

Au

tho

r P

ro

of

UNCORRECTEDPROOF

492nature of the decomposition errors. First, the neural

493drive estimate is based on the population firing

494behavior of composite spike trains from multiple MUs.

495Even though the timing of discharge in individual MUs

496may be inaccurate due to an incorrect random shift in

497discharge timings, a comparable number of false pos-

498itive and false negative errors from individual MUs can

499be averaged out in the average window. Second, the

500decomposition error can also arise from incorrect

501assignment of the discharge events. For instance, the

502spike of a MU is incorrectly assigned to a different

503MU, which would lead to two errors. However, the

504errors are cancelled out in the composite spike train,

505and would not affect the neural drive estimations.

506Nevertheless, to ensure an accurate estimation of the

507neural drive, a high decomposition accuracy is neces-

508sary and can provide confidence on the neural drive

509estimations.

FIGURE 6. The relation between SNR and R2 for bothapproaches. Each circle represents one individual trial.

FIGURE 7. (a)–(d) The mean values of the two estimates in different conditions. (e) The marginal mean values for the twoestimates on different conditions after averaging across different fingers. The bars represent the standard errors. The asteroidsrepresent significant differences (p < 0.05). The average SNR of the EMG signals in each condition are also shown.

TABLE 4. The summary of four two-way (finger 3 estimation method) repeated measures ANOVAs.

Condition Finger Estimation method

Sine 20% F(3,21) = 0.095, p = 0.962 F(1,7) = 12.590, p= 0.009

Sine 50% F(3,21) = 1.777, p = 0.182 F(1,7) = 10.677, p= 0.014

Trapezoid 20% F(3,21) = 2.837, p = 0.063 F(1,7) = 17.733, p= 0.004

Trapezoid 50% F(3,21) = 2.307, p = 0.106 F(1,7) = 2.823, p = 0.137

The values are marked in bold when significance was found.




Au

tho

r P

ro

of

UNCORRECTEDPROOF

510 Consistent with earlier studies,4,24 the simulation

511 results also showed that the SNR of the EMG signals

512 and the force level (i.e., the number of MUs in the

513 EMG signals) can affect the decomposition results. A

514 large number of MUs can increase the degree of

515 superposition and pose challenges to the source sepa-

516 ration. The simulation results can provide a reference

517 for the decomposition of experimental data. Namely, a

518 high SNR and a low-to-moderate effort is desirable in

519 order to ensure accurate decomposition results.

520 Implications of Experimental Results

521 The experimental results show that the neural-drive-

522 based approach was consistently superior to the EMG-

523 based approach for force estimation. The neural-drive-

524 based approach is not affected by different MUAP

525 features, such as amplitude, duration, or cancellation

526 due to superposition, as in the EMG-based approach.

527 For example, at low contraction levels, the EMG sig-

528 nals tend to be sparse, and the EMG-based approach

529 has large estimation biases, which could be even more

530 prominent in clinical populations, because the EMG

531 signals are often sparse and unstable due to disordered

532 control of MU activations.14,17

533 In addition, external factors, including high baseline

534 noise or variations in electrode–skin contact, can fur-

535 ther decrease the reliability of the EMG-based

536 approach. Although the simulation results show that

537 SNR can affect the estimation performance of the

538 neural-drive-based approach to some degree, the

539 EMG-based approach is more sensitive to SNR with

540 high estimation errors at low SNR levels (Fig. 5).

541 These findings indicate that the neural-drive-based

542 approach is more robust to low signal quality, com-

543 pared with the EMG-based estimates. The less strin-

544 gent requirement on the signal quality of the neural-

545 drive-based approach can facilitate wide clinical

546 applications for human–machine interactions, because

547 the quality of the signals obtained from clinical pop-

548 ulations tend to vary depending on pathology.26

549 The force estimation showed comparable perfor-

550 mance across individual fingers. These outcomes have

551 critical implications for clinical applications involving

552 human–machine interactions. First, assistive/rehabili-

553 tation strategies focusing on fine control of individual

554 finger movement are still a challenge in the field. The

555 ability to estimate individual finger forces reliably

556 shown in this study can help improve the performance

557 of human–machine interactions involving fine motor

558 control of finger movement. Together with the devel-

559 opment of advanced control frameworks, this

560 approach can help better utilize the high degrees of

561 freedom in exoskeleton or prosthetic hand, and further

562 improve the functional outcomes for individuals with

563neuromuscular disorders. Second, the robust perfor-

564mance across different task conditions involving steady

565grip or dynamic force variations can also facilitate

566applications during daily activities. However, the

567varying forces were still produced in isometric condi-

568tions. Additional studies involving dynamic finger

569movements are needed to further evaluate the perfor-

570mance in different dynamic movements with muscle

571fibers shifting substantially beneath the recording

572electrodes. Nonetheless, these findings suggest that the

573neural-drive-based estimation on individual finger

574force can be a promising approach for robust control

575of hand exoskeletons, prosthetic hands, or neuro-

576prostheses, which can help restore individual finger

577control and eventually could facilitate the utility of

578advanced devices in disabled individuals.

579The signal conditions may affect the EMG- and

580neural-drive- based approaches in different ways. For

581the neural-drive-based approach, no significant differ-

582ence was found across the two force levels. Although

583the signals from the 50% force level had a higher SNR

584than that from the 20% level, more superposition from

585more MUs is expected at higher forces. These two

586contrasting factors can balance out the effect on the

587decomposition performance. On the other hand, the

588EMG-based estimate showed a better performance at

589higher force level. The EMG envelop tended to be

590smoother at higher forces, which decreased the varia-

591tion of EMG-based estimate. In addition, the trajec-

592tory also influenced the two estimations, largely

593because the small variation of force estimate during the

594steady contraction can potentially decrease the R2

595values. The overall R2 of the neural-drive-based

596approach on simulation results exceeded the values

597shown in the experimental data. Because all the char-

598acteristics of EMG signals cannot be fully captured in

599the simulation, which can lead to higher decomposi-

600tion errors. Specifically, the EMG signals obtained

601from experiments may have sporadic action potential

602variations in amplitude and/or duration, and external

603factors (shift of electrode locations or motion artifacts)

604can also alter the signal properties.

605Limitations and Future Work

606The current study has several limitations. First, the

607subjects were instructed to only extend the designated

608finger in each trial. However, the subjects may still

609inevitably perform co-contractions due to finger

610enslaving,30 especially for the ring finger extension.

611The EMG signals from muscle co-contraction can

612potentially bias the force estimations. Second, the fin-

613ger activations were not classified. However, previous

614works6,18 have showed that the muscle activation of

615individual finger movement was localized and distin-



DAI et al.

Au

tho

r P

ro

of

UNCORRECTEDPROOF

616 guishable for both finger extension or flexion using

617 high-density EMG grid. The EMG activities of indi-

618 vidual finger contraction can be separated using pat-

619 tern recognition techniques. Third, the linear

620 regression was performed using the steady contraction

621 trial within each force level, but still exhibited a good

622 performance when tested on the varying force trials.

623 Since the mean firing rate of the decomposed MU pool

624 at each force level is not linearly related with the force

625 level, the regression was performed at individual force

626 levels. Future work will investigate whether the

627 regression performed on a single force level (e.g., close

628 to the maximum force) can be generalized across dif-

629 ferent force levels. Fourth, this study does not allow us

630 to dissect the specific levels of the central nervous

631 system in contribution to the neural drive estimations.

632 For example, different cortical regions, cerebellum,

633 basal ganglia, and brainstem can all play a role for the

634 neural drive estimation. Lastly, the force estimation is

635 performed through post-processing, and the compu-

636 tation load is a factor during real-time estimation. The

637 EMG-based approach is suitable for real-time esti-

638 mation. The neural-drive-based approach, however,

639 requires more computation time, largely at the EMG

640 decomposition step. Real-time decomposition has been

641 investigated previously.15 The strategy of online neu-

642 ral-drive-based approach is to obtain the separation

643 matrix for the extraction of neural drive information

644 using an initial batch (approximately 5-10 s). Then, the

645 separation matrix can be multiplied to the incoming

646 data to obtain the neural drive information in real-

647 time. The separation matrix will need to be updated

648 periodically during long-term use.

649 CONCLUSIONS

650 The current study shows that the neural-drive-based

651 approach out-performances the EMG-based approach

652 in predicting individual finger forces under different

653 scenarios. The boundary conditions for reliable neural

654 drive estimation were also provided. These findings can

655 potentially provide an intuitive and robust neural

656 interface for future studies that investigate the appli-

657 cations of rehabilitative or assistive devices for clinical

658 populations, with a goal to promote their indepen-

659 dence of living and enhance their quality of life. Future

660 work focusing on the real-time implementation of this

661 approach is needed for future clinical translation.

662

663

664 CONFLICT OF INTEREST

665 The authors have no financial relationships that

666 may cause a conflict of interest.

667REFERENCES

6681Al-Timemy, A. H., G. Bugmann, J. Escudero, and N.669Outram. Classification of finger movements for the dex-670terous hand prosthesis control with surface electromyog-671raphy. IEEE J. Biomed. Health. Inform. 17:608–618, 2013.6722Boretius, T., J. Badia, A. Pascual-Font, M. Schuettler, X.673Navarro, K. Yoshida, and T. Stieglitz. A transverse674intrafascicular multichannel electrode (TIME) to interface675with the peripheral nerve. Biosens. Bioelectron. 26:62–69,6762010.6773Callier, T., E. W. Schluter, G. A. Tabot, L. E. Miller, F. V.678Tenore, and S. J. Bensmaia. Long-term stability of sensi-679tivity to intracortical microstimulation of somatosensory680cortex. J. Neural Eng. 12:56010, 2015.6814Chen, M., and P. Zhou. A novel framework based on682FastICA for high density surface EMG decomposition.683IEEE Trans. Neural Syst. Rehabil. Eng. 24:117–127, 2016.6845Clancy, E. A., and N. Hogan. Probability density of the685surface electromyogram and its relation to amplitude686detectors. IEEE Trans. Biomed. Eng. 46:730–739, 1999.6876Dai, C., and X. Hu. Extracting and classifying spatial688muscle activation patterns in forearm flexor muscles using689high-density electromyogram recordings. Int. J. Neural690Syst. 29:1850025, 2019.6917Dai, C., H. Shin, B. Davis, and X. Hu. Origins of common692neural inputs to different compartments of the extensor693digitorum communis muscle. Sci. Rep. 7:13960, 2017.6948Dai, C., Y. Zheng, and X. Hu. Estimation of muscle force695based on neural drive in a hemispheric stroke survivor.696Front. Neurol. 9:187, 2018.6979Davoodi, R., C. Urata, M. Hauschild, M. Khachani, and698G. E. Loeb. Model-based development of neural prosthe-699ses for movement. IEEE Trans. Biomed. Eng. 54:1909–7001918, 2007.70110De Luca, C. J., and R. Merletti. Surface myoelectric signal702cross-talk among muscles of the leg. Electroencephalogr.703Clin. Neurophysiol. 69:568–575, 1988.70411Farina, D., L. Mesin, S. Martina, and R. Merletti. A sur-705face EMG generation model with multilayer cylindrical706description of the volume conductor. IEEE Trans. Biomed.707Eng. 51:415–426, 2004.70812Farina, D., I. Vujaklija, M. Sartori, T. Kapelner, F. Negro,709N. Jiang, K. Bergmeister, A. Andalib, J. Principe, and O.710C. Aszmann. Man/machine interface based on the dis-711charge timings of spinal motor neurons after targeted712muscle reinnervation. Nat. Biomed. Eng. 1:25, 2017.71313Fuglevand, A. J., D. A. Winter, and A. E. Patla. Models of714recruitment and rate coding organization in motor-unit715pools. J. Neurophysiol. 70:2470–2488, 1993.71614Gemperline, J. J., S. Allen, D. Walk, and W. Z. Rymer.717Characteristics of motor unit discharge in subjects with718hemiparesis. Muscle Nerve 18:1101–1114, 1995.71915Glaser, V., A. Holobar, and D. Zazula. Real-time motor720unit identification from high-density surface EMG. IEEE721Trans. Neural Syst. Rehabil. Eng. 21:949–958, 2013.72216Hu, X., W. Z. Rymer, and N. L. Suresh. Reliability of spike723triggered averaging of the surface electromyogram for724motor unit action potential estimation. Muscle Nerve72548:557–570, 2013.72617Hu, X., A. K. Suresh, W. Z. Rymer, and N. L. Suresh.727Altered motor unit discharge patterns in paretic muscles of728stroke survivors assessed using surface electromyography.729J. Neural Eng. 13:46025, 2016.




Au

tho

r P

ro

of

UNCORRECTEDPROOF

730 18Hu, X., N. L. Suresh, C. Xue, and W. Z. Rymer. Extracting731 extensor digitorum communis activation patterns using732 high-density surface electromyography. Front. Physiol.733 6:279, 2015.734 19Hyvarinen, A., and E. Oja. Independent component anal-735 ysis: algorithms and applications. Neural Netw. 13:411–430,736 2000.737 20Keenan, K. G., D. Farina, K. S. Maluf, R. Merletti, and R.738 M. Enoka. Influence of amplitude cancellation on the739 simulated surface electromyogram. J. Appl. Physiol.740 98:120–131, 2005.741 21Kuiken, T. A., G. A. Dumanian, R. D. Lipschutz, L. A.742 Miller, and K. A. Stubblefield. The use of targeted muscle743 reinnervation for improved myoelectric prosthesis control744 in a bilateral shoulder disarticulation amputee. Prosthet.745 Orthot. Int. 28:245–253, 2004.746 22LeFever, R. S., A. P. Xenakis, and C. J. De Luca. A pro-747 cedure for decomposing the myoelectric signal into its748 constituent action potentials-part II: execution and test for749 accuracy. IEEE Trans. Biomed. Eng. 29:158–164, 1982.750 23Merletti, R., and P. Di Torino. Standards for reporting751 EMG data. J Electromyogr. Kinesiol. 9:3–4, 1999.752 24Negro, F., S. Muceli, A. M. Castronovo, A. Holobar, and753 D. Farina. Multi-channel intramuscular and surface EMG754 decomposition by convolutive blind source separation. J.755 Neural Eng. 13:26027, 2016.756 25Richard, P. D., R. E. Gander, P. A. Parker, and R. N.757 Scott. Multistate myoelectric control: the feasibility of 5-758 state control. J. Rehabil. R&D 20:84–86, 1983.

75926Santello, M., and C. E. Lang. Are movement disorders and760sensorimotor injuries pathologic synergies? When normal761multi-joint movement synergies become pathologic. Front.762Hum. Neurosci. 8:1050, 2015.76327Thompson, C. K., F. Negro, M. D. Johnson, M. R.764Holmes, L. M. McPherson, R. K. Powers, D. Farina, and765C. J. Heckman. Robust and accurate decoding of766motoneuron behaviour and prediction of the resulting force767output. J. Physiol. 596:2643–2659, 2018.76828van Beek, N., D. F. Stegeman, J. C. Van Den Noort, D. H.769E. J. Veeger, and H. Maas. Activity patterns of extrinsic770finger flexors and extensors during movements of instructed771and non-instructed fingers. J. Electromyogr. Kinesiol.77238:187–196, 2018.77329Yao, W., R. J. Fuglevand, and R. M. Enoka. Motor-unit774synchronization increases EMG amplitude and decreases775force steadiness of simulated contractions. J. Neurophysiol.77683:441–452, 2000.77730Zatsiorsky, V. M., Z.-M. Li, and M. L. Latash. Enslaving778effects in multi-finger force production. Exp. Brain Res.779131:187–195, 2000.

Publisher’s Note Springer Nature remains neutral with re-

gard to jurisdictional claims in published maps and institu-

tional affiliations.



DAI et al.

Au

tho

r P

ro

of

Documents

Author Proof CTED UNCORRE - University of North Carolina