2.1.

Subjects and CIMT

Six subjects with hemiplegic CP (two female and four male; $10.2\pm 2.1$ years old) were included in this study. Two subjects (Subject 4 and Subject 8) were right hemiparetic and the other four subjects were left hemiparetic. Two subjects (Subject 1 and Subject 4) were classified as MACS 1 and the other four subjects (Subject 3, 6, 7, and 8) as MACS 2. Subjects with light impairment of their affected hand that was often hardly noticeable were classified as MACS 1. Subjects classified as MACS 2 had impaired use of their affected hand, but could perform life daily activities adequately through compensatory use of their unaffected hand. Five healthy children (two female and three male; $9.8\pm 1.3$ years old) were also included as controls. All the controls were right handed. All children with CP included in this study had a successful MRI scan that identified a single subcortical/cortical lesion affecting their motor area in one of the two cerebral hemispheres. In addition, Subject 3 had undergone a right functional peri-insular hemispherectomy due to intractable epileptic seizures and had a physical shift of 1.3 cm to the right of the brain midline separating the two hemispheres, as verified by the anatomical MRI. The children with CP participated in a 2-week, pirate themed, CIMT camp that took place 5 days a week, 6 h a day, at the Texas Scottish Rite Hospital for Children in Dallas. During CIMT camp, the children took part in group and individual activities such as ball games and painting, focused on improving gross and fine motor skills and increasing independence with activities of daily living under the supervision of occupational therapists. The unaffected arm of each child was immobilized by a removable splint during camp hours, which forced the use of their affected arm during play activities. FNIRS measurements, MACS⁹ classification, and clinical assessments using the Melbourne assessment of upper limb function,¹² and AHA¹³ of bimanual limb function were performed on each child with CP before, immediately after, and 6 months after therapy. Only Subject 4 missed one fNIRS measurement at 6 months after therapy. The occupational therapists inquired about the children’s affected arm activity levels after therapy and although these varied between subjects no additional therapies or intensive manual training were performed by any subject in the 6 months after CIMT. Control subjects were also measured by fNIRS at the same time points, but no clinical assessments were performed on them. The study was approved by the University of Texas Southwestern Medical Center at Dallas (UTSW) Institutional Review Board protocol (IRB No. 042007-064).

2.2.

Experimental Setup and Protocol

A continuous-wave fNIRS brain imager (CW-6, Techen Inc., Milford, Massachusetts) was used to map the $\mathrm{\Delta }\mathrm{HbO}$ and $\mathrm{\Delta }\mathrm{Hb}$ induced by sensorimotor cortical activation during a finger-tapping task. As $\mathrm{\Delta }\mathrm{Hb}$ dynamics are highly correlated with those of $\mathrm{\Delta }\mathrm{HbO}$³² and have significantly lower amplitudes than the latter making them susceptible to cross talk from $\mathrm{\Delta }\mathrm{HbO}$³³ and to interference from physiological artifacts, we have focused on the analysis of $\mathrm{\Delta }\mathrm{HbO}$ dynamics only. The source-detector probe geometry of fNIRS is shown in Fig. 1. There were 16 sources and 32 detector fiber bundles to cover most of the sensorimotor cortex. The rows of sources and detectors were centered around the Cz position of the EEG International 10/20 system³⁴ and attached onto the subjects’ heads by perforated Velcro straps. Near-infrared light at wavelengths of 690 and 830 nm was delivered from the source fiber bundles simultaneously at each source location as the CW-6 system enables all laser sources to be on at the same time with distinct modulation frequencies (6.4 to 12.6 kHz, with an increment of 200 Hz). The back-reflected light was sampled at the frequency of 25 Hz. Each source had up to six detectors with 3 cm separation, and each detector received light from up to three sources, which resulted in a total of 84 source-detector pairs for each wavelength. This setup enabled monitoring cortical activation over an $11\mathrm{cm}\times 20\mathrm{cm}$ field of view. Additionally, eight short-distance (1.5 cm) source-detector pairs, indicated by yellow lines in Fig. 1, were used as references to subsequently filter out superficial background hemodynamics unrelated to activation (Sec. 2.3).

Fig. 1

The fNIRS measurement geometry (sources: blue squares; detectors: green circles) spanned a field of view covering the premotor cortex (PMC), supplementary motor area (SMA) and M1/S1 (primary motor/sensory cortex) areas indicated by black ovals. Yellow lines connect the short distance source-detector pairs sampling the scalp hemodynamics. Dark gray triangles indicate the EEG International 10/20 system Cz, C3, C4, F3, and F4 locations.

EEG International 10/20 system Cz, C3, C4, F3, and F4 anatomical measurements³⁴ were used as reference points to ensure the optical probe setup was placed over the sensorimotor cortex. The optical probe setup was centered on the Cz reference point and covered a large area of sensorimotor cortex such that the probe set extended a fixed distance of 5 cm anterior to the C3 and C4 positions to map the primary motor/sensory area (M1/S1) and was $0.21\pm 0.13\mathrm{cm}$ posterior of the measured F3 and F4 positions covering the supplementary motor area (SMA) and premotor cortex (PMC) of both hemispheres, as indicated in Fig. 1. Anatomical measurements were performed on the head of each subject at all three visits and the standard deviations of the distances between Cz and any one of C3, C4, F3, and F4, across all visits, did not exceed $\sim \pm 1.5\mathrm{mm}$.

After subjects were set up for fNIRS measurements they were asked to perform a finger-tapping task, once with their left hand and once with their right hand. Subjects were instructed to tap with four fingers in unison (excluding the thumb) and keep their wrists on the table, while the other hand was to remain at rest as much as possible. An animation made on Adobe flash (Adobe Systems Incorporated, San Jose, California) guided the subjects to tap at a frequency of 1 Hz. The experimental protocol for finger tapping by each hand started with a 3-min rest period, followed by eight 40-s blocks that started with 15 s of tapping followed by 25 s of rest.

2.3.

Signal Processing and Image Analysis

In addition to detecting hemodynamic changes due to evoked cortical activation, fNIRS is also sensitive to cerebral hemodynamic fluctuations of systemic origin, which can be caused by cardiac pulsation (0.8–2.0 Hz), respiration (0.1–0.33 Hz) and Mayer waves (0.1 Hz or lower).^35,36 The cortical hemodynamic response to the motion activation protocol can be found in the frequency range of 0.01–0.4 Hz, so a bandpass filter could only effectively remove cardiac pulsation, but not the respiration or Mayer waves that overlap in frequency with the activation-induced hemodynamic fluctuations. In order to filter out these physiological interferences, recent studies have used component analyses,^37,38 state-space estimation,^35,39 adaptive filtering,⁴⁰ linear regression,⁴¹ or some combination of these.^42,43

This study used a combination of bandpass filtering, adaptive filtering, and principal component analysis (PCA) to filter the fNIRS signals as we have previously reported.³¹ In brief, the first step was to use a 0.01–0.4 Hz bandpass filter to remove cardiac pulsation signals. Subsequently, global fluctuations due to respiration and Mayer waves were removed by a combination of an adaptive least-mean-square (LMS) filter⁴⁰ and PCA.³⁷ One short source-detector pair of the eight available was used as the adaptive filter noise reference for the removal of global hemodynamic fluctuations from the fNIRS signals.⁴⁰ The reference channel chosen had the least amount of motion artifacts of the eight, which translated to having either zero or one motion-generated spikes in the time-series reference data. The spike in the reference channel time-series data was manually removed and the missing data were interpolated by the spline cubic spline function in MATLAB® (Mathworks, Natick, Massachusetts). Moreover, any block of data with obvious motion artifacts in any tapping/rest period was manually excluded from further analysis.

Reconstruction and visualization of fNIRS images from the acquired reflectance data were performed by the open-source HomER software implemented in MATLAB.⁴⁴ In this software, activation images were reconstructed by the Tikhonov perturbation solution to the photon diffusion equation, which employs a regularized Moore-Penrose inversion scheme.^45,46 The resulting $\mathrm{\Delta }\mathrm{HbO}$ maps were reconstructed $20\times 21$ pixel images ($0.55\mathrm{cm}\times 0.95\mathrm{cm}$ for each pixel) for every 0.04 s time interval.

The determination of image pixels with activation was based on a two-step process. First, the image pixels with temporal hemodynamic patterns that correlated significantly with evoked cortical activation were identified by a general linear model (GLM) similar to previous fMRI studies.^47–49 A value of $p<0.0001$ based on Bonferroni’s correction was used as the threshold to identify pixels with activation. Second, a $k$-means clustering algorithm⁵⁰ was used to calculate the signal-to-noise ratio (SNR) per image pixel. This was done to exclude any pixels with low SNR that the GLM still found to significantly correlate with the model-based hemodynamic response function⁵¹ calculated for the finger-tapping protocol being used in this study. The algorithm considered the entire temporal responses for each pixel after GLM and divided the pixel values into three clusters: activation, baseline, or deactivation.³¹ The pixel values in the baseline cluster were regarded as noise. The SNR of activation for each pixel was then calculated as the mean signal amplitude of the activation cluster over the standard deviation of the hemodynamic fluctuation amplitudes of the noise cluster in that pixel. Pixels with $\mathrm{SNR}\ge 2$ were considered to be part of the regions of activation when the $\mathrm{\Delta }\mathrm{HbO}$ amplitude equaled or exceeded two standard deviations of the amplitude of background hemodynamic fluctuations. The selected SNR threshold corresponded to a probability of 95%, or more, for the fluctuation amplitudes of these pixels to be different from those due to background hemodynamics.

The sequence of postprocessing steps described above is shown in Fig. 2. More details can be found in a prior publication from our group.³¹

Fig. 2

Flowchart of postprocessing steps.

2.4.

FNIRS-Based Metrics Derived from Activation Images and Detector Level Resting-State Data

The laterality index and time-to-peak/duration metrics were derived from the analysis of activation images. The time series data from each subject were first analyzed to find all image pixels with significant activation, as above described in Sec. 2.3. Subsequently, the laterality index metric was calculated using Eq. (1) and the time-to-peak metric was computed for each pixel with activation and was then averaged across pixels. The resting-state functional connectivity between sensorimotor regions was computed from source-detector channel data that were first averaged within each region indicated in Fig. 1.

3.1.

Laterality Index

Activation in M1/S1 cortical regions only was used for the laterality index metric calculations following previously reported fMRI and fNIRS studies.^16,18,60 Figure 3 shows the mean and standard deviation of the laterality index for controls and children with CP. The three visits for the controls were averaged into a single group of values as the laterality index was not found to change significantly over time for these subjects ($p=0.84$). As shown in Fig. 3, controls had preferential activation in the contralateral sensorimotor cortex during finger tapping. In contrast, children with CP showed bilateral activation on average before therapy. At the second visit immediately after therapy, the laterality index for children with CP increased and was closer to that of controls. At 6 months after therapy only one child with CP retained a high laterality index while all others relapsed. As a result, the mean laterality index relapsed close to the values before treatment.

Fig. 3

Laterality index for controls tapping with their dominant hand and children with cerebral palsy (CP) tapping with their affected hand across the three visits (white: controls, gray; children with CP). Error bars indicate ± one standard deviation. Asterisks indicate statistically significant differences in laterality index ($p<0.01$).

A high statistical significance was found ($p=0.007$, $1-\beta =0.91$) for the mean value difference in laterality index between controls and children with CP before therapy. Moreover, the laterality index of children with CP immediately after therapy was significantly different ($p=0.006$, $1-\beta =0.98$) from that of before therapy. No significant difference was observed between 6 months after therapy and prior visits due to the larger standard deviation of the group’s laterality index 6 months after therapy.

3.2.

Resting-State Functional Connectivity

Changes in the resting-state functional connectivity between five cortical regions (Fig. 1) were explored as a metric of response to CIMT. Based on the aforementioned $\mathrm{SL}\ge 0.6$ criterion, the two MACS 1 subjects showed significant pair-wise connectivity between all sensorimotor regions that did not change between the three visits. The four children with CP that did not show connectivity changes, all MACS 2, were compared to the average connectivity pattern from five controls in this part of the study. The thickness of lines in Fig. 4 connecting any two sensorimotor regions represents the percentage of MACS 2 subjects for whom a given pair of activation regions was connected.

Fig. 4

Group-wise resting-state functional connectivity in controls (a) and children with CP at the three assessment time points: (b) before therapy, (c) immediately after therapy, (d) 6 months later after therapy. Black line thickness represents the percentage of subjects with significant connectivity between a given pair of sensorimotor centers.

The frequency of connections between any two regions for controls was no more than 60%. In contrast, the frequency of pair-wise connection between regions for MACS 2 subjects before therapy was much higher, even reaching 100% for some connections between the SMA and other cortical regions. Interestingly, immediately after therapy MACS 2 subjects had weaker connectivity frequencies, especially between the SMA and other cortical regions, which was similar to what was observed in controls. Therefore, these results imply a certain level of sensorimotor network normalization occurring immediately after CIMT. However, at the 6 months after therapy time point, the connectivity frequencies between almost all cortical regions reversed back toward values found before therapy.

3.3.

Activation Time-to-Peak/Duration

Figure 5 shows the mean and standard deviation of the time-to-peak/duration metric for controls and the three visits of children with CP. As was done for the laterality index analyses above, all pair-wise group comparisons between visits of children with CP for the time-to-peak/duration metric were performed by paired $t$ tests with one exception: The comparison of children with CP before therapy versus immediately after therapy. This latter case did not pass the normality test (Sec. 2.4.4) and the Wilcoxon signed-rank test was used instead. No significant difference was observed for controls between the three visits ($p=0.69$) and they were averaged into a single group. A significant difference was found in time-to-peak/duration between controls and children with CP before therapy ($p=0.001$, $1-\beta =0.97$). Most subjects with CP had a smaller time-to-peak/duration metric before therapy, which increased significantly immediately after therapy ($p=0.03$, $1-\beta =0.99$) and persisted at significantly higher values ($p=0.006$, $1-\beta =0.97$) 6 months after therapy. The change of time-to-peak/duration toward higher values indicates a normalization of temporal activation patterns in all five sensorimotor regions after CIMT that persisted even 6 months later.

Fig. 5

Time-to-peak/duration of activation for controls tapping with their dominant hand and children with CP tapping with their affected hand across the three visits (white: controls; gray: children with CP). Error bars indicate ± one standard deviation. Asterisks indicate statistically significant differences in time-to-peak/duration ($p<0.05$).

3.4.

Comparisons with Clinical Assessment Scores

3.4.1.

Changes in Melbourne and AHA scores between visits

The Melbourne and AHA scales were employed to evaluate the unimanual and bimanual performance, respectively, of children with CP before, immediately and 6 months after therapy. Figure 6 illustrates changes in Melbourne scores versus changes in AHA scores for children with CP between visits. The thresholds for clinically significant improvements in Melbourne and AHA scores were 14 and 5, respectively.^12,13 These threshold values are depicted in Fig. 6 as dashed lines.

Fig. 6

Change in clinical scores for six subjects with CP between visits (V1: before therapy, V2: immediately after therapy, V3: 6 months after therapy). (a) Change in scores between before and immediately after therapy (V2-V1). (b) Change in scores between before and 6 months after therapy (V3–V1). Thresholds for clinically significant improvements in Melbourne and AHA scores were 14 (vertical dashed lines) and 5 (horizontal dashed lines), respectively. Data for MACS1 subjects are inside ovals.

Changes in clinical scale scores varied greatly between immediately after therapy [Fig. 6(a)] and 6 months after therapy [Fig. 6(b)]. Nevertheless, all four MACS2 subjects improved significantly on Melbourne or AHA scores, or both, immediately after therapy and 6 months after therapy. On the contrary, the two MACS 1 subjects included in this study, Subjects 1 and 4 [circled with black ovals in Figs. 6(a) and 6(b)], did not show significant improvements on either assessment scale. As a result, the discussion below on the comparisons between clinical scale scores and fNIRS metrics focuses on the four MACS 2 subjects, who had significant changes in motor function after CIMT.

3.4.2.

Changes in Melbourne and AHA scores versus changes in laterality index

Figures 7(a) and 7(b) illustrate the comparison between the change in Melbourne scores and the change in laterality index between visits for the four MACS 2 subjects. It is seen that for all subjects expect for Subject 3, an increase in the laterality index representing a shift of cortical activation from the ipsilateral to the contralateral hemisphere was accompanied by significant improvements in unimanual ability immediately after therapy, as reflected by the increased Melbourne scores [Fig. 7(a)]. The fact that Subject 3 appears to be an outlier is not surprising given that he had previously had brain surgery due to intractable epileptic seizures (Sec. 2.1). The laterality index change of 2 for this subject indicates a complete switch of activation between hemispheres toward a more normal laterality pattern that, however, did not result in significant motor gains [Fig. 7(a)]. At 6 months after CIMT [Fig. 7(b)] Subject 6 had relapsed but all other subjects maintained and even slightly improved unimanual ability compared to before therapy. Even though there are not enough subjects to perform formal statistical analyses, one can see a positive trend between the laterality index and Melbourne score changes in Fig. 7(b).

Fig. 7

Change in clinical scores versus change in laterality index for MACS 2 subjects between visits (V1: before therapy, V2: immediately after therapy, V3: 6 months later after therapy). Change in Melbourne scores versus change in laterality index immediately after therapy (a) and (b) 6 months after therapy. Corresponding changes in AHA scores immediately after therapy (c) and (d) 6 months after therapy.

In contrast to Melbourne score changes, an overall negative trend between laterality index and AHA score changes could be discerned both immediately after therapy [Fig. 7(c)] and 6 months after therapy [Fig. 7(d)]. An exception to this trend was Subject 8 who had improvements in both Melbourne and AHA scores that were maintained 6 months later, but at the cost of a significant relapse in laterality index 6 months after therapy that was even more abnormal than the values before therapy.

Another noteworthy observation was that Subject 7 was the only one in the MACS 2 group who retained significant improvements in unilateral motor function, as assessed by the Melbourne scale, while also maintaining a normalized laterality index [Fig. 7(b)]. Interestingly, Subject 7 also had the highest average connectivity strength between sensorimotor centers ($\mathrm{SL}=0.70$) compared to the MACS2 group average ($\mathrm{SL}=0.64$). Nevertheless, the connectivity pattern for Subject 7 was not different from that of the group average shown in Fig. 4(d).

Overall, the above results showed that each child had a unique way of responding to CIMT and attained some motor function improvements by increasing the involvement of the ipsilateral sensorimotor cortex (Subject 8), increasing the connectivity strength between sensorimotor centers (Subject 7), or by some combination thereof (Subjects 3 and 6). A positive trend between increases in Melbourne scores and increases in time-to-peak/duration, as well as a negative trend between increases in AHA scores and increases in time-to-peak/duration were also noticed. All results are compiled in Table 1.

Table 1

Clinical scores and functional near-infrared spectroscopy (fNIRS) metrics of subjects with cerebral palsy (CP) and controls before, immediately after and 6 months after therapy (V1: before therapy, V2: immediately after therapy, V3: 6 months after therapy).

Subject	Clinical scores						fNIRS metrics
	Melbourne			AHA			Laterality index			Time-to-peak/duration
	V1	V2	V3	V1	V2	V3	V1	V2	V3	V1	V2	V3
S1	110	111	113	60	62	62	0.5	1	0.4	0.92	1.15	1.13
S3	74	76	81	34	41	39	$-1$	0.9	$-0.6$	0.62	0.81	0.86
S4	117	117	117	84	84	84	0	1	—	0.80	1.22	—
S6	45	59	47	47	48	52	0.1	0.9	$-0.6$	0.62	0.85	0.81
S7	54	69	69	64	68	65	$-0.1$	0.6	0.9	0.52	0.73	0.86
S8	58	72	74	44	52	52	0	0.6	$-0.8$	0.56	0.71	1.04
Controls (Averaged)	—	—	—	—	—	—	0.69	0.71	0.69	0.97	0.99	1.00