2.2.1.

Experimental setup

Each participant was seated on a chair facing an LCD monitor. The distance between the participant’s eyes and the monitor was $\sim 70\mathrm{cm}$. All visual stimuli presented on the monitor were programmed in MATLAB (MathWorks, Natick, Massachusetts) using Cogent Toolbox software (University College London, London, United Kingdom²⁶). Healthy participants performed the reaching movements using their right hand (task hand). In the stroke group, the unaffected hand was used as the task hand for two patients with moderate hemiplegia (P07 and P10: individual FMA scores are shown in Table 2). Other stroke patients performed the reaching movements with their affected hand. Participants were asked to hold a digitizing pen on a drawing tablet (Intuos4 PTK-1240/K0, Wacom, Japan) with their task hand. A vibration motor attached to the index fingertip of the task hand was used to present tactile stimuli. As shown in Fig. 1(a), the hand position was digitized using the drawing tablet and the monitor displayed a hand cursor (filled circle) as real-time visual feedback of hand movement. The hand cursor position was recorded using the Cogent Toolbox with sampling at 60 Hz. The monitor also displayed two markers indicating the start and target positions for the reaching movements (open circles). The distance between the start and target markers was 12 cm on the monitor. The ratio of the actual hand movement distance to the hand cursor movement distance was defined as the visual cursor gain (VG). In this classification task, the hand cursor moved 1.2 cm on the monitor for a 1.0-cm hand movement ($\mathrm{VG}=1.2$); thus, the desired amplitude of the reaching movements on the drawing tablet was 10 cm.

Fig. 1

Experimental design. (a) Experimental setup. During performing the classification tasks, the participant’s task hand was occluded by a small rack. Thus, the participant could not see their hand while they performed the experimental tasks. The positions of the start and target markers were the same in all trials. (b) Time sequence of one trial. This case indicates the EF condition with visual start cue.

2.2.2.

Procedure

Figure 1(b) shows the time sequence of a trial. Before starting each trial, the monitor displayed written instructions to remind participants of the attentional strategy to use during the reaching movements. First, participants moved the hand cursor to the start marker when it appeared on the monitor. After a random delay period (1 to 2 s), a start cue was presented that was either a tactile stimulus delivered by the vibration motor or a brief color change of the hand cursor. The stimulus duration was 0.2 s. Participants were instructed to start the reaching movement as quickly as possible in response to the start cue. Participants were also instructed to control the hand cursor in as straight a line as possible from the start marker to the target marker. The trial ended 1 s after reaching the target marker.

We then introduced the two experimental conditions, namely IF and EF. Under the IF condition, because participants were instructed to covertly direct attention to their hand movements, trials with a tactile start cue and those with a visual start cue were defined as consistent-trials and inconsistent-trials, respectively. By contrast, under the EF condition, participants were instructed to direct their attention only to the hand cursor on the monitor. Therefore, trials beginning with the visual start cue were defined as consistent-trials and those with the tactile start cue were defined as inconsistent-trials.

The classification task comprised two sessions. The IF and EF conditions were randomly assigned for the first and second sessions for each participant. Each session had three phases: baseline, learning, and relearning.

The hand cursor moved contingently according to the participant’s hand movement (visuomotor rotation $\text{angle}=0\mathrm{deg}$). All participants performed nine consistent-trials.

Clockwise (CW) or counterclockwise (CCW) visuomotor rotation was applied to the hand cursor movements. In visuomotor rotation trials, hand cursor movement was rotated by 45 deg from the origin (start marker) in either the CW or CCW direction relative to the participant’s actual hand movement. Participants were required to correctly modify their hand movements to control the hand cursor in a straight trajectory. The CW and CCW settings were randomly assigned to the IF and EF conditions. Half of participants performed the reaching movements under the IF condition with the CW rotation and the EF condition with the CCW rotation, whereas the other participants performed the reaching movements under the IF condition with the CCW rotation and the EF condition with the CW rotation.

The young and elderly groups performed 40 trials (35 consistent-trials and 5 inconsistent-trials) while the stroke group performed 30 trials to avoid fatigue (27 consistent-trials and 3 inconsistent-trials). The inconsistent-trials were randomly inserted among the consistent-trials. The baseline and learning phase trials were run without interruption and participants were not informed when CW or CCW rotation started.

There was a 10-min break after the learning phase. All participants then performed nine consistent-trials again as the relearning phase. The visual feedback settings on the monitor, including the rotation angle, were identical to those in the learning phase.

2.2.3.

Analysis

2.2.4.

Statistical analysis

Differences in RT were assessed by three-way repeated measures analysis of variance (ANOVA) with group (young, elderly, or stroke) as a between-subjects factor and condition (IF condition or EF condition) and modality of the start cue (tactile or visual stimuli) as within-subject factors. To evaluate the degree of motor performance decline, a three-way ANOVA was also applied to the motor performance decline with group (young, elderly, or stroke) and individual dominance (IF-dominant or EF-dominant) as between-subjects factors and condition (IF condition or EF condition) as a within-subject factor. We used a significance threshold of $p<0.05$ (two-tailed) for all tests.

2.3.1.

Experimental setup

For behavioral data acquisition, we used the same experimental setup as in the first classification task (i.e., we recorded task hand movements using a drawing tablet with a monitor displaying a hand cursor as real-time visual feedback). To record prefrontal activity, we used a multichannel fNIRS system (ETG-7100, Hitachi Medical Corporation, Kashiwa, Japan) with sampling at 10 Hz. The fNIRS probes were arranged to cover the prefrontal area [Fig. 2(a)]. We used a $3\times 5$ multichannel probe holder consisting of eight laser sources emitting at 695 and 830 nm [emitters; red squares in Fig. 2(a)] and seven detecting probes [detectors; blue squares in Fig. 2(a)] arranged alternately at an interprobe distance of 3 cm. The midpoint of an emitter/detector pair was defined as a recording channel location [circles in Fig. 2(a)]. The probe holder was placed on the scalp with its lowest-row center emitter at the participant’s Fpz position, according to the standard international 10–20 system.

Fig. 2

Probe configuration for near-infrared spectroscopy. (a) Probes were placed over the prefrontal area. (b) Spatial registration of fNIRS maps onto MNI coordinate space. Each recording channel is numbered from the lower left to upper right position.

fNIRS signals reflect hemoglobin changes originating in cortical tissue due to brain activation and skin blood flow. To eliminate the influence of skin blood flow on fNIRS signals, we set eight additional short detecting probes at an interprobe distance of 1.5 cm [light blue squares in Fig. 2(a)] and applied multidistance independent component analysis (ICA).³⁰ ICA is a signal discrimination method in which independent components are extracted from mixed signals. It has been successfully used for fNIRS analysis.^31–34 Because signals from recording channels with a 1.5-cm interprobe distance primarily include skin blood flow signals in shallow tissues, based on these signals, we could discriminate between the effects of cortical tissue and skin blood flow on fNIRS signals. As it was possible to apply multidistance ICA only to the recording channels around the short detecting probes, the number of available recording channels was reduced to 15 [numbered circles in Fig. 2(a)]. For the spatial registration of fNIRS maps onto Montreal Neurological Institute (MNI) coordinate space, we measured scalp landmarks and all fNIRS recording channel positions using a three-dimensional magnetic space digitizer (FASTRAK, Polhemus). We then used an estimation tool without MRI.³⁵ Details on the spatial profiling of recording channels are provided in Fig. 2(b) and Table 3.

Table 3

Spatial profiling of each recording channel.

2.3.2.

Procedure

The fNIRS task consisted of two sessions each comprising six block sets containing alternating rest (20 s) and motor (15 s) blocks. A rest block was inserted at the end of each session. To avoid fatigue in stroke patients during successive hand movements, we applied a block design with relatively longer rest blocks than motor blocks. As in the first classification task, we introduced two attentional conditions and assigned the IF and EF conditions randomly to first and second sessions for each participant.

The monitor displayed the rest marker and hand cursor during all rest blocks. Regardless of the attentional condition, participants were required to match the hand cursor to the rest marker at the center of the monitor [Fig. 3(a)].

Fig. 3

Task configurations in the fNIRS task. (a) Visual settings in the rest block. (b) VG setting under the IF condition. In the motor block, VG gradually decreased over time. If the participant correctly directed attention to their hand according to the task instructions, the decreasing amplitude was anticipated only in hand cursor movement, as shown by the movement profiles in the right column. The amplitudes of the first and last cycles were defined by the difference between adjacent positional peaks. Blue and green arrows indicate hand and hand cursor amplitudes, respectively. The differential value between the first and last amplitudes was then calculated. (c) VG setting under the EF condition. VG gradually increased over time. If the participant correctly directed their attention to the hand cursor according to the task instructions, the decreasing amplitude was anticipated only in hand movement.

To isolate the habituation effect of the longitudinal direction movements applied in the first classification task, we required the participants to perform cyclic movements in a lateral direction in the second fNIRS task. When each motor block started, the rest marker disappeared and participants were asked to move their hand cyclically on the drawing tablet from side to side at a comfortable speed. The hand cursor on the monitor moved horizontally according to the rightward or leftward hand movements.

Under the IF condition, participants were instructed to covertly direct their attention to their hand movements. To correctly guide participants’ attention during the motor blocks, we prepared additional stimuli for cueing the turning point (direction change) of cyclic movements. If a participant’s hand position exceeded $\pm 10\mathrm{cm}$ in the $x$-axial direction from the center of the drawing tablet, tactile stimuli were provided on the index fingertip by the vibration motor. Participants were requested to keep their hand movement amplitude constant throughout the motor block based on presentation of tactile stimuli (i.e., the desired hand movement amplitude was maintained at 20 cm on the drawing tablet). Furthermore, to confirm participants’ attentional strategies, VG was gradually decreased to from 1 to 0.6. If a participant correctly maintained the desired hand movement amplitude, the hand cursor movement amplitude on the monitor gradually decreased from 20 to 12 cm [Fig. 3(b)].

By contrast, under the EF condition, participants were instructed to direct their attention only to the hand cursor on the monitor. To indicate the turning point, the color of the hand cursor changed when the hand cursor position exceeded $\pm 10\mathrm{cm}$ in the $x$-axial direction from the monitor center. Participants were also requested to keep the hand cursor movement amplitude constant throughout the motor block based on presentation of the visual stimuli (i.e., the desired hand cursor amplitude was 20 cm on the monitor). VG gradually increased from 1 to 1.5, so the amplitude of hand movement on the drawing tablet was expected to gradually decrease from 20 to 12 cm [Fig. 3(c)].

2.3.3.

Analysis

2.3.4.

Statistical analysis

Movement amplitudes were analyzed by three-way repeated measures ANOVA with group (young, elderly, or stroke) as a between-subjects factor and condition (IF or EF) and attentional target (hand or hand cursor) as within-subject factors. To explore intersubgroup differences in the prefrontal activity, we compared the beta values of oxy-Hb and deoxy-Hb signals between the IF- and EF-dominant subgroups in each recording channel using $t$-test. To adjust for multiple testing, we applied false discovery rate (FDR) correction^41,42 with the significance threshold set at $p<0.05$ (FDR-corrected). Regarding the AUC, because the sample size is insufficient for classification analysis, to alternatively assess the original highest AUC, we repeatedly calculated AUCs on the basis of a randomized group of participants (i.e., we randomly classified the participants into IF- and EF-dominant subgroups and calculated AUC 3000 times). Then, we compared the original highest AUC and the randomized data-based AUCs.

3.1.1.

Reaction time

We excluded trials in which RT did not meet our criterion ($<150$ or $>1500\mathrm{ms}$). In total, 3.0% of tactile-cued and 2.1% of visual-cued responses were excluded under the IF condition, whereas 4.3% of tactile-cued and 1.9% of visual-cued responses were excluded under the EF condition.

Comparison of RTs revealed a significant two-way interaction of $\text{condition}\times \text{modality}$ [$F(1,61)=226.73$, $p<0.000001$, ${\eta }_p^2=0.79$] and main effects of group [$F(\mathrm{2,61})=12.44$, $p=000029$, ${\eta }_p^2=0.29$], condition [$F(1,61)=11.05$, $p=0.0015$, ${\eta }_p^2=0.15$], and modality [$F(1,61)=9.04$, $p=0.0038$, ${\eta }_p^2=0.13$]. Post-hoc analysis with Bonferroni correction for the $\text{condition}\times \text{modality}$ interaction revealed that RTs for consistent-trials [tactile in the IF condition: $519.8\pm 16.5$ (SEM) ms, visual in the EF condition: $517.2\pm 15.2$ (SEM) ms] were significantly faster than those for inconsistent-trials [tactile in the EF condition: $617.9\pm 18.1$ (SEM) ms, visual in the IF condition: $678.2\pm 20.8$ (SEM) ms, $p<1.1\times {10}^{-10}$], which is consistent with the requirements of Eq. (1) or Eq. (2) and suggests that participants correctly directed their attention according to condition (to their hand in IF trials or the hand cursor in EF trials).

3.1.2.

Motor performance and individual optimal attentional strategy

Figures 4(a)–4(c) show the changes in movement errors as the consistent-trials progressed [young, Fig. 4(a); elderly, Fig. 4(b); stroke, Fig. 4(c)]. All groups showed the same trend. In the baseline phase without visuomotor rotation, the hand cursor did not deviate from the required straight trajectory. When the hand cursor movement was rotated relative to the actual movement during the learning phase (starting after the 10th trial), movement error markedly increased and then gradually decreased. Toward the end of the learning phase, the error reached a plateau, indicating that participants had adapted to the visuomotor rotation. After the 10 min break, regardless of the same visuomotor rotation disturbing hand movement in the relearning phase, participants quickly readapted.

Fig. 4

Learning curve in each group. (a)–(c) Movement errors in the (a) young, (b) elderly, and (c) stroke groups. Blue solid and green dotted lines indicate mean movement errors under the IF and EF conditions, respectively. In all groups, performance improved as trials progressed in the learning phase. After the 10-min break, the error increased and participants readapted to the same visuomotor rotation environment. (d) The degree of motor performance decline under IF and EF conditions. Labels “Y,” “E,” and “S” indicate the young, elderly, and stroke groups, respectively. Error bars denote the standard deviation. ***$p<0.001$.

By comparing the degree of motor performance decline between the IF and EF conditions as shown in Fig. 4(d), we classified participants into an IF-dominant subgroup ($\mathrm{\Delta }D>0$, young: $n=10$, elderly: $n=12$, stroke: $n=12$) and EF-dominant subgroup ($\mathrm{\Delta }D<0$, young: $n=13$, elderly: $n=11$, stroke: $n=6$). Analysis of motor performance decline revealed a significant main effect of group [$F(\mathrm{2,58})=4.96$, $p=0.010$, ${\eta }_p^2=0.15$] and a significant two-way interaction of $\text{condition}\times \text{individual}$ dominance [$F(\mathrm{1,58})=70.04$, $p=1.48\times {10}^{-11}$, ${\eta }_p^2=0.55$]. The other factors did not reach the level of statistical significance ($F\mathrm{s}<2.73$, $p\mathrm{s}>0.074$). Post-hoc analysis with Bonferroni correction found that motor performance decline in the elderly group was significantly larger than that in young group ($p=0.011$). Furthermore, Bonferroni correction revealed that motor performance declines under the IF condition were significantly smaller than those under the EF condition in the IF-dominant subgroup ($p=1.20\times {10}^{-8}$). Conversely, the declines under the EF condition were significantly smaller than those under the IF condition in the EF-dominant subgroup ($p=5.46\times {10}^{-5}$). The intersubgroup differences under the same attentional condition also reached significance in both conditions (IF condition: $p=8.14\times {10}^{-6}$, EF condition: $p=6.76\times {10}^{-6}$).

3.2.1.

Movement amplitude

Analysis of amplitude change revealed a significant three-way interaction of $\mathrm{group}\times \mathrm{condition}\times \text{attentional target}$ [$F(\mathrm{2,61})=3.57$, $p=0.034$, ${\eta }_p^2=0.10$]. Further, we confirmed two important aspects of this three-way interaction by multiple comparison using Bonferroni correction. First, the amplitude changes of attentional target {${\mathrm{\Delta }\mathrm{Amp}}_{\mathrm{H}}$ under the IF condition [$8.89\pm 1.57$ (SEM) cm] and ${\mathrm{\Delta }\mathrm{Amp}}_{\mathrm{C}}$ under the EF condition [$7.91\pm 1.32$ (SEM) cm]} were significantly smaller than those of the nonattentional target {${\mathrm{\Delta }\mathrm{Amp}}_{\mathrm{C}}$ under the IF condition [$26.4\pm 3.01$ (SEM) cm] and ${\mathrm{\Delta }\mathrm{Amp}}_{\mathrm{H}}$ under the EF condition [$31.2\pm 3.79$ (SEM) cm]}, respectively (all $p\mathrm{s}<0.0088$). Therefore, we can assume that participants correctly directed their attention to their hand in the IF condition or the hand cursor in the EF condition according to the experimental instructions. Second, the amplitude changes of the hand and hand cursor did not show any significant differences among the young, elderly, and stroke groups (all $p\mathrm{s}=0.99$). These results indicate that prefrontal activity was not affected by individual motor performance itself.

3.2.2.

Prefrontal activity

Based on the individual optimal attentional strategy evaluated in the first classification task, we compared the prefrontal activities between the IF-dominant subgroup ($n=34$) and EF-dominant subgroup ($n=30$).

Figures 5(a) and 5(b) present the temporal mean profiles of oxy-Hb signals (upper rows) and deoxy-Hb signals (lower rows) across the motor blocks of the same attentional condition in the (a) IF-dominant and (b) EF-dominant subgroups, respectively. The profiles in Fig. 5 show oxy-/deoxy-Hb signals that mainly reflect the effect of the cortical tissue (i.e., the signals after removing the component of the skin blood flow using multidistance ICA). Several channels in both subgroups showed similar trends. First, oxy-Hb signals increased whereas deoxy-Hb signals decreased in several areas. Second, changes in oxy-Hb signals were stronger than those in deoxy-Hb signals. These characteristics are typical hemoglobin dynamics reflecting neural activity. Notably, for oxy-Hb signal under the IF condition, we observed a marked intersubgroup difference in the boundary area including the left DLPFC and left frontopolar cortex (FPC [ch.8: squares in Figs. 5(a) and 5(b)]. More specifically, IF-dominant individuals showed a gradual increase in oxy-Hb signal during performing the cyclic motor task, whereas EF-dominant individuals did not show oxy-Hb signal changes [Fig. 5(c)]. As shown in Fig. 5(d), compared with EF-dominant individuals, IF-dominant individuals showed stronger oxy-Hb signal, especially after the latter half of the motor block.

Fig. 5

Temporal characteristics of oxy-Hb and deoxy-Hb signals. Zero on the horizontal axis indicates the start timing of the motor block. (a, b) Upper row: Under the IF condition, oxy-Hb signal in ch.8 showed a marked difference between the IF- and EF-dominant subgroups. Lower row: Deoxy-Hb signals showed neither strong temporal change nor intersubgroup difference. (c) Temporal profiles of oxy-Hb and deoxy-Hb signals in ch.8 under the IF condition. Red solid and blue dotted lines represent the time courses of oxy-Hb and deoxy-Hb signals, respectively. The lighter colored regions around the time course data indicate the standard deviation. Note that the profiles of oxy-Hb and deoxy-Hb signals show only upper red and lower blue standard deviation regions, respectively. (d) The differential mean profiles between the IF- and EF-dominant subgroups (EF-dominant subgroup minus IF-dominant subgroup). Positive values on the longitudinal axis indicate a stronger oxy-/deoxy-Hb signal in the EF-dominant subgroup than in the IF-dominant subgroup, whereas negative values indicate the opposite magnitude relation.

Figure 6(a) shows the spatial profiles of the $t$-values of the $t$-test between the IF- and EF-dominant subgroups for each recording channel’s beta value. Only oxy-Hb signal in the boundary area including the left DLPFC and left FPC (8ch) showed a significant intersubgroup difference under the IF condition [Fig. 6(a)]. In this channel, the IF-dominant subgroup showed higher beta values of oxy-Hb signal than the EF-dominant subgroup [Fig. 6(b) ($p=0.0016$, FDR-corrected)]. Incidentally, weak intersubgroup differences of oxy-Hb signals were observed in the left FPC [ch.4 under the IF condition ($p=0.012$, uncorrected)], in the boundary area including the left DFLPC and left FPC [ch.8 under the EF condition ($p=0.0063$, uncorrected)], and in the left DLPFC [ch.12 under the IF condition ($p=0.020$, uncorrected) and EF condition ($p=0.024$, uncorrected)]. Regarding deoxy-Hb signals, we observed weak intersubgroup difference only in the boundary area including the right DFLPC and right FPC [ch.10 under the EF condition ($p=0.039$, uncorrected)] and the left DLPFC [ch.13 under the IF condition ($p=0.037$, uncorrected)].

Fig. 6

Spatial configuration in the prefrontal area reflecting individual optimal attentional strategy. (a) Only the beta value of oxy-Hb signal in ch.8 under the IF condition showed a significant difference between IF- and EF-dominant subgroups. (b) Beta values in ch.8 under the IF condition. The positive and negative values indicate an increase and a decrease in oxy-Hb/deoxy-Hb signals, respectively. The beta value of oxy-Hb signal in the IF-dominant subgroup was significantly larger than that in the EF-dominant subgroup ($p=0.0016$; $t$-test, FDR-corrected). *$p<0.05$.

3.2.3.

Prefrontal activity-based classification

Using the estimated beta values of oxy-Hb signal in ch.8 under the IF condition, we then attempted to predict an individual’s optimal attentional strategy. Because the number of channels with a significant intersubgroup difference was 1, we applied the individual’s beta value as the CI. In the repeated ROC analysis based on leave-one-out cross-validation, the highest AUC was 0.74 (Fig. 7) and the mean AUC was $0.74\pm 0.00027$ SD. At the optimal cut-off value, IF- and EF-dominant individuals were distinguished with a sensitivity of 0.71 and specificity of 0.73 (open circle in Fig. 7). In addition, we confirmed that the original highest AUC was relatively higher than the randomized data-based AUCs (AUC range: 0.729 to 0.500; ratio of 0.7 or more: 0.37%). Furthermore, the classification accuracy using the optimal cut-off value was 71.9% (46 of 64 for all participants). The classification accuracies in each group were 73.9% [young group: 17 of 23 (females, 9 of 13; males, 8 of 10)], 69.6% [elderly group: 16 of 23 (females, 6 of 7; males, 10 of 16)], and 72.2% [stroke group: 13 of 18 (females, 3 of 3; males, 10 of 15)].

Fig. 7

ROC curve based on the beta value of left prefrontal activity (ch.8). The open circle indicates the optimal sensitivity and specificity when we applied a cut-off value of the highest AUC.