Neural predictors of gait stability when walking freely in the real-world

Eighteen right-handed healthy young adults (age mean ± standard deviation (SD) = 25 ± 3, 7 male/11 female] with no previous history of neurological, musculoskeletal or gait disorders, volunteered for the study by giving written informed consent. The study was approved by the University of East London Ethics Committee (UREC_1415_29) and all experiments were conducted in accordance with the Declaration of Helsinki. Data of three subjects were discarded because of technical failures during data acquisition (two males, one female) and of one subject (female) because of poor data quality, leaving a total of 14 subjects (age mean ± SD = 26 ± 3, 5 male/9 female].

Subjects were first prepared in a laboratory room and then guided outside into the campus garden [28]. During this period, no signals were recorded and subjects were instructed to become familiar with the setup and communicate to the experimenter if anything was not properly set. Once outside, subjects were shown the predefined walking path (200 m) and were instructed to walk at their preferred natural speed, as previous studies have shown that gait behaviour and patterns are optimized when walking at the natural speed [21, 26]. Experiments consisted of three conditions during which subjects walked along the predefined path naturally (Single-Task, ST), while conversing with the experimenter (Dual-Task1, DT1) or texting with their smartphone (Dual-Task2, DT2). The dual-task conditions were randomized across subjects in order to avoid bias in gait behaviour and recordings. The dual-task conditions were designed to represent real-life situations and, to standardise them, conversations during DT1 were based on a set of standard questions, whereas in DT2 subjects read and replied to a standard email. Subjects rested for a period of 5 min after each condition in order to avoid fatigue. The experimenter followed the subjects during each condition recording videos of gait behaviour. Experiments were carried out only during dry days free from strong winds and/or rain.

The experimental setup, illustrated in Fig. 1, is fully mobile and allows the recording of physiological and behavioural data during walking (or any other mobile situation). Brain activity (EEG; μV) was continuously recorded via a high-density 64 channel Waveguard cap (ANT Neuro, Enschede, Netherlands) with an EEGoPro amplifier (ANT Neuro, Enschede, Netherlands) at a sampling frequency of 1 kHz and raw signals filtered between 0.1 and 500 Hz. Impedances were kept below 5 kΩ for the whole duration of the experiment and data were referenced to the FCz channel. The recording EEG amplifier was carried by the subject within a backpack. A Samsung Galaxy S4 mini smartphone was fixed at the subject’s lower back with an elastic belt and data from its internal accelerometers and gyroscope were recorded through the AndroSensor app at a frequency of 200 Hz. The lower back position is the currently most preferred and reliable location to observe changes in gait patterns across different conditions and populations [22]. Two digital force sensing resistor sensors (FSRs) were employed as contact switches and fixed underneath the subject’s heels to detect times of heel strikes. Data were recorded at 1 kHz by a 14 bit analog-to-digital converter (DataLog MWX8, Biometrics Ltd., Newport, UK) fixed at the subject’s hip by an elastic belt. These sensors return a digital binary output where the active edge is set 1-to-0, i.e., 0 when the heels make contact with the ground. A digital button (1-to-0 active edge) was also connected to the converter and pressed by the subject for 5 seconds at the beginning and at the end of each condition to define time points of start and finish. Elastic bands were placed around the subject thighs to fix the cables of contact switches to avoid uncomfortable situations and prevent the subject from falling/stepping on them. To synchronize data from the digital sensors representing important time points (i.e., start, heel strikes, end) with physiological variables, a common train of 12 consecutive TTL pulse was simultaneously sent to both the DataLog MWX8 converter and the EEGoPro amplifier at the beginning and at the end of the experiment. The matching of start and end points of each single pulse between the recordings were checked offline and eventually used as milestones for realigning the signals’ time axes.

Statistical analyses were all run with SPSS 23 software (IBM).

Subjects walked along the predefined path of 200 m in an average time of 225 ± 25 s during single-task (ST) walking and significantly more slowly when walking while conversing (DT1; 235 ± 28 s) and when walking while texting (DT2; 260 ± 41 s; RMANOVA F = 21.660, p < 0.001). Performing any secondary tasks significantly increased the walking time with respect to ST (ST vs. DT_i with i = 1,2 all t < − 3.906, all p < 0.002) and walking while texting required a longer time compared to walking while conversing (DT2 vs. DT1, t = − 3.993, p = 0.002). Velocity was significantly slower (RMANOVA F = 34.215, p = 0.001) in both DT1 (t = 4.199, p = 0.001) and DT2 (t = 6.847, p = 0.001) compared to ST and slower in DT2 vs DT1 (t = 4.991, p = 0.001; see Table 1). Gait measures of stride duration, ver-RMS and ap-RMS were statistically different across conditions (RMANOVA F > 12.165, p < 0.001), specifically between DT2 vs. ST (t > (−) 3.503, p < 0.004) and DT2 vs. DT1 (t > (−) 3.793, p < 0.002). Gait measures of ml-RMS and ap-step regularity significantly differed across conditions (RMANOVA F > 4.559, p < 0.043), specifically between DT2 vs. ST (t > 2.772, p < 0.016). Gait measures of ver-RMSR and ap-RMSR also changed significantly across conditions (RMANOVA F > 3.990, p < 0.031), specifically between DT2 vs. DT1 (t > 3.618, p < 0.003). No significant differences were found for measures of mean step length, ml-RMSR and ver-step regularity (RMANOVA F < 3.099, p > 0.05). Measures of acceleration RMS/RMSR in each direction across the three conditions are represented in Fig. 3.

A regression model was successfully created for ver-RMSR (R-squared = 0.725, p = 0.001) for which all assumptions were met (no multicollinearity, no auto-correlation, no homoscedasticity). Only the IVs significantly contributing to the prediction of ver-RMSR were entered stepwise into the model, which predicted the DV based on Velocity (B = 0.355, p = 0.001) and left-parietal theta PSD (B = 0.009, p = 0.026) as shown in Fig. 4. The final model equation is reported in Table 2. Multiple linear regression analysis did not provide statistically significant models for ml-RMSR or ap-RMSR during single-task walking.

A regression model was successfully created for ver-RMSR (R-squared = 0.727, p = 0.001) for which all assumptions were met (no multicollinearity, no auto-correlation, no homoscedasticity). Only the IVs significantly contributing to the prediction of ver-RMSR were stepwise entered into the model, which predicts the DV based on Velocity (B = 0.029, p = 0.003) and left-parietal alpha PSD (B = 0.021, p = 0.020) as shown in Fig. 5. The final model equation is reported in Table 2. Multiple linear regression analysis did not provide statistically significant models for ml-RMSR and ap-RMSR when walking while conversing.

A regression model was successfully created for ml-RMSR (R-squared = 0.434, p = 0.010) for which all assumptions were met (no multicollinearity, no auto-correlation, no homoscedasticity). Only the IVs significantly contributing to the prediction of ml-RMSR were stepwise entered into the model, which predicts the DV based only on left-parietal beta PSD (B = − 0.055, p = 0.010), as shown in Fig. 6. The final model equation is reported in Table 2. Multiple linear regression analysis did not provide statistically significant models for ver-RMSR and ap-RMSR when walking while texting.

We observed that region- and frequency- specific brain activation could predict gait stability during free self-paced walking and walking during dual-tasking. Two walking conditions (i.e., single-task walking and walking while conversing) were characterised by a positive relationship between ver-RMSR, velocity and left parietal PSD in θ and α frequency bands respectively. On the other hand, walking while texting featured a negative relationship between ml-RMSR and left parietal PSD in the β frequency band. Thus the left PPC appears to play a general role in walking in the real-world and the type of secondary task may be correlated to frequency-specific coding in the left PPC.

We tested whether prefrontal and bilateral posterior parietal spectral activity could relate to gait behaviour as it has been shown that these are the most modulated cortical areas during dual-task free walking [6, 11, 37, 38] and in dual-task laboratory-based experiments [39]. No linear relationship was found between the PFC and gait stability, despite the great interest in this brain region when monitoring multitasking ambulatory activities [6, 7, 37, 40‐43]. On the contrary, the left PPC was shown to be related to gait behaviour regardless of the secondary task type undertaken. The left PPC has been extensively studied in laboratory-based experiments both on animals and humans [44]. Currently, left PPC is thought to act as the sensorimotor integrator and rapid online updater of movement planning [45], integrating spatial information of the surroundings and sensory feedback with motor planning and executive commands [44]. First evidence of the involvement of the PPC in inter-limb coordination during walking was obtained through intra-cortical recordings of visually-guided walking cats [46]. It was postulated that PPC neurons were involved in the regulation of inter-limb coordination during locomotion requiring visual guidance, thus playing a role in processes of sensorimotor integration (i.e., visuomotor integration) [47]. In humans, PPC is believed to be involved in the integration of sensory information with motor plans, which is severely impaired in PD patients with FoG [48]. Indeed, the spatial navigation deficits seen after damage of bilateral PPC as in Balint’s syndrome [49] or neglect [50] arise because of the inability to integrate spatial orientation with current/future planning of the voluntary movement needed to accomplish the end goal. If the PPC is in general believed to play such a complicated role, then each hemispheric area has been characterized with specific roles. The right-PPC, in connection with frontal regions, is strongly involved in multiple types of attention [12, 51]. On the other hand, the left PPC has been recently shown to play a major role as a sensorimotor integrator as PD patients that exhibit FoG display reduced functional connectivity between left PPC and multiple brain regions such as the somatosensory and auditory areas [52]. Left PPC is therefore suggested to work as a sensorimotor integrator during online movement planning and monitoring, whereas the right PPC actively engages with different attentional networks. In summary, we suggest that left PPC plays a primary active role in monitoring and planning the walking movement regardless of any secondary conditions undertaken in parallel. However, the different relationships between neural activity in this brain region and changes in gait stability across walking conditions are encoded by different frequency-specific activity.

When walking naturally, a positive linear predictive relationship could be identified between left PPC activity in the θ frequency band and ver-RMSR and velocity. Ver-RMS(R) has been shown to depend linearly on gait speed and to represent the quality of gait, with higher values symbolic of a more stable and rapid walk (typical of healthy young adults) and lower values symbolic of reduced flexibility and bent postures (typical of older adults) [22, 23, 26]. We confirmed the positive relationship between ver-RMSR and velocity, but added a predictive neural component (left PPC). When walking naturally without engaging in any secondary tasks, left PPC has been shown to be active in the θ range (4–7 Hz) [53‐55] and is believed to be connected to deeper structures such as the hippocampus actively engaged in the navigation process [56].

When walking while conversing, a positive predictive relationship could be identified between left PPC activity in the α frequency band and ver-RMSR and again velocity. As θ neural oscillations likely engage in memory retrieval and organization of ‘thoughts’ when subjects are engaged in a conversation [57, 58], is it likely that higher frequency oscillations (i.e., α) took over the duty of sensorimotor integrator and monitor during walking in the present study. This hypothesis is further supported by recent studies showing involvement of α (8–12 Hz) oscillations in spatial navigation [55, 59] and sensorimotor integration during walking speed adaptation to an external pace cue [11].

A significant negative predictive relationship was found between left PPC spectral activity in the β frequency band (15–30 Hz) and ml-RMSR when walking while texting. ml-RMSR has been recently validated as a marker of gait abnormality and recovery, where abnormally high values of trunk acceleration decrease alongside recovery [26, 27]. Moreover, stronger β desynchronization is required in arduous conditions in order to maintain the ‘status quo’ and to promote the voluntary action in a more challenging dual-task context [9‐11, 60]. The negative relationship suggests that those subjects showing higher gait variability in the medio-lateral direction required stronger β desynchronization in order to accomplish the simultaneous tasks. This is in line with previous studies that showed that β PSD in the posterior parietal and occipital areas could be used to reliably classify and detect events such as FoG and GIF in Parkinson’s disease patients [14, 15], thus confirming the involvement of this brain region in postural and movement online control via sensorimotor integration. Of note, ml-RMSR did not correlate with velocity in this study as previously reported in the literature [26].

In summary, spectral activity in the left PPC can predict stability of gait during different walking conditions through frequency-specific neural mechanisms. As the healthy population sample was young and small in number, the study could be expanded to include healthy older adults and PD patients. The inclusion of healthy older adults and PD patients would enable further testing of the hypothesis that the left PPC is crucially involved in the underlying mechanisms of successful walking in the presence of complex secondary tasks. This would be important in order to observe how specific neural activities can predict altered gait patterns due to cognitive decay and/or neurological impairments.

The analyses performed in this study were limited to the sensor-level (i.e., electrode data) and in the absence of subject-specific anatomical MRI data and care was taken not to over-stretch data interpretations towards the sources of neural activity. Future studies would benefit from anatomical MRI data for each subject or patient in order to reliably source the origins of neural activity especially as brain anatomy varies with age and stage of PD [61]. Lastly, both dual-task conditions required the subjects to share part of their attention to the ambulatory task with a second task, and difficulties in reallocating attentional resources could have contributed to the decrease in gait stability [39]. Future studies should therefore identify changes in level of attention as potential covariates in the predictive relationship between brain activations and gait behaviour using recent developments in MOBI of locomotion in health and disease [62, 63].