Protocol, set-up and biomechanical data processing

Patients were investigated in the morning at least 12 h after their last dose of antiparkinsonian medication and 2 h after pausing the stimulation (i.e. medication OFF/stimulation off condition). Patients walked overground and barefoot at a self-selected speed over a 15 m path that included passing through one turning door (1 m wide) in the gait laboratory and two common doors (1.2 m and 1.6 m wide) outside the gait laboratory (Fig. 2A). We chose this pathway to mirror a daily-life situation of adapting gait to different environmental conditions. All patients performed at least four trials (range 4–8), according to their clinical conditions. Due to the severity of akinetic-rigid symptoms, Patient nwk01 performed only recordings in the gait laboratory.

Experimental set-up, kinematic identification of gait freezing, power spectral densities and β-burst identification analysis. (A) Experimental set-up. The walking path inside and outside the gait laboratory consisted of walking through a turning door (inside the gait laboratory) and two common doors outside, where a representative freezing episode took place (red dot and figure). (B) Kinematic representation of one freezing episode. Representative traces of the ankle angular velocity relative to the medial-lateral axis during (effective) walking and gait freezing. We identified five time frames: (effective) walking is shown as light grey boxes (1.5-s time epochs free of gait freezing), F_PRE and F_STOP are the yellow boxes (1.5-s time epochs preceding and following a freezing episode, respectively), and F_START and F_STOP are shown as red boxes (the first and the last 1.5 s of freezing, respectively). (C) Cortical and STN power spectral densities. The cortical LFPs in the selected regions of interest (SMA, M1 and PC) displayed a bimodal distribution with two distinct activity peaks in the θ- and α-frequency bands. The STN power spectra also showed a bimodal distribution with a small peak at 11 Hz and a prominent peak in the β-frequency band. Shaded areas represent the group level variance computed using the bootstrapping technique (20 repetitions, resampling with replacement) and estimating the confidence intervals between the 5th and 95th percentiles of the bootstrap distributions. The background colour indicates the frequency ranges used for further analyses. (D) β-burst identification. Pearson’s correlation coefficient between average β-amplitude and number of β-peaks above the threshold computed in all 1.5-s walking epochs is reported for the two STN separately (STN+ and STN−). Solid lines are the average correlation curves across subjects. Dashed lines represent the standard error computed with the bootstrap technique. Red lines identify the values used as threshold for β-burst detection. (E) Top: A segment of the wavelet real part (blue line) derived from the wavelet transformed LFPs in the β-peak frequency (20 Hz) of a representative subject is reported. Middle: The wavelet amplitude was z-scored and the β-burst peaks were identified (black dots) and sorted according to their amplitude. We then identified the burst duration with the FWHM method. Bottom: A close view on the identification of burst duration. Starting from the higher peak (peak I°) we found the closest points (blue circles) in which the z-scored wavelet amplitude goes below the peak half amplitude. The time difference between these two points determined the burst duration. Since peak II° was located inside the burst duration of peak I°, we eliminated peak II° and considered these two peaks part of the same burst. STN+ or STN− refers to the side with more and less striatal dopaminergic innervation, respectively.

Throughout the entire walking path, kinematics of lower limbs was measured using two inertial recording units (IMU, Opal, APDM), with a sampling rate of 128 Hz, placed on the outer anklebones. A representation of the complete set-up is shown in Fig. 2A. To detect gait freezing episodes, we computed the wavelet spectrum of the ankle angular velocity around the medial-lateral axis with respect to the walking direction (Fig. 2B). Gait freezing was identified by a switch to higher frequency compared to (effective) walking as in Moore et al. (2008). In particular, we defined for each time t a ‘freezing index’ (FI) as the ratio between the square of the area under the power spectra in the ‘freezing’ band (3–8 Hz) and in the ‘locomotion’ band (0.5–3 Hz), calculated in a 6-s window centred in t. For each subject, a specific freezing gait threshold was defined as the mean + 1 SD of the peak FI from volitional standing (Moore et al., 2008).

Walking trials were also video-recorded by two synchronized cameras (VIXTA) and two independent raters (N.G.P. and I.U.I.) clinically verified all freezing episodes (Shine et al., 2014).

We selected five time frames of 1.5 s each: (i) (effective) gait (walking, time epochs free of gait freezing); (ii) pre-freezing (F_PRE, the 1.5-s time epoch immediately preceding a freezing episode); (iii) freezing start (F_START, first 1.5 s of a freezing episode); (iv) freezing stop (F_STOP, last 1.5 s of a freezing episode); and (v) post-freezing (F_POST, the 1.5-s time epoch after the resolution of a freezing episode) (Fig. 2B). The epochs were defined based on the shortest freezing episode, which lasted 3 s, and never overlapped. The total time of (effective) walking was 592 s and it was analysed with 395 epochs of 1.5 s. These epochs were recorded in the same environmental settings, therefore controlling for its difficulties. In addition, we compared gait freezing epochs directly with epochs of successful passing through doors (32 epochs of 1.5 s each) and voluntary stop (34 epochs of 1.5 s each).

In the gait laboratory, we measured the kinematics of body segments during (effective) steady-state linear walking (reached before approaching the turning door) using an optoelectronic system (SMART-DX400, BTS), which computed the 3D coordinates of 29 spherical retro-reflective markers (15 mm diameter) fixed to anatomical landmarks (Palmisano et al., 2019). The marker coordinates were low-pass filtered (cut-off frequency of 7 Hz) and interpolated. Kinematic parameters were automatically extracted by custom scripts developed in MATLAB® ambient (MATLAB 2017b, The MathWorks, Inc., Natick, Massachusetts, USA) and then checked by visual inspection. We computed the stride length, duration, and velocity (expressed as percentage of subject’s height) and the stance and double-support duration. Temporal parameters (i.e. stance and double-support) were time-normalized as a percentage of the stride duration. For each subject and condition, all variables were averaged over the trials (Arnulfo et al., 2018). These findings were compared with data obtained with the same experimental set-up from 11 healthy control subjects (nine males, two females, mean age: 58 ± 5 SD years, range: 50–66 years) matched for age and anthropometric measurements (Table 2).

Anthropometric and kinematic measurements

Stance and double-support duration are expressed as the percentage of the duration of the stride (i.e. the interval between two subsequent heel strikes of the same foot). The stride length and the stride velocity were calculated as a percentage of the body height of each subject (%BH). Data are shown as mean ± SD.

^aStatistical significance (P < 0.05).

ASIS = anterior-superior iliac spines; BH = body height; BMI = body mass index.