Brain–computer interface based on functional near-infrared spectroscopy integrated with plantar load simulator in post-stroke rehabilitation: a pilot study

Full Text

Introduction

Gait disorders represent a prevalent and severe consequences of stroke and are a key reason for the loss of autonomy, increased risk of falls, and reduced quality of life [1–4]. It is estimated that over 44% of patients in the acute [3] and about 33% in the chronic [4] phase of stroke have significant gait disorders, with the average walking speed (0.25–0.60 m/s) being considerably lower than the threshold required for independent daily activity (0.80 ± 0.15 m/s) [1]. Impaired lower limb function negatively affects mobility and safety, increases the risk of falls, limits independence, and is exacerbated by concomitant emotional and cognitive impairments, which complicates social adaptation and integration. Therefore, gait restoration is one of the primary goals of post-stroke rehabilitation, as it promotes the return of independence, increases social activity, and reduces the frequency of falls [1, 2].

A brain–computer interface (BCI) is a system that measures brain activity and converts it in real time into functionally useful output to replace, restore, enhance, supplement, and/or improve the brain’s natural output¹. In post-stroke rehabilitation, BCIs are used to provide biological feedback during motor imagery training. This technology enables the integration of motor intention or movement imagery with sensorimotor feedback, which helps stimulate neuroplasticity and supports the patient’s involvement in restoring motor functions [5].

Whereas the use of BCIs for restoring upper limb movements after stroke is supported by numerous studies, systematic reviews, and meta-analyses [6–8], the application of this technology for gait recovery remains insufficiently studied [7]. This imbalance in research focus is primarily explained by the technical challenges of decoding neural patterns associated with locomotion and gait phases when recorded from the scalp surface [9].

In BCI-based rehabilitation systems, various actuators are used to provide feedback — visual feedback [10], functional electrical muscle stimulation [11], pedaling systems [12], or exoskeletons [13]. The Korvit (Russia) plantar support load simulator (pneumatic orthosis) is used in space medicine and post-stroke rehabilitation to simulate the physiological sequence of foot contact with the surface and stimulate plantar receptors [14]. Such stimulation generates an afferent flow that promotes the restoration of sensorimotor integration and the activation of central structures controlling movement [15, 16]. Integrating the Korvit device into the BCI loop allows for the temporal synchronization of cortical activation and peripheral response. To our knowledge, a design combining a BCI and a pneumatic gait simulator has not been previously used or studied in clinical practice.

For recording brain activity signals in the BCI loop, electroencephalography (EEG) is used in the vast majority of cases. As an alternative, functional near-infrared spectroscopy (NIRS) can be used, which is based on recording the cortical hemodynamic response [17] using infrared light sources and detectors placed on the scalp surface. Functional NIRS captures local changes in blood flow that reflect the activation of brain cortical areas during stimulation or when a person performs a specialized task. Compared to EEG, NIRS records changes in two parameters simultaneously (concentrations of oxygenated and deoxygenated hemoglobin), exhibits high resistance to motion artifacts and electromagnetic interference, and its application does not require applying conductive gel under the sensors [18, 19].

Despite the promising prospects of using NIRS as a control signal for a BCI, there is currently insufficient data on whether the level of hemodynamic response recorded by this method in patients with cerebrovascular diseases is sufficient for effective system control. Furthermore, the influence of a gait simulator on signal quality and BCI system performance during rehabilitation training remains unclear. Technical features of the Korvit device are associated with the inability to rapidly change pressure in the plantar section after the active phase, which limits the full reproduction of the rest phase in BCI-assisted training. During rest periods, a gradual release of air occurs, causing tactile stimulation of the feet, which may affect the accuracy of the BCI classifier, making it difficult to differentiate between states of active motor imagery of the legs and rest. The potential impact of interference of the Korvit pneumatic orthosis on the quality of the NIRS signal during operation should be also studied. These issues determine the need for studies aimed at assessing the functional and technical feasibility of using NIRS-BCI with the Korvit device in the context of post-stroke gait rehabilitation.

The aim of the study is to evaluate the possibility of clinical application of NIRS-BCI in combination with the Korvit pneumatic orthosis in the comprehensive rehabilitation of post-stroke patients.

Materials and methods

Study design

The study was conducted at the Russian Center of Neurology and Neurosciences. The study protocol was approved by the Local Ethics Committee of the Russian Center of Neurology and Neurosciences (protocol No. 9-6/22 dated October 19, 2022). All participants provided informed consent.

In addition to the standardized course of in-hospital rehabilitation procedures (therapeutic exercises with an instructor, massage, physiotherapy, robotic mechanotherapy), the patients included in the study underwent motor imagery training of walking controlled by a NIRS-BCI system with feedback from a pneumatic orthosis. The course consisted of 7–13 training sessions, averaging 23 minutes each (one training session per day, comprising two approximately 11.5-minute runs), over a 3-week hospitalization period.

Eligibility criteria

Inclusion criteria:

• Male or female patients who provided informed consent to participate in the study;
• Clinically stable condition;
• Age 18–80 yrs;
• Primary or recurrent supratentorial ischemic or hemorrhagic stroke;
• Time since stroke from 7 days to 24 months inclusive;
• Mild-to-moderate post-stroke motor impairment of the lower limbs (10+ points on the Fugl-Meyer Assessment scale for the lower extremity).

The non-inclusion criteria were:

• Absence of lower limb motor impairment as assessed by the Fugl-Meyer Assessment scale (34 points on the lower extremity section);
• Patient’s refusal to participate in the study;
• Cognitive impairments preventing the ability to follow the physician’s instructions, including speech disorders (sensory aphasia, severe motor aphasia);
• Severe visual impairment preventing the ability to see visual instructions on a computer screen;
• Lower limb motor impairments of a nature other than post-stroke;
• Tissue contractures in the affected lower limb;
• Acute illnesses, exacerbations of chronic diseases, or acute life-threatening conditions after inclusion in the study;
• Inability to maintain a sitting position due to pain syndrome or other reasons.

The analysis included data from patients who completed at least 7 training sessions (14 training runs).

Clinical Assessment

To assess lower limb motor function before and after the rehabilitation, the following were used: the lower limb section of the Fugl-Meyer Assessment scale (maximum possible score — 34 points [20]), the 10-Meter Walk Test [21], and the Timed Up and Go test [22].

Furthermore, after each training session, the patient was interviewed about their well-being and any possible adverse events.

Training Procedure Using the NIRS-BCI-Pneumoorthosis System

Motor imagery training was conducted using a non-invasive BCI based on recording the hemodynamic response in the cerebral cortex via NIRS. Feedback on the brain signal recognition corresponding to motor imagery was provided visually on a monitor screen and tactilely via the Korvit pneumoorthosis.

The NIRS-BCI-pneumoorthosis system (Fig. 1) included:

• the NIRSport2 hardware-software complex (NIRx Medical Technologies), containing 16 light sources and 16 detectors;
• software for classifying brain activity and providing real-time feedback;
• a 22-inch monitor for displaying tasks and feedback;
• a pneumoorthosis for reproducing weight-bearing load on the leg — the Korvit system (VIT).

 

Fig. 1. Experimental setup during a session: positioning of the participant, monitor with instructions, and pneumoorthosis.

 

During the procedure, a NIRSport cap equipped with light sources and detectors. The patient was seated, and their lower limbs were secured in pneumatic orthoses of the Korvit system. To control the device via BCI using Bluetooth, a command system provided by the manufacturer after a firmware update was used.

A circle for gaze fixation was displayed in the center of the monitor screen, with two arrows positioned above and below it that changed color to indicate instructions. The patient was instructed to kinesthetically imagine walking (down arrow) or to relax (up arrow) while maintaining their gaze fixed on the center of the screen. The instructions to imagine locomotion and to relax alternated with each other and were cued by a change in color of one of the two arrows. Upon correct recognition of the signals corresponding to the movement imagination (walking), the marker in the center of the screen changed color to green, and the pneumatic orthosis applied a supportive load to the feet with an increased frequency, simulating an acceleration of the step. If the task was recognized incorrectly, the marker color did not change, and the orthosis operated in the basic mode. During the presentation of the “relax” instruction, no feedback was provided, and the orthosis operated in the basic mode. The duration of each instruction randomly varied from 13 to 23 seconds. Each instruction was preceded by a 3-second preparatory interval. In total, the session consisted of 16 pairs of instructions (rest + movement imagination).

From all source–detector pairs of the NIRS system, 51 channels were selected (Fig. 2). The channel positions corresponded to the positions of fp1, af7, nz, fp2h, fp2, fp1h, afpz, aff1, af3, af4, af8, f6h, ffc3h, f5h, fc3, aff2h, ffc4h, fz, fc2h, fcc5h, c5, fc1h, fcc3h, fccz, c1h, fc4, fcc4h, c4, fcc6h, ccp6h, c6, c2h, ccpz, ccp4h, cp2h, cp4, p2, ppo4, ppo2h, pooz, po2, c3, ccp5h, ccp3h, cp3, cp1h, p1, pz, ppo1h, ppo3, po1 of the 10-5 system. The sampling frequency was 27.1 Hz.

 

Fig. 2. Arrangement of NIRS channels on the cap projection. Top — frontal part, bottom — occipital. Red circles — sources, blue — detectors, black numbers — channel numbers.

 

The BCI system used to control the walking simulator was implemented in the MATLAB 2022b software environment. During the experiment, signals for the two wavelengths were converted into local concentrations of HbO and HbR, followed by frequency filtering. A Chebyshev type I filter with passband edges of 0.018 and 0.031 Hz was used for filtering. The mean concentration values from each channel, obtained over a one-second segment, were combined into a single feature vector. Feature vectors were calculated every 100 ms. After each pair of instructions was presented, the classifier was updated. Regularized linear discriminant analysis was used for classification. Data from already recorded sessions of each participant were used to retrain the classifier, which allowed feedback to be initiated immediately after the start of each session, except the first one. The threshold for brain signal classification accuracy was considered to exceed the level of random guessing at values above 50%, since the task involved distinguishing between two states. The signal processing and classification method is described in detail in [23].

Statistical Analysis

Statistical analysis was performed using the Statistica v. 6 software package (StatSoft, Inc.). Due to the small sample size, non-parametric methods were used for statistical analysis of the results (the Wilcoxon test and Spearman’s correlation analysis). The level of statistical significance was set at p < 0.05. Quantitative data are presented as median, 25th and 75th percentiles.

Results

Patient characteristics

The study included 17 post-stroke patients.

Two participants withdrew from the study prematurely. Patient 1: female, 75 years old, cortical ischemic stroke with a duration of 4 months. Baseline measures: Fugl-Meyer Assessment scale for the lower extremity — 20 points; maximum barefoot walking speed on the 10-meter walk test — 0.63 m/s; Timed Up and Go test — 17 s. Patient 2: male, 61 years old, cortical ischemic stroke with a duration of 9 months. Baseline measures: Fugl-Meyer Assessment scale for the lower extremity — 19 points; maximum barefoot walking speed on the 10-meter walk test — 0.91 m/s; Timed Up and Go test — 21 s.

Fifteen patients (6 women and 9 men) completed the course with at least 7 training sessions using the fNIRS-BCI-pneumoorthosis technology. Their median age at inclusion was 58.0 [47.0; 64.0] years (range 37–71 years), and the median time since stroke was 6.0 [3.0; 9.0] months (range 1–11 months). Two patients had a recurrent stroke, while the others had a primary stroke; in 8 (53.3%) patients, the lesion was located in the right hemisphere, in 6 (40.0%) — in the left hemisphere, and in 1 (6.7%) — bilaterally with one focus in each hemisphere. Cortical lesion was present in 9 (60.0%) patients, subcortical — in 4 (26.7%), and cortico-subcortical — in 2 (13.3%). Eleven (73.3%) patients had an ischemic stroke, and 4 (26.7%) had a hemorrhagic stroke.

No statistically significant differences in baseline characteristics were found between those who withdrew and those who completed the study.

Indicators of the NIRS-BCI–pneumatic orthosis system control

A total of 147 training sessions (292 runs) were conducted for patients using the NIRS-BCI–pneumatic orthosis technology; the median number of training days was 10.0 [9.0; 10.0]. The variability in the number of training sessions was low: 11 out of 15 patients completed 9–10 training sessions, while 1 patient each completed 7, 8, 11, and 13 sessions, respectively. The median total exposure per patient was 229.42 [209.56; 232.12] min, the median total time of feedback presentation via the pneumatic orthosis per course was 80.80 [74.14; 82.00] min, or 8.08 [7.82; 8.19] min per single training session.

The median BCI classifier recognition rate for patients’ mental states was 54.93% [53.10; 69.70]. The median of the maximum achieved recognition rates was 75.57% [67.54; 87.14]. These values varied among different patients (Fig. 3). For individual patients, the best control attempts occurred on different days, and no clear dynamics of their change were observed. At the same time, 10 participants (participants 1, 3, 4, 6–10, 12, 14 in Fig. 3) managed to demonstrate BCI control accuracy of 70% and above. No correlation was found between the patient’s age or disease duration and BCI control accuracy.

 

Fig. 3. BCI control accuracy in individual patients. Dots represent data from individual training sessions. Boxes indicate medians, whiskers indicate the 25th–75th percentiles.

 

No pronounced group dynamics in BCI control accuracy were observed during training (Fig. 4).

 

Fig. 4. BCI control accuracy on individual training days. Dots represent data for individual patients. Boxes indicate medians, whiskers indicate the 25th–75th percentiles.

 

The control signals for the fNIRS-BCI were an increase in HbO concentration and a decrease in HbR concentration during motor imagery (Fig. 5). Fluctuations in HbO and HbR concentrations, reflecting the activation and subsequent deactivation of cortical regions, were most pronounced in the supplementary motor (channels 8, 16, 18) and premotor (channels 12, 14) cortex areas, as well as near the leg motor representation area of the primary motor cortex (channels 19, 22). For these channels, a typical curve shape was characteristic: an increase in HbO and a synchronous decrease in HbR in the post-stimulus period, which corresponds to the physiological hemodynamic response to cortical activation.

 

Fig. 5. Hemodynamic response in NIRS channels in the frontoparietal (8, 12, 14, 16, 18) and primary motor (19, 22) cortical regions in all patients during BCI control sessions. On each graph, the period to the right of the vertical lines represents walking imagery, and to the left represents rest. The blue curve indicates the relative change in HbO concentration, and the red curve indicates HbR.

 

Changes in the lower limb motor function indicators over time

A significant (p < 0.05) improvement in scores was observed on the Fugl-Meyer Assessment scale and the Timed Up and Go test during the course. According to the 10-meter walk test data, no statistically significant increase in walking speed was found after the rehabilitation course with BCI control training (Table 1).

 

Table 1. Changes in the lower limb motor function indicators before and after the rehabilitation course supplemented NIRS-BCI-pneumatic orthosis training

Parameter	Before the course (n = 15)	After the course (n = 15)	p
Fugl-Meyer Assessment scale for lower extremity, points	19.0 [16.0; 24.0]	24.0 [20.0; 25.0]	0.001
Comfortable speed, with shoes, m/s	0.68 [0.56; 0.85]	0.77 [0.63; 0.93]	0.112
Maximum speed, without shoes, m/s	0.81 [0.71; 1.25]	1.05 [0.84; 1.22]	0.730
Comfortable speed, without shoes, m/s	0.77 [0.52; 1.00]	0.77 [0.59; 1.02]	0.600
Maximum speed, without shoes, m/s	0.90 [0.67; 1.15]	1.00 [0.83; 1.25]	0.096
Timed Up and Go, s	17.61 [14.21; 22.34]	16.06 [13.00; 17.41]	0.041

Note. After applying the Bonferroni correction for multiple comparisons (6 tests), the adjusted significance threshold was 0.0083.

 

When applying the Bonferroni correction for multiple comparisons (6 analyzed indicators, adjusted significance level p < 0.0083), statistical significance was retained only for the Fugl-Meyer Assessment scale (p = 0.001). Differences on the Timed Up and Go test (p = 0.041) did not reach the adjusted significance threshold.

Correlation analysis did not show a statistically significant relationship between the number of training sessions conducted and the magnitude of change in clinical indicators.

Tolerability and safety of NIRS-BCI-pneumatic orthosis training

Two participants experienced adverse events leading to premature discontinuation: a 75-year-old female patient (ischemic stroke 4 months prior) reported discomfort in the healthy leg during pneumatic orthosis operation and refused further training after the first session, and a 61-year-old male patient (ischemic stroke 9 months prior) discontinued participation after 3 -sessions due to significant difficulties with concentration and falling asleep during the procedures.

One female patient discontinued participation after the 7th training session due to significant discomfort from the NIRS cap; however, her data were included in the analysis. Another female patient reported fatigue during the sessions but completed the full course.

The remaining 11 patients satisfactorily tolerated training with the NIRS-BCI-pneumatic orthosis, reporting no complaints.

Discussion

When managing the NIRS-BCI-pneumatic orthosis system in post-stroke patients, high recognition accuracy rates were achieved: the median of the maximum achieved recognition scores exceeded 75%. Changes in HbO and HbR concentrations in the frontoparietal and motor cortical regions were used as the control signal for the BCI. The obtained data confirm the feasibility of the proposed technology for patients with post-stroke gait impairment.

To date, the use of NIRS-BCI for gait restoration after stroke has been studied in only one randomized controlled trial [24] and one pilot study [25]. In the study [24], feedback was provided only visually, without the use of technical devices to affect the limbs. In the pilot study [25], 7 patients underwent only 1 session each of NIRS-BCI control with a cycle ergometer integrated into the system. Neither of these studies utilized a motor imagery paradigm for BCI control: in the study [24], the system was based on classic neurofeedback using brain activity signals, while in the study [25], signals associated with movement preparation were recorded.

Upon completion of the rehabilitation course incorporating training with the NIRS-BCI–pneumatic orthosis, improvement in lower limb motor function was noted on the Fugl-Meyer Assessment scale and the Timed Up and Go test. Repeated analysis using the Bonferroni correction confirmed the reliability of the improvement on the Fugl-Meyer Assessment scale (p = 0.001); however, the results of the Timed Up and Go test should be interpreted with caution, as they lost statistical significance after correction. Changes in the 10-Meter Walk Test did not reach the level of statistical significance. This discrepancy may be related to the short observation period, limited number of training sessions, as well as the differential sensitivity of the scales used: the Fugl-Meyer Assessment scale evaluates individual motor elements, the Timed Up and Go test assesses dynamic balance, while the 10-Meter Walk Test reflects the complex skill of walking. Given the pilot nature of the study and the small sample size, the obtained data require confirmation in larger studies.

The absence of a statistically significant correlation between the number of training sessions and the clinical effect may be related to the narrow range of variability in the number of sessions (11 out of 15 participants had 9 or 10 sessions). Determining the adequate number of training sessions within the standard duration of an inpatient rehabilitationshould become one of the priority tasks for subsequent controlled studies with a larger sample.

Although the maximum individual control accuracy for the BCI reached 75.57%, the median accuracy for the group was 54.93%, which is only slightly above the level of random guessing (50% for binary classification). We attribute these results to the high complexity of the experimental task: participants were required to consistently maintain a state of kinesthetic motor imagery over a long interval — from 13 to 23 seconds, which is due to both the technical features of the BCI system and NIRS characteristics. In the configuration used, changes in walking speed cannot be instantaneous, which forced subjects to perform several consecutive steps within a single attempt of motor imagery. Nevertheless, 10 out of 15 participants demonstrated control accuracy at the level of 70% and above. These findings support a favorable assessment of the developed NIRS-BCI–pneumatic orthosis system and suggest its potential for further clinical application.

This study expands the existing body of scientific evidence on the clinical application of BCIs by demonstrating the feasibility of a training protocol that integrates NIRS-BCI technology with a pneumatic orthosis. The system stimulates foot loading based on the recognition of brain signals associated with motor imagery. Although vertical training is preferable for gait rehabilitation, as it better reproduces the functional conditions of movement, motor imagery training using BCI technology can also be conducted in a sitting position for patients with severe balance disorders or an inability to stand independently, since these systems do not require body weight support devices.

According to the results of this study, the mean total activation time of the feedback—and, consequently, of the pneumatic orthosis — during successful motor imagery recognition was 8 minutes per training session, or approximately 80 minutes over a 10-session course. Whether this duration of exposure is sufficient to provide an advantage of BCI-controlled training over passive pneumatic orthosis stimulation administered for the full 26-minute session remains an open question that requires further comparative studies. Extending the total session duration is challenging due to patient fatigue during motor imagery tasks — a persistent obstacle to the clinical application of post-stroke BCIs [6]. Given the technology’s good tolerability and drawing on fundamental principles of neurorehabilitation, a viable solution for future studies may be to increase the total number of training sessions, rather than the per-session duration. Nevertheless, a randomized placebo-controlled trial by Mihara et al. demonstrated the superiority of NIRS-BCI technology for gait restoration with a feedback duration of approximately 5 minutes per session over a course of 6 sessions [24].

The main limitations of the present study are the small sample size and the absence of a control group, which preclude a definitive assessment of the technology’s clinical efficacy. However, the aim of this study was to evaluate the clinical applicability and technical feasibility of the NIRS-BCI system integrated with a pneumatic orthosis as part of a comprehensive post-stroke rehabilitation program. The successful implementation of this system within an inpatient rehabilitation setting provides a rationale for future controlled trials to assess the efficacy of this approach.

Conclusion

This study confirmed the feasibility of recording and classifying cortical activity signals in post-stroke patients using NIRS within the NIRS-BCI–pneumatic orthosis system. Analysis of hemodynamic signals obtained via NIRS revealed characteristic physiological activation patterns, including increased oxygenated hemoglobin (HbO) and decreased deoxygenated hemoglobin (HbR) concentrations in the motor and associative regions of the cerebral cortex.

Following a 10-session training course using the NIRS-BCI–pneumatic orthosis system as an adjunct to a standard motor rehabilitation program, patients showed improvements in the lower extremity subsection of the Fugl–Meyer Assessment scale and the Timed Up and Go test. Most patients tolerated the procedures well; 2 out of 17 patients withdrew from the study.

A randomized controlled trial is warranted to evaluate the clinical efficacy of the NIRS-BCI–pneumatic orthosis technology. Given the limitations of the functional outcome scales used in the present study, future research should incorporate more sensitive assessment methods, such as video-based gait analysis and measurement of changes in serum neurotrophic factor concentrations. Furthermore, given the high interindividual variability in BCI control performance with the NIRS-BCI–pneumatic orthosis system, future studies should identify factors that predict successful control.

 

¹ BCI society. URL: https://bcisociety.org/bci-definition (data of access: 15.12.2025)..

About the authors

Olesya A. Mokienko

Russian Center of Neurology and Neuroscience

Author for correspondence.
Email: lesya.md@yandex.ru
ORCID iD: 0000-0002-7826-5135

Cand. Sci. (Med.), researcher, Group of neural interfaces, Institute of Neurorehabilitation and Restorative Technologies

Russian Federation, Moscow

Svatlana A. Pak

Russian Center of Neurology and Neuroscience

Email: lesya.md@yandex.ru
ORCID iD: 0009-0009-7188-9365

junior researcher, Group of neural interfaces, Institute of Neurorehabilitation and Restorative Technologies

Russian Federation, Moscow

Ekaterina S. Ikonnikova

Russian Center of Neurology and Neuroscience

Email: ikonnikovaes@list.ru
ORCID iD: 0000-0001-6836-4386

junior researcher, Group of neural interfaces, Institute of Neurorehabilitation and Restorative Technologies

Russian Federation, Moscow

Roman Kh. Lyukmanov

Russian Center of Neurology and Neuroscience

Email: xarisovich@gmail.com
ORCID iD: 0000-0002-8671-5861

Cand. Sci. (Med.), Head, Group of neural interfaces, Institute of Neurorehabilitation and Restorative Technologies

Russian Federation, Moscow

Mikhail R. Isaev

Pirogov Russian National Research Medical University

Email: shycmympuk@yandex.ru
ORCID iD: 0000-0002-3907-5056

junior researcher, Department of brain-computer interfaces

Russian Federation, Moscow

Anastasiia N. Cherkasova

Research Center of Neurology and Neuroscience

Email: lesya.md@yandex.ru
ORCID iD: 0000-0002-7019-474X

junior researcher, Group of neural interfaces, Institute of Neurorehabilitation and Restorative Technologies

Russian Federation, Moscow

Natalia A. Suponeva

Russian Center of Neurology and Neuroscience

Email: suponeva@neurology.ru
ORCID iD: 0000-0003-3956-6362
SPIN-code: 3223-6006

Dr. Sci. (Med.), Corr. Member of RAS, Director, Institute of Neurorehabilitation

Russian Federation, Moscow

Pavel D. Bobrov

Pirogov Russian National Research Medical University

Email: bobrov_pd@mail.ru
ORCID iD: 0000-0003-2566-1043

Cand. Sci. (Med.), senior researcher, Department of brain-computer interfaces

Russian Federation, Moscow

References

Supplementary files