**1. Introduction**

Gait and balance disturbances are common and important clinical manifestations of Parkinson's disease (PD), leading to mobility impairment and falls [1]. Current treatments (pharmacological and deep brain stimulation (DBS)) provide only partial benefits in gait derangements in PD, with a wide variability in outcomes [2–5].

Despite detailed testing, specific factors that are critical to predicting locomotor deterioration in PD remain elusive [6–9]. Beside subtle onset and clinical heterogeneity [10], technical limitations have hampered the timely and direct recording of supraspinal locomotor derangements in these patients. Only recently have advances in portable electroencephalography systems [11,12] and new DBS devices capable of on-demand recording using chronically implanted electrodes (e.g., Activa PC+S and Percept PC (Medtronic PLC)

**Citation:** Haufe, S.; Isaias, I.U.; Pellegrini, F.; Palmisano, C. Gait Event Prediction Using Surface Electromyography in Parkinsonian Patients. *Bioengineering* **2023**, *10*, 212. https://doi.org/10.3390/ bioengineering10020212

Academic Editor: Christina Zong-Hao Ma

Received: 3 January 2023 Revised: 31 January 2023 Accepted: 3 February 2023 Published: 6 February 2023

**Copyright:** © 2023 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

or AlphaDBS (Newronika Srl)) [13–15] enabled the recording of ongoing brain activity during actual gait in PD to be performed [16–18].

The precise assessment of gait dynamics should account for its context dependency. New study setups employing fully immersive virtual reality (VR) or augmented reality allow gait assessment (with optoelectronic systems, force plates, etc.) to be conducted in environments that deliver patient-specific triggers of gait impairment (e.g., [19]). These setups could facilitate the identification of biomarkers for the fine-tuning of therapy delivery, e.g., adaptive DBS programming and so-called VR Exposure Therapy [20].

An open challenge is the continuous monitoring of gait parameters in laboratory as well as real-world environments. Technically, parameters such as the timings of heel strike and toe-off events, which define swing and stance phases and provide valuable information about cadence patterns, etc., can be assessed with optoelectronic systems and force plates. Both systems, however, are expensive, require qualified personnel, and do not offer monitoring in ecological settings. Video-based analyses of gait have also been proposed [21], although it is unclear whether these could reach the required precision to identify individual events within a gait cycle, especially for clinical applications and in ecological settings. Wearable motion sensors such as inertial measurement units (IMUs) are another viable option to capture gait events in natural environments with high temporal accuracy [22,23]. However, they do not contain further neurophysiological information that may be crucial to understanding and predicting gait derangements [24]. Surface electromyography (EMG) provides the missing link between neural signals and kinematics that makes the comprehensive characterization of pathological gait possible. EMG measurements have been used to predict lower-limb motion in advance [25,26] for real-time control of a prosthesis [27–29] or adaptive DBS devices [16–18,25,26,30,31]. EMG profiles of the gait cycle have also been shown to anticipate specific gait derangements in PD, such as freezing of gait [32], a sudden episodic inability to produce effective stepping despite the intention to walk. The combined use of IMU and EMG signals would make the description of the motor actions and intentions underlying gait kinematic features and alterations possible.

However, some practical limitations should be considered when applying additional sensors on severely ill patients, especially when performing recordings after suspension of medications. For example, in patients with PD, the overnight suspension of dopaminergic drugs is fundamental to evoke and study PD-related symptoms but greatly reduces the time window available for experimental recordings. Limiting the preparation period by limiting the number of sensors may help considerably in this regard. In addition, an excessive number of sensors may alter the natural behavior of subjects, undermining the advantages of working in ecological environments. Another crucial aspect is the cost of multiple sets of sensors. Considering that probes comprising both IMUs and EMG are generally more expensive than standalone solutions, the need for IMUs and EMG in the fine-grained evaluation of gait may be a limiting factor for many laboratories and applications in clinical routine. The use of multiple devices may also not be practical in clinical routine, as synchronization or different recording software may be needed.

Considering this, the development of novel technologies that can extract multiple types of signals from the same set of sensors is highly desirable. While the same kinematics can be produced by different muscular patterns, lower-limb kinematics can be inferred using analysis of EMG [33]. The idea of detecting gait events directly using EMG signals, circumventing additional IMUs or force plates, is gaining traction [34–38]. Ziegier and colleagues [38] reported high accuracy in classifying stance and swing phases during human gait based on EMG recordings. They first extracted a weighted signal difference that exploits the difference in EMG activity between corresponding muscles of the two legs and then trained a support vector machine to classify the gait phases. Using a deep learning approach, Morbidoni and colleagues [35] were also able to classify stance and swing phases and predict foot–floor contacts under natural walking conditions in healthy subjects. Other studies showed similar results in learned and unlearned subjects [37], and using intra-subject training only [34]. This would not only simplify future recording setups

but also permit the re-analysis of EMG datasets recorded without IMUs or in cases of data loss due to technical problems with the IMU to be performed. This second scenario is particularly problematic when recordings cannot be repeated due to the patient's clinical condition. Additionally, the extraction and prediction of gait events using lower-limb EMG activity is of fundamental importance for the development of an EMG-driven prosthesis, where predicting the subsequent gait phase using muscular signals increases prosthesis efficiency and responsiveness [33].

Previous approaches aimed to predict discrete gait events (i.e., heel contact and toeoff), and little attention has been paid to reconstructing the time course of the relevant kinematic variables. Brantley and colleagues were able to predict knee and ankle kinematics, but with varying accuracy across trials and the subjects included [33]. In addition, most previous studies focused on treadmill walking, which may not capture the variability of ground walking [36–38]. Lastly, we are not aware of any approach that has been tested on unmedicated PD patients during long periods of continuous walking.

In the present study, we explore the possibility of identifying fundamental gait events from surface EMG in parkinsonian patients using a machine learning approach. Compared with previous studies, we did not frame the problem as one of detection (i.e., to identify the timings of a fixed set of events) or classification (i.e., to segment the data into contiguous gait phases). Instead, we used an innovative regression approach to approximate continuous angular velocity profiles as measured by IMUs. We consider this approach strictly more powerful and flexible than previous approaches, as access to the predicted IMU time series allows us not only to extract predetermined types of gait events but also biomechanical quantities such as joint angular velocity and further parameters on which our model has not been trained. Our study is further set apart from published work in that we focused on a clinical cohort rather than healthy participants. To our knowledge, our study is the first to demonstrate the feasibility of accurate gait parameter estimation using EMG in such a population. Remarkably, our approach accounts for the substantial across-patient variability observed in gait patterns of clinical populations, allowing it to be applied without any patient-specific calibration.

#### **2. Materials and Methods**
