Towards Ultrasound Imaging-based Closed-loop Peripheral Nerve Stimulation for Tremor Suppression

Abstract

Despite several decades of research investigating the use of peripheral electrical stimulation (ES) for tremor suppression in an upper limb, ES design for effective tremor suppression remains elusive. The article reviews sensing approaches to measure limb tremors and existing musculoskeletal models of tremor and their use in closed-loop suppression control. We also motivate a case for incorporating ultrasound (US) imaging into the closed-loop control for increased tremor suppression efficacy. When combined with wearable US transducers, the novel approach could be a promising technique to advance musculoskeletal models that investigate tremor mechanisms and new ES closed-loop techniques with personalized stimulation parameters for tremor suppression.

Introduction.

Tremors are involuntary oscillations of muscles in the hands and arms, affecting over 11 million people in the United States alone [1]. Neurological disorders such as Essential Tremor (ET), Parkinson’s disease (PD), dystonia, and cerebellar ataxia cause such tremors, making activities of daily living (e.g., buttoning a shirt, shaving, preparing a meal, drinking a glass of water, etc.) challenging and increasing one’s reliance on care providers. Social implications, combined with the often-progressive nature of the underlying neurological disorder, increase their risk for depression and other negative impacts on mental health. This review discusses data-driven musculoskeletal modeling and closed-loop control via US imaging as a new approach to suppress tremors via peripheral nerve stimulation.

Tremor Suppression Approaches.

Existing approaches for tremor suppression are lifestyle modifications, medication, neurosurgery, and wearable devices such as Cala Trio [2]. Lifestyle modifications with weighted utensils and assistive writing devices usually address mild tremors [3]. Medications and brain surgery become necessary treatments for functionally disabling tremors. Initial response rates of the current tremor suppression methods vary, while frequent treatment tolerance plagues chronic benefits. Most medications reduce tremor amplitude by about 50% but can have side effects, including nausea, sedation, loss of appetite, and low blood pressure [4]. The most common brain surgery to help with tremor suppression is Deep Brain Stimulation (DBS), in which a surgically inserted electrode stimulates the thalamus to control tremor output. DBS has drawbacks, including its invasive nature and cost and stimulation-induced side effects like gait and speech abnormalities. Adverse effect rates from DBS, including the risk of cognition, mood, speech, and gait disturbances, range from 14–38% in patients with Parkinson’s disease [5]. Due to the adverse effects of DBS, noninvasive means for brain neuromodulation, such as through ultrasound, are under investigation [6].

Our review focuses on external devices’ use to suppress tremor as a potential alternative for patients who do not want or are ineligible for DBS surgery. Robotic exoskeletons that mechanically change joint impedance have tremor suppression efficacy of 30–98% [7]. Weight and bulkiness remain the main criticisms of robotic exoskeletons. Lightweight, compact, and minimally intrusive orthotic mechanisms while providing desired tremor suppression are suggested for future research [7]. Alternatively, electrical stimulation of tremoring muscles, which applies low-level electrical currents via adhesive transcutaneous electrodes (or intramuscular wires), is a non-cumbersome and effective intervention to suppress tremor with high user acceptance [8, 9]. Electrical muscle stimulation is applied in a phase-locked loop [10] or simultaneously to antagonist muscles to increase the joint stiffness [9, 11]. Both strategies use high pulse intensity to activate efferent nerve fibers and elicit muscle contractions. However, this approach has several drawbacks, such as a rapid onset of stimulation-induced fatigue, potential discomfort due to high-intensity stimulation, and interference with voluntary movements. These issues hinder the long-term effectiveness of the stimulation.

Recent studies suggested stimulating afferent fibers can be equally effective for tremor suppression instead of efferent fibers [12]. The muscles were stimulated below the threshold of direct muscle activation to avoid stimulation of efferent fibers [8, 13–20] and target spinal afferent pathways for tremor suppression. The strategy is potentially fatigue resistant and comfortable due to low stimulation intensity. However, results on tremor suppression using this technique differ across patients and neurological conditions that cause tremors. On average [14], Hao et al. [19] reported 61.5% of tremor suppression across all assessed joints. Dideriksen et al. [13] reported 52% of average optimal suppression across all subjects; however, tremor increased in some patients. Jitkritsadakul et al. [20] obtained an average percentage of improvement in tremor suppression of 43.81%. Dosen et al. [8] reported 42% tremor suppression using afferent stimulation. Finally, Heo et al. [16] showed 40% of tremor reduction at the wrist joint during stimulation.

Inconsistent tremor suppression results are likely due to a lack of consensus on an afferent stimulation strategy, which can be random, continuous, or out-of-phase to the tremor. Further complicating stimulation strategies is the current differing pathophysiological mechanisms proposed for the various etiologies of tremor. Moreover, the heuristically chosen stimulation parameters may need more personalization to deliver consistent performance.. Systematically determined stimulation parameters through musculoskeletal tremor models may help attain effective desired suppression performance.

Quantifying Limb Tremor-related Muscle Activity via US Imaging.

The first step towards designing an effective closed-loop suppression is to choose the most appropriate sensing technology that aids in modeling afferent mechanisms and controlling tremor suppression. Afferent fibers conduct information on muscle velocity and position back to the spinal cord. Therefore, it will be important to use sensing technology to measure muscle velocity and position changes and facilitate the modeling of afferent stimulation mechanisms. Different instrumentation/techniques have been used in tremor suppression control, namely accelerometry or inertial measurement units (IMU) [8], 3D gyrosensors [17] and 3D motion capture s [18, 19]. Cala Trio (CalaHealth, Inc.) is a recent FDA-approved Essential Tremor therapy device that uses accelerometry to detect oscillatory limb displacements[21]. However, accelerometry measures a summated limb displacement (tremor effect), not actual muscle tremor frequency, which is key muscle-specific information for effecting antagonistic muscle stimulation. Limb displacement oscillations at a joint being mechanically coupled to other limb joints are affected by a tremorgenic torque input from a distant joint [22, 23]. As the tremor originates from the muscles, not from joints, measuring tremor characteristics directly from the tremoring muscles is preferable in a closed-loop suppression design. Therefore, current research studies of afferent stimulation envisage using electromyography (EMG) for a closed-loop system [8, 13]. EMG can detect tremorgenic inputs to a muscle, and its analysis can determine tremor frequency, which is especially useful as one can design a proper timing of afferent stimulation. EMG, though more advantageous than accelerometry, has drawbacks. First, electrical stimulation is turned-off during the EMG recording phase to prevent stimulation artifact interference. The recording phase is critical for measuring tremor characteristics such as tremor phase and frequency. These measurements inform the stimulation timing during the stimulation phase during which EMG recording is turned off. This type of recording and stimulation is discontinuous, and does not facilitate continuous tremor suppression. Thus, EMG-based tremor suppression works in an on-off-on-off fashion [8]. Further, crosstalk from neighboring muscles in the EMG signal is difficult to decouple. Lastly, both EMG and accelerometry cannot directly measure muscle velocity and position, which are correlates of afferent feedback that may play a significant role in tremor propagation.

Alternative sensing technology, such as U.S. imaging, may be more useful in modeling the afferent mechanisms. As a sensing approach, U.S. imaging provides direct visualization of superficial and deeper muscles, is unaffected by signal interference from neighboring muscles and stimulation artifacts. Thus, it may be a more suitable sensing modality to locate and identify tremor-causing muscles. Most importantly, US imaging has been shown to measure muscle features or muscle contractility that include tissue displacement in axial and lateral directions or muscle contractility [24–26]. Thus, as in Fig. 1, US feedback can be used to model afferent fiber activity that provides feedback of muscle velocity and position changes to the spinal cord.

Figure 1:

Wearable ES and ultrasound for Tremor Investigation and Suppression

Our research group recently performed a feasibility study on characterizing tremor activity with US-derived muscle contractility signals [27]. We recruited two subjects with no tremor, one subject with Parkinsonian tremor, and one subject with Essential Tremor. Subjects grasped a solid object and held the grasp until asked to release. A clinical linear transducer (L7.5SC Prodigy Probe, S-Sharp, Taiwan) connected to the ultrasound system (Prodigy, S-Sharp, Taiwan) imaged the contraction of the muscle, as shown in Fig. 2. A 2-D speckle-tracking algorithm was implemented to track frame-to-frame pixel displacement, and strain in a region of interest (ROI) was analyzed [24]. In Fig. 2, the results show a frequency of 2 Hz, which is closer to the actual grasp movement frequency in a subject with no tremor (S1). A higher frequency of 6–7 Hz is visible in the hold portion of the data obtained from a subject with tremor (S3). The frequency is between 4–12 Hz, reported as the dominant tremor frequencies across a spectrum of tremor-causing disorders [28]. These feasibility studies show that ultrasound determines tremor frequency.

Figure 2.

A tremor subject was holding a solid object while the forearm muscle was imaged with a commercial ultrasound transducer. A frequency spectrum of the strain-time graphs is plotted for the later phase of one of the experimental trials when the subjects hold the object. In the subject with tremor (S3), a tremor frequency during the hold phase can be observed, which is absent in the able-bodied subject trial (S1).

Table 1 and Fig. 3 compare and summarize the advantages and disadvantages of EMG and IMU sensors with US imaging. US imaging enables direct visualization of targeted muscles. Unlike IMU, EMG and US imaging (Fig. 3) can provide muscle-specific tremor frequency information. Moreover, due to direct visualization of the muscle, one can more confidently extract muscle-specific information than in the case of EMG, which can be affected by the signals from neighboring muscles and artifacts. US imaging, though, currently lacks wearability when compared to EMG and IMU sensors. The following section discusses an emerging area of wearable ultrasound sensors.

Table 1:

Comparison of advantages and disadvantages of two main popular sensors when compared to US imaging

Sensor/Characteristics	EMG	US Imaging	IMU
Tremor frequency	Muscle Specific	Muscle Specific	Summated Frequency
Stimulation Artifacts	Affected	Unaffected	Unaffected
Muscle Specificity	Low-to moderate	High due to direct visualization	Lacking
Wearability	Feasible	Currently lacking	Feasible

Figure 3:

Comparison of three sensing modalities: IMU, EMG, and US imaging during the No Stimulation and Stimulation On conditions. These preliminary results were obtained from participants with Parkinsonian Tremor. a) IMU measurements provide summated wrist angular velocity, unaffected by stimulation artifacts. b) Stimulation artifacts clearly affect E.M.G. measurements as the signal completely saturates when the stimulation is on. c) Ultrasound measurements of tissue displacements obtained from both wrist flexor and extensor are unaffected by stimulation artifacts.

Wearable Ultrasound Sensing

Ultrasound has been widely adopted for biomedical applications due to its non-invasive nature, accuracy, and real-time capabilities. It is capable of measuring variations in blood pressure depending upon changes in artery deformation [29], evaluating muscle fatigue during functional electrical stimulation (FES) [30], and identifying brain activation based on variations in blood volume [31]. Nevertheless, conventional US transducers are rigid and do not conform to the shape of a limb, making them unsuitable for biomedical applications demanding wearability.

Advances in flexible devices, such as electrical circuits, substrates, and transducers, have enabled the development of wearable US biomedical applications. Consequently, these advancements pave the way for future long-term disease diagnosis and health monitoring[32]. In contrast to traditional US transducers, flexible/wearable US devices may be easily adapted to areas of the body that are curved [33]. It allows continuous data collection and real-time analysis without restricting target tissue movement or transducer shifting. Some studies used wearable US sensing to detect upper limb muscle features, including muscle activity monitoring [34], wrist/hand motion prediction and pattern recognition [35], and prosthetic control [36].

Despite the many potential advantages of wearable ultrasound, some challenges in its implementation remain. An acoustic coupling medium, such as a US-gel-based coupling medium, is often required between the transducer and the body surface. Due to its tendency to dry out over time, this is unsuitable for long-term wear. Alternatively, a dry couplant can also be integrated to design a skin-friendly wearable US transducer. Regarding acoustic impedance, it has a similar value to water, has low acoustic attenuation, and is durable [37]. However, the inherent flexibility of the transducer can lead to acoustic deformation, resulting in inaccurate measurements and decreased performance. A novel technique to reduce acoustic deformation, called deep learning time delay estimation, was reported in [38] to address these concerns. This method effectively reduces distortion in B-mode imaging and fixes incorrect time delays. Additionally, phase correction algorithms further improve image reconstruction on curved surfaces [39]. Finally, integrating US sensing technology with other sensing modalities, such as a laser scanning system to track the position of each element, may provide a more comprehensive and accurate assessment of the beam profile[40]. The potential of wearable US sensing technology to provide more precise and real-time data for tremor model development is promising. Further investigation is necessary to identify and address any potential issues associated with this technology.

Musculoskeletal Models of Tremor Can Aid Closed-loop Design

The agonist and antagonist muscles of pathological tremor often follow an out-of-phase activation pattern, in which the muscles alternate contractions at the tremor rate [41, 42]. Measured tremor frequency help inform open-loop sequential stimulation of peripheral nerves [43, 44]. Nevertheless, since open-loop algorithms are based on pre-calibrated stimulation patterns, they might not be appropriate for varying tremor phases and frequencies. They may have limited efficacy compared to closed-loop strategies[45]. Trigger-based closed-loop systems are achievable, which mainly turn-on-off stimulation when the tremor is detected.

Musculoskeletal models of tremor may play an important role in closed-loop systems because they can describe tremorgenic movements and voluntary movements by analyzing the mechanical properties of muscles and tendons, which are controlled by neural circuits and electrical stimulation. These models are instrumental in predicting limb tremor features such as tremor frequency, especially when direct measurement may be unavailable or in hypothesizing the mechanisms of tremor.

In the existing literature, the real-time tremor models are described as time-varying harmonic models, time-variant estimation models, and time-delay models, including the Weighted-Frequency Fourier Linear Combiner [46], Band-limited Multi-frequency Fourier Linear Combiner[47] and Autoregressive models (AR)[48] [49]. However, the models, as mentioned earlier, are usually simple oscillator models and can only provide millisecond-level tremor predictions [49]. In addition, these models only provide time series information about tremors without describing the tremor mechanism. Therefore, these models may be unsuitable for designing tremor suppression strategies.

Furthermore, neural oscillator models [50], neuromuscular models [51], and neurocomputational models [52] have been developed based on physics. Models of neural oscillators are usually composed of coupled neurons with self-inhibition, tonic input, and sensory feedback. According to [51], a bi-muscular arm model can be modeled by topologically pooling motor, sensory, and interneuronal spiking cells. Further incorporating propriospinal neurons that act as intermediary gating agents between cortical oscillatory signals and peripheral muscles [53, 54] can enhance this neuromuscular model. These models could track the recorded data and produce dynamically consistent neural spike patterns, muscle forces, and movement kinematics. Neuromuscular and musculoskeletal modeling can be used to study tremor dynamic and suppression control. These models can simulate neurophysiological tremor data and enable the exploration of afferent feedback loop interactions with cortical oscillatory signals and proprioceptive feedback. Using neuromuscular and musculoskeletal modeling can provide insight into the underlying mechanisms of tremor, improve our understanding of tremor suppression, and facilitate the development of more effective and efficient tremor suppression strategies.

Nevertheless, in practice, physics-informed models are difficult to model accurately due to the nonlinear and time-varying nature of tremor dynamics. Additionally, each subject would have different tremor model parameters. The difficulty of applying first principle approaches in tremor modeling and model parameter variations among patients motivates the development of a data-driven modeling method for tremor suppression controllers. The data-driven approach does not require prior knowledge of tremor dynamics (Fig. 4). Since the advent of miniature or wearable sensing technology, the feasibility of gathering a large volume of data has tremendously increased. Therefore, it is possible to identify unknown musculoskeletal dynamics coupled to supraspinal tremor circuitry by using data-driven modeling. These models personalize the approach of diagnostic acquisition and therapeutic delivery to each subject, regardless of tremor etiology.

Figure 4:

Tremor modeling and control using a data-driven approach.

An strategy based on a model-free but data-driven approach, such as Support Vector Machines (SVMs)[55], and a recurrent neural network (RNN)[56], was used to predict tremor motions. Notably, since neural network models are packaged as black boxes, which map inputs to outputs, they provide only limited insight into the development of tremor suppression control techniques. Moreover, obtaining the desired results from neural networks requires considerable tuning and computational resources due to the number of hidden layers and activation functions. Therefore, we assert that the data-driven approaches that create a dynamical mapping would be more insightful in designing tremor suppression control strategies. These approaches could capture nonlinear tremor neuromuscular dynamics from the available experimental sensory data such as EMG, and US-derived muscle features (see Fig. 4) during the tremoring process. They can help determine effective stimulation parameters and facilitate a model-based afferent stimulation control.

Our recent work proposed a Koopman-based method for system identification and a model predictive control (MPC) scheme to suppress tremors [57, 58]. The results of preliminary research demonstrate that US image data collected from ET or PD patients during grasping tasks, as shown in Fig. 5 (a) and (b), can be used to develop a data-driven approach for approximating the dynamic model of tremors and implementing the MPC scheme for tremor suppression. To address the issue of model prediction accuracy, we proposed a low-complexity recursive least squares (RLS) algorithm for online model updating.

Figure 5.

Data-driven tremor modeling and suppression control: a preliminary study. (a) Experimental setup for experiments, a participant with tremor holding a mug while the US transducer and IMU are strapped to the forearm to capture images of the FCR and wrist motion. (b) B-mode image with location of FCR marked. (c) Koopman-based tremor modeling prediction with experimental data. Figure adapted from [53, 54].

The effectiveness of the proposed online updating method for tremor modeling was verified using experimental data collected from an IMU sensor attached to a participant’s (essential tremor) wrist during grasping motion. Fig. 5 (c) compares the experimental data, the prediction with a fixed model, and the prediction with the proposed online updating method. The proposed method significantly improved the accuracy of model predictions, with normalized root mean square errors (NRMSE) for wrist angle and wrist angular velocity of 3.76% and 3.16%, respectively, compared to 17.78% and 17.43% for a one-step modeling.

Conclusion

Previous studies used IMUs and EMGs for closed-loop tremor suppression. While feasible, their efficacy remains limited due to many factors, including the discontinuous suppression to avoid stimulation artifacts while recording EMG for determining tremor frequency. We argue that adding wearable US sensors to obtain muscle tremor data via muscle fascicle length and velocity changes enables high-fidelity computation models. Of course, the feasibility of models requires advancements in wearable US sensing technology and data-driven methods to acquire tremor diagnostics. Improvements in wearable and body-conformable US sensors and data-driven models with direct information on muscles’ tremor activity hold significant potential for an effective closed-loop ES.