About

Levi Hargrove is a Professor of Physical Medicine and Rehabilitation and a Professor of Biomedical Engineering (by courtesy) at Northwestern University. His research interests include signal processing, pattern recognition, and myoelectric control of powered prostheses. Dr. Hargrove focuses on the research and development of clinically realizable myoelectric control systems with the goal of making these systems available to amputees in the near-term. His work involves developing innovative control systems for prosthetic devices, aiming to improve functionality and integration for users. He holds a BScE in Electrical Engineering, an MScE in Electrical Engineering, and a PhD in Electrical Engineering from the University of New Brunswick, completed in 2003, 2005, and 2008 respectively.

Research topics

• Computer Science
• Artificial Intelligence
• Medicine
• Simulation
• Physical medicine and rehabilitation
• Embedded system
• Psychology
• Biomedical engineering
• Human–computer interaction
• Mathematics
• Neuroscience
• Surgery
• Engineering
• Computer vision
• Statistics
• Computer hardware

Selected publications

• Simplifying rehabilitation control of lower-limb exoskeletons in five ambulation modes via dataset-driven state-machine calibration

  Journal of NeuroEngineering and Rehabilitation · 2026-03-20

  articleOpen access

  Lower-limb exoskeletons are a useful tool in rehabilitation settings as they can provide customized assistance to individuals during functional exercises. These approaches typically rely on state-machine-based control with impedance controllers tailored to different locomotion phases, ensuring appropriate assistance across various activities and environments. However, these methods necessitate lengthy calibration procedures, as many impedance parameters need to be fine-tuned to provide appropriate assistance for various activities (e.g., overground walking, ramps, and stairs). This study presents three contributions: (1) a state-machine-based control strategy for partial assistance lower-limb exoskeletons, (2) a computational method to extract reference trajectories from a benchmark dataset (Camargo et al. in J Biomech 119:110320, 2021), enabling the identification of state-machine controller parameters and simplifying calibration procedures and (3) a dataset of 19 healthy individuals walking in five walking conditions (overground walking, upstairs, downstairs, up ramps, and down ramps) using either the state-machine approach or a transparent controller. The state-machine controller produced in average more negative interaction power ($$-2.6\times 10^{-2}$$ W/kg) compared to transparent control ($$0.8\times 10^{-2}$$ W/kg), indicating greater user assistance. Preferred walking speed was notably faster with the state-machine controller, particularly on level ground, ramps and stairs ascent (25–32% increase). Kinematic analysis revealed closer alignment to able-bodied gait patterns with the state-machine controller, suggesting improved gait quality. At the same time, the dataset of the collected locomotion activities (dataset link) will constitute a new benchmark dataset for locomotion. In this work, we presented and evaluated a novel state-machine-based control strategy for partial-assistance lower-limb exoskeletons. In this approach, reference trajectories are extracted from a benchmark dataset, simplifying calibration procedures. Additionally, we provide a dataset of 19 healthy individuals using two exoskeleton controllers. The proposed controller will be applied to patient populations, while the dataset will serve as a valuable resource for advancing robust and effective control mechanisms through machine learning techniques.

  PublisherDOI

• Learning Therapist Policy from Therapist-Exoskeleton-Patient Interaction

  arXiv (Cornell University) · 2026-03-02

  preprintOpen access

  Post-stroke rehabilitation is often necessary for patients to regain proper walking gait. However, the typical therapy process can be exhausting and physically demanding for therapists, potentially reducing therapy intensity, duration, and consistency over time. We propose a Patient-Therapist Force Field (PTFF) to visualize therapist responses to patient kinematics and a Synthetic Therapist (ST) machine learning model to support the therapist in dyadic robot-mediated physical interaction therapy. The first encodes patient and therapist stride kinematics into a shared low-dimensional latent manifold using a Variational Autoencoder (VAE) and models their interaction through a Gaussian Mixture Model (GMM), which learns a probabilistic vector field mapping patient latent states to therapist responses. This representation visualizes patient-therapist interaction dynamics to inform therapy strategies and robot controller design. The latter is implemented as a Long Short-Term Memory (LSTM) network trained on patient-therapist interaction data to predict therapist-applied joint torques from patient kinematics. Trained and validated using leave-one-out cross-validation across eight post-stroke patients, the model was integrated into a ROS-based exoskeleton controller to generate real-time torque assistance based on predicted therapist responses. Offline results and preliminary testing indicate the potential of their use as an alternative approach to post-stroke exoskeleton therapy. The PTFF provides understanding of the therapist's actions while the ST frees the human therapist from the exoskeleton, allowing them to continuously monitor the patient's nuanced condition.

  PublisherDOI

• Proximal powered knee placement: a case study

  arXiv (Cornell University) · 2026-02-19

  articleOpen accessSenior author

  Lower limb amputation affects millions worldwide, leading to impaired mobility, reduced walking speed, and limited participation in daily and social activities. Powered prosthetic knees can partially restore mobility by actively assisting knee joint torque, improving gait symmetry, sit-to-stand transitions, and walking speed. However, added mass from powered components may diminish these benefits, negatively affecting gait mechanics and increasing metabolic cost. Consequently, optimizing mass distribution, rather than simply minimizing total mass, may provide a more effective and practical solution. In this exploratory study, we evaluated the feasibility of above-knee powertrain placement for a powered prosthetic knee in a small cohort. Compared to below-knee placement, the above-knee configuration demonstrated improved walking speed (+9.2% for one participant) and cadence (+3.6%), with mixed effects on gait symmetry. Kinematic measures indicated similar knee range of motion and peak velocity across configurations. Additional testing on ramps and stairs confirmed the robustness of the control strategy across multiple locomotion tasks. These preliminary findings suggest that above-knee placement is functionally feasible and that careful mass distribution can preserve the benefits of powered assistance while mitigating adverse effects of added weight. Further studies are needed to confirm these trends and guide design and clinical recommendations.

  PublisherOA PDF

• Learning Therapist Policy from Therapist-Exoskeleton-Patient Interaction

  ArXiv.org · 2026-03-02

  articleOpen access

  Post-stroke rehabilitation is often necessary for patients to regain proper walking gait. However, the typical therapy process can be exhausting and physically demanding for therapists, potentially reducing therapy intensity, duration, and consistency over time. We propose a Patient-Therapist Force Field (PTFF) to visualize therapist responses to patient kinematics and a Synthetic Therapist (ST) machine learning model to support the therapist in dyadic robot-mediated physical interaction therapy. The first encodes patient and therapist stride kinematics into a shared low-dimensional latent manifold using a Variational Autoencoder (VAE) and models their interaction through a Gaussian Mixture Model (GMM), which learns a probabilistic vector field mapping patient latent states to therapist responses. This representation visualizes patient-therapist interaction dynamics to inform therapy strategies and robot controller design. The latter is implemented as a Long Short-Term Memory (LSTM) network trained on patient-therapist interaction data to predict therapist-applied joint torques from patient kinematics. Trained and validated using leave-one-out cross-validation across eight post-stroke patients, the model was integrated into a ROS-based exoskeleton controller to generate real-time torque assistance based on predicted therapist responses. Offline results and preliminary testing indicate the potential of their use as an alternative approach to post-stroke exoskeleton therapy. The PTFF provides understanding of the therapist's actions while the ST frees the human therapist from the exoskeleton, allowing them to continuously monitor the patient's nuanced condition.

  PublisherOA PDF

• Proximal powered knee placement: a case study

  Open MIND · 2026-02-19

  preprintSenior author

  Lower limb amputation affects millions worldwide, leading to impaired mobility, reduced walking speed, and limited participation in daily and social activities. Powered prosthetic knees can partially restore mobility by actively assisting knee joint torque, improving gait symmetry, sit-to-stand transitions, and walking speed. However, added mass from powered components may diminish these benefits, negatively affecting gait mechanics and increasing metabolic cost. Consequently, optimizing mass distribution, rather than simply minimizing total mass, may provide a more effective and practical solution. In this exploratory study, we evaluated the feasibility of above-knee powertrain placement for a powered prosthetic knee in a small cohort. Compared to below-knee placement, the above-knee configuration demonstrated improved walking speed (+9.2% for one participant) and cadence (+3.6%), with mixed effects on gait symmetry. Kinematic measures indicated similar knee range of motion and peak velocity across configurations. Additional testing on ramps and stairs confirmed the robustness of the control strategy across multiple locomotion tasks. These preliminary findings suggest that above-knee placement is functionally feasible and that careful mass distribution can preserve the benefits of powered assistance while mitigating adverse effects of added weight. Further studies are needed to confirm these trends and guide design and clinical recommendations.

  DOI

• A Deep Learning Framework With Domain Generalization and Few-Shot Learning for Locomotion Mode Classification Across Users, Sessions, and Prostheses

  IEEE Transactions on Medical Robotics and Bionics · 2025-09-04

  articleOpen accessSenior author

  Transfemoral amputees don and doff their prostheses at least daily, making inter-session classification performance important for clinical implementation of locomotion mode classification algorithms. Here, we present a deep-learning framework based on domain-adversarial training and few-shot learning fine-tuning to classify locomotion modes in unseen sessions or subjects' data across different prosthesis models. We validated the approach with a leave-one-session-out analysis repeated five times and made comparisons to a prosthesis-specific classifier. The dataset was created by merging data from two different prosthesis models (Vanderbilt University, VU, Gen 2 and Gen 3 powered knee-ankle prostheses), for a total of 31 sessions acquired across multiple days from 11 subjects. Subjects performed five locomotion tasks: level walking, incline and decline walking, and stair ascent and descent. Since transitions between different locomotion modes happen at different gait events, the analyses have been repeated for both heel-strike (HS) and toe-off (TO) events. At HS events, the proposed approach achieves a median f1-score of 99.12% and 92.41% on VU Gen 2 and Gen 3 prostheses respectively. At TO events, the proposed approach reaches a median f1-score of 96.83% with VU Gen 2 and 94.36% with VU Gen 3. The proposed framework is a promising solution for locomotion classification on data of previously unseen sessions or subjects, allowing classification on multiple prosthesis models.

  PublisherDOI

• IntelliMan_WP4_Adaptive shared autonomy_T4_3. Human intent detection for autonomy arbitration_(Un)supervised (Co)adaptation

  Zenodo (CERN European Organization for Nuclear Research) · 2025-08-25

  datasetOpen access

  This dateset repository includes tabular data from the paper "(Un)supervised (Co)adaptation via Incremental Learning for Myoelectric Control: Motivation, Review, and Future Directions" with the doi: 10.1109/TNSRE.2025.3602397

  PublisherDOI

• Human-in-the-Loop Myoelectric Pattern Recognition Control of an Arm-Support Robot to Improve Reaching in Stroke Survivors

  IEEE Transactions on Neural Systems and Rehabilitation Engineering · 2025-01-01 · 2 citations

  articleOpen accessSenior author

  The objective of this study was to assess the feasibility and efficacy of using real-time human-in-the-loop pattern recognition-based myoelectric control to control vertical support force or vertical position to improve reach in individuals with chronic stroke. This work attempts to move proven lab-based static arm support paradigms towards a controllable wearable device. A machine learning (linear discriminant analysis)-based myoelectric pattern recognition system based on movement intent as determined by real-time muscle activation was used to control incremental changes in either vertical position or vertical support force during a reach and retrieve task, with the goal of improving reaching function. Performance under real-time control of both options was compared to two unchanging static-support conditions (current gold standard) and a no-support condition. Both real-time control paradigms were successfully implemented and resulted in greater forward-reaching performance as demonstrated by increased elbow extension and horizontal shoulder adduction compared to no-support and was not different from the current gold standard static support paradigms. Muscle activation levels with real-time support were lower than the no-support condition and similar to those observed during the static support paradigms. Real-time detection of user intent was successful in controlling both vertical position and vertical support force and enabled greater reaching distance than without it demonstrating both its feasibility and efficacy albeit with some limitations.

  PublisherDOI

• Evaluating the impact of channel count and feature set on online pattern recognition control of a virtual arm

  2025-07-14 · 1 citations

  articleSenior author

  Pattern recognition is a commercially available strategy for transradial prosthesis control. Control performance depends on how many channels of electromyographic (EMG) signal are collected, as well as on the set of features extracted from the signal. Prior work has established common selections for channel count and feature set, importantly suggesting that there are diminishing returns in performance beyond eight EMG channels. However, these results are largely based on offline analyses of classifier error rate, rather than online control performance. This study aimed to evaluate the impact of EMG channel count and extracted feature set on online control of a virtual arm in a randomized, double-blind fashion. The primary metric of interest was Median Target Achievement Control Test Completion Time (TAC-CT), which is a measure of the time it takes a participant to guide a virtual avatar's wrist and hand into a target position. Thus, lower scores indicate better control. It was found that channel count had a significant impact on Median TAC-CT (p<0.001), but feature set did not (p=0.056). Across all feature sets, the 8-channel condition resulted in a clinically important reduction in Median TAC-CT over the 4-channel condition (-3.36s). The 16-channel condition also resulted in a clinically important reduction in Median TAC-CT over the 8-channel condition (-1.50s). Across channel count conditions, the effects of expanding the feature set were mixed. Overall, this work suggests that offline error analysis is, on its own, insufficient to assess online prosthesis control and that the addition of more than eight EMG channels may result in improved control. The authors aim to continue this line of research to further explore these phenomena.Clinical Relevance- This preliminary study demonstrates an improvement in online transradial pattern recognition control when using 16 EMG channels (versus eight EMG channels).

  PublisherDOI

• Measuring metabolic energy expenditure with short duration walking tests for individuals with lower limb amputation

  PLoS ONE · 2025-03-31 · 1 citations

  articleOpen accessSenior authorCorresponding

  INTRODUCTION: Metabolic assessment of prosthetic gait is useful when comparing devices, interventions, or populations. However, the standard requirement to walk continuously for six minutes or more to reach steady state (SS) is difficult for many individuals with lower limb amputation. Our goal was to assess the concurrent validity of metabolic outcomes from shorter duration walking tests with those from the standard six-minute walk, in persons with transfemoral or transtibial amputation. METHODS: Thirty participants (amputation: 10 transfemoral, 10 transtibial, 10 none) performed three walking tests while data were collected with a wearable metabolic system: 1) two-minute treadmill walk plus 10-minute recovery, 2) six-minute treadmill walk, and 3) overground two-minute walk test (2MWT). Three different analyses were performed to correlate SS metabolic outcomes from minutes 5-6 of the six-minute treadmill walk with: 1) total oxygen uptake from the two-minute treadmill walk, incorporating excess post-exercise oxygen consumption (EPOC), 2) minute interval outcomes from minutes 1-4 of the six-minute treadmill walk, and 3) outcomes during minutes 1 and 2 of the 2MWT. RESULTS: Strong correlations were found between total oxygen uptake of the two-minute treadmill walk plus EPOC and SS oxygen uptake (Pearson r 0.86 to 0.94). Likewise, there were strong correlations between minute interval outcomes of minutes 2, 3, and 4 of the six-minute treadmill walk and SS outcomes (Pearson r 0.82 to > 0.99). Fewer significant correlations were observed when comparing 2MWT outcomes with SS outcomes (Pearson r 0.41 to 0.78). CONCLUSION: Strong correlations between metabolic outcomes of shorter duration walking tests with SS outcomes suggest that treadmill walking tests as short as two minutes may be acceptable to compare energy expenditure between conditions in individuals with lower limb amputation for circumstances where longer duration tests would not be possible. Additionally, these shorter tests would be more similar to real-life activities and more accessible for those with lower limb amputation.

  PublisherOA PDFDOI

Recent grants

• Improved function performance in individuals with transradial limb difference using cloud based health system and deep learning

  NIH · $5.3M · 2014–2028

• NRI: Small: Modeling, Quantification, and Optimization of Prosthesis-User Interface

  NSF · $1000k · 2013–2018

• The Functional Importance of Powered Wrist Flexion/Extension and Simultaneous Control for Upper Limb Prostheses

  NIH · $3.0M · 2018–2023

• NRI: Collaborative Research: Unified Feedback Control and Mechanical Design for Robotic, Prosthetic, and Exoskeleton Locomotion

  NSF · $303k · 2015–2019

• Intuitive Control of a Hybrid Prosthetic Leg During Ambulation

  NIH · $4.3M · 2014–2023

Frequent coauthors

• Todd Kuiken

  177 shared

• Ann M. Simon

  Shirley Ryan AbilityLab

  168 shared

• Eric J. Perreault

  Shirley Ryan AbilityLab

  70 shared

• Aaron J. Young

  Georgia Institute of Technology

  53 shared

• José L. Pons

  Shirley Ryan AbilityLab

  50 shared

• Yue Wen

  Beijing Institute of Technology

  43 shared

• Emek Barış Küçüktabak

  Shirley Ryan AbilityLab

  42 shared

• Matthew R. Short

  41 shared