📄 bioengineering

Synergy Feedback Control Predicts Walking Across Multiple Cycles

This study demonstrates that predictive simulations of post-stroke walking using a personalized neuromusculoskeletal model require a minimum level of feedforward synergy control combined with feedback mechanisms to accurately reproduce experimentally observed, dynamically consistent gait patterns.

Original authors: Williams, S. T., Li, G., Fregly, B. J.

Published 2026-03-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Williams, S. T., Li, G., Fregly, B. J.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine your body's ability to walk is like a highly sophisticated self-driving car. To drive smoothly, the car needs two things:

A Pre-Planned Route (Feedforward): A map telling it where to go next.
Real-Time Sensors (Feedback): Cameras and radar that say, "Whoa, there's a pothole! Steer left!" or "The road is slippery, slow down!"

For people with neurological conditions like stroke, this system gets glitchy. Sometimes the map is wrong, and sometimes the sensors are too sensitive (causing stiff, jerky movements) or not sensitive enough (causing falls).

This paper is about building a digital twin of a real human who has had a stroke to figure out how to fix their "self-driving" system.

The Experiment: Building the Digital Twin

The researchers took data from a real 79-year-old man who had a stroke. They collected:

Motion Capture: Like putting him in a movie with 300 cameras to track every joint.
Force Plates: To see exactly how hard his feet hit the ground.
Muscle Signals (EMG): Like plugging into his muscles to hear the electrical "whispers" telling them to move.

They used this data to build a personalized computer model of his body. This wasn't a generic "average person" model; it was his model, with his bone lengths, his muscle strengths, and his specific walking quirks.

The Puzzle: The Recipe vs. The Taste Test

The researchers wanted to see if they could predict how he would walk in the future using a computer simulation. They tried a new method using something called "Muscle Synergies."

Think of Muscle Synergies like a recipe. Instead of telling every single one of your 600+ muscles what to do individually (which is too complicated), your brain groups them into teams.

Team A: "Lift the foot."
Team B: "Push off the ground."

The researchers tried to find the perfect recipe (Feedforward) and the perfect set of sensors (Feedback) to make the digital twin walk exactly like the real man.

They tested six different versions of this recipe:

0% Recipe, 100% Sensors: The car has no map; it only reacts to the road.
100% Recipe, 0% Sensors: The car follows the map blindly, ignoring the road.
The Mix: Various combinations in between.

The Big Discovery: You Need a Map!

Here is the surprising result: The "Pure Sensor" approach failed.

When the researchers tried to make the model walk using only feedback (sensors) and no pre-planned recipe (feedforward), the simulation crashed. The digital walker couldn't figure out how to start or keep moving without a basic plan. It was like trying to drive a car in the dark with no headlights, hoping the sensors would tell you where the road is.

The Sweet Spot:
The model worked best when it had a strong pre-planned recipe (100% Feedforward) with just a tiny bit of sensor correction (Feedback).

The 100% Recipe Model: This was the winner. It walked almost exactly like the real man. It had a solid plan, and the sensors only made tiny, necessary adjustments.
The Low-Recipe Models: When they reduced the recipe to 0%, 25%, or 50%, the model could only walk if they let it start in a weird, unnatural position. Once it started, it took giant, clumsy steps (like a drunk person trying to walk a straight line) to stay upright.

Why Does This Matter?

This study teaches us a crucial lesson about how our brains work, especially after an injury like a stroke:

We aren't just reactive: We don't just react to the world step-by-step. We have a strong internal plan (a "feedforward" command) that tells our muscles what to do before we even move.
The "Glitch" in Stroke: For people with stroke, this internal plan might be weak or broken. If we try to treat them by only focusing on their reflexes (feedback), we might miss the bigger picture.
Better Treatments: This method allows doctors to test treatments in a computer before trying them on a patient. They can say, "If we strengthen this specific muscle group (the recipe), the patient will walk better," without risking a fall in the real world.

The Bottom Line

Think of walking as a dance. You need the choreography (the feedforward plan) to know the steps, and you need balance (the feedback) to not fall over if someone bumps you.

This paper found that for a person with a stroke, the choreography is the most important part. If you take away the choreography and only rely on balance, the dance falls apart. To help them walk again, we need to help them rebuild that internal choreography, not just fix their balance.

By using real data to build these digital twins, scientists are one step closer to designing personalized "dance lessons" (treatments) that can help people with neurological disorders walk with confidence again.

1. Problem Statement

Neural feedback is critical for healthy movement control, yet its disruption is a hallmark of neurological disorders such as stroke, cerebral palsy, and Parkinson's disease. While researchers have developed computational neuromusculoskeletal models controlled by simulated neural feedback, these models typically:

Represent generic populations rather than specific human subjects.
Fail to incorporate actual experimental data (EMG, kinematics, kinetics) from individual patients.
Lack the ability to predict dynamic, subject-specific motions under new conditions.

Consequently, these models have limited practical application in designing patient-specific treatments. The core challenge addressed by this study is developing a personalized, predictive neuromusculoskeletal model that uses real experimental data to simulate walking motions for a specific post-stroke individual, balancing feedforward (FF) and feedback (FB) control mechanisms.

2. Methodology

The study utilized the Neuromusculoskeletal Modeling (NMSM) Pipeline to create a subject-specific 3D model and evaluate a novel synergy-based control strategy.

A. Data Collection & Subject

Subject: A 79-year-old male with right-sided hemiparesis post-stroke.
Data: 30 seconds of treadmill walking (0.5 m/s) captured via motion capture (16 markers/leg), ground reaction forces (split-belt treadmill), and EMG (16 muscles/leg, using surface and fine-wire electrodes).
Dataset Split: 13 gait cycles were identified. 8 cycles were used for calibration (5 for fitting, 3 for tracking optimization), and 3 cycles were withheld for predictive testing.

B. Model Personalization (NMSM Pipeline)

Joint & Ground Contact: A generic OpenSim model was scaled and personalized using a "Joint Model Personalization" tool to optimize joint axes and body scaling. A "Ground Contact Personalization" tool calibrated viscoelastic foot-ground contact parameters to match experimental ground reaction forces (GRF).
Torque-Driven Tracking: An optimal control problem (Tracking Optimization) generated dynamically consistent torque-driven walking motions for the 8 calibration cycles, ensuring joint moments and GRFs matched experimental data.
Muscle-Tendon Personalization: Muscle-tendon parameters (e.g., optimal fiber length, tendon slack length) and time-varying activations were solved to match the torque-driven moments and EMG data.
Neural Control Personalization:
- Synergy Extraction: Muscle synergies were extracted from the 5 calibration cycles.
- Feedforward (FF) Calculation: Nominal FF commands were derived by averaging the synergy commands across the 5 cycles.
- Feedback (FB) Fitting: FB weights were fitted to reproduce cycle-specific synergy commands using joint positions, velocities, and moments as surrogates for muscle states.
- FF/FB Ratio Testing: Six models were created by scaling the nominal FF commands to 0%, 25%, 50%, 75%, 100%, and 125%. The remaining gap in reproducing cycle-specific activations was filled by the fitted FB component.

C. Predictive Simulations

The six FF+FB models were used in Synergy-Driven Predictive Simulations (using Tracking Optimization) to reproduce the 3 withheld testing cycles.
Constraints ensured dynamic consistency (muscle moments within 1 Nm of ID loads) and near-periodicity.
For low FF models (0–50%), initial condition constraints were relaxed to allow convergence, as they failed to converge with strict experimental initial values.

3. Key Contributions

Subject-Specific Predictive Modeling: Demonstrated a pipeline capable of generating dynamically consistent, predictive walking simulations for a specific post-stroke patient using real experimental data, moving beyond generic population models.
Synergy-Based FF+FB Control: Introduced a novel control architecture where a baseline feedforward synergy command is adjusted by a linear feedback controller. This allows for the systematic investigation of the balance between pre-planned movement and reactive correction.
Identification of Minimum FF Requirements: Established that successful predictive simulation of real-world gait requires a significant baseline of feedforward control; purely feedback-driven models (0% FF) failed to reproduce natural gait without diverging significantly from experimental initial conditions.
Data-Driven Personalization: Integrated dense EMG, kinematic, and kinetic data into a single personalized model, overcoming the limitation of previous studies that often ignored EMG or used simplified 2D models.

4. Key Results

Model Fitting: The synergy-based models successfully reconstructed the 5 calibration cycles with high accuracy (Mean Variance Accounted For: 88.46 ± 1.05%). Feedback weight fitting achieved $R^2$ scores $\ge$ 0.90 for most cases.
Predictive Performance:
- 100% FF Model: Produced the most accurate reproduction of the testing cycles (joint angles, moments, GRFs) while maintaining initial conditions close to experimental values.
- 75% & 125% FF Models: Successfully generated near-periodic walking motions consistent with experimental initial conditions.
- 0%, 25%, & 50% FF Models: Failed to converge when constrained to experimental initial values. When constraints were relaxed, they produced walking motions characterized by wide, short strides (increased base of support), indicating a significant movement impairment rather than healthy gait.
Sensitivity: Models with 6 synergies (initially tested) failed to generalize due to poor weight distribution and high sensitivity. The 5-synergy solution proved more robust.
Data Volume: Fitting models to a single gait cycle resulted in poor predictive performance compared to fitting to 5 cycles, highlighting the need for diverse training data to ensure dynamic consistency.

5. Significance and Limitations

Significance:
This study provides a repeatable framework for creating patient-specific digital twins for gait analysis. By demonstrating that a minimum level of feedforward control is necessary for predictive simulations to match real-world data, the findings suggest that therapeutic interventions for neurological disorders may need to target the restoration of feedforward planning capabilities, not just reactive reflexes. The methodology bridges the gap between theoretical control models and clinical application.

Limitations:

Partial Control: The feedback model did not control hip rotation or toe joints, limiting full leg control.
Lack of Muscle-Specific Feedback: The model used joint-level surrogates rather than direct muscle-specific feedback (e.g., stretch reflexes), which may be crucial for fine motor control.
Model Simplifications: The model used rigid tendons (ignoring short-range stiffness) and instantaneous feedback (ignoring neural time delays).
Generalizability: The study focused on a single subject; future work is needed to validate the approach across a broader patient population.

In conclusion, the paper establishes that predictive simulations of human gait using real experimental data require a substantial feedforward component to maintain dynamic consistency and reproduce natural walking patterns, offering a new tool for optimizing treatments for movement impairments.