Gait event prediction from surface electromyography in parkinsonian patients

Stefan Haufe, Ioannis U. Isaias, Franziska Pellegrini, Chiara PalmisanomedRxiv 2023 · DOI

TL;DR

This paper demonstrates that surface electromyography (EMG) can be used to accurately reconstruct continuous motion profiles and predict gait events in Parkinson’s disease (PD) patients. By using a patient-agnostic machine learning approach, the researchers found that just two EMG sensors placed on the thigh muscles are sufficient to identify gait phases with high temporal precision, potentially eliminating the need for separate motion sensors in clinical and research settings.

What Problem Does This Paper Address?

Analyzing how people with Parkinson’s disease walk is critical for tailoring treatments like medication or Deep Brain Stimulation (DBS). Currently, researchers use two main tools: Inertial Measurement Units (IMUs) to track motion (kinematics) and surface EMG to track muscle activity (neurophysiology).

However, using both systems simultaneously is cumbersome. It increases setup time, adds physical discomfort for patients, and creates technical headaches regarding data synchronization. While IMUs are great at identifying gait events (like when a heel hits the ground), they don't explain why a movement is happening at a muscular level. Conversely, previous attempts to use EMG alone for gait detection usually relied on "black-box" classification (e.g., "is the patient in the swing or stance phase?"). This paper seeks a middle ground: using EMG to reconstruct the continuous angular velocity signals typically provided by IMUs, thereby getting the best of both worlds from a single sensor type.

Methods

The researchers moved away from simple classification and instead used an innovative regression approach.

  1. Temporal Embedding: Instead of looking at a single snapshot of muscle activity, the model looks at a 1-second window of EMG data (500ms before and after the target time point) to capture the dynamics of muscle activation.
  2. Linear Regression: They used multiple linear regression to map these EMG signals to the angular velocity of the ankles. Despite the complexity of PD gait, a relatively simple linear model proved robust.
  3. Rule-Based Extraction: Once the IMU-like signal was reconstructed from the EMG data, the researchers applied standard biomechanical rules to identify specific events: Swing Peak Velocity (SWP), Heel Contact (HC), and Toe-Off (TO).
  4. Cross-Validation: To ensure the model wasn't just "memorizing" specific patients, they used a leave-one-patient-out strategy. The model was trained on five patients and tested on a sixth, proving it could generalize to new individuals without recalibration.

Datasets

The study involved a specialized clinical dataset:

  • Participants: 6 patients with idiopathic Parkinson’s Disease.
  • Conditions: Patients were recorded in a "meds-off" state (overnight withdrawal of dopaminergic medication) to capture clear PD-related gait characteristics.
  • The Task: Continuous walking for at least three minutes on an elliptical path.
  • Data Volume: 26 walking trials totaling 105 minutes and 5,253 full gait cycles.
  • Hardware: 10 surface EMG probes (bilateral leg muscles) and 2 IMUs (placed on the ankles).

Key Insights and Ideas

  • The "Vastus Lateralis" Surprise: The researchers tested 31 different combinations of muscle pairs. While they initially hypothesized that lower leg muscles (like the calves or shins) would be best, the vastus lateralis (thigh muscle) actually provided the most accurate predictions.
  • Reconstruction vs. Classification: By reconstructing the continuous angular velocity profile rather than just predicting "swing" or "stance," the model provides a signal that clinicians can visually inspect. If a sensor is noisy or misplaced, it’s immediately obvious in the reconstructed trace.
  • Patient-Agnostic Utility: The model accounts for the substantial variability in gait patterns across different PD patients. This suggests that a universal "EMG-to-Kinematics" filter could be developed that doesn't require tedious per-patient calibration.

Key Results

The EMG-based system performed remarkably well compared to the "ground truth" IMU data:

  • High Correlation: The Pearson correlation (r) between measured and reconstructed IMU traces was between 0.80 and 0.86.
  • Temporal Precision: Gait events were detected with a median error of less than 50 milliseconds.
  • Reliability: The model missed fewer than 2% of gait events and had a false-positive rate of less than 0.1%.
  • Phase Accuracy: The F1 score for determining swing vs. stance phases was approximately 0.89, indicating high fidelity.

Limitations and Future Work

While promising, the study has a few caveats:

  • Small Cohort: The model was validated on only six patients. Larger studies are needed to confirm these results across the broader spectrum of PD phenotypes.
  • Homogeneous Speed: Patients walked at their self-selected "natural" speed. It is currently unclear how the model handles significant changes in walking speed or "freezing of gait" episodes.
  • The "Gold Standard" Question: The study used IMUs as the ground truth. While IMUs are standard for mobile studies, they are not as precise as laboratory-grade force plates or optoelectronic motion capture.

The authors suggest that future work should focus on adapting the model to varying walking speeds and integrating this "sensor-lite" approach into real-time applications, such as controlling neuroprosthetics or adjusting adaptive DBS settings on the fly.

Summary updated

Abstract

Abstract Gait disturbances are common manifestations of Parkinson’s disease (PD), with unmet therapeutic needs. Inertial measurement units (IMU) are capable of monitoring gait, but they lack neurophysiological information that may be crucial for studying gait disturbances in these patients. Here, we present a machine-learning approach to approximate IMU angular velocity profiles, and subsequently gait events from electromyographic (EMG) channels. We recorded six parkinsonian patients while walking for at least three minutes. Patient-agnostic regression models were trained on temporally-embedded EMG time series of different combinations of up to five leg muscles bilaterally (i.e., tibialis anterior, soleus, gastrocnemius medialis, gastrocnemius lateralis, and vastus lateralis). Gait events could be detected with high temporal precision (median displacement <50 msec), low numbers of missed events (<2%), and next to no false positive event detections (<0.1%). Swing and stance phases could thus be determined with high fidelity (median F1 score ∼0.9). Interestingly, the best performance was obtained using as few as two EMG probes placed on the left and right vastus lateralis. Our results demonstrate the practical utility of the proposed EMG-based system for gait event prediction while allowing the simultaneous acquisition of an electromyographic signal. This gait analysis approach has the potential to make additional measurement devices such as IMU and force plates less essential, and thereby to reduce financial and preparation overheads and discomfort factors in gait studies.
BibTeX
@misc{haufe2023gait,
  author = {Stefan Haufe and Ioannis U. Isaias and Franziska Pellegrini and Chiara Palmisano},
  title = {{Gait event prediction from surface electromyography in parkinsonian patients}},
  year = {2023},
  doi = {10.1101/2023.01.13.22282375},
  howpublished = {medRxiv},
}
Figure