Gait Characterization in Duchenne Muscular Dystrophy (DMD) Using a Single-Sensor Accelerometer: Classical Machine Learning and Deep Learning Approaches

Albara Ah Ramli, Huanle Zhang, Jiahui Hou, Rex Liu, Xin Liu, Alina Nicorici, Daniel Aranki, Corey Owens, Poonam Prasad, Craig M. McDonald, Erik K. Henricson

Abstract

Differences in gait patterns of children with Duchenne muscular dystrophy (DMD) and typically-developing (TD) peers are visible to the eye, but quantifications of those differences outside of the gait laboratory have been elusive. In this work, we measured vertical, mediolateral, and anteroposterior acceleration using a waist-worn iPhone accelerometer during ambulation across a typical range of velocities. Fifteen TD and fifteen DMD children from 3-16 years of age underwent eight walking/running activities, including five 25 meters walk/run speed-calibration tests at a slow walk to running speeds (SC-L1 to SC-L5), a 6-minute walk test (6MWT), a 100 meters fast-walk/jog/run (100MRW), and a free walk (FW). For clinical anchoring purposes, participants completed a Northstar Ambulatory Assessment (NSAA). We extracted temporospatial gait clinical features (CFs) and applied multiple machine learning (ML) approaches to differentiate between DMD and TD children using extracted temporospatial gait CFs and raw data. Extracted temporospatial gait CFs showed reduced step length and a greater mediolateral component of total power (TP) consistent with shorter strides and Trendelenberg-like gait commonly observed in DMD. ML approaches using temporospatial gait CFs and raw data varied in effectiveness at differentiating between DMD and TD controls at different speeds, with an accuracy of up to 100%. We demonstrate that by using ML with accelerometer data from a consumer-grade smartphone, we can capture DMD-associated gait characteristics in toddlers to teens.

Introduction

Duchenne Muscular Dystrophy (DMD) is a progressive muscle-wasting disease that significantly alters a child's gait. While clinicians can visually identify these changes, quantifying them typically requires expensive, specialized gait laboratories. These facilities are often inaccessible to many families and only provide a "snapshot" of performance in a sterile, controlled environment.

There is an urgent need for low-cost, remote monitoring tools capable of tracking disease progression and the efficacy of new therapeutics in a child’s natural environment. The primary challenge lies in extracting meaningful clinical signatures from the noisy data produced by a single, low-cost accelerometer across a wide range of walking and running speeds.

Methodology

The researchers utilized an iPhone 11 fastened to the waist, near the center of mass, to collect triaxial acceleration data at 100 Hz via a custom application called Walk4Me. The study involved 30 male children: 15 with DMD and 15 typically developing (TD) peers, aged 3–16 years.

Sensor placement on the waist and the defined triaxial coordinate system (X: Vertical, Y: Mediolateral, Z: Anteroposterior).

Participants performed eight distinct activities, including five speed-calibration tests (ranging from slow walking to running), a 6-minute walk test (6MWT), a 100-meter fast-walk/run, and a free walk. For clinical anchoring, participants were also evaluated using the Northstar Ambulatory Assessment (NSAA), a 34-point scale for functional mobility.

The study compared two distinct computational pipelines for classification:

  1. Classical Machine Learning with Clinical Features (CML-CF): Eight "temporospatial" features were manually extracted, including normalized speed, step length, step frequency, total power, and the percentage of power distributed across the vertical, mediolateral, and anteroposterior axes. Six classifiers (e.g., Random Forest, SVM, KNN) were tested alongside dimensionality reduction techniques like PCA and LDA.
  2. Deep Learning with Raw Data (DL-RAW): Raw, time-windowed accelerometer signals were fed directly into a Convolutional Neural Network (CNN). This approach avoids manual feature engineering but typically requires larger datasets to converge.
Overview of the two computational pipelines: (1) Deep Learning with raw data and (2) Classical Machine Learning with clinical features.

Validation was performed using Leave-One-Subject-Out (LOSO) cross-validation to ensure the models could generalize to new patients.

Results and Discussion

The analysis revealed several key clinical signatures and performance metrics:

  • The "Waddle" Signature: The models confirmed a classic clinical observation: children with DMD exhibit significantly higher mediolateral power (side-to-side movement). This aligns with the "Trendelenburg-like" gait used to compensate for hip-girdle weakness.
  • The "Double-Off" Failure: At higher speeds, TD children develop a "double-off" running pattern to increase step length. The data showed that DMD children fail to develop this pattern, leading to a significant divergence in step length compared to peers as velocity increases.
  • Feature Engineering vs. Deep Learning: The CML-CF approach outperformed the DL-RAW approach. This suggests that for medical studies with smaller sample sizes, domain-specific features (like step length and power distribution) are more robust than automated feature discovery via deep learning.
  • Classification Accuracy: The CML-CF approach (using Linear Discriminant Analysis) achieved 100% accuracy in differentiating DMD from TD children during self-selected walking speeds. It maintained over 96% accuracy for the 6-minute walk test and the 100-meter run/walk. The DL-RAW approach reached a maximum accuracy of 86.67%.
  • Clinical Correlation: Extracted step lengths were highly correlated with NSAA scores, suggesting that smartphone sensors can serve as a reliable digital proxy for traditional clinical assessments.

Conclusion

This study demonstrates that consumer-grade hardware, such as a smartphone, can provide clinical-grade monitoring for progressive neuromuscular diseases. While the small sample size (30 participants) is a limitation for complex deep learning models, the high accuracy achieved through classical machine learning with engineered features is promising.

Future work will focus on expanding the study to larger cohorts and other mobility-related conditions. Additionally, researchers aim to improve the "explainability" of deep learning models to help clinicians understand the specific signal components used for diagnosis. Transitioning to "free-living" data collection will also be a key step in developing a truly remote monitoring tool.

Summary updated

Abstract

Differences in gait patterns of children with Duchenne muscular dystrophy (DMD) and typically-developing (TD) peers are visible to the eye, but quantifications of those differences outside of the gait laboratory have been elusive. In this work, we measured vertical, mediolateral, and anteroposterior acceleration using a waist-worn iPhone accelerometer during ambulation across a typical range of velocities. Fifteen TD and fifteen DMD children from 3-16 years of age underwent eight walking/running activities, including five 25 meters walk/run speed-calibration tests at a slow walk to running speeds (SC-L1 to SC-L5), a 6-minute walk test (6MWT), a 100 meters fast-walk/jog/run (100MRW), and a free walk (FW). For clinical anchoring purposes, participants completed a Northstar Ambulatory Assessment (NSAA). We extracted temporospatial gait clinical features (CFs) and applied multiple machine learning (ML) approaches to differentiate between DMD and TD children using extracted temporospatial gait CFs and raw data. Extracted temporospatial gait CFs showed reduced step length and a greater mediolateral component of total power (TP) consistent with shorter strides and Trendelenberg-like gait commonly observed in DMD. ML approaches using temporospatial gait CFs and raw data varied in effectiveness at differentiating between DMD and TD controls at different speeds, with an accuracy of up to 100%. We demonstrate that by using ML with accelerometer data from a consumer-grade smartphone, we can capture DMD-associated gait characteristics in toddlers to teens.
BibTeX
@misc{ramli2021gait,
  author = {Albara Ah Ramli and Huanle Zhang and Jiahui Hou and Rex Liu and Xin Liu and Alina Nicorici and Daniel Aranki and Corey Owens and Poonam Prasad and Craig M. McDonald and Erik K. Henricson},
  title = {{Gait Characterization in Duchenne Muscular Dystrophy (DMD) Using a Single-Sensor Accelerometer: Classical Machine Learning and Deep Learning Approaches}},
  year = {2021},
  eprint = {2105.06295},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2105.06295},
}
Figure