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

Albara Ah Ramli, Xin Liu, Kelly Berndt, Erica Goude, Jiahui Hou, Lynea B. Kaethler, Rex Liu, Amanda Lopez, Alina Nicorici, Corey Owens, David Rodriguez, Jane Wang, Huanle Zhang, Daniel Aranki, Craig M. McDonald, Erik K. Henricson

TL;DR

This paper explores using smartphone accelerometers to characterize gait differences in children with Duchenne Muscular Dystrophy (DMD). By applying classical machine learning (CML) and deep learning (DL) techniques to extracted gait features and raw sensor data, the study demonstrates high accuracy in differentiating DMD from typically developing (TD) children, paving the way for remote, community-based mobility monitoring.

What Problem Does This Paper Address?

Duchenne Muscular Dystrophy (DMD) is a severe neuromuscular disorder that progressively weakens muscles, impacting gait. While gait abnormalities are observable, quantifying these changes outside of specialized labs has been challenging. This is crucial for early diagnosis, monitoring disease progression, and evaluating the effectiveness of new therapies. Traditional gait analysis requires specialized equipment and controlled environments, limiting its accessibility for frequent, real-world assessments. The authors aim to bridge this gap by leveraging ubiquitous smartphone accelerometers and advanced machine learning to capture clinically relevant gait characteristics in a more accessible manner.

Methods

The study employed two primary machine learning approaches:

  • Classical Machine Learning with Extracted Clinical Features (CML-CF): This method involved extracting eight temporospatial gait clinical features (CFs) from accelerometer data. These features included speed, step length, step frequency, total power, and the percentage of power in different axes (vertical, mediolateral, anteroposterior). Six different CML classifiers (Random Forest, Decision Tree, SVM, KNN, Gaussian Naive Bayes, Logistic Regression) were then trained using these CFs. Dimensionality reduction techniques like Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) were also explored to see if they improved classification accuracy.
  • Deep Learning with Raw Data (DL-RAW): This approach utilized raw accelerometer data directly. A Convolutional Neural Network (CNN) model was trained on time-windowed segments of the raw sensor data. The window size was optimized to capture sufficient gait cycle information for classification.

Both approaches were evaluated using Leave-One-Subject-Out (LOSO) cross-validation to ensure robust performance.

Datasets

The study utilized data collected from 30 male participants (15 with DMD, 15 typically developing) aged 3 to 16 years. Participants performed eight different walking and running activities, including:

  • Five speed-calibration tests (SC-L1 to SC-L5) at incrementally increasing speeds.
  • A 6-minute walk test (6MWT).
  • A 100-meter fast walk/jog/run (100MRW).
  • A free walk (FW).

The Northstar Ambulatory Assessment (NSAA) was also administered to the DMD participants as a clinical anchor for functional ability. No new benchmark datasets were introduced; rather, data was collected from participants for this specific study.

Key Insights and Ideas

  • Smartphone Accelerometers as Powerful Gait Sensors: The study strongly validates the use of single-sensor accelerometers from consumer smartphones for capturing subtle gait differences in DMD, even across a wide range of speeds and activities.
  • Extracted Gait Features Reflect Clinical Observations: The identified temporospatial gait features (e.g., reduced step length, increased mediolateral power) align with known clinical descriptions of DMD gait, such as Trendelenberg-like gait. This provides a crucial link between machine learning outputs and clinical understanding.
  • CML-CF Outperforms DL-RAW in this Context: While DL is powerful, the CML-CF approach, leveraging well-defined gait features, achieved higher classification accuracies, particularly with LDA dimensionality reduction. This suggests that for this specific problem, engineered features provide valuable discriminative information.
  • Gait Speed is a Critical Differentiator: The study highlights gait speed as a significant factor influencing classification accuracy, underscoring its importance in gait analysis for DMD.
  • Potential for Remote Monitoring: The ability to collect data remotely using a smartphone app opens doors for continuous, real-world monitoring of mobility and disease progression, reducing reliance on clinical visits.

Key Results

  • High Classification Accuracy: The CML-CF approach achieved up to 100% accuracy in classifying DMD vs. TD participants, notably in self-selected walk speed (SC-L3). Accuracies exceeded 96% for fast walk/jog/run (100MRW) and fast walk (6MWT) activities.
  • DL-RAW Performance: The DL-RAW approach achieved a maximum accuracy of 86.67% in slow walk (SC-L2) and free walk (FW) activities.
  • DMD Gait Characteristics: Participants with DMD consistently showed lower walking speed, reduced step length, and a higher percentage of mediolateral power compared to TD peers, especially at faster speeds.
  • Step Length and Functional Ability Correlation: Reduced normalized step length in DMD participants correlated with lower NSAA scores, indicating its utility as a proxy for functional decline.

Limitations and Future Work

  • Small Sample Size: The study acknowledges a limitation in the small number of participants and plans to expand the scope to include more individuals across various diagnoses and age groups.
  • Short-Term Data Collection: The current data collection was short-term. Future work should involve multi-day data collection to better assess community travel patterns and differentiate between DMD and TD groups in free-living conditions.
  • DL Explainability: While DL-RAW offers advantages, its "black box" nature makes clinical interpretation challenging. Future research could focus on improving DL explainability or transferring knowledge from DL to enhance CML-CF features.
  • Further Feature Engineering: The authors suggest that developing more correlated temporospatial gait CFs could further improve CML-CF model outcomes.
  • Broader Applications: The methodology has the potential to be extended to other mobility-related diseases and healthy aging populations.

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 Xin Liu and Kelly Berndt and Erica Goude and Jiahui Hou and Lynea B. Kaethler and Rex Liu and Amanda Lopez and Alina Nicorici and Corey Owens and David Rodriguez and Jane Wang and Huanle Zhang and Daniel Aranki 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