Data-Driven Investigation of Gait Patterns in Individuals Affected by Normal Pressure Hydrocephalus

Kiran Kuruvithadam, Marcel Menner, William R. Taylor, Melanie N. Zeilinger, Lennart Stieglitz, Marianne Schmid Daners2021 · DOI

TL;DR

This study investigates how the environment affects gait patterns in individuals with Normal Pressure Hydrocephalus (NPH), finding that real-world measurements are more effective for diagnosis than controlled lab settings. They also demonstrate that machine learning models, particularly those using parametric gait features, can accurately classify NPH strides.

What Problem Does This Paper Address?

Normal Pressure Hydrocephalus (NPH) is a progressive neurological disorder affecting primarily the elderly, characterized by gait disturbances, cognitive decline, and urinary incontinence. Diagnosing NPH is challenging, with current methods often relying on visual observation or lab-based gait analysis. However, controlled laboratory environments can alter natural walking behavior due to psychological pressure, potentially leading to biased or inaccurate assessments. This study aims to bridge this gap by comparing gait patterns in both controlled and real-world environments, and by developing data-driven approaches to improve NPH diagnosis.

Methods

The researchers utilized five wearable ZurichMOVE inertial measurement units (IMUs) attached to the ankles and chest of participants to collect kinematic data. Gait analysis involved two distinct protocols:

  • Controlled Environment: A 10-meter walk test with prescribed foot spacing and 180° turns, performed at the subject's preferred speed.
  • Real-World Environment: A continuous 72-hour recording of movement patterns in the participants' daily lives.

Key gait parameters, including stride length, gait speed, foot clearance, and swing/stance ratio, were extracted. Statistical analysis involved Wilcoxon signed rank tests to compare intra-group differences between environments and Mann-Whitney U-tests to compare inter-group differences between healthy elderly controls (EHC) and NPH patients.

To further investigate diagnostic capabilities, four machine learning classifiers were employed:

  • Support Vector Machine (SVM-stride): Based on a parametric representation of 15 gait features.
  • Neural Network (NN-stride): Also based on a parametric representation of gait features.
  • Long Short-Term Memory (LSTM): A time-series based approach that directly uses filtered kinematic data.
  • Convolutional Neural Network (CNN): Another time-series based approach, using a Gramian Angular Field encoder to transform time-series data into images.

The performance of these classifiers was evaluated using accuracy, Correct Detection Rate (CDR), and Area Under the Receiver Operating Characteristic Curve (AUROC).

Datasets

The study used data collected from three groups of participants:

  • 20 healthy young subjects (YHC)
  • 20 healthy elderly subjects (EHC)
  • 12 NPH patients

The data was collected in both controlled laboratory settings and real-world environments over 72 hours. No new benchmark datasets were introduced; rather, the study focused on analyzing existing data in novel ways.

Key Insights and Ideas

  • Environmental Influence is Significant: Gait patterns are demonstrably different between controlled lab settings and real-world environments, particularly for elderly individuals and NPH patients. This highlights the limitations of relying solely on lab-based assessments.
  • Real-World Data is More Discriminative: Measurements taken in natural, unconstrained environments provided better discrimination between NPH patients and healthy controls compared to controlled settings. This suggests that home-monitoring systems could be more effective for NPH diagnosis.
  • Parametric Features Excel in Classification: Machine learning models that utilize pre-defined gait parameters (SVM and NN) outperformed time-series based models (LSTM and CNN) in classifying strides as belonging to EHC or NPH. This suggests that established gait features, derived from expert knowledge, are highly informative.
  • NPH Gait is Distinct: NPH patients exhibit specific gait alterations, such as reduced stride length and gait speed, and altered swing/stance ratios, which are more pronounced in real-world settings.

Key Results

  • Environmental Differences: NPH patients showed significant differences in gait speed (p=0.0073), stride length (p=0.0073), foot clearance (p=0.0117), and swing/stance ratio (p=0.0098) between controlled and real-world environments.
  • Diagnostic Efficacy: Logistic regression models achieved higher accuracy in the real-world environment (95.8% for EHC, 91.6% for NPH) compared to the controlled environment (89.5% for EHC, 75.6% for NPH).
  • Machine Learning Performance:
    • SVM-stride and NN-stride models achieved high Correct Detection Rates (CDRs) for both EHC and NPH subjects, often exceeding 90%.
  • These parametric models also demonstrated superior AUROC values (around 0.976) compared to time-series models (LSTM: 0.903, CNN: 0.876).
  • The study suggests that outward rotation and step width, while often considered important for NPH, did not significantly contribute to inter-group differentiation in their analysis, potentially simplifying future diagnostic protocols.

Limitations and Future Work

  • Calibration Burden: The current system requires an initial calibration process, which can be cumbersome for clinicians and patients. The authors suggest that if outward rotation and step width are not critical for differentiation, this calibration step might be streamlined or eliminated.
  • Data Imbalance: The study acknowledges a potential issue with data imbalance, as EHC subjects generated more strides than NPH subjects. This could skew results, and the authors suggest oversampling the minority data as a future solution.
  • Hyper-parameter Optimization: The LSTM and CNN models were only partially optimized, suggesting that further tuning of their hyper-parameters could lead to improved performance and better characterization of stride types.
  • Generalizability: While the parametric models showed good generalizability, further investigation into the robustness of time-series models with more extensive hyper-parameter optimization is warranted.
  • Clinical Integration: The authors propose that their findings provide a strong basis for developing home-monitoring systems and supporting clinical decision-making for NPH diagnosis. Future work could focus on refining the IMU attitude estimation algorithms to improve accuracy for parameters like outward rotation and step width.

Summary updated

BibTeX
@misc{kuruvithadam2021data,
  author = {Kiran Kuruvithadam and Marcel Menner and William R. Taylor and Melanie N. Zeilinger and Lennart Stieglitz and Marianne Schmid Daners},
  title = {{Data-Driven Investigation of Gait Patterns in Individuals Affected by Normal Pressure Hydrocephalus}},
  year = {2021},
  doi = {10.3390/s21196451},
}
Figure