An Inertial Sensor-Based Gait Analysis Pipeline for the Assessment of Real-World Stair Ambulation Parameters

Nils Roth, Arne Küderle, Dominik Prossel, Heiko Gassner, Bjoern M. Eskofier, Felix Kluge2021 · DOI

TL;DR

This paper introduces a novel gait analysis pipeline that uses wearable inertial sensors to accurately segment, analyze, and classify real-world stair climbing (both ascending and descending) alongside level walking. This advancement is crucial for understanding mobility impairments in conditions like Parkinson's disease, offering a more comprehensive picture of daily functional capabilities.

What Problem Does This Paper Address?

Climbing stairs is a fundamental daily activity that places significant demands on our balance and motor control, making it a critical indicator of overall mobility and health. However, existing gait analysis systems, particularly those using wearable sensors for real-world monitoring, often focus solely on level walking. This oversight misses a vital aspect of daily life and a key area where individuals with neurological or functional impairments may struggle. Stair ambulation presents unique challenges, such as altered foot-rolling behavior and the absence of a clear heel strike during initial contact, which standard gait analysis algorithms are not equipped to handle. This gap limits our ability to accurately assess fall risk and disease progression in conditions where stair negotiation is particularly difficult.

Methods

The researchers developed a comprehensive pipeline with several key components:

  • Stride Segmentation: They extended a Hidden Markov Model (HMM)-based approach, originally designed for level walking, to incorporate specific models for stair ascending and descending strides. This allows for the simultaneous segmentation of mixed gait activities.
  • Gait Event Detection: An adapted algorithm was developed to reliably detect crucial gait events like Initial Contact (IC) and Terminal Contact (TC) even when a heel strike is absent, which is common during stair climbing. This was achieved by analyzing specific signal features from the IMU data.
  • Trajectory Reconstruction: To overcome the error accumulation inherent in IMU-based trajectory estimation, they employed an Error-Tracking Kalman Filter (ETKF) guided by Zero Velocity Updates (ZUPTs). This method reconstructs the foot's path with improved accuracy.
  • Stride-Type Classification: Spatial features derived from the reconstructed trajectory, specifically stride height and inclination, were used to classify strides into level walking, stair ascending, or stair descending. Simple threshold-based rules were applied for this classification.
  • Walking Bout Assembly: A final step combines consecutive strides of the same type to form "walking bouts," ensuring that only valid stair sequences are considered for further analysis, thereby increasing the precision of stair stride classification.

Datasets

  • Evaluation Dataset: A new dataset was collected from 20 healthy participants walking on three different staircases with varying geometries and at different speeds (slow, normal, fast). This dataset includes approximately 16,000 stair strides and 13,000 level-walking strides, with ground truth provided by pressure insoles and video annotations.
  • Validation Dataset: An independent dataset of 13 Parkinson's disease (PD) patients was used to validate the pipeline's performance on individuals with gait impairments. This dataset involved a single transition between floors in a hospital setting.

Key Insights and Ideas

  • Stair Ambulation is Fundamentally Different: The research highlights that stair climbing isn't just a more demanding version of level walking; it involves distinct biomechanical patterns (e.g., forefoot contact, altered foot roll) that necessitate specialized algorithms.
  • HMMs Can Be Extended for Complex Activities: By adapting HMMs with specific sub-models for different gait types, the system can robustly segment mixed activities, a crucial step for real-world gait analysis.
  • Spatial Features are Key for Stair Classification: While temporal features are important, the paper demonstrates that stride height and inclination derived from trajectory reconstruction are highly effective discriminators for classifying stair strides.
  • Walking Bout Assembly Enhances Precision: Requiring a minimum number of consecutive strides of the same type for a "stair walking bout" significantly reduces misclassifications, especially for transition strides.

Key Results

  • Stride Segmentation: The multiclass HMM achieved an impressive average F1 score of 98.5% for stride segmentation across all tasks and speeds.
  • Gait Event Detection: The system demonstrated a high detection rate of 99.0%, with mean timing errors for IC and TC below ±10 ms, comparable to existing state-of-the-art methods.
  • Stride-Type Classification: The spatial feature-based classification achieved a balanced accuracy of 98.2%.
  • Full Pipeline Validation: The end-to-end pipeline showed good agreement with ground truth, with mean differences in temporal parameters below 0.01 s and 95% limits of agreement around ±0.05 s for healthy participants and ±0.06 s for PD patients.

Limitations and Future Work

  • Dataset Limitations: While the dataset is comprehensive, it was collected under supervised conditions. Future work should focus on validating the pipeline in truly unsupervised, free-living environments with more diverse and unpredictable conditions (e.g., varying stair geometries, uneven terrain).
  • Clinical Validation: The extracted Digital Mobility Outcomes (DMOs) require further clinical validation to confirm their utility in assessing fall risk, disease progression, and therapeutic effects.
  • Stride-Type Classification Extension: The authors suggest exploring more advanced classification methods, such as deep learning, for stride-type classification to potentially improve robustness against sensor errors and to differentiate between single and double stair step strides more granularly.
  • Spatial Feature Validation: While spatial features showed good separability, they require separate validation with more advanced systems like motion capture to assess their absolute accuracy in all dimensions.
  • Stair Clearance: The paper notes that stair clearance is a critical feature for gait safety and could be easily extracted from the reconstructed trajectory, suggesting this as a valuable future metric.

Summary updated

BibTeX
@misc{roth2021inertial,
  author = {Nils Roth and Arne Küderle and Dominik Prossel and Heiko Gassner and Bjoern M. Eskofier and Felix Kluge},
  title = {{An Inertial Sensor-Based Gait Analysis Pipeline for the Assessment of Real-World Stair Ambulation Parameters}},
  year = {2021},
  doi = {10.3390/s21196559},
}
Figure