TL;DR
This paper introduces a novel method called multi-dimensional subsequence Dynamic Time Warping (msDTW) for accurately segmenting individual strides from inertial sensor data, even during complex, real-world walking scenarios like free walks. This advancement is crucial for reliable gait analysis, enabling better monitoring of health conditions and movement disorders.
What Problem Does This Paper Address?
Accurately identifying individual strides from continuous walking data is a fundamental challenge in sensor-based gait analysis. This is essential for monitoring changes in gait patterns, which can indicate health issues, detect fall risks, and track the progression of movement disorders like Parkinson's disease. Existing methods often struggle with the variability of real-world walking (e.g., turns, stairs) and rely on simplified, controlled environments or single sensor modalities. The paper aims to bridge this gap by developing a robust method that can segment strides from complex, free-walking sequences using multi-dimensional inertial sensor data.
Methods
The core of this paper is the multi-dimensional subsequence Dynamic Time Warping (msDTW) algorithm. This approach combines two key concepts:
- Dynamic Time Warping (DTW): A technique for finding the optimal alignment between two time series that may vary in speed or length. This is crucial because strides naturally have variations in duration and intensity.
- Subsequence DTW (sDTW): An extension of DTW that allows finding a pattern (the stride template) within a longer, continuous sequence.
- Multi-dimensional DTW (mDTW): This aspect integrates data from multiple sensor axes (accelerometer and gyroscope) simultaneously, providing a richer representation of movement.
The msDTW algorithm works by:
- Template Generation: Creating a representative stride template from manually labeled strides (specifically from the gyroscope's sagittal plane data, GZ, which showed good performance).
- Data Normalization: Scaling accelerometer and gyroscope data to a common range [-1, 1] for effective combination.
- Distance Matrix Calculation: Computing the similarity (using Euclidean distance) between each point of the stride template and every point in the continuous movement sequence.
- Accumulated Cost Matrix: Building a cost matrix that represents the cumulative cost of warping the template to the sequence, using dynamic programming to manage computational complexity.
- Warping Path Calculation: Identifying the path of minimum cost through the accumulated cost matrix, which represents the aligned stride within the movement sequence. Boundary conditions and a threshold are used to define valid stride segments.
The paper also implements a peak detection method as a baseline for comparison, using only the gyroscope's GZ axis.
Datasets
The study utilized data collected from 70 subjects:
- Template Generation: 25 elderly controls.
- Algorithm Evaluation:
- 15 elderly controls
- 15 patients with Parkinson's disease (PD)
- 15 geriatric patients
Two measurement protocols were used:
- 40 m Walk: Subjects walked a straight 40m course with 180-degree turns every 10m.
- Free Walk: Subjects walked for two minutes in a hospital environment, including straight walking, turns, navigating through doors, sit-to-stand movements, and stair climbing. This protocol was designed to mimic natural daily activities.
The data was collected using Shimmer 2R inertial measurement units (IMUs) worn on both feet, recording tri-axial accelerometer and gyroscope data at a sampling rate of 102.4 Hz.
Key Insights and Ideas
- The Power of Multi-Dimensionality: Combining data from multiple axes of both accelerometers and gyroscopes significantly improves stride segmentation accuracy, especially in complex free-walk scenarios. This suggests that a holistic view of movement is more robust than relying on a single sensor or axis.
- Gyroscope's Sagittal Plane is Key (but not sufficient): The gyroscope's z-axis (GZ), representing sagittal plane rotation, is highly informative for stride segmentation, as it captures the foot's roll-over motion. However, for more complex movements, additional information is needed.
- msDTW Handles Variability: The msDTW approach is inherently suited to handle the natural variations in stride length and timing, making it more robust than methods relying on fixed-length templates or simple peak detection.
- Free Walks are a Real Challenge: Accurately segmenting strides during free walks, which include diverse activities like stair climbing, is significantly harder than in controlled 40m walks. This highlights the need for advanced algorithms like msDTW for real-world applications.
- Template Appropriateness Matters: The performance of the msDTW algorithm, particularly for the geriatric patient group, was sensitive to the template used. A template derived from healthy elderly controls may not be optimal for older individuals with gait impairments.
Key Results
The msDTW algorithm demonstrated strong performance across various groups and conditions:
- 40 m Walk:
- Controls: Achieved an F-measure of 98% with the combined accelerometer and gyroscope data (AXAYAZGYGZ).
- PD Patients: Achieved an F-measure of 97% with the combined sensor data.
- Geriatric Patients: Achieved an F-measure of 94% with the combined sensor data.
- Free Walk:
- Controls: Achieved an F-measure of 96% with the combined sensor data.
- PD Patients: Achieved an F-measure of 97% with the combined sensor data.
- Geriatric Patients: Achieved an F-measure of 95% using only the gyroscope's GYGZ axes, with the combined sensor data yielding 83%.
Compared to the peak detection method, msDTW showed 5% to 15% higher F-measures, particularly in precision, while maintaining high recall. The peak detection method struggled with non-gait movements like stair climbing, misinterpreting them as strides.
Limitations and Future Work
- Template Optimization: The study identified that the template's suitability is crucial, especially for specific patient groups. Future work should explore creating individualized or group-specific templates (e.g., age-matched, gender-matched) to improve performance.
- Threshold Selection: An optimal fitting threshold needs to be identified for each group and sensor combination. The paper suggests that a general high threshold might be a compromise, potentially increasing recall at the cost of precision.
- Turning and Stair Climbing: While the msDTW algorithm showed robustness, specific analysis of turning movements or stair climbing as distinct activities could be further investigated by defining appropriate templates for these actions.
- Sensor Fusion: While combining accelerometer and gyroscope data improved results for controls and PD patients, it didn't significantly benefit geriatric patients. Further research could explore optimal sensor fusion strategies and axis combinations for different populations.
- Adaptive Learning: Incorporating adaptive learning techniques could help automatically select the best fitting template and threshold for individual datasets, enhancing practical implementation.
- Left and Right Sensor Integration: Simultaneously using data from both left and right foot sensors could improve the identification of alternating strides.
Summary updated