TL;DR
This study investigates how different parameter choices for state-space reconstruction affect the classification accuracy of individuals with Multiple Sclerosis (MS) using the Local Divergence Exponent (LDE) as a gait stability biomarker. The key finding is that using a fixed embedding dimension (dE=5) and time delay (τ=10) with norm acceleration data from a sternum-placed sensor provides the best classification accuracy, suggesting a standardized approach for LDE implementation in MS research and clinical settings.
What Problem Does This Paper Address?
The Local Divergence Exponent (LDE) has emerged as a promising tool for quantifying walking stability, particularly in individuals with Multiple Sclerosis (pwMS). Studies consistently show that pwMS exhibit impaired gait stability, even in early stages of the disease, and this impairment is linked to neurological damage. However, a significant challenge in utilizing LDE is the lack of standardization in its calculation. Various methodological choices, including sensor placement, data type (acceleration vs. angular velocity), and crucially, the parameters used for state-space reconstruction (embedding dimension 'dE' and time delay 'τ'), can significantly influence LDE values. This variability hinders direct comparisons between studies and complicates the adoption of LDE as a reliable clinical biomarker for disease monitoring and intervention assessment. This paper aims to address this by systematically evaluating how different parameter selection strategies for state-space reconstruction impact the ability to classify pwMS from healthy controls (HCs).
Methods
The study employed a systematic approach to compare different parameter selection methods for state-space reconstruction in LDE calculation. Participants, including 55 pwMS and 23 HCs, walked along a 20-meter corridor for 5 minutes. Inertial Measurement Units (IMUs) were attached to the sternum, lumbar spine, and feet to capture 3D acceleration data.
The core of the methodology involved calculating the LDE using three distinct approaches for determining the embedding dimension (dE) and time delay (τ):
- Individual (I): dE and τ were calculated for each individual subject and each specific direction (vertical, mediolateral, anteroposterior, and norm) using the Global False Nearest Neighbors (GFNN) and Average Mutual Information (AMI) methods, respectively.
- Median (M): The median dE and τ values across all subjects were used for each direction and sensor.
- Fixed (F): Predefined, fixed values for dE and τ were used, based on previous literature. Specifically, dE=5 and τ=10 were used for 3D data, and for directional data, fixed values of dE=5 and τ=10 were also applied.
The LDE was calculated for vertical (VT), mediolateral (ML), anteroposterior (AP) accelerations, the norm (N) of these accelerations, and 3D data from both sternum (STR) and lumbar (LUM) sensors. Quadratic Discriminant Analysis (QDA) was used to classify pwMS from HCs based on the calculated LDE values. The classification accuracy was evaluated, and a secondary analysis considered walking speed and age as covariates.
Datasets
The study utilized data collected from 55 individuals with relapsing-remitting multiple sclerosis (RRMS) and 23 healthy controls. The pwMS group had a median Expanded Disability Status Scale (EDSS) of 2.0, indicating no or minor gait impairments. The data consisted of 3D acceleration measurements from IMUs worn during a 5-minute walking task. The dataset was randomly split into training and testing sets (2:1 ratio) for the QDA classifier. The authors did not introduce a new benchmark dataset but rather analyzed existing data with novel parameter selection strategies.
Key Insights and Ideas
- Parameter Sensitivity: The choice of parameters for state-space reconstruction (dE and τ) significantly impacts the LDE's ability to classify pwMS. This underscores the need for standardization.
- Norm Acceleration Superiority: Using the norm (N) of acceleration, which combines directional components, yielded better classification accuracy than using individual directional components (VT, ML, AP). This suggests that a holistic measure of acceleration captures gait dynamics more effectively for classification.
- Sternum Sensor Advantage: Sensors placed on the sternum consistently provided higher classification accuracy compared to those on the lumbar spine. This might indicate that sternum-based LDEs better reflect compensatory strategies employed by pwMS to maintain balance, potentially due to central nervous system reallocations.
- Fixed Parameters Simplify and Improve: Using fixed parameters (dE=5, τ=10) for state-space reconstruction, particularly with sternum norm acceleration, not only simplified the calculation but also achieved the highest classification accuracy. This is a crucial finding for practical implementation.
- Speed as a Covariate: Incorporating walking speed as a covariate significantly improved the classification performance of most LDE calculations, highlighting its importance in gait stability analysis.
Key Results
The study found that the highest classification accuracy of 84% was achieved when using the LDE derived from the norm acceleration data from a sternum-placed sensor with fixed parameters (dE=5 and τ=10), and considering speed as a covariate.
- Fixed Parameters: Using fixed dE=5 and τ=10 for the sternum sensor's norm acceleration data resulted in the best classification accuracy (0.826 when speed was included).
- Directional vs. Norm: Sternum LDEs generally achieved higher accuracy (>0.700) than lumbar LDEs (ranging from 0.679 to 0.717). The norm acceleration from the sternum sensor with fixed parameters outperformed directional sternum LDEs.
- Speed's Impact: Controlling for speed improved classification accuracy for most calculated LDEs. Speed alone achieved an accuracy of 0.710.
- Lumbar vs. Sternum: Lumbar LDEs were consistently less accurate than sternum LDEs.
Limitations and Future Work
The authors acknowledge several limitations:
- Study Population: The study focused on individuals with relapsing-remitting MS (RRMS) with no or minor gait impairments (EDSS < 4.0). The generalizability of findings to other MS phenotypes or subgroups with more severe pyramidal signs needs further investigation.
- Data Length: While 150 gait cycles were used, some research suggests over 300 gait cycles might be needed for optimal de estimation. However, this is often impractical in clinical settings due to fatigue.
- Sensor Placement: The study focused on sternum and lumbar placement. Other sensor locations might yield different results.
- Intrinsic Factors: The impact of factors like fatigue and mood on motor performance and stability was not explored due to dataset limitations. Future work could examine these in contexts with longer walking tasks and IMU assessments.
- Controlled Environment: The study used a controlled corridor walking task. Future research could explore how these findings translate to daily life walking, which may reveal different dynamics.
The authors suggest that future research should explore population-specific parameters for LDE calculation, especially for monitoring progressive neurodegenerative diseases. They also emphasize the need for further investigation into the hypothesis that sternum-based LDEs reflect compensatory balance strategies in pwMS.
Summary updated