Feature Selection and Predictors of Falls with Foot Force Sensors Using KNN-Based Algorithms

Shengyun Liang, Yunkun Ning, Huiqi Li, Lei Wang, Zhanyong Mei, Yingnan Ma, Guoru Zhao2015 · DOI

TL;DR

This paper introduces a novel method for predicting fall risk in the elderly using ground reaction force (GRF) data from foot force sensors. By quantifying GRF time series with sample entropy and employing KNN-based classifiers, they achieved 100% accuracy in classifying individuals as fallers or non-fallers, highlighting specific GRF features as key predictors.

What Problem Does This Paper Address?

Falls are a significant health concern for the aging population, leading to reduced quality of life and increased healthcare costs. While various fall risk assessment tools exist, many rely on subjective clinical scores or are difficult to standardize. There's a pressing need for objective, reliable, and easily implementable methods to identify individuals at high risk of falling. This paper aims to address this by leveraging objective biomechanical data – specifically ground reaction forces – to predict fall risk.

Methods

The core of this research lies in a two-pronged approach: feature extraction and classification.

  • Feature Extraction: The study utilizes ground reaction force (GRF) data collected from both walking and sit-to-stand (STS) movements. To capture the complex dynamics of these movements, the researchers quantified the time series signals of eight specific GRF features (e.g., superior-inferior GRF for left foot during walking, vertical GRF during STS) using sample entropy. Sample entropy is a nonlinear measurement that quantifies the regularity of a time series and is known for its robustness to noise. Before calculating sample entropy, GRF data was normalized by body weight and standardized.
  • Classification: Three K-Nearest Neighbor (KNN)-based algorithms were employed for classification:
  • Local Mean-based K-Nearest Neighbor (LMKNN): Calculates a local mean vector for the k nearest neighbors in each class and classifies based on the distance to these mean vectors.
  • Pseudo-Nearest Neighbor (PNN): Assigns weights to the k nearest neighbors based on their distance to the test sample, giving more weight to closer neighbors.
  • Local Mean Pseudo-Nearest Neighbor (LMPNN): An extension that first calculates local mean vectors for neighbors and then applies weighted distances, similar to PNN.

The classification performance was evaluated using leave-one-out cross-validation (LOOCV), a rigorous method for estimating classifier accuracy. Feature selection was performed by evaluating all possible combinations of the eight extracted features (255 combinations) and selecting the subset that yielded the best classification performance, prioritizing sensitivity and accuracy.

Datasets

The study used data collected from 38 participants aged 65 years and older, comprising 23 self-reported fallers and 15 non-fallers. The data was acquired through functional movement tests:

  • 3-meter walking test: GRF data was collected using multi-axis force platforms.
  • Sit-to-stand (STS) test: Vertical GRF time series data was recorded using the MatScan® system.

No new benchmark datasets were introduced; the data was collected specifically for this study.

Key Insights and Ideas

  • Sample Entropy as a Powerful Feature: The study demonstrates that sample entropy, a measure of time series regularity, can effectively capture subtle differences in GRF patterns between fallers and non-fallers, even where linear methods might struggle. This highlights the value of nonlinear dynamics in biomechanical analysis for fall prediction.
  • Specific GRF Components are Key Predictors: The research identified that specific GRF components, particularly the superior-inferior GRF for the left foot during walking, and medial-lateral and anterior-posterior GRFs for the right foot during walking, along with vertical GRF for the left foot during STS, are significant predictors of fall risk. This suggests that localized biomechanical deficits are more indicative of fall risk than overall GRF magnitude.
  • LMPNN's Superior Performance: The LMPNN classifier consistently outperformed LMKNN and PNN, achieving a perfect 100% accuracy, sensitivity, and specificity. This indicates that its approach of combining local mean vectors with weighted distances is particularly effective for this type of classification problem.
  • Objective Measures Outperform Subjective Scales: The study's objective GRF-based approach achieved significantly higher performance metrics (100% accuracy) compared to the median sensitivity and specificity scores (85% and 78%) reported for traditional functional assessment tools.

Key Results

  • 100% Classification Accuracy: The LMPNN classifier, using a selected subset of features, achieved perfect classification rates: 100% accuracy, 100% sensitivity, and 100% specificity.
  • Optimal Feature Subset: The most effective features identified were:
  • Superior-inferior GRF for the left foot during walking (L_SI_F)
    • Medial-lateral GRF for the right foot during walking (R_ML_F)
    • Anterior-posterior GRF for the right foot during walking (R_AP_F)
    • Vertical GRF for the left foot during STS (L_V_F)
  • Significant Differences in GRF: Wilcoxon rank sum tests revealed statistically significant differences between fallers and non-fallers for the sample entropies of L_SI_F, R_ML_F, R_AP_F, L_V_F, and R_V_F.
  • Moderate Correlations: Spearman correlation analysis showed moderate correlations between certain GRF features, such as L_SI_F and R_AP_F, and L_V_F and R_V_F, suggesting some interdependencies in how these movements are performed.

Limitations and Future Work

  • Sample Size and Generalizability: The study acknowledges that the 100% performance might not be maintained when applied to different participants or in real-world scenarios with new users. The sample size of 38 participants, while sufficient for initial validation, could be expanded for greater generalizability.
  • Optimal Test Duration: The study used a fixed number of walking and STS tests. Future work could investigate the optimal number of repetitions for statistical processing to ensure robust data collection.
  • Sample Density Dependence: The authors note a limitation in nearest neighbor classification related to sample density. Future work will involve adding new samples at runtime to verify the superiority of the selected features and to develop a more robust fall risk assessment system.
  • Real-World Application: While promising, the study's findings need to be validated in more diverse and realistic environments to confirm their practical utility.

Summary updated

BibTeX
@misc{liang2015feature,
  author = {Shengyun Liang and Yunkun Ning and Huiqi Li and Lei Wang and Zhanyong Mei and Yingnan Ma and Guoru Zhao},
  title = {{Feature Selection and Predictors of Falls with Foot Force Sensors Using KNN-Based Algorithms}},
  year = {2015},
  doi = {10.3390/s151129393},
}
Figure