Elsevier

Journal of Biomechanics

Volume 71, 11 April 2018, Pages 37-42
Journal of Biomechanics

Machine learning algorithms based on signals from a single wearable inertial sensor can detect surface- and age-related differences in walking

https://doi.org/10.1016/j.jbiomech.2018.01.005Get rights and content

Abstract

The aim of this study was to investigate if a machine learning algorithm utilizing triaxial accelerometer, gyroscope, and magnetometer data from an inertial motion unit (IMU) could detect surface- and age-related differences in walking. Seventeen older (71.5 ± 4.2 years) and eighteen young (27.0 ± 4.7 years) healthy adults walked over flat and uneven brick surfaces wearing an inertial measurement unit (IMU) over the L5 vertebra. IMU data were binned into smaller data segments using 4-s sliding windows with 1-s step lengths. Ninety percent of the data were used as training inputs and the remaining ten percent were saved for testing. A deep learning network with long short-term memory units was used for training (fully supervised), prediction, and implementation. Four models were trained using the following inputs: all nine channels from every sensor in the IMU (fully trained model), accelerometer signals alone, gyroscope signals alone, and magnetometer signals alone. The fully trained models for surface and age outperformed all other models (area under the receiver operator curve, AUC = 0.97 and 0.96, respectively; p ≤ .045). The fully trained models for surface and age had high accuracy (96.3, 94.7%), precision (96.4, 95.2%), recall (96.3, 94.7%), and f1-score (96.3, 94.6%). These results demonstrate that processing the signals of a single IMU device with machine-learning algorithms enables the detection of surface conditions and age-group status from an individual’s walking behavior which, with further learning, may be utilized to facilitate identifying and intervening on fall risk.

Introduction

Falling during walking is a significant risk to people’s health and safety across the age span (Verma et al., 2016). In the United States, falls are the leading cause of death due to unintentional injury for older adults (56.0% in 2014) (Centers for Disease Control and Prevention, 2017). Falls in community-dwelling older adults can occur both inside and outside the home. Some studies report a greater prevalence of falls in outdoor settings (Bergland et al., 1998, Li et al., 2006, Oloughlin et al., 1994), which are predominantly associated with hazards such as uneven surfaces that are present in the built-environment (Li et al., 2006), either by design or degradation of walking infrastructure.

Previous work has shown that gait adaptations occur on uneven, compared to flat, surfaces and that older adults may not be as able as young adults to produce the required adaptive strategies to improve stability and decrease fall risk (Marigold and Patla, 2008, Menz et al., 2003, Thies et al., 2005, Zurales et al., 2016). Therefore, investigating gait biomechanics in more challenging situations like uneven as compared to flat surfaces appears to be warranted to better understand fall risk in outdoor environments.

Moreover, with the increased attention on the surveillance of physical activity for enhanced health promotion, wearable sensors play a growing role in supporting identified goals (Fulton et al., 2016). The feasibility and reliability of using wearable IMU sensors for human motion research have been confirmed from previous literature (Moe-Nilssen, 1998, Patel et al., 2012). Research and development of wearables that can detect surface conditions associated with incidents of falling are needed in order to better monitor, classify, and interpret human movement. Analytical tools are also needed to better understand circumstances of falls and the efficacy of fall prevention interventions (Pannurat et al., 2014). Previous work has shown that metrics of movement quality derived from the acceleration signal of a trunk-mounted IMU, such as step and stride regularity, can accurately detect surface differences during walking (Dixon et al., 2017) and running (Schutte et al., 2015). One primary limitation of such methods is that a significant amount of expertise and human input are needed to process the raw data and extract meaningful features for analyses. Therefore, it may be advantageous to develop advanced machine learning approaches that do not require a priori derivation on outcome measures.

Machine learning methods have been applied to a range of populations and technologies in the context of gait research to classify adults into fall risk categories, individuals with and without health conditions, and to improve estimation of minimum toe clearance. Deep learning networks (neural networks with multiple layers) are a category of machine-learning algorithms that possess multiple, non-linear, processing layers. This method has outperformed other machine-learning algorithms in many areas such as computer vision (Krizhevsky et al., 2012), audio classification (Lee et al., 2008) and even achievement of human expert level of performance (Silver et al., 2016) if given enough data and trained properly. Therefore, application of the deep learning networks warrants investigation in other contexts such as biomechanics and human motion analysis. Previous literature has used multiple machine learning algorithms for on-body sensor-based activity recognition such as Support Vector Machine (SVM) and Decision Trees (Chavarriaga et al., 2013). In that paper, machine learning algorithms showed reasonably good prediction performance (up to 0.89 f-score and 0.90 area under the curve). However, it remains unknown if deep learning networks can contribute to the human motion recognition task. In addition, it remains unclear whether such deep learning methods, coupled with simple IMU setups, can detect relatively subtle gait differences such as those induced by walking surface changes in different age groups.

Therefore, the primary aim of this study was to investigate if a deep learning network with long short term memory (LSTM) units relying on IMU input data could be used to detect surface- and age-related effects on walking. The secondary aim of this study was to determine which combination of sensors (accelerometer, gyroscope, or magnetic) provides the best classification performance. We hypothesized that (1) our deep learning network with LSTM algorithms could correctly classify trials according to surface and age and (2) the fusion of all sensor sources would result in the greatest classification performance.

Section snippets

Participants

Forty-six healthy community-dwelling young (18–35 years) and older (65+ years) adults volunteered for this study. Eleven (two young and nine old) participants were excluded after a medical screen identified neurological disorders, vertigo, dizziness, lower-limb musculoskeletal abnormalities (joint replacement, chronic or acute pain, fractures within the last two years), diabetes, obesity, or poor age-normative balance scores (O'Hoski et al., 2014). Thus, seventeen older (71.5 ± 4.2 years,

Walking surface model

The model trained with all nine channels as inputs elicited the best performance in predicting surface type with an accuracy of 96.3% (158/164 trials correctly classified), precision and recall of 96.4% and 96.3%, respectively and an f1-score of 96.3%. ROC analysis indicated that the fully trained model achieved an AUC of 0.97.

Models using only 3 accelerometer, 3 gyroscope, or 3 magnetic channels achieved AUCs of 0.92, 0.80, and 0.71, respectively (Fig. 3a). The confusion matrix of the

Discussion

This study investigated the use of machine learning algorithms to detect surface- and age-related differences in walking using signals from a single trunk-mounted IMU. In agreement with our first hypothesis, the results show that the machine learning algorithms discriminated between flat and uneven surfaces as well as between young and older adults at high levels of accuracy, precision, recall, and f1-score. Moreover, in agreement with our second hypothesis, utilizing all IMU sensors

Acknowledgements

The second author of this work was supported by the Fonds de Recherche Québec – Santé post-doctoral training award (Dixon # 33358). The authors would like to thank Amanda Rivard, Niall O’Brian, and Jacob Banks for help with the collection and initial processing of the data.

Conflict of interest

This is to declare that we had no financial or personal relationships with other people or organizations that could inappropriately influence (bias) our work submitted herein.

References (42)

  • Centers for Disease Control and Prevention, 2017. Web–based Injury Statistics Query and Reporting System (WISQARS)....
  • P.-T. de Boer et al.

    A tutorial on the cross-entropy method

    Ann. Oper. Res

    (2005)
  • E.R. DeLong et al.

    Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach

    Biometrics

    (1988)
  • P.C. Dixon et al.

    Gait adaptations of older adults on an uneven brick surface can be predicted by age-related physiological changes in strength

    Gait Post

    (2017)
  • J.E. Fulton et al.

    Strategic priorities for physical activity surveillance in the United States

    Med. Sci. Sports Exerc.

    (2016)
  • R.H. Hahnloser et al.

    Digital selection and analogue amplification coexist in a cortex-inspired silicon circuit

    Nature

    (2000)
  • S. Hochreiter et al.

    Long short-term memory

    Neural Comput.

    (1997)
  • Karpathy, A., Johnson, J., Li, F., 2015. Visualizing and understanding recurrent networks. CoRR,...
  • Kingma, D.P., Ba, J., 2014. Adam: a method for stochastic optimization. CoRR,...
  • J. Klucken et al.

    Plus One

    (2013)
  • A. Krizhevsky et al.

    ImageNet Classification with Deep Convolutional Neural Networks

  • Cited by (69)

    • Assessing exposure to slip, trip, and fall hazards based on abnormal gait patterns predicted from confidence interval estimation

      2022, Automation in Construction
      Citation Excerpt :

      Since the developed model predicted the 6th gait from the previous 5 gaits, at least 10 s of consecutive gait data were required for gait abnormality analysis. The total length of the segmented walking activities was 4.9 h. Any walking activity data shorter than 10 s were discarded; these accounted for 0.38 h. Because previous studies have already proposed walking detection from IMU sensor data [63–67], walking activity detection was outside the scope of this study. For the segmented walking activity, a total of 28,305 steps were detected, and each step was labeled as a WOSTF gait or a WSTF gait by construction safety experts.

    • Generalizability of deep learning models for predicting outdoor irregular walking surfaces

      2022, Journal of Biomechanics
      Citation Excerpt :

      Importantly, studies have also implemented machine learning to differentiate more subtle differences. Hu et al. (2018) employed a long short-term memory (LSTM) model to detect minor differences in gait patterns over flat and uneven brick surfaces (accuracy 96.3%). Similarly, Dixon et al. (2019) employed gradient boosting and convolutional neural networks (CNNs) to identify running over concrete, woodchip, and synthetic rubber surfaces via accelerometry data (> 96% accuracy).

    View all citing articles on Scopus
    View full text