Machine learning for identifying patterns in human gait
Zhou, Yuhan
DOI:
10.33612/diss.159240405
IMPORTANT NOTE: You are advised to consult the publisher's version (publisher's PDF) if you wish to cite from
it. Please check the document version below.
Document Version
Publisher's PDF, also known as Version of record
Publication date:
2021
Link to publication in University of Groningen/UMCG research database
Citation for published version (APA):
Zhou, Y. (2021). Machine learning for identifying patterns in human gait: Classification of age and clinical
groups. University of Groningen. https://doi.org/10.33612/diss.159240405
Copyright
Other than for strictly personal use, it is not permitted to download or to forward/distribute the text or part of it without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license (like Creative Commons).
Take-down policy
If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately and investigate your claim.
Downloaded from the University of Groningen/UMCG research database (Pure): http://www.rug.nl/research/portal. For technical reasons the number of authors shown on this cover page is limited to 10 maximum.
CHAPTER
EVOLUTION OF DATA SCIENCE
Advancements in technology are intended to increase comfort of human life. The intended goals can only be achieved if the vast amount of data produced by technology can be appropriately recorded, processed, and stored. Having evolved as an interface between technology-generated products and end-users, data science can process data produced by individual science disciplines but more so by interdisciplinary areas, such as natural language processing, marketing, and health science [1]. Data science is the field of study that combines domain expertise, programming skills, and mathematical and statistical knowledge to extract meaningful information from data [2]. The terminology “data science” was firstly proposed by J.W. Tukey in 1962. He established the relationship between statistics and data analysis, heralding the future of data science [3]. The field has evolved considerably over the last 50 years [4]; today, data science allows us to extract and synthesize information by recognizing patterns embedded in vast troves of information [5]. Technology companies relied more heavily on academic data scientists in earnest in the mid-1990s, as the quantity of data was so vast that it had become impossible to analyze and interpret it using existing methods [6].
Data science has become diversified and versatile in the 21st century thanks to the evolution of
sophisticated and complex data analytical methods from a variety areas of research [7]. In the 2000s, academic journals began to consider data science as a novel and independent discipline [8]. Moreover, in 2005, the National Science Board advocated for establishing expert positions in data science [4], [9]. Around 2010, the rapid development of diverse computational algorithms has led to data science becoming indispensable [4]. This process of diversification has manifested itself through the emergence of data scientists from key disciplines, including health informatics, computer science, and medicine [5].
EVOLUTION OF DATA SCIENCE IN THE MEDICAL SCIENCES
Traditionally, in clinical practice, advice and recommendations for treatment are indispensable but might also be subjective and lack accuracy, sensitivity, and specificity. The correct implementation of data science has the potential to support precision of clinical diagnoses and treatment selection. Data science approaches have now been adopted in medical imaging, drug discovery, genetics, and predictive diagnostics [10]. More than 80% of healthcare consumers are willing to use sensors and other technologies to track their lifestyle, physical activity and vital signs [11], while data science methods and advanced computational methods are gaining popularity in healthcare and medical science [12]. Since the early publications that employed explicitly data science methods in medical science (~1962), such studies in PubMed now number more than 200,000 as of 2020 (Fig. 1A).
Data science methods have been used most often in medical imaging [13], [14]. Computational approaches can be used to process medical imaging data to interpret X-ray, computed tomography,
magnetic resonance imaging (MRI) and other types of images; through identification of patterns in data, these methods can detect tumors, arterial stenosis, organ abnormalities and many more crucial pieces of information [15]. Machine learning algorithms are nowadays often used in medical sciences to increase diagnostic accuracy through analysis of previous cases [16]. The number of studies in PubMed using machine learning approaches has increased to 16,000 since the first publication appeared in 1965 (Fig. 1B).
EVOLUTION OF DATA SCIENCE IN HUMAN MOVEMENT SCIENCE
As well as in medical image processing, data science methods have also become popular in human movement science [17]. The primary application of data science has been the analysis of physical activity data across the lifespan, with more than 4,000 papers appearing in PubMed in the decades since the first such study was published in 1967 (Fig. 1C).
Historically, human movement science has employed statistical methods that make a priori assumptions about the distribution of the data, orthogonality among variables, sample sizes, and effect sizes. Wearable devices during locomotion produce large volumes of time-series data, which are high-dimensional and heterogeneous. Traditional statistical analyses are not well suited to recognize underlying patterns in this type of data. Machine learning approaches, however, can model the underlying relationships among variables and identify unique features that differentiate group members [18]. While the number of studies that applied data science approaches in human movement sciences was still small before 2000, this has increased rapidly in recent years, totaling 766 by 2020 (Fig. 1(D)). Specifically, the application of machine learning approaches to identify age-effects on movement performance and physical activity has become an important and challenging research topic in the past decade.
MACHINE LEARNING APPROACHES IN DATA SCIENCE
Machine learning is a core branch of artificial intelligence that gives computers the ability to learn without being explicitly programmed [19]. An example can be the development of computer algorithms through learning the relationships between the input and output data. There are machine learning approaches that make use of unsupervised and supervised methods [20]. Unsupervised machine learning algorithms analyze the data without prior knowledge of the labels, using algorithms such as clustering. The aim of clustering is to discover the inherent groupings in the data, assign specific gait patterns to specific neurological patients [20]. For instance, K-means clustering was used to recognize gait patterns in children with cerebral palsy who used orthoses [21]. The supervised machine learning algorithm takes a set of input data with the known classes (labels) and trains a model based on an automatic adaptable algorithm. The dataset is then split into a training set and a testing set so that the algorithm can learn to predict the label in the testing set based on the distribution of the data in the training set [20].
Figure 1. Development of data science and machine learning methods employed in medical and human
movement sciences in academia and published by the PubMed database. (A) and (B) show the histogram of publications related to investigating medical science by using data science methods or the more specific method of machine learning over the last 60 years. (C) and (D) illustrate the number of publications related to human movement science with data science methods from 1967 to October 2020, and with machine learning methods from 1994 to October 2020, respectively. The data of this figure extracted from PubMed. Supervised learning algorithms can be grouped into algorithms related to classification and algorithms related to regression problems. Classification predicts a discrete class label, such as fallers or non-fallers. For the regression approach, the output variable is a real value, such as gait speed. Examples of commonly used supervised machine learning algorithms are Random Forest (RF), Support Vector Machine (SVM) and Artificial Neural Network (ANN) methods. RF constructs multiple decision trees randomly based on the data correlations [22]. SVM determines support vectors to maximize the separation (margin) between different classes [23]. ANN calculates the correlations and interactions between neurons in hidden layers that weigh
their inputs according to activation functions in order to output the predicted labels and enable comparisons with the corresponding original labels [24]. The evaluation metrics of machine learning classification include accuracy, sensitivity, specificity, the area under the receiver operating characteristic curve (AUC) and the model validation F1 score.
In order to improve machine learning classification performance, data processing methods such as kernel PCA have been widely implemented to reduce redundancies in datasets and generate new and independent variables, while including the original data in its entirety [25]. For example, when the input data is non-linear, such as dynamic gait outcomes, the radial basis function kernel or polynomial kernel can be used to transfer the non-linear data structure to linear-separated data structure, then PCA was used to create a set of orthogonal bases that identify the directions of maximum variance and the uncorrected expansion coefficients, deducing the dimension of these new linear data. The data processed by Kernel PCA in advance to reduce the redundancy and dimensions, so that the SVM could take the data with less noise as inputs and run with a quicker speed, therefore the accuracy was improved [26]. Another novel data pre-processing method is the Path Signature method, which is designed to solve rough path by the generation of unique geometric features from original non-linear features [27], [28].
Several constraints must be taken into account when employing machine learning approaches in human movement science for clinical purposes. For instance, more variables might be selected than the number of subjects as input to a machine learning classifier; this would result in a good classification performance for the training model, but the generalizability of the trained model to new data will be very low, because of an excessively complicated machine learning model, a problem is known as ‘overfitting’ will occur due to overtraining on a dataset [18]. Another consideration is the “black box” problem for some machine learning approaches, such as SVM. It means that the model has limited explanation capability in terms of identifying critical variables that are highly associated with target population groups [29]. Therefore, the selection of a machine learning method needs to consider these limitations of small data size and model transparency. The SVM model can process small size clinical data, and the RF model provides a transparent computing procedure for classification.
In the present study, we compared several supervised machine learning methods to identify gait patterns for the classification of age groups, fallers, or geriatric patients with cognitive impairment. Gait was assessed by collecting data with wearable inertial sensors from which spatial-temporal gait or signal-based dynamic gait variables were calculated. These variables can be highly correlated and contain redundant information. For instance, gait speed is strongly correlated with stride time [30], and gait regularity relates closely to gait symmetry [31]. In addition to the linearly correlated spatial-temporal gait variables [32], dynamic gait outcomes might be interrelated in a non-linear fashion [33], which requires the optimal machine learning classification model to analyze and interpret the complex gait variables that have the most outstanding contribution in distinguishing groups.
QUANTITATIVE GAIT MEASURES
Recent advancements in technology have enabled the development of wearable sensors to assist clinicians in gait analysis [34]. Inertial measurement units (IMU) are one such device which is widely used in the laboratory to analyze a person’s movement because IMUs are portable, inexpensive, and can be used in a real world setting [35]. Accelerometers record linear accelerations in 3D during daily activities, and these signals can be mathematically transformed into variables that can characterize gait.
Gait speed has been used in previous studies as a predictor of geriatric status, survival and life expectancy [36] [37]; however, since geriatric patients have many comorbidities, gait speed alone cannot represent all of the changes caused by aging and disorders [38]. In this case, additional gait variables should be considered as part of a comprehensive geriatric assessment.
A commonly used gait variable set for geriatric assessment involves spatial-temporal gait variables, derived from IMU recordings. Such data capture the characteristics of gait such as the coefficients of variation of step velocity, stride length, swing time and cadence [39]. Spatial-temporal gait variables can be used to classify different neurological conditions among older adults, to identify the specific neurological gait patterns by machine learning methods [40]. Another set of gait variables are variables that quantify how gait evolves over time, so-called dynamic gait outcomes [41], [42]. A variety of measures are employed in analyzing the accelerometer signals to categorize the dynamics of gait into several domains such as pace, variability, stability, regularity, predictability, smoothness, synchronization and symmetry [43], [44]. Previous studies have used dynamic gait outcomes to classify geriatric patients with or without the risk of falls and cognitive impairment [33], [41].
MACHINE LEARNING APPROACHES APPLIED TO GAIT CLASSIFICATION
The extracted features from accelerometer data recorded by wearable sensors while walking can be processed by the supervised machine learning classification approaches in order to classify different population groups based on clinical and gait variables [45]. Previous studies have successfully utilized machine learning approaches to recognize gait changes in different groups of populations [46]. For instance, based on the sensor’s gait features, 12 older women with Alzheimer’s disease and 12 healthy older women were successfully classified by SVM with an accuracy of 91% [47]. In another study, the foot drop gait disorder was distinguished from the normal walking gait pattern using the wrapper feature selection technique in combination with Random Forest (RF) based on time series gait data, to obtain an accuracy of 93% [48]. Using the spatial-temporal gait features, long short-term memory neural networks (LSTM) successfully identified neurological disorder patients with or without fall risk, achieving a satisfactory classification accuracy of 92.1% [49]. Furthermore, Parkinson’s disease (PD) patients with or without mild cognitive impairment (MCI) have been classified by using PCA in combination with the SVM model: based on spatial-temporal gait parameters, Time Up and Go and jump tests, these
methods obtained a successful classification performance with 91.67% accuracy [50]. In another study, spatial-temporal parameters were used in isolation to classify PD patients with and without MCI, by using KNN with a high AUC of 0.9 under the cognitive dual-task [51].
CLINICAL IMPLICATION OF GAIT
Large proportions of older adults continue to experience abnormal aging, such as an excessive decline of various bodily functions leading to an increased risk of falls, morbidity and mortality. The co-occurrence of aging and disease gives rise to geriatric syndromes [52]. For instance, geriatric patients suffer from sarcopenia, cognitive impairment, osteoporosis, weight loss, frailty, and of course high chronological age [52], [53]. Human gait depends on the complex interaction between neural, musculoskeletal and cardiorespiratory function. The prevalence of gait impairment increases with age so that older people walk with shorter and wider steps compared with younger adults, as well as having increased step time and variability [54], [55]. In addition to the gait features occurring during natural aging, geriatric patients have further modifications in gait, including even slower speed, shuffling, and a cautious walking pattern. Thus, an individual’s development of a characteristically old and geriatric gait could be comprised of premature and subtle changes that untrained eyes would not be able to accurately identify. Moreover, the traditional clinical assessment of geriatric gait, such as visual observation with or without the help of gait lab procedures, might be time-consuming and subjective [56]. Data science techniques such as machine learning approaches can be potentially helpful in speeding up and improving the accuracy of the diagnostic process.
Gait is not only an automated motor activity, but also an activity related to cognitive function. Gait control is associated with several cognitive domains, such as memory resources, executive attentional function or visuospatial abilities; gait-cognition relationships actually arose from neuropsychological studies [57]. It is also known that psychological factors can have a significant impact on gait: depression can slow gait, and anxiety can produce an overly cautious gait [58]. Severe cognitive impairment is also associated with an impaired gait [59]. For example, poorer executive function and attention performance are associated with increased gait variability [60], while gait stability relates closely to visuospatial abilities [61].
Concurrent with cognitive decline, gait disorders frequently occur within older adults who have high fall risk [62]. Falls are the main cause of injuries in geriatric patients, and the risk of falls increases when geriatric patients have an impaired gait [63], [64]. Gait characteristics might therefore be used for the classification of different groups, such as people with different age, older adults with and without cognitive impairment and/or with or without fall risk. In addition, by identifying changes in gait characteristics, specific disorders such as dementia can be identified in an early stage [65], [66]. Machine learning approaches can assist in identifying the gait characteristics that are changed due to for instance age and changes in cognitive function.
OBJECTIVES AND OUTLINE OF THE THESIS
The overall objective of this thesis was to classify groups of individuals based on age, fall history, or cognitive status using machine learning approaches. The hypothesis was that if gait characteristics differ between age groups, between fallers and non-fallers, these groups could be classified accurately based on gait variables as inputs to the classification models. Additionally, geriatric patients with or without cognitive impairment present with comorbidities, therefore, using a combination of gait variables and standard clinical tests could accurately classify these groups rather than inputting only gait or only clinical variables to the models. Optimal machine learning models were identified by comparing classification performance. The overarching idea is to identify features from gait that can eventually assist clinicians in their diagnosis of gait and cognitive impairments and screen for patients with fall risk. The overall structure of Chapter 2 to Chapter 5 is shown in Figure 2.
Firstly, in Chapter 2, the aim is to classify young-middle age adults, healthy older adults and geriatric patients without cognitive impairment based on dynamic gait outcomes calculated from accelerometer signals by comparing three machine learning classification models: RF, ANN and Kernel PCA in combination with SVM. The most accurate classification model has the potential to determine the specific dynamic gait outcomes that were sensitive to aging. The machine leaning models RF and ANN were employed in Chapter 3, to classify geriatric patients with or without cognitive impairment based on the same dynamic gait outcomes that were used in Chapter 2, with additional standard clinical tests: including Charlson Comorbidity Index (CCI), Geriatric depression scale (GDS), frailty criteria (Frail), Timed-Up-and-Go (TUG), hand grip strength (HandGrip), the number of medications used (NumMed), Drug Burden Index (DBI) and Body mass index (BMI). The classification models were established based on clinical assessments only, dynamic gait outcomes only or combined clinical and gait data, to examine the interactions between variables which were hypothesized to improve classification accuracy. Besides, the contribution of the variables to the most accurate classification model were determined. Based on 3D accelerometers, 27 spatial-temporal gait characteristics under single task and two dual-task gait conditions (walking while checking boxes on a paper sheet, and walking while serially subtracting a number for 7s) were quantified and used as input in Chapter 4, the aim of this chapter is to classify patients with multiple neurological disorders into fallers and non-fallers based on patients’ retrospective falls, using Partial Least Square Discriminant Analysis. The model identified gait characteristics highly associated with falls but irrespective of the specific clinical signature of neurodegenerative disorder. However, the accuracy in this study was average. Therefore, the aim in Chapter 5 was to compare no data pre-processing, PCA or the Path Signature method in combination with six machine learning methods. These machine learning classifiers are from distance-based, linear-based, tree-based and prior-based domains, to classify fallers and non-fallers based on the same group of neurological patients and the same spatial-temporal gait characteristics under a single task walking used in Chapter 4. The study showed that the Path Signature method greatly improved the machine learning classification of fallers and non-fallers.
Finally, Chapter 6 provides a summary of the main results, a general discussion and a conclusion including possible clinical implications and suggestions for future directions.
Figure 2. Machine learning algorithms include unsupervised methods and supervised methods. The present
study used Random Forest (RF), Artificial Neural Network (ANN) and the Kernel Principal Component Analysis (PCA) in combination with Support Vector Machine (SVM) classified healthy young-middle (Y/M) age adults, healthy older adults and geriatric patients based on dynamic gait outcomes. Another pre-pro-cessing method Path Signature in combination with six machine learning classification methods is able to perfectly classify fallers and non-fallers, based on spatial-temporal gait variables. Without any data pre-pro-cessing methods, the classification methods ANN and RF were used to classify geriatric patients with or without cognitive impairment according to dynamic gait outcomes and clinical assessments. Finally, Partial Least Square Discriminant Analysis (PLS-DA) was directly used to classify fallers and non-fallers based on spatial-temporal gait variables. LR: Linear Regression; LDA: Linear Discriminant Analysis; KNN: K-Nearest Neighbors; NB: Naïve Bayes.
REFERENCE
1. A. Karpatne et al., “Theory-guided data sci-ence: A new paradigm for scientific discov-ery from data,” IEEE Trans. Knowl. Data
Eng., vol. 29, no. 10, pp. 2318–2331, Oct. 2017.
2. V. Dhar, “Data science and prediction,”
Commun. ACM, vol. 56, no. 12, pp. 64–73,
Dec. 2013.
3. J. W. Tukey, “The Future of Data Analysis,”
Ann. Math. Stat., vol. 33, no. 1, pp. 1–67, 1962.
4. D. Donoho, “50 Years of Data Science,” J.
Comput. Graph. Stat., vol. 26, no. 4, pp. 745–
766, Oct. 2017.
5. G. Peyré and M. Cuturi, “Computational op-timal transport,” Found. Trends Mach. Learn., vol. 11, no. 5–6, pp. 1–257, 2019.
6. L. Cao, “Data science: A comprehensive over-view,” ACM Computing Surveys, vol. 50, no. 3. Association for Computing Machinery, 01-Jun-2017.
7. F. Murtagh and K. Devlin, “The Development of Data Science: Implications for Education, Employment, Research, and the Data Revolu-tion for Sustainable Development,” Big Data
Cogn. Comput., vol. 2, no. 2, p. 14, Jun. 2018.
8. R. A. Irizarry, “The Role of Academia in Data Science Education,” Harvard Data Sci. Rev., vol. 2, no. 1, Jan. 2020.
9. W. S. Cleveland, “Data Science: an Action Plan for Expanding the Technical Areas of the Field of Statistics,” Int. Stat. Rev., vol. 69, no. 1, pp. 21–26, Apr. 2001.
10. S. Consoli, D. R. Recupero, and M. Petković,
Data science for healthcare: Methodologies and applications. Springer International
Publish-ing, 2019.
11. A. Phaneuf, “Wearable Tech in Healthcare: Smart Medical Devices & Trends in 2020 - Business Insider,” Jan. 2020.
12. H. M. Krumholz, “Big data and new knowl-edge in medicine: The thinking , training , and tools needed for a learning health system,”
Health Aff., vol. 33, no. 7, pp. 1163–1170, 2014.
13. M. I. Razzak, S. Naz, and A. Zaib, “Deep learning for medical image processing: Over-view, challenges and the future,” in Lecture
Notes in Computational Vision and Biome-chanics, vol. 26, Springer Netherlands, 2018,
pp. 323–350.
14. C. Parmar, J. D. Barry, A. Hosny, J. Quack-enbush, and H. J. W. L. Aerts, “Data Analysis Strategies in Medical Imaging,” Clin. Cancer
Res., vol. 24, no. 15, pp. 3492–3499, Aug. 2018.
15. D. S. W. Ting, Y. Liu, P. Burlina, X. Xu, N. M. Bressler, and T. Y. Wong, “AI for medical imaging goes deep,” Nat. Med., vol. 24, no. 5, pp. 539–540, May 2018.
16. B. J. Erickson, P. Korfiatis, Z. Akkus, and T. L. Kline, “Machine learning for medical imag-ing,” Radiographics, vol. 37, no. 2, pp. 505–515, Mar. 2017.
17. S. Dodge, “A Data Science Framework for Movement,” Geogr. Anal., vol. 0, p. gean.12212, Jun. 2019.
18. E. Halilaj, A. Rajagopal, M. Fiterau, J. L. Hicks, T. J. Hastie, and S. L. Delp, “Machine learning in human movement biomechanics: Best prac-tices, common pitfalls, and new opportuni-ties,” J. Biomech., vol. 81, pp. 1–11, Nov. 2018. 19. A. L. Samuel, “Some studies in machine learn-ing uslearn-ing the game of checkers,” IBM J. Res.
Dev., vol. 44, no. 1.2, pp. 206–226, 2000.
20. M. Alloghani, D. Al-Jumeily, J. Mustafina, A. Hussain, and A. J. Aljaaf, “A Systematic Review on Supervised and Unsupervised Ma-chine Learning Algorithms for Data Science,” Springer, Cham, 2020, pp. 3–21.
21. J. Choisne et al., “An Unsupervised Data-Driv-en Model to Classify Gait Patterns in ChildrData-Driv-en with Cerebral Palsy,” J. Clin. Med., vol. 9, no. 5, p. 1432, May 2020.
22. W.-Y. Loh, “Variable Selection for Classi-fication and Regression in Large p, Small n Problems,” in Probability Approximations
and Beyond, 2012th ed., D. S. Andrew
Barbo-ur, Hock Peng Chan, Ed. Springer New York, 2012, pp. 133–157.
23. P. P. Matykiewicz and J. Pestian, “Effect of small sample size on text categorization with support vector machines,” in BioNLP:
Biomedi-cal Natural Language Processing, 2012, pp.
193–201.
24. C. Bircanoglu and N. Arica, “A comparison of activation functions in artificial neural networks,” in 2018 26th Signal Processing and
Communications Applications Conference (SIU), 2018, pp. 1–4.
25. S. García, S. Ramírez-Gallego, J. Luengo, J. M. Benítez, and F. Herrera, “Big data preprocess-ing: methods and prospects,” Big Data Anal., vol. 1, no. 1, pp. 1–22, Dec. 2016.
26. J. Wu, J. Wang, and L. Liu, “Feature extraction via KPCA for classification of gait patterns,”
Hum. Mov. Sci., vol. 26, no. 3, pp. 393–411,
Jun. 2007.
27. P. J. Moore, T. J. Lyons, and J. Gallacher, “Using path signatures to predict a diagnosis of Alzheimer’s disease,” PLoS One, vol. 14, no. 9, Sep. 2019.
28. I. Chevyrev and A. Kormilitzin, “A primer on the signature method in machine learning,”
arXiv Prepr. arXiv1603.03788, Mar. 2016.
29. D. S. Watson et al., “Clinical applications of machine learning algorithms: Beyond the black box,” BMJ, vol. 364, Mar. 2019. 30. O. Beauchet et al., “Walking speed-related
changes in stride time variability: effects of decreased speed.,” J. Neuroeng. Rehabil., vol. 6, p. 32, Aug. 2009.
31. A. Tura, M. Raggi, L. Rocchi, A. G. Cutti, and L. Chiari, “Gait symmetry and regularity in transfemoral amputees assessed by trunk ac-celerations,” J. Neuroeng. Rehabil., vol. 7, no. 1, pp. 1–10, Jan. 2010.
32. Y. Zhou et al., “Classification of Neurologi-cal Patients to Identify Fallers Based on Spa-tial-Temporal Gait Characteristics Measured by a Wearable Device,” Sensors 2020, Vol. 20,
Page 4098, vol. 20, no. 15, p. 4098, Jul. 2020.
33. L. H. J. Kikkert, N. Vuillerme, J. P. Van Campen, B. A. Appels, T. Hortobágyi, and C. J. C. Lamoth, “The relationship between gait dynamics and future cognitive decline: A pro-spective pilot study in geriatric patients,” Int.
Psychogeriatrics, vol. 30, no. 9, pp. 1301–1309,
Sep. 2018.
34. A. Muro-De-La-Herran, B. Garcia-Zapirain, and A. Mendez-Zorrilla, “Gait analysis meth-ods: An overview of wearable and non-wear-able systems, highlighting clinical applica-tions,” Sensors, vol. 14, no. 2, pp. 3362–3394, Feb. 2014.
35. L. Donath, O. Faude, E. Lichtenstein, C. Nuesch, and A. Mündermann, “Validity and reliability of a portable gait analysis system for measuring spatiotemporal gait characteristics: Comparison to an instrumented treadmill,”
J. Neuroeng. Rehabil., vol. 13, no. 1, pp. 1–9,
Jan. 2016.
36. N. M. Peel, S. S. Kuys, and K. Klein, “Gait speed as a measure in geriatric assessment in clinical settings: A systematic review,”
Journals of Gerontology - Series A Biological Sciences and Medical Sciences, vol. 68, no. 1.
Oxford Academic, pp. 39–46, 01-Jan-2013. 37. S. Studenski et al., “Gait speed and survival in
older adults,” JAMA - J. Am. Med. Assoc., vol. 305, no. 1, pp. 50–58, Jan. 2011.
38. L. H. J. Kikkert, N. Vuillerme, J. P. van Campen, T. Hortobágyi, and C. J. Lamoth, “Walking ability to predict future cogni-tive decline in old adults: A scoping review,”
Ageing Res. Rev., vol. 27, pp. 1–14, May 2016.
39. E. P. Washabaugh, T. Kalyanaraman, P. G. Adamczyk, E. S. Claflin, and C. Krishnan, “Validity and repeatability of inertial mea-surement units for measuring gait param-eters,” Gait Posture, vol. 55, pp. 87–93, Jun. 2017.
40. C. Pradhan et al., “Automated classification of neurological disorders of gait using spa-tio-temporal gait parameters,” J.
Electromyo-gr. Kinesiol., vol. 25, no. 2, pp. 413–422, Apr.
2015.
41. L. H. J. J. Kikkert et al., “Gait dynamics to op-timize fall risk assessment in geriatric patients admitted to an outpatient diagnostic clinic,”
PLoS One, vol. 12, no. 6, p. e0178615, Jun. 2017.
42. N. M. Kosse, N. Vuillerme, T. Hortobágyi, and C. J. Lamoth, “Multiple gait parameters de-rived from iPod accelerometry predict age-re-lated gait changes,” Gait Posture, vol. 46, pp. 112–117, May 2016.
43. D. Hamacher et al., “Kinematic measures for assessing gait stability in elderly individuals: a systematic review,” J. R. Soc. Interface, vol. 8, no. 65, pp. 1682–1698, Dec. 2011.
44. J. J. Kavanagh and H. B. Menz, “Accelerom-etry: A technique for quantifying movement patterns during walking,” Gait Posture, vol. 28, no. 1, pp. 1–15, Jul. 2008.
45. A. Phinyomark, G. Petri, E. Ibáñez-Marce-lo, S. T. Osis, and R. Ferber, “Analysis of Big Data in Gait Biomechanics: Current Trends and Future Directions,” J. Med. Biol. Eng., vol. 38, no. 2, pp. 244–260, Apr. 2018.
46. J. Figueiredo, C. P. Santos, and J. C. Moreno, “Automatic recognition of gait patterns in human motor disorders using machine learn-ing: A review,” Med. Eng. Phys., vol. 53, pp. 1–12, Mar. 2018.
47. M. Seifallahi, H. Soltanizadeh, A. Hassani Mehraban, and F. Khamseh, “Detection of Alzheimer’s Disease in Elder People Using Gait Analysis and Kinect Camera,” J. Heal.
Biomed. Informatics, vol. 6, no. 3, pp. 178–196,
Dec. 2019.
48. S. Sharif Bidabadi, I. Murray, G. Y. F. Lee, S. Morris, and T. Tan, “Classification of foot drop gait characteristic due to lumbar radic-ulopathy using machine learning algorithms,”
Gait Posture, vol. 71, pp. 234–240, Jun. 2019.
49. C. Tunca, G. Salur, and C. Ersoy, “Deep Learning for Fall Risk Assessment with Iner-tial Sensors: Utilizing Domain Knowledge in Spatiooral Gait Parameters,” IEEE J. Biomed.
Heal. Informatics, vol. 24, no. 7, pp. 1994–
2005, Jul. 2020.
50. P. H. Chen, C. W. Lien, W. C. Wu, L. S. Lee, and J. S. Shaw, “Gait-Based Machine Learning for Classifying Patients with Different Types of Mild Cognitive Impairment,” J. Med. Syst., vol. 44, no. 6, pp. 1–6, Jun. 2020.
51. C. Ricciardi et al., “Machine learning can detect the presence of Mild cognitive im-pairment in patients affected by Parkinson’s Disease,” in IEEE Medical Measurements and
Applications, MeMeA 2020 - Conference Pro-ceedings, 2020.
52. S. K. Inouye, S. Studenski, M. E. Tinetti, and G. A. Kuchel, “Geriatric syndromes: Clinical,
research, and policy implications of a core geriatric concept,” Journal of the American
Geriatrics Society, vol. 55, no. 5. NIH Public
Access, pp. 780–791, May-2007.
53. A. Marengoni et al., “Aging with multimor-bidity: A systematic review of the literature,”
Ageing Res. Rev., vol. 10, no. 4, pp. 430–439,
Sep. 2011.
54. A. Aboutorabi, M. Arazpour, M. Bahramiza-deh, S. W. Hutchins, and R. Fadayevatan, “The effect of aging on gait parameters in able-bod-ied older subjects: a literature review,” Aging
Clin. Exp. Res., vol. 28, no. 3, pp. 393–405,
Jun. 2016.
55. J. H. Hollman, E. M. McDade, and R. C. Pe-tersen, “Normative spatiotemporal gait pa-rameters in older adults,” Gait Posture, vol. 34, no. 1, pp. 111–118, May 2011.
56. H. J. Woodford and J. George, “Cognitive as-sessment in the elderly: A review of clinical methods,” QJM, vol. 100, no. 8. Oxford Aca-demic, pp. 469–484, 01-Aug-2007.
57. M. Amboni, P. Barone, and J. M. Hausdorff, “Cognitive contributions to gait and falls: Evi-dence and implications,” Movement Disorders, vol. 28, no. 11. Mov Disord, pp. 1520–1533, 15-Sep-2013.
58. T. C. Brandler, C. Wang, M. Oh-Park, R. Holtzer, and J. Verghese, “Depressive symp-toms and gait dysfunction in the elderly,”
Am. J. Geriatr. Psychiatry, vol. 20, no. 5, pp.
425–432, 2012.
59. M. B. Van Iersel, W. Hoefsloot, M. Munneke, B. R. Bloem, and M. G. M. Olde Rikkert, “Sys-tematic review of quantitative clinical gait analysis in patients with dementia,” Zeitschrift
fur Gerontologie und Geriatrie, vol. 37, no. 1. Z
Gerontol Geriatr, pp. 27–32, Feb-2004. 60. S. Springer, N. Giladi, C. Peretz, G. Yogev, E.
S. Simon, and J. M. Hausdorff, “Dual-tasking effects on gait variability: The role of aging, falls, and executive function,” Mov. Disord., vol. 21, no. 7, pp. 950–957, Jul. 2006. 61. E. Y. Uc, M. Rizzo, S. W. Anderson, S. Qian, R.
L. Rodnitzky, and J. D. Dawson, “Visual dys-function in Parkinson disease without demen-tia,” Neurology, vol. 65, no. 12, pp. 1907–1913, Dec. 2005.
62. H. Axer, M. Axer, H. Sauer, O. W. Witte, and G. Hagemann, “Falls and gait disorders in geriatric neurology,” Clinical Neurology and
Neurosurgery, vol. 112, no. 4. Clin Neurol
Neurosurg, pp. 265–274, May-2010.
63. M. E. Taylor, K. Delbaere, S. R. Lord, A. S. Mikolaizak, H. Brodaty, and J. C. T. Close, “Neuropsychological, physical, and function-al mobility measures associated with ffunction-alls in cognitively impaired older adults,” Journals
Gerontol. - Ser. A Biol. Sci. Med. Sci., vol. 69,
no. 8, pp. 987–995, 2014.
64. J. Verghese, A. LeValley, C. B. Hall, M. J. Katz, A. F. Ambrose, and R. B. Lipton, “Epidemiol-ogy of gait disorders in community-residing older adults,” J. Am. Geriatr. Soc., vol. 54, no. 2, pp. 255–261, Feb. 2006.
65. O. Beauchet et al., “Poor Gait Performance and Prediction of Dementia: Results From a Meta-Analysis,” Journal of the American
Med-ical Directors Association, vol. 17, no. 6.
Else-vier Inc., pp. 482–490, 01-Jun-2016.
66. L. Bahureksa et al., “The Impact of Mild Cognitive Impairment on Gait and Balance: A Systematic Review and Meta-Analysis of Studies Using Instrumented Assessment,”
Gerontology, vol. 63, no. 1. S. Karger AG, pp.
67–83, 01-Dec-2016.
