Community-Based and Everyday Life Gait Analysis: Approach to an Automatic Balance Assessment and Fall Risk Prediction in the Elderly

*Britam Arom Gómez Arias, Sebastián Gonzalo Chávez Orellana, Paulina Cecilia Ortega-Bastidas and Pablo Esteban Aqueveque Navarro*

## **Abstract**

This chapter discusses the potential of wearable technologies in predicting fall risks among older adults, a demographic susceptible to falls due to age-related walking ability decline. We aimed to explore the feasibility of portable body sensors, mobile apps, and smartwatches for real-time gait analysis in non-clinical, everyday settings. We used classification models like Random Forest, Support Vector Machine with a radial basis function kernel, and Logistic Regression to predict fall risks based on gait parameters. Notably, both Random Forest and Support Vector Machine models demonstrated over 72% accuracy, underscoring the critical role of feature selection and model choice in fall risk prediction. These technologies can enhance older adults' quality of life by predicting fall risks. However, future developments should focus on technologies adapted to nonclinical environments, predictivity, and high-risk group usability. The integration of these features may enable more efficient fall risk assessment systems.

**Keywords:** gait analysis, fall risk prediction, wearable technology, elderly care, community-based assessment

## **1. Introduction**

Aging is associated with a decline in ambulatory skills as a result of a variety of factors, which have been linked to an increase in the incidence of falls in older adults over 65, being the leading cause of traumatic injury in this population [1, 2].

At least one-third of older adults fall once a year [3, 4]. Alterations in gait and falls are relevant issues in this population, potentially causing limitations in daily life and associating with a range of adverse outcomes, including death [5, 6]. These alterations have a prevalence that progressively increases with age, reaching more than 60% above 80 years old [7].

Importantly, the majority of falls occur during ambulation in the community [1]. Community ambulation has been defined as locomotion outdoors that includes activities necessary for independent living such as visiting the bank, the pharmacy, and the supermarket. Maintaining independent community ambulation can be integral to the quality of life of older adults and their ability to participate socially [6, 8].

Ambulation requires a series of mechanisms that must interact in a coordinated manner for it to be safe and effective [7, 9]. An important determinant is the performance of physical skills such as gait speed and body balance [6, 10]. It has been observed that from 60 years of age, gait speed decreases by about 1% per year [4, 6]. Also, other parameters have been seen to change with age, such as cadence, stride length, angular displacement, joint torque, and power [6, 11].

Commonly, the identification of these variables and parameters in clinical contexts is carried out through scales and tests [12, 13]. Nowadays, due to the subjectivity of these tools [14], the use of instrumentation and/or technologies to increase their predictive value and objectivity has progressively been associated with their use [15, 16].

It is important to note that the majority of these tools are implemented in safe and controlled clinical contexts [14]. However, it has been observed that falls are influenced by various factors, with the environment being a significant determinant in their performance [16, 17]. While the environment alone is not a cause of falls, there is an interaction between the characteristics of the environment that can be dangerous and the intrinsic risk factors of older people [17]. Therefore, restricting evaluations to clinical contexts does not allow for the identification of real issues that may arise according to the different characteristics of the environment and community spaces [14].

The objective of this chapter is to present a novel methodology for gait analysis, concentrating on the application of technological advancements in everyday community settings. We will delve into the current state of the art, highlighting gaps in technology development and its applicability, particularly with respect to gait monitoring initiatives that are based on clinical parameters. Furthermore, we will explore a specific case study that involves an approach based on lower-back accelerometry data. This data, collected from both elderly individuals prone to falls and those not, will be analyzed to evaluate the effectiveness of automated classifiers in predicting fall risks. The overarching aim of this investigation is to expand the scope of gait analysis, bringing it closer to the reality of everyday life for older adults, thereby enhancing its predictive capacity for fall risks.

## **2. Background and related work**

#### **2.1 Methods and technologies for assessing spatiotemporal gait parameters**

A substantial body of research has delved into various techniques and devices for assessing crucial spatiotemporal gait parameters such as cadence, step time, and gait asymmetry. The focus of such research often involves practical applications in everyday settings.

The literature reveals wearable body sensors, including accelerometers and gyroscopes, as commonly used tools for this purpose [18–21]. **Figure 1** illustrates some commercially available systems [22–24].

*Community-Based and Everyday Life Gait Analysis: Approach to an Automatic Balance… DOI: http://dx.doi.org/10.5772/intechopen.112873*

#### **Figure 1.**

*Commercial systems to assess gait parameters in uncontrolled or community environments. Images from BTS Gwalk system website [22], QMUV website [23], and APDM mobility lab website [24].*

**Figure 2.** *Mobile application that use in-built sensors to assess gait. Image from ref. [27].*

A more controlled approach involves instrumented walkway systems [25], capturing footfall information to derive extensive gait insights.

Mobile technologies have also entered the field, with smartphone applications designed for gait analysis [26–28]. Utilizing the phone's built-in sensors, these applications deliver real-time metrics and feedback (**Figure 2**).

In addition to these, instrumented insoles and shoes [25, 29–32] provide a seamless solution for gait evaluation. Equipped with pressure and motion sensors, these devices offer critical insights into foot loading and movement patterns. **Figure 3** shows a commercial instrumented insole and **Figure 4** illustrates a capacitive instrumented insole to assess gait automatically.

Smartwatches and other commercially available wearables, as mentioned in Refs. [18, 21, 34, 35], offer an attractive, non-intrusive solution. These devices provide continuous monitoring and recording of motion data, allowing for real-time and longitudinal gait analysis.

#### **Figure 3.**

*Commercial systems to assess gait parameters in uncontrolled or community environments. Image from motion sensor insoles system website [33].*

**Figure 4.** *Capsens SPA instrumented insole system to assess gait automatically. Image from ref. [32].*

### **2.2 Identifying gaps in existing literature**

While the variety of gait analysis tools is extensive, certain drawbacks limit their application, particularly in non-clinical, everyday settings. Many of the present technologies require specific conditions for optimum outcomes, restricting their utility for routine use.

A prevalent focus of many tools is to capture existing gait parameters, with little emphasis on predicting future instability or fall risk [18–21]. Additionally, there is a notable scarcity of user-friendly systems that can be easily employed by high-risk groups, such as the elderly [26, 27].

The intricate nature of gait, involving environmental challenges, cognitive distractions, or medication-induced changes, is often overlooked [25, 30, 31].

*Community-Based and Everyday Life Gait Analysis: Approach to an Automatic Balance… DOI: http://dx.doi.org/10.5772/intechopen.112873*

These observations underline the demand for versatile, easy-to-use gait analysis devices that can be integrated into everyday use. Importantly, these devices should integrate predictive models to spot potential anomalies and elevate fall risk. Catering to these demands could substantially strengthen fall prevention approaches, potentially enhancing the safety and quality of life for high-risk groups.

## **3. Our approach**

Spatiotemporal parameters of gait refer to measures that describe how a person walks in relation to space and time. They are a fundamental aspect of gait analysis, providing valuable insights into a person's locomotion ability.

Estimating the spatiotemporal parameters of gait typically involves identifying specific gait events, which are distinct points in the gait cycle. Here are the primary gait events and how they relate to the estimation of different parameters:


For more information see **Figure 5**.

While this proposal considers any technological system capable of extracting spatiotemporal parameters from the gait cycle in an uncontrolled or relevant environment for older adults (including smartwatches, smartphones, instrumented shoes/ insoles, low back accelerometry, chest accelerometry, in-home video analysis, and more), in this chapter we extracted the necessary parameters for fall risk estimation by analyzing the Physionet "Long Term Movement Monitoring Database" [36]. This

database contains mobility data from 71 elderly individuals (mean age = 78.36�4.71 years; age range 65 to 87 years) during their everyday activities. Within this repository, there are low back accelerometry measurements using a belt with inertial sensors (three-axis accelerometer and three-axis gyroscope) while performing daily life activities, separated into a control group and a fall risk group. These groups were classified based on each volunteer's fall history over the past 12 months as provided by the repository.

The entire process of characterization, feature selection, and training of the classification algorithms was carried out using Python 3.11 with the scikit-learn 1.2.2 package. The development was conducted within the DataSpell 2023.1 Integrated Development Environment (IDE) on a computer running Windows 11 Enterprise (version 22621.963). The system specifications of the computer were as follows: 64-bit operating system, x64-based processor (Intel(R) Xeon(R) CPU E5–2678 v3 @ 2.50GHz, 2.50 GHz), 63.9 GB of installed RAM, and an AMD Radeon RX 5802048SP GPU.

### **3.1 Feature extraction: Sample analysis**

From the Physionet repository, we manually identified and extracted the segments of the measurement that corresponded to independent gait, whether it was in a straight line or otherwise. Of the total test subjects, 69 were evaluated with the method proposed by González et al. [37] for the processing of triaxial lower back accelerometry for the estimations of gait cycle events.

The method consists of processing the signals from each accelerometer axis, which will be mentioned below as: antero-posterior acceleration signal (A–P), vertical acceleration signal (V), and medio-lateral acceleration signal (M–L). This is done to avoid tying the sensor orientation to one and present a general algorithm.

During gait execution, a characteristic cyclic waveform appears from the A–P, V, and M–L accelerometer signals, which can be used to identify the IC and FC events as follows:

	- a. If valid, continue.
	- b. If not valid, return to step 1.

*Community-Based and Everyday Life Gait Analysis: Approach to an Automatic Balance… DOI: http://dx.doi.org/10.5772/intechopen.112873*

**Figure 6.** *Gait cycle events identified in a fall-risk subject using a three-axis accelerometer in the lower back [38].*


The result of applying the above algorithm is illustrated in **Figure 5**. Once the gait cycle events are identified, calculations of the spatiotemporal indices characterizing it can be performed (**Figure 6**).

#### **3.2 Gait parameter estimation**

To calculate the gait cycle index, it is necessary to identify at least 1 gait cycle for each limb (right and left), which corresponds to the identification of 4 Initial Contact (IC) events and 4 Final Contact (FC) events. If the first detected event corresponds to a right IC, use 1. Otherwise, use 2.

$$\text{circle}\_{i} = \left\{ \mathbf{IC}\_{\mathcal{R}\_{i}}, \mathbf{FC}\_{L\_{i}}, \mathbf{IC}\_{L\_{i}}, \mathbf{FC}\_{\mathcal{R}\_{i}}, \mathbf{IC}\_{\mathcal{R}\_{i+1}}, \mathbf{FC}\_{L\_{i+1}}, \mathbf{IC}\_{L\_{i+1}}, \mathbf{FC}\_{\mathcal{R}\_{i+1}} \right\} \tag{1}$$

$$\text{cide}\_{i} = \left\{ \text{IC}\_{L\_{i}}, \text{FC}\_{R\_{i}}, \text{IC}\_{R\_{i}}, \text{FC}\_{L\_{i}}, \text{IC}\_{L\_{i+1}}, \text{FC}\_{R\_{i+1}}, \text{IC}\_{R\_{i+1}}, \text{FC}\_{L\_{i+1}} \right\} \tag{2}$$

From Eqs. (1) and (2), *R* corresponds to a right event, *L* corresponds to a left event and the subscript *i* is an iterator that goes from 1 to the maximum number of identified cycles. Consider all these events in seconds.

Once at least 1 gait cycle has been obtained, the following temporal indices can be calculated:

• Cadence: Number of steps in 1 minute (see Eq. (3)).

$$\text{Cadence}\_{R} = \frac{60}{\left(IC\_{R\_{i+1}} - IC\_{L\_i}\right)}; \text{Cadence}\_{L} = \frac{60}{\left(IC\_{L\_{i+1}} - IC\_{R\_i}\right)}\tag{3}$$

• Stride Time: Time it takes to perform a stride, defined as the time between two ipsilateral IC events (see Eq. (4)).

$$\text{StrideTime}\_{R} = \left( \text{IC}\_{R\_{i+1}} - \text{IC}\_{R\_i} \right); \text{StrideTime}\_{L} = \left( \text{IC}\_{L\_{i+1}} - \text{IC}\_{L\_i} \right) \tag{4}$$

• Step Time: Time it takes to perform a step, defined as the time between two contralateral IC events (see Eq. (5)).

$$\text{StepTime}\_{R} = \left( \text{IC}\_{R\_{i+1}} - \text{IC}\_{L\_i} \right); \text{StepTime}\_{L} = \left( \text{IC}\_{L\_{i+1}} - \text{IC}\_{R\_i} \right) \tag{5}$$

• Single Support: Total time in which only one foot is in contact with the ground, defined as the time between an ipsilateral FC and IC events (see Eq. (6)).

$$\text{SingleSupport}\_{R} = (\text{IC}\_{L\_i} - \text{FC}\_{L\_i}); \text{SingleSupport}\_{L} = \left(\text{IC}\_{R\_{i+1}} - \text{FC}\_{R\_i}\right) \tag{6}$$

• Double Support: Time in which both feet are in contact with the ground, defined as the time between contralateral IC and FC events (see Eq. (7)).

$$DoubleSupport\_R = (FC\_{L\_i} - IC\_{L\_i}); DoubleSupport\_L = \left(FC\_{R\_{i+1}} - IC\_{L\_i}\right) \tag{7}$$

In addition, parameters can be extracted that immediately identify the general state of the gait:

• Swing Phase: General phase of gait where one foot remains in the air. It is calculated as a percentage of the total of 1 gait cycle (see Eq. (8)).

$$\text{96Swing}\_R = \frac{\mathbf{100}(FC\_{R\_i} - IC\_{R\_i})}{IC\_{R\_{i+1}} - IC\_{R\_i}}; \text{96Swing}\_L = \frac{\mathbf{100}(FC\_{L\_{i+1}} - IC\_{L\_i})}{IC\_{L\_{i+1}} - IC\_{L\_i}} \tag{8}$$

• Support Phase: General phase of gait where one foot remains in contact with the ground. It is calculated as a percentage of the total of 1 gait cycle (see Eq. (9)).

$$\% \text{Support}\_R = \frac{100 \left(IC\_{R\_{i+1}} - FC\_{R\_i}\right)}{IC\_{R\_{i+1}} - IC\_{R\_i}}; \% \text{Support}\_L = \frac{100 \left(IC\_{L\_{i+1}} - FC\_{L\_i}\right)}{IC\_{L\_{i+1}} - IC\_{L\_i}} \tag{9}$$

• Symmetry Factor: Knowing the duration or percentage of the duration of each phase of the gait (swing and support) for each limb (left and right), a factor can be extracted to observe the symmetry of the gait. Its value lies between 0 and 1. If the factor is close to 0, it indicates asymmetry, where 0 is perfect asymmetry. If it is close to 1, it indicates symmetry, where 1 is perfect symmetry. The calculation is illustrated in Eq. (10).

$$F\_{\text{symmetry}} = \frac{\min\left\{\text{96}\text{Support}\_R, \text{96}\text{Support}\_L\right\}}{\max\left\{\text{96}\text{Support}\_R, \text{96}\text{ Support}\_L\right\}} = \frac{\min\left\{\text{96}\text{Swing}\_R, \text{96}\text{Swing}\_L\right\}}{\max\left\{\text{96}\text{Swing}\_R, \text{96}\text{Swing}\_L\right\}}\tag{10}$$

Two subjects did not have the data quality necessary to correctly calculate the gait cycle indices, so they were not included in this analysis.

The gait cycle indices that could be calculated from this database using the proposed method were:

*Community-Based and Everyday Life Gait Analysis: Approach to an Automatic Balance… DOI: http://dx.doi.org/10.5772/intechopen.112873*


## **3.3 Pre-processing and characterization of spatiotemporal parameters**

The database exhibits a balanced distribution, as illustrated in **Figure 7**. Out of the 69 subjects included in the study, 31 are classified as "fallers," while the remaining 38 are categorized as "non-fallers."

In examining the variance of each feature, it becomes evident that the feature "Double Support" demonstrates higher variance compared to its closest characteristic, "Simple Support," as illustrated in **Figure 8**.

As a result, the application of a data scaling method became necessary. By employing the scikit-learn framework, various normalization methods available within its library were assessed, including:


**Figure 7.**

*The pie chart depicts the balanced distribution of subjects in the database categorized by their risk of falling.*

**Figure 8.** *Representation of the variance of each characteristic using the unscaled data.*

#### **Figure 9.**

*The sub-figures in the top row show the variance for each feature applying different normalization techniques. The sub-figures in the bottom row show the information gained for each feature according to the type of normalization.*


In **Figure 9**, the four methods employed are depicted. The min-max normalization method is observed to generate the least dispersion among the features. Additionally, when considering the estimated mutual information, it is generally apparent that the various types of data scaling do not influence the information contribution from the features to the binary classifier used, with the exception of the Normalizer method.

*Community-Based and Everyday Life Gait Analysis: Approach to an Automatic Balance… DOI: http://dx.doi.org/10.5772/intechopen.112873*

**Figure 10.**

*Histogram for each feature, classified according to class type. The solid lines in the figure represent the kernel density (KDE) for each feature according to the color code of the classifier.*

The findings suggest that cadence contributes the most information to the classification problem, followed by stride time, with step time coming in third place.

To determine if there are potential outliers and enhance the interpretability of the data, the distribution of each characteristic was analyzed, as shown in **Figure 10**. From the figure, it can be observed that there are no outliers. To check for normality in the data distribution, the Shapiro-Wilks test was employed with a significance level of 0.05. In seeking to determine the null hypothesis that the data come from a normal distribution, the resulting p-values were evaluated. It was found that characteristics such as cadence, single support, stance duration, and swing duration have a normal distribution, while the rest of the characteristics do not.

In order to investigate the linear dependency between the features, the Pearson correlation coefficient was estimated, and the results were visualized in the correlation matrix, as depicted in **Figure 11**. The findings indicate that cadence exhibits a strong negative correlation with step time, stride time, and single support in comparison with the other features. Conversely, weaker correlations are observed between cadence and double support, symmetry and swing duration, single support and double support, and symmetry with both step time and stride time.

#### **3.4 Training of classification algorithms**

To determine the optimal classification model, an exhaustive search was conducted over specific parameter values for an estimator.

For model selection, the GridSearchCV class available in the scikit-learn library was utilized. This class implements an exhaustive search on a grid of hyperparameter values for a given estimator, aiming to identify the best hyperparameter values to maximize the model's performance through cross-validation.

To operationalize this class, a parameter grid was defined in the form of a dictionary, with the hyperparameter names as keys and lists of values to be tested as the

## **Figure 11.**

*Matrix of correlations as a heat map. The values represent the Pearson correlation coefficient between the features. A value close to 1 indicates a strong positive connection, a value close to 1 indicates a strong negative connection and a value close to 0 indicates a weak or no connection.*

corresponding values. This systematic exploration allowed the evaluation of numerous combinations of hyperparameters, subsequently identifying the optimal configuration for the model.

Regarding the strategy for evaluating the model's performance on the test set, the precision metric was chosen, given the balanced nature of the dataset. This choice is suitable for classification problems when there is an equitable distribution of samples across each class.

For cross-validation, the dataset is configured to be divided into five equal parts. This approach facilitates training and splitting the model into different training and test sets.

The classification models used in this research include:


*Community-Based and Everyday Life Gait Analysis: Approach to an Automatic Balance… DOI: http://dx.doi.org/10.5772/intechopen.112873*


#### **Table 1.**

*Classification results for different models. The table shows the ranking scores for each model with different combinations of features.*

discovery of a hyperplane that effectively separates the classes—even when linear separation in the original feature space is unattainable. The RBF kernel calculates the similarity between two data points in this higher-dimensional space based on their distance.

• Random forest classifier (RFC): Composed of a large number of decision trees, this classifier allows the trees to complement each other independently. When a sample is introduced to the classifier, it is categorized according to the collective voting results of the individual trees.

The results of the classification algorithms using the proposed methodology are presented in **Table 1**.

Next, we will present the details of the training process and feature selection for each classifier.

#### *3.4.1 K-nearest neighbors*

For this model, the parameter grid used includes a test list to evaluate the number of neighbors, specifically considering 3, 5, and 7 neighbors. The weights of the function used in the prediction take into account uniform weights and weights that are inversely proportional to their distance. Lastly, a power parameter (*p*) for the Minkowski metric was considered, which corresponds to using both the Manhattan and Euclidean distances.

The performance of this model relative to others can be observed in **Table 1**. The best results in terms of model performance were achieved when considering the features of swing duration and symmetry, with Accuracy = 62.07%, F1-Score = 59.26%, and ROC-AUC = 64.14%. These results were obtained with the following parameters: a number of neighbors equal to 5, *p* equal to 1, and weights determined by the inverse of their distance. Conversely, the worst performance was recorded when considering the features of cadence and step time, with Accuracy = 43.18%, F1-Score = 33.33%, and ROC-AUC = 56.57%.

#### *3.4.2 Logistic regression*

For this model, no specific values were set in the parameter grid, thereby defaulting to the scikit-learn library's automatic settings.

The optimal combination of features for this model was achieved by considering cadence and step time, resulting in Accuracy = 72.41%, F1-Score = 69.23%, and ROC-AUC = 74.24%. Conversely, the poorest performance was observed when considering swing duration and symmetry, yielding Accuracy = 37.93%, F1-Score = 50.00%, and ROC-AUC = 46.46%.

#### *3.4.3 Support vector classification with RBF*

The initial attempts to develop the model involved configuring a parameter grid, where an exhaustive search was conducted using four kernels (linear, sigmoid, rbf, and poly), and the following values for the regularization parameter (C): 2, 4, 3, 5, 6, and 8. Additionally, variations in the kernel coefficient (gamma) were evaluated at 0.1, 0.3, and 0.5, and for the polynomial kernel, the following degrees were examined: 2, 3, 4, 5, 6, and 7. This method facilitated the narrowing of the search for the best classifier, with observations indicating a higher score for the model with the RBF kernel. Consequently, a new exhaustive search was undertaken using a refined parameter grid, employing only the RBF kernel, with C parameters of 1, 2, 4, 3, 5, 6, and 8, and finally, a scaled gamma, the latter being the default parameter used by the class in the scikit-learn library.

The optimal combination of features for this model was found by considering stride time and unique support, resulting in Accuracy = 72.41%, F1-Score = 76.75%, and ROC-AUC = 72.47%. The optimal C parameter was identified as 6, and **Figure 12**

#### **Figure 12.**

*The decision surface of the SVC model with RBF kernel, with a cost function of 6 and a scaled kernel coefficient. The decision boundary, represented by the solid black line, separates the two classes, with the dashed lines representing the margin boundaries. The support vectors, marked with green dotted line circles, are the crucial data points that define the decision limit. The color map differentiates the two classes, where blue corresponds to the subjects that do not fall and red correspond to those that fall.*

*Community-Based and Everyday Life Gait Analysis: Approach to an Automatic Balance… DOI: http://dx.doi.org/10.5772/intechopen.112873*

illustrates the decision surface of the algorithm. The least effective performance was observed when considering step time, stride time, single support, double support, stance duration, and swing duration, yielding Accuracy = 34.48%, F1-Score = 34.48%, and ROC-AUC = 36.62%, with a C parameter of 8.

#### *3.4.4 Random forest classifier*

In training the random forest classifier (RFC), the parameter grid was configured to test various numbers of estimators: 5, 10, 50, 100, and 200. The maximum depth was set at either 5 or 10, and the minimum samples required to split an internal node were set at 2, 3, and 4.

The optimal feature combination for this model was identified by considering single support, double support, stance duration, and swing duration, resulting in an Accuracy of 72.41%, an F1-Score of 75.00%, and an ROC-AUC of 74.24%. The corresponding parameters were found to be a maximum depth of 5, a minimum of 2 samples required to split, and 50 estimators. **Figure 13** illustrates the decision surface for the first tree of the model. Conversely, the worst performance was observed when considering cadence and step time, yielding an Accuracy of 41.38%, an F1-Score of 26.09%, and an ROC-AUC of 38.64%. In this case, the parameters were a maximum depth of 5, a minimum of 3 samples required to split, and 10 estimators.

## **4. Discussion**

Aging brings about a decline in ambulation skills, increasing the incidence of falls in adults over 65 years old. The community represents an environment where the

#### **Figure 13.**

*The decision surfaces of the first decision tree obtained from the RFC, which has been trained with a maximum depth of 5 and 50 estimators. From this tree, the pairs of characteristics used in the decision nodes were obtained, allowing visualization of the decision edges through six subgraphs. Each subplot presents colored training points, where the colors correspond to the non-fallers and fallers classes of the training set. The shaded areas represent the decision regions of the decision tree for each feature pair. The legend provides a key to the classes and colors used in the chart.*

majority of these falls occur, making community ambulation critical for the independent and social life of older adults [1]. As such, technology has come to play an essential role in the evaluation of gait cycle variables, such as step time, cadence, and gait symmetry [15, 16].

Portable body sensors, mobile apps, instrumented insoles, and smartwatches have all been used for this purpose [18–35]. Despite the variety of tools, there remain challenges in their application in non-clinical settings and in predicting future instability [18–21].

Body sensors, mobile applications, and smartwatches offer an attractive and nonintrusive solution [22–25, 29–32, 34, 35]. These devices can monitor and record realtime motion data, enabling continuous gait analysis. The availability of commercial systems illustrates the feasibility of these tools in everyday settings. However, the current demand underscores the need for versatile, easy-to-use devices that can be integrated into daily use. These devices should incorporate predictive models to detect potential anomalies and increased fall risk, potentially improving the safety and quality of life for high-risk groups [26, 27].

Based on the results shown in **Table 1**, the various classification models exhibited a wide variation in performance. Random forest (RFC), support vector machine with a radial basis function kernel (SVC-RBF), and logistic regression (LR) models showed outstanding performance in terms of accuracy, F1-Score, ROC-AUC, precision, and recall.

The RFC model with the best features achieved an accuracy of 72.41%, an F1-Score of 75%, and an ROC-AUC of 74.24%. Similarly, the SVC-RBF with the best features demonstrated comparable performance with an accuracy of 72.41%, an F1-Score of 76.75%, and an ROC-AUC of 72.47%. The LR with the best features also showed solid performance with an accuracy of 72.41%, an F1-Score of 69.23%, and an ROC-AUC of 74.24%. In contrast, the k-nearest neighbors (KNN) model had considerably lower performance with an accuracy of 62.07% and an F1-Score of 59.26%.

The diversity in the performance of these models suggests that feature selection and model choice are critical factors to consider when developing fall risk assessment systems. However, it should be noted that the application of these models in a community or everyday life setting can present unique challenges.

For example, although RFC, SVC-RBF, and LR models have shown promise in terms of accuracy and other performance metrics, their practicability in a community setting may be limited by their computational complexity, the need to gather a substantial amount of data for training, and sensitivity to environmental changes and individual variability.

Compared to other technologies and approaches for balance and gait assessment, our approach stands out for its emphasis on data collection in community or everyday life settings, which can provide a more realistic view of fall risk. However, other approaches, such as wearable sensors, mobile phone applications, and instrumented insoles, have also proven valuable for gait assessment in similar settings. These approaches could complement our methodology, potentially providing a more comprehensive and accurate assessment of fall risk.

The evidence presented here suggests several future directions for improving the accuracy and utility of automatic fall risk assessment through gait analysis in community or everyday life settings. There is an urgent need to develop technologies tailoredto non-clinical environments, considering environmental challenges, cognitive distractions, or medication-induced changes. This could include optimizing existing devices, improving human activity recognition algorithms or developing new tools.

*Community-Based and Everyday Life Gait Analysis: Approach to an Automatic Balance… DOI: http://dx.doi.org/10.5772/intechopen.112873*

## **5. Conclusions and future directions**

This chapter underscores the significant role of technology in assessing gait parameters and predicting fall risks among the elderly in community or everyday settings. The use of portable body sensors, mobile applications, and smartwatches has proven invaluable in monitoring real-time motion data, enabling continuous gait analysis. The accuracy and utility of these technologies have been demonstrated through random forest, support vector machine with a radial basis function kernel, and logistic regression models, which showed commendable performance in their predictive capacities. However, the need for accessible and user-friendly devices is evident, particularly for high-risk groups such as the elderly.

Looking forward, there is immense potential to enhance the utility and accuracy of automatic fall risk assessment through gait analysis in non-clinical settings. The development of technologies tailored to these environments, taking into account factors such as environmental challenges, cognitive distractions, and medicationinduced changes, is imperative. As we move forward, a combination of advanced technological tools and improved predictive models can lead to more efficient and effective fall risk assessment, thereby enhancing the quality of life for older adults in community and everyday settings.

## **Author details**

Britam Arom Gómez Arias1 \*, Sebastián Gonzalo Chávez Orellana†, Paulina Cecilia Ortega-Bastidas2,3† and Pablo Esteban Aqueveque Navarro†

1 Faculty of Engineering, Biomedical Engineering, Universidad de Santiago de Chile, Santiago, Chile

2 Kinesiology Department, Faculty of Medicine, Universidad de Concepción, Concepción, Chile

3 Health Sciences PhD Programme, International Doctoral School, Universidad Rey Juan Carlos, Madrid, Spain

4 Faculty of Engineering, Electrical Engineering Department, Universidad de Concepción, Concepción, Chile

\*Address all correspondence to: britam.gomez@usach.cl

† These authors contributed equally.

© 2023 The Author(s). Licensee IntechOpen. This chapter is distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/3.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

## **References**

[1] Sloot LH, Malheiros S, Truijen S, Saeys W, Mombaur K, Hallemans A, et al. Decline in gait propulsion in older adults over age decades. Gait & Posture. 2021;**90**:475-482

[2] Beck Jepsen D, Robinson K, Ogliari G, Montero-Odasso M, Kamkar N, Ryg J, et al. Predicting falls in older adults: An umbrella review of instruments assessing gait, balance, and functional mobility. BMC Geriatrics. 2022;**22**(1):1-27

[3] Lusardi MM, Fritz S, Middleton A, Allison L, Wingood M, Phillips E, et al. Determining risk of falls in community dwelling older adults: A systematic review and meta-analysis using posttest probability. Journal of Geriatric Physical Therapy (2001). 2017;**40**(1):1

[4] Huijben B, Van Schooten K, Van Dieën J, Pijnappels M. The effect of walking speed on quality of gait in older adults. Alphen aan den Rijn, Netherlands, Philadelphia, United States: Gait & Posture. 2018;**65**:112-116

[5] Marini K, Mahlknecht P, Schorr O, Baumgartner M, De Marzi R, Raccagni C, et al. Associations of gait disorders and recurrent falls in older people: A prospective population-based study. Gerontology. 2022;**68**(10): 1139-1144

[6] Nascimento MDM, Gouveia ÉR, Gouveia BR, Marques A, Martins F, Przednowek K, et al. Associations of gait speed, cadence, gait stability ratio, and body balance with falls in older adults. International Journal of Environmental Research and Public Health. 2022;**19**(21): 13926

[7] Ronthal M. Gait disorders and falls in the elderly. Medical Clinics. 2019;**103**(2): 203-213

[8] Brown CJ, Bradberry C, Howze SG, Hickman L, Ray H, Peel C. Defining community ambulation from the perspective of the older adult. Journal of Geriatric Physical Therapy. 2010;**33**(2): 56-63

[9] Shumway-Cook A, Woollacott MH. Control Motor de la Investigación a la práctica Clnica. Alphen aan den Rijn, Netherlands (global) and Philadelphia, United States (corporate): Wolters Kluwer; 2019

[10] Middleton A, Fritz SL, Lusardi M. Walking speed: The functional vital sign. Journal of Aging and Physical Activity. 2015;**23**(2):314-322

[11] Hollman JH, McDade EM, Petersen RC. Normative spatiotemporal gait parameters in older adults. Gait & Posture. 2011;**34**(1):111-118

[12] Montero-Odasso M, van der Velde N, Martin FC, Petrovic M, Tan MP, Ryg J, et al. World guidelines for falls prevention and management for older adults: A global initiative. Age and Ageing. 2022;**51**(9):afac205

[13] Park S-H. Tools for assessing fall risk in the elderly: A systematic review and meta-analysis. Aging Clinical and Experimental Research. 2018;**30**(1):1-16

[14] Young F, Mason R, Morris RE, Stuart S, Godfrey A. Iot-enabled gait assessment: The next step for habitual monitoring. Sensors. 2023;**23**(8):4100

[15] Ridao-Fernández C, Pinero-Pinto E, Chamorro-Moriana G, et al. Observational gait assessment scales in patients with walking disorders: Systematic review. BioMed Research International. 2019;**2019**:15. Article ID: 2085039

*Community-Based and Everyday Life Gait Analysis: Approach to an Automatic Balance… DOI: http://dx.doi.org/10.5772/intechopen.112873*

[16] Ortega-Bastidas P, Gómez B, Aqueveque P, Luarte-Martínez S, Canode-la Cuerda R. Instrumented timed up and go test (itug)—More than assessing time to predict falls: A systematic review. Sensors. 2023;**23**(7):3426

[17] Pereira CB, Kanashiro AMK. Falls in older adults: A practical approach. Arquivos de Neuro-Psiquiatria. 2022;**80**: 313-323

[18] Werner C, Hezel N, Dongus F, Spielmann J, Mayer J, Becker C, et al. Validity and reliability of the apple health app on iphone for measuring gait parameters in children, adults, and seniors. Scientific Reports. 2023;**13**(1): 5350

[19] Betteridge C, Mobbs RJ, Ho D. Proposed objective scoring algorithm for walking performance, based on relevant gait metrics: The simplified mobility score (smos™)—Observational study. Journal of Orthopaedic Surgery and Research. 2021;**16**(1):1-7

[20] Clay L, Webb M, Hargest C, Adhia DB. Gait quality and velocity influences activity tracker accuracy in individuals post-stroke. Topics in Stroke Rehabilitation. 2019;**26**(6):412-417

[21] Paul S, Tiedemann A, Hassett L, Ramsay E, Kirkham C, Chagpar S, et al. Validity of the fitbit activity tracker for measuring steps in community-dwelling older adults. BMJ Open Sport & Exercise Medicine. 2015;**1**(1):e000013. DOI: 10.1136/bmjsem-2015-000013

[22] Bioingeneering B. G-Walk: Wearable Inertial Sensor for Motion Analysis. [Online]. 2023. Available: https://www.b tsbioengineering.com/products/g-walk/

[23] SPA N. Qmuv - el futuro de la rehabilitación. [Online]. 2020. Available: https://qmuv.cl/

[24] Technologies AW. All Day Activity Tracker and More - Apdm Wearable Technologies. [Online]. 2023. Available: https://apdm.com/continuous-record ing/

[25] Cramer LA, Wimmer MA, Malloy P, O'Keefe JA, Knowlton CB, Ferrigno C. Validity and reliability of the insole3 instrumented shoe insole for ground reaction force measurement during walking and running. Sensors. 2022; **22**(6):2203

[26] Shema-Shiratzky S, Beer Y, Mor A, Elbaz A. Smartphone-based inertial sensors technology–validation of a new application to measure spatiotemporal gait metrics. Gait & Posture. 2022;**93**: 102-106

[27] Hammoud A, Duchêne J, Abou-Ghaida H, Mottet S, Goujon J-M, Hewson DJ. Validation of a smartphone gait analysis system. In: 6th European Conference of the International Federation for Medical and Biological Engineering: MBEC 2014, 7–11 September 2014. Dubrovnik, Croatia: Springer; 2015. pp. 910-913

[28] Su D, Liu Z, Jiang X, Zhang F, Yu W, Ma H, et al. Simple smartphone-based assessment of gait characteristics in parkinson disease: Validation study. JMIR mHealth and uHealth. 2021;**9**(2): e25451

[29] Ngueleu AM, Blanchette AK, Maltais D, Moffet H, McFadyen BJ, Bouyer L, et al. Validity of instrumented insoles for step counting, posture and activity recognition: A systematic review. Sensors. 2019;**19**(11):2438

[30] el Achkar CM, Lenoble-Hoskovec C, Paraschiv-Ionescu A, Major K, Büla C, Aminian K. Instrumented shoes for activity classification in the elderly. Gait & Posture. 2016;**44**:12-17

[31] Solanki D, Lahiri U. Design of instrumented shoes for gait characterization: A usability study with healthy and post-stroke hemiplegic individuals. Frontiers in Neuroscience. [Internet]. 2018;**12**:459. [Cited 2020 May 2]

[32] Aqueveque P, Germany E, Osorio R, Pastene F. Gait segmentation method using a plantar pressure measurement system with custom-made capacitive sensors. Sensors. 2020;**20**(3):656

[33] Moticon. Moticon Sensor Insoles - Enhancing Sports Health. [Online]. 2023. Available: https://moticon.com/

[34] Huang Y, Xu J, Yu B, Shull PB. Validity of fitbit, jawbone up, nike+ and other wearable devices for level and stair walking. Gait & Posture. 2016;**48**:36-41

[35] Farmer C, van den Berg ME, Vuu S, Barr CJ. A study of the accuracy of the fitbit zip in measuring steps both indoors and outdoors in a mixed rehabilitation population. Clinical Rehabilitation. 2022; **36**(1):125-132

[36] Weiss A, Brozgol M, Dorfman M, Herman T, Shema S, Giladi N, et al. Does the evaluation of gait quality during daily life provide insight into fall risk? A46novel approach using 3-day accelerometer recordings. Neurorehabilitation and Neural Repair. 2013;**27**(8):742-752

[37] González RC, López AM, Rodriguez-Uría J, Alvarez D, Alvarez JC. Real-time gait event detection for normal subjects from lower trunk accelerations. Gait & Posture. 2010;**31**(3):322-325

[38] Aqueveque P, Gomez B, Ortega-Bastidas P, Pena G, Retamal G, Cano-de-la-Cuerda R. Predicting risk of falls in elderly using a single Inertial Measurement Unit on the lower-back by estimating spatio-temporal gait parameters. In: 2022 44th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC). IEEE. July 2022. pp. 2390-2394

## *Edited by Manuel Domínguez-Morales and Francisco Luna-Perejón*

This book introduces the reader to the deeper details of the human gait process. To this end, it takes a multidisciplinary approach, covering all the fields involved in the gait pattern: biomechanics, neuroscience, and even technology (from the point of view of detecting anomalies and improving them). From the basics of walking to the latest scientific advances, the aim is to bring together the knowledge of experts in the field to provide a multidisciplinary view of this area of study. The target audience for this book is not only the scientific community but also clinicians and readers interested in unlocking the secrets of the walking process. Through this book, readers will discover how in-depth knowledge of human gait can change their perception of rehabilitation, sports performance, or even their own everyday mobility. On behalf of the editors of this book, we recommend that you take a look at all that *Human Gait - Recent Findings and Researc*h has to offer. The works contained in this book are the result of careful selection and evaluation, making this book an essential resource for those who want to learn about the human gait pattern.

*Robert Koprowski, Biomedical Engineering Series Editor*

Published in London, UK © 2024 IntechOpen © blackdovfx / iStock

Human Gait - Recent Findings and Research

IntechOpen Series

Biomedical Engineering, Volume 24

Human Gait

Recent Findings and Research

*Edited by Manuel Domínguez-Morales* 

*and Francisco Luna-Perejón*