Classifying activities of daily living based on hybrid features from vertical ground reaction force data

Abstract

Leg posture plays a crucial role in performing activities of daily living (ADL) by influencing balance, mobility, and overall quality of life. The objective of this study was to classify various instrumental activities of daily living (IADL) leg postures using 1D force data using machine learning techniques. The study collected data from a SENSIX force platform, which measured the vertical ground reaction force (VGRF) and center of pressure (COP) data from 25 participants performing four static postures. These postures included double-leg standing, toe standing, single-leg standing, and squat standing. The study used five statistical and six nonlinear time series features from the VGRF data and COP to classify IADL postures. The artificial neural network performed the best, with an accuracy of approximately 98%, in organizing postures for everyday life activities.

Introduction

Innovative assistive technologies, like those found in smart homes, rehabilitation¹, health support^2,3, skill assessment⁴, or industrial settings⁵, are based on the recognition of human activities. Numerous real-world applications exist for human activity recognition (HAR), including indoor navigation, gaming, healthcare, fitness, and military tactical applications. Fitness trackers and fall detection devices are two examples of available basic activity-aware systems. Despite the economy’s march toward automation, many individuals still have to execute physically demanding and often dangerous duties as part of their everyday jobs. Prominent instances include the heavy metal and service industries, including emergency services, logistics, and nursing care⁶.

Maintaining proper posture in daily life is crucial since it lowers the risk of chronic diseases and promotes healthy blood circulation⁷. Healthcare monitoring systems should monitor an individual’s posture and daily activity level, for instance, in rehabilitation settings^8,9. By enabling both the patient and the healthcare provider to evaluate the state and the results of any actions, this type of monitoring helps to prevent injuries like those caused by falls. It enhances the patient’s physical state¹⁰. Healthcare professionals require postural categorization to better understand a patient’s body posture and movement patterns. It provides a uniform way to describe postural disorders, aiding in making diagnoses, planning treatments, and monitoring patient improvement. Postural classification is also helpful for researchers looking into the biomechanics and health effects of various postures. Figure 1 shows two main categories of HAR systems: sensor-based HAR systems that rely on wearable sensors affixed to the human body, wherein the subject’s activity is converted into patterns for recognition.

Fig. 1.

Approaches employed in human activity recognition.

Proper posture is crucial to maintain balance and prevent undue pressure on the muscles, ligaments, and joints. Unusual lower limb movement patterns have been linked to noncontact anterior cruciate ligament injuries¹¹, as well as other musculoskeletal issues such as acetabular labral tears¹² and patellofemoral discomfort. The early diagnosis of disorders, including scoliosis, kyphosis, and lordosis, can be assisted by identifying postural patterns. Evaluating lower limb movement patterns may help identify individuals at risk for future injury or musculoskeletal pain and direct treatment of current musculoskeletal pain issues. Several occupations necessitate extended periods spent in uncomfortable positions¹³, which raises the possibility of musculoskeletal problems, sick days, and early retirement¹⁴. HAR’s foundation is that particular bodily motions correspond to distinctive patterns in sensor signals, which may be detected and categorized by machine learning methods.

A study¹⁵ used four inexpensive strain sensors positioned on the human body as the basis. With a single voltage value serving as the sensor output, they implemented three classifiers using long-short-term memory (LSTM) networks. Initially, the data was categorized into two dynamic activities (walking and running) and three static postures (laying, sitting, and standing). Subsequently, they identified motionless stances accompanied by haphazard arm gestures, and ultimately, they categorized the information into two distinct slouched postures and sitting. Harms et al.¹⁶ used wearable accelerometers to classify ten postures that would be helpful for rehabilitation. While the current study employs a machine-learning (ML) approach to construct the classes instead of hard-coding the threshold values, ÓLaighin¹⁷ used threshold values for the joint angles to classify distinct postures.

Al-Faris et al.¹⁸ proposed a literature study on computer vision-based methods for HAR. A novel algorithm has been proposed by Ding et al.¹⁹ in which the rule learning method combined with bagging and random space, including multiple features like angle and distance features were proposed. The algorithm was evaluated on four human pose datasets, proving to be more efficient than the traditional CNN method. In another study, the images of the poses of human activity were acquired through the processed video formats at regular intervals. These were trained with a CNN deep learning model that provided 100% accuracy in classification²⁰. CNN variants with hyperparameter optimization techniques have been used with image augmentation methods for classifying human poses. Among these, AlexNet provided an accuracy of about 91.2%²¹. However, gait recognition with an accuracy of 97.67% was achieved by using multiple discriminant analysis (MDA)²². A smart watch using tri-axis accelerometer and stacked denoising autoencoder (SDAE) model was made to find the human recognition pattern²³. Moreover, Basketball players movement to identify fouls and goals were achieved by utilizing machine vision techniques²⁴. Yu had proposed a human gesture recognition and pattern which had the combined technique of You Only Look Once Version version 5 (YOULOVv5) and hybrid attention mechanism²⁵. With these rapid developments in computerized computational techniques, we can extract, analyze, and understand the postures of the human body. Many studies have recently been conducted to recognize human body postures. Additionally, artificial intelligence has been augmented with the conventional techniques for estimating human postures. Wearing several sensors can be tiresome or challenging for the user. This can be made worse if a physical obstacle makes connecting the sensor to the body complex, such as a motor impairment or a design requiring skill. In addition, misplacing or losing sensors during the study can complicate the sensor data processing. So, to overcome these drawbacks, an attempt was made to extract one-dimensional (1D) data from the force platform without attaching the sensors to the body. The proposed work aims to use 1D force data to recognize and categorize various human positions.

• To collect/record real-time force and centre of pressure (COP) data using a force platform for four different standing postures.

• To extract the features from data acquired, such as vertical ground reaction force (VGRF) and COP.

• To classify various postures using machine learning techniques based on the statistical, non-linear time domain, and COP features that were retrieved.

• To compare the state-of-the-art results and evaluate the performance of machine learning models with hybrid features (statistical, non-linear, and COP).

This study is based on the following three hypotheses:

H1: Would a 1D force-platform-based approach provide reliable posture classification with accuracy comparable to or better than traditional methods using video or IMU-based sensors?

H2: Would hybrid features (statistical and COP) improve model performance for distinguishing different postures?

H3: Would ANN perform better than all other models in handling high-dimensional data such as the combination of vertical GRF and COP features in this study?

Proposal model

The proposed model is described with system architecture, participant and protocol adopted, data acquisition, data processing and feature extraction and classification algorithms.

System architecture

Figure 2 illustrates the proposed methodology: acquisition of 1D force and COP pre-processing, learning algorithm deployment, and comparative analysis to classify four distinct standing postures of the human body utilizing a force platform.

Fig. 2.

Proposed methodology to classify human body postures.

The primary objective of this research was to classify leg postures during standing or quasi-static activities, where vertical GRF and COP data are the most critical for determining postural alignment and balance²⁶. The vertical GRF is closely associated with weight-bearing and postural control, which makes it a more suitable input for classifying different leg postures in ADLs²⁷. The horizontal GRFs are more relevant for dynamic movements, such as walking or running, where they influence acceleration and movement direction²⁸. However, in static or quasi-static poses (as seen in many ADL tasks), the horizontal forces are relatively small and less informative for the classification of standing postures. Moreover, using horizontal GRFs would also require more sophisticated sensor arrangements and computational models to account for the directional components of these forces. Hence, the horizontal GRFs were not used in the current models. This study aimed to offer a practical and real-time solution for classifying complex leg postures in ADLs, limiting the model inputs to vertical GRF and COP data helped ensure simplicity and robustness without compromising accuracy.

The input data is preprocessed using the preprocessing method, and the necessary features from VGRF and COP were extracted. Specifically, COP is computed based on the distribution of the VGRF and can be susceptible to noise, especially in real-time measurements or when there are variations in force distribution due to posture changes. In this study, a force plates system, a gold standard approach was used to record VGRF to have higher accuracy when compared the foot insole sensor or any other device. In addition, the COP errors were minimized through non-linear re-calibration techniques. The correlation analysis was done, after which learning algorithms like logical regression (LR), artificial neural network (ANN), decision tree (DT), support vector machine (SVM), Naïve Bayes (NB), Ensemble (EN), and K-Nearest Neighbours (K-NN), and were used to classify postures based on the input data. Then, a comparative analysis is done to select the best classification algorithm. Finally, the best classifier is incorporated to analyze the input data and perform the prediction process effectively.

Participants and protocol adopted

A total of 25 participants volunteered for the study. There were eight female and seventeen male participants of college students, ages 18 to 22. The average age, height and weight of the participants were 19.96 ± 0.2 years, 166.6 ± 9.7 cm, and 60.4 ± 9.5 kg, respectively. Every participant who took part in this study was in good health and had no history of musculoskeletal conditions that could impair their ability to balance, such as neck or low back pain. Randomly, each subject had two trials of each posture, lasting 15 s. Subjects who could not complete the study were also debriefed, and the causes had been noted. The criteria for exclusion were any injuries or surgeries in the past year, as well as neurological-related disorders. All experiments were performed at the Biomechanics Lab, Bioengineering Division of our institution. The study adhered to the protocol set by the institutional committee of our institution. All experimental procedures were approved by the institutional ethics committee at SASTRA Deemed University, Thanjavur, Tamil Nadu, India, ensuring compliance with ethical guidelines and participant safety. After informing each participant about the inquiry’s purpose and necessity, their informed consent was obtained.

Experimental protocol

The experiment was carried out on human activities of daily living (ADL), such as leg standing postures. The force platform used in this examination was SENSIX, which has an integrated device mounted onto a dedicated platform of length 600 mm, width 400 mm, and a vertical position of 2.4 mm concerning the top of the plate. Before the data collection, all the volunteers were instructed to warm up for 5 s to become accustomed to the various standing positions and prevent cramping. Subjects were given four distinct standing postures (activities of daily living) to try twice each on the force platform. The data was recorded for 15 s for each posture with a 1-minute break between postures and a sampling frequency of 1000 Hz. Four postures were used: toe stand (TS) (Fig. 3(A)), double leg stand (DS) (Fig. 3(B)), squat stand (Fig. 3(C)) and single leg stand (SS) (Fig. 3(D)). The four postures were administered to subjects in a randomized trial fashion. Every posture was required to be held for 15 s. For each of the four ADL postures, two trials were recorded for each participant, with each trial lasting 15 s. For static postures, particularly when considering the vertical component of GRF, relatively low-frequency content, usually below 20 Hz, is more dominant²⁹. Hence, the raw analogue data from the force plates was filtered using a low-pass digital Butterworth filter with a cut-off frequency of 20 Hz that was second-order with zero phase lag to eliminate measurement noise³⁰.

Fig. 3.

Postures chosen for the study: (A) Toe stand (TS), (B) Double leg stand (DS), (C) Squat stand, and (D) Single leg stand (SS).

Standing positions on the force platform are used to collect the associated force and torque data (Fx, Fy, Fz & Tx, Ty, Tz) and COP data (Cx and Cy) in the anterior-posterior (AP) and medial-lateral (ML) directions. The classifier in MATLAB received the linear and non-linear time series features retrieved from VGRF and COP features and performed the analysis. The classification algorithm was used to identify human activities of daily living (ADL) standing positions.

Data processing and feature extraction

Under static conditions, posture control involves a period of stabilization, during which the body adjusts to a steady state. The consistent nature of posture signals usually needs a few seconds to stabilize³¹. The 2-second window is consistent with previous research, such as Beelen et al.³², who used a 2-second window to assess postural stabilization during cryotherapy, and Banos et al.³³, who found that a time window of approximately 1.25–3.25 s is optimal for recognizing postures and activities involving leg and trunk movements.

Regarding signal processing, the raw data were collected for 15 s per pose at a sampling frequency of 1000 Hz. To eliminate any initial bias, the first second of each data set was discarded. The remaining 14 s of data were then divided into non-overlapping 2-second intervals using a rectangular window function, resulting in seven 2-second segments per participant. This approach ensured that each segment contained 2000 data points for force platform signals, providing a stable and representative sample for analysis.

The utilization of the time-windowing procedure had various advantages. For the machine learning model to make an accurate prediction, additional data must be collected. Secondly, shorter signal data were acquired thanks to the time-windowing procedure. Time-windowing techniques can be used for analyzing time series data in machine learning studies, including improved handling of temporal dependencies, feature extraction, computational efficiency, and enhanced interpretability.

Using these pre-processed VGRF in the z-direction (Fz) and the COP in the AP and ML directions, a total of 22 features were extracted from the VGRF and COP signals. These comprised five statistical features—mean, standard deviation, kurtosis, variance and skewness; six nonlinear time series (NLTS) features—fuzzy entropy, approximate entropy, sample entropy, Higuchi fractal dimension, Katz fractal dimension and Lyapunov exponent³⁴; eleven COP features— Anterior-Posterior sway (APsway), medial-lateral sway (MLsway), RMS distance (RDIST), mean distance (MDIST), RMS distance-AP (RDISTAP), RMS distance-ML (RDISTML), total excursion (TOTEX), Total displacement of sway (DOT), mean velocity (MVELO), MVELOAP, and MVELOML³⁵ were chosen to categorize the four postures.

When it comes to how motor behavioural changes over time, nonlinear measures can capture the temporal component of the variation in VGRF. As a result, regularity, environmental adaptation, stability, and complexity can all be measured using these metrics. Non-linear features, such as fractal dimension (FD), entropy families, and Lyapunov exponent (LE), referencing the work of Adalarasu et al.³⁴, were extracted from each 2-second signal window after the pre-processing steps which involved filtering the signals to remove high-frequency noise and ensuring that the data were stationary for accurate analysis. Once the features were extracted, they were used to characterize the dynamics of postural control and were subsequently fed into the machine learning model to enhance classification accuracy.

The fractal dimension (FD), entropy families, and the Lyapunov exponent (LE) are nonlinear tools that assess the aforementioned postural control properties. The generated signal is known as a stabilogram, and it is frequently investigated using either its two-dimensional trajectory or one-dimensional fluctuations in the ML or AP direction. It explains the participant’s postural sway diagram in each stance. The COP trajectory shows that for the participant standing in a pose, the forces applied to the feet are in the upward direction or centre of the feet or all directions, both AP and ML. From this, one can conclude that using 1D force data (a force acting on the feet), one can classify the human body leg poses without video capturing information.

The ML model classification was performed using the MATLAB Classifier Learner App, a built-in tool in MATLAB for training and testing. The following MATLAB functions were used to calculate the various statistical and non-linear features.

Statistical features

• Mean: ME = mean( ).

• Standard Deviation: SD = std( ).

• Kurtosis: KUR = kurtosis( ).

• Skewness: SKE = skewness( ).

• Variance: VA = var( ).

Non-Linear features

• Sample Entropy: saen = SampEn( ).

• Approximate Entropy: approxEnt = approximateEntropy( ).

• Fuzzy Entropy: entr = fuzzyen( ).

• Higuchi Fractal Dimension: [HFD] = Higuchi_FD( ).

• Katz Fractal Dimension: [KFD] = Katz_FD( ).

• Lyapunov Exponent: lyapExp = lyapunovExponent( ).

Classifier algorithms

In this study, the ML models were chosen for posture classification as the models provide a robust and interpretable method for analyzing a relatively small dataset of labelled posture samples. While deep learning models can offer powerful capabilities for large datasets, they face several challenges, including the need for a large amount of training data, interpretability issues, and computational complexity, as outlined by Alzubaidi et al.³⁶. Given the relatively smaller dataset in this study, the ML approach was selected for its ability to effectively classify postures without the need for massive training data. Machine learning-based leg posture classification entails building a model to differentiate between various leg postures based on input features. For posture classification, a diverse set of features were extracted to capture both the statistical and dynamic characteristics of the data. The features used for training and testing the ML models include: 5 statistical features from the VGRF signal (mean, standard deviation, kurtosis, variance and skewness), 6 non-linear time series features (fuzzy entropy, approximate entropy, sample entropy, Higuchi fractal dimension, Katz fractal dimension and Lyapunov exponent), and 11 COP features (Anterior-Posterior sway (APsway), medial-lateral sway (MLsway), RMS distance (RDIST), mean distance (MDIST), RMS distance-AP (RDISTAP), RMS distance-ML (RDISTML), total excursion (TOTEX), Total displacement of sway (DOT), mean velocity (MVELO), MVELOAP, and MVELOML). Furthermore, hybrid features have been used as input to the classifiers and prediction was also done, as shown in Fig. 4. Selecting the best classifier is contingent upon several variables, such as the data type, the dataset size, and the particular demands of the classification. For the model validation, the hold-out validation method was employed, which is a common approach in machine learning for estimating model performance. In this method, 75% of the dataset was used for training, and the remaining 25% was used for testing. Table 1 shows the key settings and parameters used for seven adopted ML models, logical regression (LR), artificial neural network (ANN), decision trees (DT), support vector machine (SVM), Naïve Bayes (NB), Ensemble (EN), and K-Nearest Neighbours (K-NN).

Fig. 4.

Hybrid feature-based classification.

Table 1.

Model settings and parameters.

Model	Logical Regression (LR)	Artificial Neural Networks (ANN)	Decision Tree (DT)	Support Vector Machines (SVM)	Naïve Bayes (NB)	Ensemble (EN)	K-Nearest Neighbors (K-NN)
Preset	Efficient Logistic Regression	Bilayered Neural Network	Fine Tree	Cubic SVM	Kernel Naive Bayes	Bagged Trees	Weighted K-NN
Maximum number of splits	-	-	100	-	-	999	-
Split criterion	-	-	Gini’s diversity index	-	-	-	-
Surrogate decision splits	-	-	Off	-	-	-	-
Kernel function	-	-	-	Cubic	Gaussian	-	-
Kernel scale	-	-	-	Automatic	-	-	-
Box contract level	-	-	-	1	-	-	-
Multi class coding	-	-	-	One vs. one	-	-	-
Number of fully connected layer	-	2	-	-	-	-	-
First layer size	-	10	-	-	-	-	-
Second layer size	-	10	-	-	-	-	-
Activation	-	ReLU	-	-	-	-	-
Iteration limit	-	1000	-	-	-	-	-
Regularization strength (Lambda)	Auto	> 0	-	-	-	-	-
Number of neighbours	-	-	-	-	-	-	10
Distance metric	-	-	-	-	-	-	Euclidean
Distance weight	-	-	-	-	-	-	Squared inverse
Support		-	-	-	Unbounded		-
Learner	Logistic regression	-	-	-	-	-	-
Solver	Auto	-	-	-	-	-	-
Regularization	Auto	-	-	-	-	-	-
Relative coefficient tolerance (Beta tolerance)	0.0001	-	-	-	-	-	-
Ensemble method	-	-	-	-	-	Bag	-
Learner type	-	-	-	-	-	Decision tree	-
Number of learners	-	-	-	-	-	30	-

Results

The results of the ML models used to classify four activities of daily living (ADL) postures based on VGRF data and COP are presented in this section. The dataset, consisting of 1000 observations and 22 features, which were analyzed using multiple classifier models including LR, ANN, DT, SVM, NB, EN and K-NN. Both performance of these models were evaluated before and after applying feature selection methods.

Initially, the models fed with an individual set of features, namely, statistical, non-linear time series and COP features, was analyzed using hold-out validation method. The average classification accuracy was calculated across all four postures, where each posture includes the average accuracy of 2 trials. The ANN performed the best compared to other models with 51.2% accuracy using statistical features, 76.4% accuracy using the non-linear time series features, and 93% accuracy with COP features, as depicted in Fig. 5. Similarly, K-NN and EN are the models competing with ANN model. K-NN provided 94% accuracy with COP features, and EN provided 93.5% accuracy, which is higher when compared to the accuracy provided by the ANN model. However, K-NN and EN accuracy in terms of statistical (36% and 41%) and NLTS features (55% and 70%) are less compared to the one provided by the ANN model. The result demonstrated that COP features improve the classifier accuracy in most models.

Fig. 5.

Performance indicators for distinct features (DT - Decision Tree; SVM- Support Vector Machine (SVM); LR - Logical Regression; KNN - K-Nearest Neighbours; EN – Ensemble; ANN - artificial neural network; NLTS – non-linear time series; COP - centre of pressure).

Figure 6 show the AUC values for the ANN models using statistical and non-linear time series features. The statistical features fair discrimination capability between the postures as shown in Fig. 6(A). For ANN model except DS (AUC-0.82), the other three poses have less than 0.8 AUC value (SS – 0.71; Squat – 0.68; TS-0.77). This shows that statistical features are unable to distinguish between positive and negative classes. In the non-linear time series features of the ANN model, all posture AUC values are greater than 0.8, as shown in Fig. 6(B). The-fore, these results show that the non-linear time series features have good discrimination capability in classifying the four ADL poses. One can infer that the double support achieved a better AUC of 0.925 compared to other postures.

Fig. 6.

AUC for ANN: (A) Statistical features, and (B) Non-linear time series features (DS - double leg stand; SS - single leg stand; TS - toe stand).

Figure 7 shows the AUC value of ANN model using COP features. For ANN model with COP features all posture have AUC value of 0.95, it shows that COP features have excellent discrimination capability for classify ADL poses.

Fig. 7.

AUC for ANN for COP features (DS - double leg stand; SS - single leg stand; TS - toe stand).

To envisage the effect of collective features fed to the classifiers, hybrid features, as discussed in Fig. 4 (a few combinations of features), have been applied to the model, namely, statistical and non-linear time series features, non-linear time series and COP features, and statistical and COP features to improve model accuracy. The performance metrics of the ML models are assessed in terms of accuracy and are displayed in Table 2, demonstrating that ANN performs better with non-linear and COP features (accuracy – 92.0%) when compared to other features combination such as statistical and nonlinear time series, statistical and COP features. Meanwhile, the ensemble model has best competent with all other model with statistical and COP features with accuracy of 95.6%, and hypothesis H2 was proved.

Table 2.

Performance metrics for combined features.

Model	Accuracy % (Hold-out Validation)
	Statistical and Non-linear Time Series Features	Non-linear Time Series and COP Features	Statistical and COP Features
Logical Regression (LR)	30.8	34.8	47.2
Artificial Neural Network (ANN)	80.4	92.0	91.6
Decision Tree (DT)	55.2	85.2	88.0
Support Vector Machine (SVM)	74.8	93.6	88.8
Naïve Bayes (NB)	47.6	62.4	54.0
Ensemble (EN)	71.2	92.0	95.6
K-Nearest Neighbours (K-NN)	54.4	89.6	68.4

Best feature selection

In ML, identifying the relevant features is essential for improving detection performance. It emphases on what important features that carry significant information from the dataset³⁷. The significant ranking of each attributes was determined by applying Analysis of Variance (ANOVA), a statistical approach for analyzing experimental data in which one or more response variables are evaluated under various situations and are identified by one or more classification variables. High F-value and low p-value features are regarded as the most significant. Features can be selected using p-value thresholds and F-values. The performance of the ML models has been recognized with all 22 features, and the 17 best features such as skewness, fuzzy entropy, approximate entropy, sample entropy, Higuchi fractal dimension, Katz fractal dimension, Anterior-Posterior sway (APsway), medial-lateral sway (MLsway), RMS distance (RDIST), mean distance (MDIST), RMS distance-AP (RDISTAP), RMS distance-ML (RDISTML), total excursion (TOTEX), Total displacement of sway (DOT), mean velocity (MVELO), MVELOAP, and MVELOML in terms of accuracy were selected, as shown in Table 3.

Table 3.

Performance metrics using all and selected features.

Model	Accuracy % (Hold-out Validation)
	All Features	Best Selected Features
Logical Regression (LR)	46.4	34.4
Artificial Neural Network (ANN)	89.2	98.0
Decision Tree (DT)	85.6	86.8
Support Vector Machine (SVM)	93.2	95.2
Naïve Bayes (NB)	62.0	62.8
Ensemble (EN)	92.4	94.4
K-Nearest Neighbours (K-NN)	76.0	89.2

Table 3 compares the performance of models using all features to those using only the selected 17 features. Following feature selection, the classifiers demonstrated improved performance and higher accuracy for most machine learning models. For instance, the ANN model achieved an accuracy of 98.0% with the selected features, which was significantly higher than the accuracy obtained with all features. This finding supports the validity of hypothesis H3. Similarly, the K-NN model showed improved performance in terms of accuracy with selected features, from 76.0% to 89.2%.

Figure 8 provides the actual and predicted class of ANN for all features and best features selected for four ADL poses using the confusion matrix. The best features selection method ensures the 100% expected and actual values for the double leg stand (DS). For all features of the ANN model, all four poses predicted error of DS at 4.8%, SS at 7.9%, squat at 22.2%, and TS at 8.1%, as shown in Fig. 8(A). The classification error for single-leg stands at 1.4%, and the squat and toe stand at 3.2%, as shown in Fig. 8(B). The ANN model demonstrated the lowest misclassification rate among all four ADL poses; these matrices illustrate the classifiers’ ability to differentiate between various classes and highlight the degree of separability for best feature selection techniques.

Fig. 8.

Confusion matrix for neural network with hold-out validation: (A) All features, and (B) Best features (DS - double leg stand; SS - single leg stand; TS - toe stand; TPR – True positive rate; FNR – False negative rate).

Figure 9 shows the ROC curve for the ANN classifier using all features and selected features. The ANN provides the highest AUC values of 0.96 for double standing posture for all features, and about 0.99 for the best features using hold-out validation. This evaluates the effectiveness of the ANN’s accuracy in predicting and classifying the different leg postures.

Fig. 9.

AUC for neural network with hold-out validation: (A) All features, and (B) Best features (DS - double leg stand; SS - single leg stand; TS - toe stand).

Table 4 shows the overall classification report of ANN model performance regarding precision, recall, specificity, accuracy, and F1 score using statistical and COP features, non-linear time series, hybrid features, and best-selected features. In this study, the artificial neural network (ANN) model that utilized the best feature selection method demonstrated impressive performance metrics. It achieved a specificity value of 0.99 and a precision value of 0.98, outperforming other feature combinations. This indicates that the ANN effectively identifies positive cases with fewer false negatives and is often accurate when predicting a positive outcome. Additionally, the ANN model attained the highest F1 score of 0.98. This suggests that the ANN model with the optimal features is superior in classifying activities of daily living (ADL) postures.

Table 4.

Activities of daily living classification performance of ANN model using features combination, hybrid and best selected features.

Parameters	Statistical and COP Features	Non-linear Time Series and COP Features	Hybrid Features	Best Selected Features
Precision	0.92	0.92	0.89	0.98
Recall	0.91	0.92	0.89	0.98
Accuracy (%)	91.2	92.0	89.2	98.0
Specificity	0.97	0.97	0.96	0.99
F1Score	0.91	0.92	0.89	0.98
False Positive Rate	0.03	0.03	0.04	0.01

Discussion

The purpose of this study was to classify four distinct leg postures using machine learning models based on features extracted from VGRF and COP data. The findings indicate that COP features were the most significant for improving classification accuracy. The ANN achieved the highest accuracy of 98.0% using the best-selected features and outperformed other classifiers, particularly with hybrid feature sets. The experimental results confirmed that all three hypotheses were supported by the data.

According to the literature assessment, several researchers have used numerous methods for identifying human body postures based on learning algorithms and utilizing photographs, video clips, or motion recordings. The researchers combined many image processing techniques, including picture segmentation, feature extraction, and classification, using 2D and 3D methodologies. To classify various postures like sitting, standing, bending, etc., these algorithms first detect certain postural features and patterns using RGB and depth data analysis. Instead of a camera and wearable sensors in the proposed method, we used a force platform and 1D force data. We tried to gather data for the four positions because only 1D data (foot force) is now available in the literature to evaluate the leg postures. With a force platform to identify the four body postures, we collected the posture dataset in our indoor Biomechanics Laboratory for the current experiment. We used pre-existing intelligent models to classify them. Direct comparison with literature is impossible since 1D postural datasets are not available.

Additionally, the research in the literature involves processing and feature extraction, and the datasets are 2D photos or time-consuming 3D video sequences. Instead, the proposed work adopted a novel methodology that categorizes postures using 1D force moment and COP data. This shows that even though the postures have been classified using video clips, photos and wearable sensors (IMU), 1D force and moment data may be used as a HAR. Using COP and VGRF data that can distinguish the lower body postures, the force data can help the researchers identify the postures. Moreover, the features extracted from the centre of pressure (COP) data seem more effective in classifying the postures than other feature sets. The anticipated applications of this study include the evaluation of different lower body postures, automated assessment of postures in the shop floor industry, and musculoskeletal disorder assessment.

Traditional statistical models often assume linear relationships between the features and the target variable. These models, such as linear regression, can struggle to capture complex, non-linear patterns within the data. The data collected in this study, including VGRF and COP signals, are inherently non-linear due to the bending of the force platform³⁸ and the complex nature of postural control. In addition, the relationships between the multiple features (such as statistical and non-linear parameters) and the target postural conditions are not simple and cannot be fully described using linear techniques. The artificial neural network (ANN) model outperformed the traditional statistical model due to its ability to handle non-linear data, capture complex feature interactions, and work effectively with high-dimensional datasets. This is consistent with findings from previous studies, such as Ngoh et al.³⁹, who demonstrated the success of neural networks in estimating VGRF during running using shoe-mounted IMUs. The non-linear nature of force platform signals and the complexity of postural control are better captured by neural networks, which is why the NN models showed superior correction rates in this study.

The current study did not involve wearable sensors (IMU), complex research of skeletal information, or the relationships between joints to identify human postures. Instead, the subjects were requested to stand on a force platform, and features were extracted from the gathered data. These characteristics include variables like the trajectory of the centre of pressure (COP), the force distribution, or other pertinent metrics. The testing results of the suggested methodology showed that using the features extrapolated from the force platform data, the system could correctly recognize the four leg postures (double leg stand, single leg stand, toe stand and squat stand). This implies that the features retrieved from the force platform data provided helpful information that made it possible to classify the postures accurately. The proposed 1-D force data employed with machine learning models allowed for recognizing human postures with similar accuracy to traditional approaches that use video images, biomarkers, or IMG sensors. The comparison between the proposed method and the most advanced methods currently in use is shown in Table 5. It is noted that the proposed approach outperforms previous models and leverages real-time data collection to provide 98.0% accuracy, proving hypothesis H1.

Table 5.

Comparison of the proposed method with existing state-of-the-art.

Author (Year)	Sensor	Algorithms Used	Highest Accuracy (%)
Lee et al. (2022)	UAV’s camera - human body posture recognition and tracking	Deep neural network model (YOLO) - fuzzy logic	95.2
Kim et al. (2018)	Pressure	Naïve Bayes classifier (NB), decision tree (DT), neural network (NN), and support vector machine (SVM)	95.3 (CNN)
Lee et al. (2020)	IMU	Machine learning and deep learning algorithms	80.9 (Deep learning)
Proposed Model	Force platform	DT, EN, SVM, ANN, K-NN, and LR	98.0 (ANN)

By adding more characteristics and subjects or selecting the right features for the posture classification, the accuracy of the classification system might further be enhanced. In this study, as the first 1 s of data was discarded for the analysis, the stabilization effects might not have been properly taken into consideration. Since initial instability may have an impact on data quality and classification accuracy, a longer exclusion period may be beneficial in providing a more consistent and reliable baseline, particularly for individuals who may require more time to stabilize on the force platform. Furthermore, future research would consider screening for and reporting spinal conditions to evaluate their potential influence on balance and posture control.

Conclusion and future scope

The study used a force platform approach to classify postures for four different leg poses. Frame rate, video resolution, backdrop lighting, and image blurring are some significant drawbacks of 2D and wearable sensors (IMU) techniques that the proposed method attempted to address. The characteristics retrieved from the VGRF and COP were used as input features to the classifier models, with different feature combinations to categorize the leg postures of 25 people across two trials. The ANN, Ensemble and K-NN models provided superior classification performance with several feature combinations. With the best feature selection techniques, ANN surpassed the other classifiers in classifying the leg postures, providing 98.0% accuracy.

The model’s accuracy might be improved by adding more characteristic subjects or selecting the correct hyperparameters for the classifiers. Human postural identification has several uses and can reveal necessary information about behaviour, mobility, and health. As technology develops, there is potential for future advancements in this area, making it possible to analyze posture in real time to avoid musculoskeletal disorders.

Author contributions

K. Adalarasu: Conceptualization, Methodology, Data curation, Formal analysis, Writing – original draft. M Jagannath: Conceptualization, Supervision, Validation, Writing - review & editing. K. Ghousiya Begum: Data curation and writing. Sannasi Ganapathy: Supervision, Formal analysis.

Funding

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

Data availability

The data and materials will be made available upon request to the corresponding author.

Declarations

Consent for publication

The authors affirm that human research participants provided informed consent for publication of the images in Fig. 3.

Competing interests

The authors declare no competing interests.