INTRODUCTION
Ensemble learning has only recently become a reality due to the ease of implementing detection through internet of things (IoT) devices (e.g. smart tools like bracelets and fitness trackers) or smartphone devices (Zhou et al., 2019). Smartbeds, for instance, incorporate embedded sensors to comprehend the surroundings, identify user behavior, and pinpoint the patient’s position (Carpineti et al., 2018). The study of patient mode detection is a form of patient activity identification and has attracted a lot of interest (Kwapisz et al., 2011; Jahangiri and Rakha, 2015). The physical methods of tracking and monitoring patients by hospital staff require all-time alertness from the hospital staff, and they are practically less efficient. Thus, the ensemble learning-based method substitutes for the conventional approach, which is viewed as being constrained, less efficient, and more costly (Widhalm et al., 2012). Moreover, fitness monitors (Lee and Kwan, 2018) can be used to monitor human health and physical activity. These software programs can also show everyday tasks and calorie expenditure. Environmental danger exposure can also be tracked, and actions that have an impact on the ecosystem can be approximated. Presently, three distinct kinds of approaches are used to identify patients’ transit modes: sensor-based methods (Bedogni et al., 2016) and efforts based on cellular networks (Reddym et al., 2010). Additionally, a number of study articles (Feng and Timmermans, 2013; Bhattacharya and Lane, 2016; Su et al., 2016) have made an effort to develop the patterns of the patients’ state to enhance the precision of these methodologies.
There has been widespread sensing in recent years, which primarily concentrates on extending knowledge using the data collected from these omnipresent devices (Wang et al., 2019). With the potential to be a key component of numerous applications, including video security (Shao et al., 2012), healthcare (Osmani et al., 2008), participatory gameplay, smart homes, and general tracking systems (Reyes-Ortiz et al., 2016), ensemble learning has caught the interest of both businesses and academics.
The state of the art contains many intricate and understandable models, and these techniques have shown substantial capacity to meet all essential requirements in feature engineering (Wang et al., 2019). One of the most popular deep neural network (DNN) research techniques is the convolutional neural network (CNN). It uses consecutive data analysis and has been proven to have the dual benefits of learning from and effectively depicting the data (Münzner et al., 2017). The recurrent neural network model is flexible enough to adapt to changes in the input size, in contrast to CNN networks, which operate on a fixed-size input. Although the auto-encoder is flexible, it significantly relies on the pricey and time-consuming data labeling (Zhang et al., 2022). While the spatial feature restriction is seen in long short-term memory (LSTM), CNN also has a limited ability to derive time features (Ghosal et al., 2022).
One strategy is ensemble learning, which allows you to benefit from all of the aforementioned models. It uses several machine learning (ML) algorithms to obtain a mediocre forecast result based on the characteristics retrieved, which can then be combined to produce results that are superior to those obtained using individual algorithms (Dong et al., 2020). Additionally, the current structure of deep neural methods has limitations, and the data are naturally chaotic and have varied traits. As the model has many hyperparameters, the separation of information necessitates layering numerous levels in deep learning (DL), which ultimately increases intricacy as well as ambiguity. These versions are also relatively steady and less chaotic. Additionally, as more models are used, they become more beneficial for both linear and nonlinear data, and the data vary because of the various sensor device technologies used for efficient data retrieval (Sewell, 2008).
The proposed method has the ability to identify four types of activities: sleeping, standing, sitting, walking, and emergency states. All techniques for figuring out the average patient state are based on the same universal general principle: collect raw data from smartbed sensors, prepare the data for feeding into the classification model, and for training purposes, have the classifier use the pre-processed data. Later, the remaining pre-processed data are used for prediction and generalization.
The research work that has been suggested to help increase the utility of transit mode detection methods still has some drawbacks, despite the fact that these methods have been successful in many uses. First off, it is challenging to evaluate the reported findings because there is no standard sample.
Additionally, many plans have employed binary classification, where the classifier only needs to differentiate between two forms of transit; for instance, taking standing and walking as means of transit of the same type into account (Krieg et al., 2018). A range of approaches have been used to improve the precision of a single machine learning model used to detect the patient’s state. A singular ML method may perform well on certain sets of samples, but over- or underfitting may become apparent on other sets of samples. Therefore, when using one ML method, a maximum limit of precision may be achieved even with parameter tuning. This flaw is fixed by combining different ML methods (ensemble learning) (Ma et al., 2018). Recently, several ensemble methods have been investigated to identify the patient’s condition, including extreme gradient boosting (XGBoost), gradient boosting, and random forests (RFs). The only ML methods that can be combined in these ensemble approaches are decision trees and other analogous ones.
To identify the patient’s condition using smartbed devices, an ensemble learning model is also demonstrated. This model greatly increases the detection accuracy over the traditional approach.
In this paper, we propose to use an ensemble learning method-based ML algorithm that can address the issue of patients’ state using smartbed sensors. We have developed a novel ensemble technique to distinguish between the transitory states of patients, as suggested in light of the findings from this research. Compared to other methods, this work yields encouraging outcomes.
The rest of this paper is structured as follows: The next section provides the literature review followed by the proposed method, testing conditions, results and discussion, and finally, the conclusion.
LITERATURE REVIEW
Sensor-based techniques make use of unprocessed data obtained using smartbed devices. The recorded data in the form of series reading are pre-processed, and the complexity of the captured data may be reduced to improve the efficiency or streamline the method. The pre-processed data are then used as the training input for an ML algorithm. In order to forecast the method of transit recorded by the smartbed sensor, the duration and data quality of the training period were significant factors. The ML methods have two benefits: their ability to help identify motion and their low energy usage, making them the most probable smartbed-integrated sensor (Hemminki et al., 2013; Shafique and Hato, 2017; Urbancic et al., 2018; Wang et al., 2018). To extend a multiclass technique capable of identifying the patient’s state, ML methods can be used. These methods include the following classes: strolling, standing, sitting, walking, and emergency states. Support vector machines (SVMs), deep learning (DL), k-nearest neighbors (KNNs), bagging, and RFs are some of the ML techniques that have been proposed by researchers. Smartbed-integrated devices were used to collect raw data. The out-of-bag error and k-fold cross-validation are used during the training and testing stages to choose the model that works best. Features that have a detrimental impact on the performance of the classifier features were eliminated using a feature selection technique called minimal redundancy maximum relevance.
It was observed that the information gathered by smartbed-integrated devices offers useful patterns that differentiate between various forms of the patient’s state. After using a feature selection technique, the authors contrasted the performances of the evaluated ML algorithms and discovered that the proposed ensemble approach and SVMs outperformed the others.
In order to distinguish the active method of state detection, Shafique and Hato (Shafique and Hato, 2017) looked at four ML algorithms, specifically adaptive boosting (AdaBoost), RFs, DTs, and SVMs. A deep learning-based method was put forth by (Xiao et al., 2017) to successfully create a nonlinear structure between named and consecutive data that were collected using sensors. This method created two different feature groups using a DNN.
With the emergence of a hyperconnected world where people and objects are linked by networks such as information and communication technology and as sensing technologies advance, the smart device business is expanding quickly. Wearable technology has the benefit of being able to communicate with the user while being worn and constantly collecting data on the user and their surroundings. This research makes use of these benefits by focusing on user identification technologies based on biosignals that can be generated as the output by smartbeds from the sensor devices.
The ML techniques for patient care rely on publicly displayed physiological and bodily types of knowledge, and user collaboration is necessary. Researchers are presently working on user identification studies that make use of sensory signals that can be monitored with regard to the patient’s position. SVM (Shafique and Hato, 2015), KNN (Fang et al., 2017), and RF (Mehta and Lingayat, 2008) are just a few examples of the ML techniques used by existing user identification technologies that use heart rate and blood pressure. Body position can have a significant effect on heart rate and blood pressure. Presently, academics are looking at a deep learning-based user identification technology, which instantly captures features while learning without the need for a distinct feature extraction procedure. Because a single network cannot acquire all the data that are challenging to identify, previous neural networks made up of just one network showed performance constraints. Overfitting happens in the learning process, and performance declines if all difficult-to-recognize data are learned (Zhao and Zhang, 2005). In order to address the issues of overfitting, which happens in single networks currently in use, and a decreased identification rate because comparable characteristics show between classes, this article suggests an ensemble network. The system is made to merge outstanding characteristics generated by distinct neural networks with various parameter values, followed by relearning.
The suggested deep learning-based ensemble network uses sound, horizontal motion, vertical motion, circular motion, motion, direction of motion, speed of motion, and heart rate and blood pressure data.
Electrocardiograms, which have traits that are particular to each person due to the electrical variables of the heart (heart position, size, and body circumstances), have been the subject of numerous studies on user identification. Deep learning, which exhibits exceptional performance in a variety of technological areas such as identification, categorization, and prediction, is currently being studied by academics in the field of user recognition techniques. Extraction of useful features is a key success indicator in current, conventional ML techniques. As a result, it is essential to carry out a procedure whereby specialists with prior expertise directly separate characteristics. On the other hand, deep learning conducts learning without a distinct feature extraction procedure while autonomously extracting features.
Smartbed Systems and Disabilities
An all-inclusive strategy for providing personalized care and support for patients with disabilities: smartbed systems: a comprehensive approach
It has recently come to light that smartbed systems are a potentially useful tool for supporting individuals with impairments in a variety of elements of their day-to-day life. These systems are able to offer a complete and individualized method of patient care and support because they include cutting-edge sensor technology, data processing capabilities, and ML algorithms.
Real-time monitoring of patients’ mobility statuses, such as whether they are sleeping, standing, sitting, walking, or in an emergency scenario, is possible with the use of smartbed systems. Carers and healthcare professionals are able to acquire useful insights into the patient’s current health conditions, identify prospective problems at an early stage, and offer support that is both timely and suitable when they precisely document these activities. In addition, the use of these technologies may assist people with impairments in continuing to live independently while also improving the overall quality of life they have. The use of ML algorithms and ensemble learning methods has the potential to significantly improve the accuracy and resilience of the smartbed system. This will ensure that patients get the highest possible level of care and assistance that is specifically catered to meet their individual requirements.
Patients with disabilities gaining independence through improved mobility adaptive assistance and state detection via the use of smartbed systems
The use of smartbed systems has the ability to completely transform the manner in which disabled people are cared for as well as given more autonomy. These systems give patients with a greater degree of personalized care by using cutting-edge technology and powerful algorithms. As a result, patients are able to lead lives that are more independent and rewarding.
The capacity of smartbed systems to reliably identify the many mobility states of patients, such as whether they are sleeping, standing, sitting, walking, or in an emergency, is one of the most important qualities of these beds. This information may be put to use in the creation of individualized care plans and adaptive support techniques that are tailored to meet the specific requirements of each individual patient. For instance, smartbed systems may detect when a patient is having difficulty standing up or moving and send an alarm to carers so that they can offer the patient prompt assistance if necessary.
In addition, smartbed systems are able to monitor and assess the movement patterns of patients over time. This provides the medical personnel with the ability to recognize patterns and make choices about treatment, medicine, and assistive equipment based on the data collected from the patients. Smartbed systems may help contribute to a more adaptable and efficient care environment by continually monitoring and adapting care plans depending on the patient’s development and changing requirements.
In a nutshell, smartbed systems have the potential to revolutionize the manner in which people with disabilities are cared for as well as given more autonomy. These technologies have the potential to dramatically enhance the quality of life of patients by offering precise mobility status detection, personalized care plans, and adaptive support while also easing the strain on the shoulders of carers and healthcare professionals.
Enhanced patient care for people with disabilities via the use of the ensemble approach: a comprehensive overview
In this part, we give a more complete description of the ML methods that are utilized in our ensemble learning approach. This approach has been tuned to the unique needs of improving patient care for those who have impairments within the context of smartbed systems.
SVMs: the classification method for mobility states
Support vector machines, or SVMs for short, are a kind of supervised learning technique that is often used to solve problems involving classification and regression. For the purpose of our research, we categorize the different mobility states of people with disabilities using SVMs. These mobility states include sleeping, standing, sitting, walking, and emergency scenarios. Because SVMs are able to process both linear and nonlinear data, the algorithm is able to identify intricate patterns in the sensor data that have been acquired by the smartbed system.
RF for improved accuracy in predictive models
RF is a form of ensemble learning that combines the output of numerous decision trees that have been constructed using the RF method. This results in a forecast that is more accurate and stable. In the context of smartbed systems for people with disabilities, RF delivers improved prediction accuracy for a variety of mobility states by lowering the correlation between trees and giving a more diversified collection of models. This is accomplished by minimizing the amount of overlap that exists between the two.
KNNs: adaptive state recognition
KNN is a method for the classification of problems that are instance-based, non-parametric, and used to find the neighbors of a given point. In our ensemble method, KNNs play a role in the process of adaptively recognizing the mobility statuses of impaired patients based on the similarity of their sensor data. KNNs enable the categorization of the patients’ mobility statuses based on the most common class among their neighbors. This is accomplished by first locating the K data points in the training dataset that are physically closest to one another.
Bagging: an improved model stability
Bagging, also known as bootstrap aggregating, is a method of ensemble learning that was developed with the objective of enhancing the consistency and precision of ML algorithms, in particular decision trees. Bagging is used to construct several training sets (TSs) in our smartbed system for people with disabilities. These sets are created using bootstrapped samples of the original data. This results in a model that is more reliable and accurate for classifying persons’ mobility states.
XGBoost
Extreme gradient boosting, or XGBoost for short, is an efficient and scalable version of the popular gradient boosting method. This method is yet another instance of an ensemble approach. XGBoost is used in the context of smartbed systems for people with impairments in order to improve the predictive performance of the system. One way to achieve this goal is by the sequential development of a set of decision trees, with each tree making an effort to fix the shortcomings of the one before it.
Multilayer perceptron
In artificial neural networks, a multilayer perceptron (MLP) is a subtype consisting of multiple layers of neurons connected by weighted edges. MLPs may learn very intricate, nonlinear relationships between input and output variables. Our smartbed system for the disabled makes use of MLPs that can decipher complex patterns in the sensor data and accurately categorize the user’s mode of movement.
Ensemble learning approach: enhanced patient care of individuals with disabilities
To increase the overall accuracy and robustness of our model in forecasting the patient’s mobility status, our suggested ensemble learning technique incorporates the predictions from numerous ML algorithms, such as SVM, RF, KNN, bagging, and XGBoost. These ML algorithms include SVM, RF, KNN, and bagging. Our goal is to improve patient care for people with disabilities by precisely monitoring their actions and giving help that is both timely and appropriate. We want to do this by capitalizing on the strengths of a variety of algorithmic approaches. In this scenario, an MLP neural network acts as a meta-learner, and its role in the generation of the final output of the ensemble model (EM) is to aggregate the individual predictions made by other components of the model.
THE PROPOSED METHOD
The suggested structure in Figure 1 shows that ensemble learning has been used in conjunction with the ML method to distinguish between the patient’s states. The ML algorithms are used to make a forecast using the ensemble technique in order to more effectively extend the predictions. The ensemble’s forecasts are then combined with two other ML techniques, bagging (Khazaee and Zadeh, 2014), XGBoost, AdaBoost (Li et al., 2012), linear support vector machine (LSVM), and KNNs (Hawkins, 2004), to create layered learning, which ultimately employs the MLP neural network design (Ho, 1998) to determine the patient’s state. The goal is to identify which combination of Machine Learning (ML) algorithms in a stack produces the most accurate forecast. Therefore, we select base estimators because the final patient state is determined most accurately when all component predictions are combined. Figure 1 shows the suggested layered approach.
Formally, let’s symbolize the activity data recorded by the subject smartbed’s integrated sensors for model training, and let’s use the symbol to signify the target class designating that activity.
The entire TS is represented by the symbols:
Additionally, the feature space F (also known as the vertical representation of the dataset) is made up of a number of traits that reflect smartbed-integrated sensor readings that have been applied to all of the dataset’s examples. For instance, the first feature in the collection, F 1, may indicate the average values from one sensor’s measured value; the second feature, F 2 indicates the average of the standard deviation value, and so on.
By repeatedly splitting the dimension space, the multiclass groups that exist in the classifier and have the same labels are denoted as (for data samples), (for target vector), such that n = 1…N. The result, which is in the shape of 0, 1, … L − 1, for a particular target node d, as Td which is written as . Here, we use an indicator function which is denoted as I(.) whose value becomes 1 if the statement is true and 0 if the statement is false. As a result, the function I(Tn ) takes the value 1 if this statement is true and 0 if the statement is false (which means the patient’s state is recorded correctly). Td is the target node vector for d nodes.
Following that, an ML method that satisfies the criteria of the general significant ensemble prediction model is applied to the examples and their names. In order to perform better, the EM draws lessons from the errors of the prior branch. The EM that fits the target data is denoted by E 1(s) = T and the records of previous data are given as E 1(s) = T − E 0(s). This process keeps going until something stops it. After finishing several stages (identical to other boosting techniques), the EM is one of the types of the additive models; here, it is considered to have k independent trees. A random differential loss function is then optimized for generalization, as shown in the adjacent equation (Cover and Hart, 1967).
In the proposed method, each branch in the ensemble technique uses random variable gathering. Using a function P(A), which has been used in order to reduce the value of loss and is denoted by a definite loss function, the objective is to anticipate the class target T. The basic estimators are used to build the ensemble predictor E(P). Using polling, the basic estimators are combined to create the ensemble prediction E(P). T′ is a collection of possible target states, and its formula is written as . The ensemble receives the forecasts from the classifier in the second step. Let . If the classifier selects the target class T then E(P) = 1. The forecast class is the one that has been classified with maximum weight. KNN is an additional ML method that is used in this stage. The feature vector’s k-closest training examples are used as the KNN algorithm’s input. The result is a member of a particular class in which a particular state is chosen. Since KNNs’ function is estimated locally and their classification is completed after deferring all calculations, they serve as an excellent illustration of passive learning or instance-based learning (Kleene, 1951). Let K(s) = t stand for the label function, which selects a class target for a training instance from among l possible labels .
The end technique generates different bootstrap instances by teaching multiple base learners using a base learning method. By using the TS in 30% and 10% ratios, the prediction is made when a particular class is dominant after basic learners have been obtained and merged using the random subspace technique.
The suggested ML algorithm is then used to build the ensemble learning algorithm from the training data in the second phase (MLA). At least three layers are used to create an MLA: an input layer, a hidden layer, and an input layer. The hidden layer is the computational component (neuron). The objective of a hidden layer is to create an outcome from a stimulus that comes from another source or another hidden layer.
TESTING CONDITIONS
A notebook with the Windows 10 operating system, an i7 processor, and 32 gigabytes (GB) of random access memory (RAM) is used to test the suggested technique. The well-known ML toolkit is made available through the use of the Anaconda version and the Python computer language (i.e. Scikit-learn).
A total of nine sensors were included in the data collection from the patient’s smartbed. All sensors are connected to the patient’s body via the smartbed. These data include signals from limbs, forearms, blood pressure, and heart rate data, which are divided into nine categories based on the instrument placement as shown in Table 1. In this case, learning data accounted for 80% of the total data and trial data for the remaining 20%. We have used the proposed method with different classifiers such as XGBoost, AdaBoost, KNN, and LSVM.
Confusion matrix for the proposed method to classify different patient states.
Sleeping | Standing | Sitting | Walking | Emergency state | |
---|---|---|---|---|---|
30% Testing data split predicted class | |||||
Sleeping | 78 | 0 | 1 | 0 | 0 |
Standing | 2 | 65 | 2 | 1 | 0 |
Sitting | 1 | 2 | 87 | 0 | 0 |
Walking | 2 | 2 | 1 | 98 | 0 |
Emergency state | 0 | 0 | 0 | 0 | 14 |
10% Testing data split predicted class | |||||
Sleeping | 108 | 0 | 1 | 0 | 0 |
Standing | 2 | 231 | 2 | 2 | 0 |
Sitting | 2 | 2 | 119 | 0 | 0 |
Walking | 2 | 3 | 1 | 276 | 0 |
Emergency state | 0 | 0 | 0 | 0 | 27 |
The findings of patient identification performance using these signals are based on ensemble learning. Table 1 shows the confusion matrix table representing the ensemble learning network for the patient’s state identification performance. The forecast values were obtained by feeding the trial data into the learned neural network model, and the findings are shown in the table. The rows represent the trained data that the suggested technique was able to identify, while the columns represent the actual data received from sensors. According to the testing findings, the ensemble network showed a maximal improvement in accuracy of 0.8%.
In terms of precision, sensitivity, and specificity, the suggested technique performed better than any of the available methods, according to the study of its performance. It also looked at how well the e-mail network performed. The ensemble network, however, performed exceptionally well at identifying all the trial data. It was able to get around the shortcomings of single networks, which were unable to identify identical patterns, by relearning the characteristics that were generated by each individual network.
The dataset that records real-time sensor readings from patients’ smartbeds is shown in Table 2. This data collection gathers sensing data from nine distinct sensors placed on the patient and the smartbed. The Android software that records the method of transit can be used without being constrained by the dataset. To mimic real-world scenarios, each sensor captures the instrument readings as an action is carried out. Each sensor is given a task to complete in order to capture the instrument readings it causes. Table 3 shows the 24-h record of patients between the different states. The patient may move between various states such as sleeping, standing, sitting, walking, and an emergency state. The dataset was created in accordance with the basic procedures and actions used by the patient.
Features of the dataset chosen.
Sensors (class category) | Features |
---|---|
Sound | Sound |
Horizontal vector sensors | Horizontal motion |
Vertical vector sensors | Vertical motion |
Rotation vector | Circular motion |
Linear acceleration | Motion |
Direction | Direction of motion |
Speed | Speed of motion |
Heart rate | Heart rate |
Blood pressure | Blood pressure |
Recording time for all patient states.
Activity | Measurement time |
---|---|
Sleeping | 06:10:19 |
Standing | 03:20:13 |
Sitting | 09:23:14 |
Walking | 05:06:13 |
Emergency state | 00:00:01 |
A program that establishes a maximum cadence of 20 hertz (Hz) then gathers the sensing data. The frequency components of bodily movements are recorded and found to be lower than 20 Hz, making the highest recording rate of 20 Hz a reasonable option (Raschka, 2018; Boehmke and Greenwell, 2019). Additionally, even when the patient is not moving, the upholstery of the sensor vibrates between 3 and 5 Hz, and the frequency of vibration changes when the patient is moving (Karantonis et al., 2006). When completing one of these tasks, the connected sensor, using a program, starts or ceases recording the data.
Once the observed state of the patient changes, some data are given, including the identity of the sensor, raw sensor data, and date. The dataset includes a total of 24 hours of recording, as shown in Table 3. This table depicts the actions taken by the patient, with each measurement being recorded. The dataset is described as follows: the dataset contains 26% of the sleeping state, 25% of the standing state, 24% of the sitting state, 20% of the walking state, and only 2% of the emergency state. The gathered information combines observations made by all sensory devices connected to the patient and the smartbed.
Sensors are placed that record sound, horizontal motion, vertical motion, circular motion, random motion, direction of motion, speed of motion, heart rate, and blood pressure. Time frames were created from the information. The time window was set by the dataset’s creators to 5 seconds for two reasons: to capture the patient’s state. It was observed that a 3-second window was the ideal time to measure the patient’s entire state. Setting a tiny window size leads to misclassification because some modes need more time than 3 seconds to show their state traits (e.g. data collection must last for at least 4 seconds for sleeping state identification). Setting a large window area causes sounds from various other activities to enter, which interferes with the intended state of the patient. Thus, a total of nine features from the sensory data were built as shown in Table 2.
RESULTS AND DISCUSSION
The values of the base estimators were adjusted in order to enhance the performance of the suggested technique in terms of the four assessment measures while maintaining a reasonable temporal complexity. By using the ensemble techniques, key hyperparameters were selected for the other basic estimators used to make the model less complicated. These hyperparameters, which include the number of hidden layers, the number of neurons in each hidden layer, the activation function, and the algorithm, were taken into account in the final stage of the suggested approach. To reduce the temporal complexity, two concealed levels were added to the input layer and the output layer. The Adam optimizer was selected over stochastic gradient descent from the two weight optimizers that were tried because it performs exceptionally well with datasets that have a sufficient amount of samples.
The outcomes of the suggested approach and the various ML methods put to test in the trials are presented in this part. Precision, memory, F1-score, and accuracy were the four metrics used to assess how well the existing techniques and the suggested strategy performed.
The datasets were divided into 30% testing and 10% testing. Using four assessment metrics—accuracy, precision, recall, and F1-score—the suggested method was assessed and contrasted with the other machine methods, as shown in Table 4. Table 5 shows the training and testing times for our experiments. The results show that the suggested method worked better than the other methods; its accuracy is 89%.
Evaluation measures of all algorithms.
Accuracy (%) | Precision (%) | Recall (%) | F1 (%) | |
---|---|---|---|---|
30% Testing data split predicted class | ||||
Proposed | 86 | 86.4 | 87 | 86 |
XGBoost | 82.1 | 83.2 | 84.9 | 82.5 |
AdaBoost | 78.5 | 79.2 | 80.2 | 77.9 |
KNN | 66.3 | 67.6 | 68.8 | 66.8 |
LSVM | 68.3 | 70.4 | 70.1 | 69.1 |
10% Testing data split predicted class | ||||
Proposed | 89 | 86.9 | 88.2 | 87.1 |
XGBoost | 83.3 | 84.8 | 85.9 | 82.9 |
AdaBoost | 79.1 | 80.6 | 81.7 | 78.3 |
KNN | 67.2 | 69.9 | 69.9 | 65.7 |
LSVM | 69.6 | 70.8 | 71.2 | 69.8 |
AdaBoost, adaptive boosting; KNN, k-nearest neighbors; XGBoost, extreme gradient boosting.
The value of measured time (in seconds) for all algorithms.
Method | 10% testing data split | 30% testing data split | ||
---|---|---|---|---|
Training | Testing | Training | Testing | |
Proposed | 3.743 | 0.002 | 4.982 | 0.002 |
XGB | 1.643 | 0.008 | 3.542 | 0.007 |
AdaBoost | 0.052 | 0.002 | 0.626 | 0.002 |
KNN | 0.012 | 0.002 | 0.011 | 0.034 |
LSVM | 120.8 | 0.026 | 378.4 | 0.045 |
AdaBoost, adaptive boosting; KNN, k-nearest neighbors.
The details of each category of the sample that was successfully classified and the sample that was wrongly classified are shown in Table 1 using an uncertainty matrix. The first set of experiments found that the samples from the sleeping class were correctly classified 94% of the time, the samples from the standing class about 90%, the samples from the sitting class about 77%, the samples from the walking class about 92%, and the samples from the emergency state class were accurately classified 96.5% of the time when the dataset was divided into 30% training.
It was discovered that the samples from the sleeping class were correctly classified 96% of the time when the first dataset was divided into 10% testing, the samples from the standing class were roughly 91% of the time, the samples from the sitting class were roughly 79% of the time, the samples from the walking class were roughly 93% of the time, and the samples from the emergency state class were correctly classified 98% of the time.
CONCLUSION
In recent years, significant progress in the field of ensemble learning has been made with sensors. The knowledge of a patient’s movement can be used to provide a range of services. Researchers’ interest in using ML to identify the method of transit is growing because there are many services that can make this possible. Using the data collected from the sensors placed on the patient and the smartbed, activity can be detected, such as sleeping, standing, sitting, walking, and being in an emergency state. To verify the suggested approach, we used 30% testing and 10% testing on a split dataset, and the high precision attained by the suggested method suggests that the sample size used to assess the suggested method is adequate.