Design and implementation of intelligent patient in-house monitoring system based on efficient XGBoost-CNN approach

Abstract

Because of the scarcity of caregivers and the high cost of medical devices, it is difficult to keep track of the aging population and provide assistance. To avoid deterioration of health issues, continuous monitoring of personal health should be done prior to the intervention. If a problem is discovered, the IoT platform collects and presents the caretaker with graphical data. The death rates of older patients are reduced when projections are made ahead of time. Patients can die as a result of minor abnormalities in their ECG. The cardiac dysrhythmia/irregular heart rate is classified with several multilayer parameters using a deep convolutional neural network (CNN) approach in this paper. The key benefit of utilizing this CNN approach is that it can handle databases that have been purposefully oversampled. Using the XGBoost approach, these are oversampled to deal with difficulties like minority class and imbalance. XGBoost is a decision tree-based ensemble learning algorithm that uses a gradient boosting framework. It uses an artificial neural network and predicts the unstructured data in a structured manner. This CNN-based supervised learning model is tested and simulated on a real-time elderly heart patient IoT dataset. The proposed methodology has a recall value of 100%, an F1-Score of 94.8%, a precision of 98%, and an accuracy of 98%, which is higher than existing approaches like decision trees, random forests, and Support Vector Machine. The results reveal that the proposed model outperforms state-of-the-art methodologies and improves elderly heart disease patient monitoring with a low error rate.

Introduction

Technology plays an important role in people's daily lives. Because living spaces are combined with other social environments, feasible-based assisted living is introduced. Many various uses for improving the efficiency of assistive parts for living things have been presented. Wearable technology (Aliverti 2017) is widely used because it allows people to track their health-related information. With advancements in technology, even elderly patients can benefit from active health monitoring through the Internet of Things (IoT) (Lee and Lee 2015; Gubbi et al. 2013). The vitals of elderly patients must be monitored on a regular basis as their health deteriorates without warning. Low-cost fitness trackers have been widely utilized in medical applications to track their everyday activities. Vital signs such as ECG, heart rate, rapid heartbeat, body temperature, breathing rate, blood pressure, and other vital parameters are monitored in bedridden and elderly patients. A wearable sensor collects data from users by identifying vital metrics on noninvasive platforms (Kim and Kim 2020).

To determine the number of inputs into an associated output, deep learning/ machine learning (Zhang et al. 2017) is used. These are low-cost and provide quick optimization services. Improving life quality by determining health conditions ahead of time can extend life expectancy. Without any additional complications, caregivers can easily check on the patients' condition. To achieve this objective, an elderly monitoring IoT application is developed in this work and the statistics graph in the app is based on linear type features. Elderly people and their caregivers can utilize the app. By statistical data, the caretaker receives a message report of the patient's health condition. If there is a problem with the patient's health, the carer receives a voice message. Early diagnosis of sickness in patients can be aided by early intervention methods, according to this report. This can help clinicians make better decisions by providing precise results.

The major goal is to use sensory data to obtain physiological data and the vitality of human activities (Neethirajan 2017; Mathew et al. 2018). Since technology is rapidly evolving, IoT has become a part of living beings. This Internet of Things comprises a framework that connects other devices, such as for monitoring purposes. The fundamental motivation for our research is to improve living conditions by increasing life expectancy. These are also used to detect the old and bedridden, as well as to monitor their vital parameters. This study focuses on non-invasive remote monitoring of elderly heart disease patients and alerts the caregivers via text messages and automated phone calls.

Non-invasive platforms are mostly used to participate in technologies such as wearable sensors and other health-care-related sensors. Large volumes of medical data are collected by these wearable devices. Low-power hardware is used in this paper to implement ECG-based authentication. This uneven heartbeat prevents the proper proportion of blood flow to the heart. These are detrimental to the heart and can induce brain malfunction. The electrocardiogram (ECG) is a tool for monitoring the heart's electrical activity. Because the signal orientation is rapid, deep learning techniques are prominent in ECG. Consecutive R peak values are used to validate heart rate from an ECG.

Diverse types of sensors can be used to investigate the various forms of cardiac diseases. Chronic patients must be regularly evaluated to see if they are suffering from a long-term illness while in the hospital. Sudden cardiac arrest, ventricular arrhythmias, and high blood pressure caused by hypertension and stress can shorten older people's lives. Because of signal noise and baseline type shift, identifying every minute heart rate signal is difficult. For chronic patients, the complex signal QRS is evaluated. Photoplethysmography (PPG) is used to quantify the periodic polarization and depolarization of the atrial and ventricles in the heart using a complex signal and provides good ECG parameter accuracy. PPG simply delivers light through the vessel, which the photodetector senses. PR, QRS, QT, and RR intervals are used to derive these.

The complexities associated with in-house patient monitoring are medical errors such as incorrect medication, wrong dosage, frequency, administration route, etc.which mainly affects the patient's life. Mainly the elderly population is the one who is often prone to these medical errors since they suffer from different medical conditions. The elderly patients are often prone to these medical errors since they have very low knowledge related to the drug or their physical illness which increases the complexity to analyze the medications. These medical errors can be analyzed efficiently by monitoring the physical signals sent by the patient and offering proper treatment at the appropriate times. In this way, the degree of adherence to the treatment is improved and the medication errors are minimized to prevent health deterioration. To overcome this issue, a novel Intelligent Patient in-house Monitoring System is proposed in this paper which is our major contribution.

Approaches based on supervised learning algorithms allow for the efficient selection of maximum coefficients (Oliver et al. 2018; Entezami et al. 2019; Rejeesh and Thejaswini 2020; Sundararaj 2016, 2019). The diagnosis of atrial fibrillation is based on a heart rate rhythm examination. Every hour, statistical data is analyzed and given to the caretaker in the form of an automatic message. If the continuous monitoring of vital parameters of elderly patients changes, an automated voice call is sent to the caregiver as a warning. Since the shallow machine learning techniques (Tamal et al. 2019: Nhu et al. 2020; Sundararaj et al. 2020;Vinu 2019; Jose et al. 2021; Gowthul Alam and Baulkani 2019; Nanjappan and Albert 2019; Nanjappan 2021; Srinivasan and Madheswari 2018; Azath et al. 2011; Sundararaj and Selvi 2021) such as Support Vector Machine (SVM), Naïve Bayes (NB), Random Forest (RT), and Decision Tree (DT) need manual training to identify the crucial patterns in the dataset, they are often error-prone and computationally expensive. The major contributions of this paper are presented below:

• Design and deployment of an elderly heart disease monitoring system using XGBoost based CNN classifier to monitor the elderly patients in their home and enhance their adherence to the treatment.

• To overcome the time-consuming manual tuning of parameters, minimize the medical errors, and reduce the computational complexity, in this paper, we have used a CNN algorithm to predict the abnormalities based on the patient's input data.

• The data imbalance and overfitting issues associated with the CNN classifier are overcome via the usage of the eXtreme Gradient Boosting (XGBoost) technique. In this way, the overall performance does not degrade with a minimal number of data classes.

• The performance of the proposed XGBoost based CNN classifier is evaluated in terms of different performance metrics such as accuracy, recognition accuracy, recall, precision, and execution time to verify its efficiency.

The rest of this paper is ordered as follows: “Review of related works” section presents the existing literary works and “Proposed XGBoost CNN model for elderly heart patient monitoring” section presents the overview of the various concepts used in the proposed methodology along with its implementation in detail. “Experimental results and discussion” section demonstrates the experiments conducted using the medical datasets by comparing the results obtained by our proposed classifier with others. This section also discusses the benefits of our proposed approach with the existing technique and the paper is concluded in “Conclusion” section.

Review of related works

Liaqat et al. (2021) used modified logistic and dynamic Newton Leipniks maps for monitoring the wandering behavior of elderly patients. This paper uses the publicly available radar spectrogram dataset to identify the large-scale body movement of elderly patients such as walking, drinking water, sitting, standing, etc. They used both machine learning and deep learning models such as random forest, k-nearest neighbors, Long Short Term Memory (LSTM), Bi-directional Long Short-Term Memory (Bi-LSTM), and Convolutional Neural Network and Support Vector Machine. The principal component analysis and data augmentation are used to preprocess the raw input images. In identifying the wandering behavior of the elderly patients this approach offers an accuracy of 95.30%.

Kim et al. (2020) monitored the elderly sleep and wake conditions by creating an unobtrusive sensing environment equipped with passive infrared sensors and high sensitive accelerometer. To access the performance of the proposed technique they are using five classifiers namely Multilayer feed-forward neural network (MFNN), naïve Bayes, AdaBoost, C4.5, and Random forest. Among these classifiers, the MFNN offered the best performance when compared to others. Vanus et al. (2017) presented a Levenberg–Marquardt algorithm (LMA) based Artificial Neural Network classifier to identify the presence of senior citizens in smart home care. This methodology also monitors humans' daily living activities and classifies them. The efficiency of their method is tested using different performance metrics such as Root Mean Square Error(RMSE), correlation analysis, MAPE (Mean Absolute Percent Error), and Black-Altman. However, the prediction results were not successful for a total of 200 neurons.

Huang et al. (2019) presented a two-stage end-to-end convolutional neural network with a data augmentation technique for wearable human activity recognition. Their technique offered two advantages such as minimized computational complexity and recognition accuracy. The existing techniques often offer average recognition accuracy even though in most cases it is acceptable, for certain activities such as climbing stairs need more recognition accuracy. The recognition accuracy is mainly affected due to low training data and complex behavior patterns. To overcome these issues, a novel XGBoost based CNN model is proposed in this work. Gao et al. (2018) provided a deep learning architecture named as recurrent 3D convolutional neural network (R3D) for extracting the spatial–temporal information for efficient human action recognition. For extracting the long-term temporal information, they are using the LSTM and 3D convolutional architecture. In this way, they are extracting the high-level abstraction of human actions, and this application is mainly designed to be implemented in smart monitoring of remote healthcare.

Kim et al. (2020) developed an indoor emergency awareness system using a deep neural network to monitor the children and elderly people at risk via mobile phones. They are extracting three types of information such as indoor location, human activity, and sound. The recognition accuracy is improved via sound and human activity. A residual Neural Network is used to identify the sound events and the deep spiking neural network is used to identify the activities related to an indoor location and human movement. However, the experiment area is small. Hosseinzadeh et al. (2020) provided an IoT-based health monitoring system to identify the behavioral and biological changes of elderly patients via smart IoT sensor technologies. For monitoring the behavioral and physiological changes they are using different techniques such as Decision Tree (J48), Sequential Minimal Optimization (SMO), Multi-Layer Perceptron (MLP), and Naïve Bayes (NB) classifiers. The results show that the decision tree classifier offers the highest performance of all. For monitoring the abnormal elderly patient movements and offering timely assistance, Youssef et al. (2021) used the overlapping gait pattern recognition method aided with regression learning. To enhance the recognition accuracy and minimize the time taken for training, the errors are recurrently analyzed. Based on the angle of deviation and direction of movement, the gait pattern is classified.

Lin et al. (2020) used wearable inertial measurement sensors to prevent elderly patients from falling in real-time and avoid the risks associated with it. The sudden rise of the acceleration and the orientation of the user's head are monitored via the accelerometer and gyroscope. Using five participants, the performance of the proposed methodology is evaluated using 450 non-falling events and 120 falling events. The proposed methodology achieves a fall detection accuracy of 95.44%.

Xie et al. (2020) proposed a new end-to-end model to integrate the extraction of features and their classification. It not only simplifies the process of data analysis but also provides better accuracy with robustness. Portable devices enable the user to monitor the cardiac status anywhere and at any time. For the daily life of patients and elderly people, their computational analysis of diagnostic gives the better potential for their health. Li et al. (2018) introduced an approach using a generic convolutional neural network that trains up the large scale of datasets without any distinction in inpatient status. In can, it takes time to monitor long-term clinical ECG status. The author proposed a new step in which accelerating the ECG classification by using original ECG considered as input to the convolutional neural network or using FFT way of representation. Use of deep CNN approach, convolutional kernels were used to improve accuracy in classifying the dataset.

Shehieb et al. (2017) developed a project to aid the patients by engaging and communicating more effectively in terms of daily life. In this approach, an Electroencephalogram (EEG) is used to facilitate communication between patients and caretakers. In this implementation, both hardware and software are used. EEG model along with 14 channels and tablet is used for hardware. In software, processing, an android application, an algorithm is used. Continuous reading of brainwaves is validated by the EEG module. It detects what the patients specifically needed. It matches the detection by using a predefined algorithm. Fuzzy logic pattern recognition algorithm determines the processing by using patients' eye movement and detection in color.

Wei et al. (2017) proposed an ensemble method approach to diagnosing diabetic patients on a large scale with an imbalanced dataset. From using this method, the author got the results in promising accuracy with high sensitivity such as mean up to 91%, respectively. Zhao et al. (2020) used a novel-based hybrid sampling ensemble algorithm with Xgboost. These algorithms are used for under-sampling of the majority class of samples. A robust type of classifier is developed by a multi-balanced sample of sets. These samples set strains up a multi-level of differentiated based sub-classifiers. It improves by classifying the minority class.

Proposed XGBoost CNN model for elderly heart patient monitoring

Ensemble learning based feature selection

From the original spaces, the subset of those features are been chosen by feature solution. It can able to remove unrelated and features that are redundant. In this feature, ensemble learning uses three different methods such as a wrapper, filter type, and embedded-based method. The wrapper uses features with a high level of classification with more accuracy, by estimating them in a specified type of learning algorithm. Constraint wrapper, the filter used to evaluate the internal characteristic of the data. It calculates with simple calculation and statistics which are computed from the empirical distribution. Wrappers usually get more achievable results than filters. Only the use of a wrapper can increase the time duration. It preprocesses steps for classification, where the classified approach is the main goal for feature selection. Feature selection features only retain important characteristics. It removes other irrelevant features. Some of the vital parameters of the human body which are featured are heart rate, irregular heartbeat, blood diastole, blood systole, oxygenated blood, movement artifacts, cardiac arrhythmia, excess blood flow, pulse dropping level.

Data preparation

To build an efficient diagnosis for elderly patients, an effective way of measuring vital parameters has to be evaluated. Firstly, the class imbalance problem has to be solved. Here, the clustering-based under-sampling method gives an effective way of solving the issue. Here, the number of clusters in the majority class is been taken which is set to be equal to the number of dataset points in terms of minority-based class. To reduce the error rate in classification, an ensemble learning type algorithm classifies different majority classes, which have a dataset of approximately 99% and a minority class of around 1%. The cluster classification training dataset is used to preprocess each data level approach. These assessments are carried out separately. The resampling approach is utilized for the preprocessing method. These are the resample imbalance training datasets that occur before the model training stage. The data samples are categorized into the same state of clusters using the clustering approach. In general, the k-mean cluster technique is used to group similar datasets and assign a Centroid to each cluster (Aggarwal and Aggarwal 2012). Figure 1 represents the way the feature is been extracted in XGBoost using the extraction technique.

Fig. 1.

Xgboost using feature extraction

Bagging (Dudoit and Fridlyand 2003) is used to reduce overall variance, whereas boosting is used to replace over-weighted data into compact train sets such as tree-based. XGBoost (Devan and NeeluKhare 2020) is an example of a boosting technique that is a regressor, whereas bagging uses a random forest regressor. XGBoost is an ensemble learning technique based on decision trees that employ a gradient boosting framework. It uses an artificial neural network to forecast unstructured input in an organized fashion. Gradient boosting machine learning is another name for XGBoost. It solves larger data sets in a more efficient and timely manner.At the residual state level, the initial state where P₀ is defined will be connected with predicting the target value of the Z model. To fit the level of residual range from the previous way steps, a new state model n₁ is used now. P₀ and n₁ are then merged to produce P₁, a boosted version of P₀.

$$P_0(x)+n_1(x)\to P_1(x)$$ 1

Similarly, in order to improve P₁'s classification and performance, we might modernize after the residual state of P₁ and generate a P₂ value as shown in the below equation.

$$P_1(x)+n_2(x)\to P_2(x)$$ 2

where ‘m' represents iteration until the residual is minimized, the evaluation can be denoted by Eqs. (1) and (2) as follows.

$$P_{m-1}(x)+n_m(x)\to P_m(x)$$ 3

The above equation represents the way of classifying the model using boosting ensemble learning technique. Keep the dataset based on the clusters after the preprocessing processes. Each clustered data set is preserved depending on the amount of data and the sample set. Medical examination data are large and complex, and they must be precise for the doctor's advice to be valid. By comparing the clustered dataset to the cloud-stored dataset, the data is evaluated and supplied as an average value. We can assess the state of the dataset by classifying it based on the characteristic, which is high, medium, low, and very low. According to the sample set, each classified object is separated and subcategorized. This strategy allows the data to be calibrated using a sampling technique. The following procedures must be taken: the model is started with $P_0(X)$, where it minimizes the loss function or Mean Square Error (MSE).

$$P_0(X)=argminv\sum _{i=1}^{n}L(Z_i,v)$$ 4

$$argmin=\sum _{i=1}^{n}L(Z_i,v)=argmin\sum _{i=1}^n{\left(Z_i-v\right)}^2$$ 5

Taking the first type of differential of the above equation with respect to ν, where it is seen that the function minimizes at mean i = 1 in,

$$P_0(X)=\sum _{i=1}^n\frac{Z_i}{n}$$ 6

where P₀(x) gives the prediction from the first stage of our model. Residual error for each instance is represented as Z_i in the above equation and n is the total number of instances. Here it uses k-Nearest Neighbor (KNN) (Soucy and Guy 2001) to substitute the mean impute missing values using a mean value obtained. This approach is evaluated by comparing the clustered dataset to the cloud-stored dataset, and an average of this data is returned as an output. The XGBoost technique is used to enhance the performance of the CNN classifier, eliminate redundant information, enhance the generalization performance, and minimize the computational complexity, Fig. 2 depicts the clustering representation diagram.

Fig. 2.

Clustering representation diagram

XGBoost

The XGBoost model is a collection of decision trees and its prediction is done using the below equation:

$$o_j^{`}=\sum _{n=1}^k{d}_{k}n$$ 7

where the total number of decision trees is represented as n, $d_k$ is the decision tree prediction, and $o_j^{`}$ is the feature vector of the jth data point. We are using a loss function to optimize the model during training and it is computed as shown below

$$L=-\frac{1}{N}\sum _j^N\left(\left(o_{j}log{d}_k\right)+(1-o_j)log(1-d_k)\right)$$ 8

The regularization process of the XGBoost model is demonstrated as shown below:

$$\varphi ={\rho }^L+\frac{1}{2}\zeta \sum _{i=1}^L{S}_i^2$$ 9

where the regularization terms are represented as $\zeta$, $\varphi$, and $\rho$, L is the number of leaves, and the leaf score is represented as $S_i^2$. The objective of the model is explained using the below equation

$$O=L+\varphi$$ 10

In Eq. (10), the L value is used to identify the predictive behavior of the model in terms of training loss, and the model's complexity is controlled using the regularization parameter which also prevents it from overfitting. The objective function is optimized using the mean and variance of the gradient descents.

Convolutional Neural Network (CNN)

The CNN architecture (Zang et al. 2020) consists of the convolutional layer, pooling layer, and fully connected layer. The CNN extracts the hidden features present in the dataset via multiple filters with the help of the pooling and convolutional layers. In the end, the hidden features are integrated via a fully connected layer. Using the convolution operation, the previous layer feature maps are convolved and the output feature map is generated using an activation function. The below equation

$$a_i^k=f\left(\sum _{j\in {\rho }_i}{a}_i^{k-1}\ast {\omega }_{\mathit{ij}}+{\beta }_i^k\right)$$ 11

where a is the ith output feature map of the kth layer, ρ represents the input feature map selection, $\omega$ is the weight value, f(.) is the activation function (Rectified Linear Unit), * represents the convolution operation and $\beta$ is the bias value. To minimize the number of CNN parameters, the pooling layer is incorporated. The pooling operation is computed as follows:

$$a_i^k=f\left(\left(\text{down}\left(a_i^{k-1}\right)\right)+{\beta }_i^k\right)$$ 12

where $a_i^{k-1}$ is the ith input feature map of the (k-1)th layer and down(.) is the max-pooling subsampling function. The final output value is computed by integrating the feature maps obtained from the different convolutional and pooling layers and the process is explained using the below equation:

$$a^k=f\left(W^k{a}_i^{k-1}+{\beta }_i^k\right)$$ 13

The weights between the (k−1)th layer and the kth layer are represented as $W^k$ and $a^k$ is the final output vector. The Output vector is mainly classified into 0 and 1 where 0 represents the sample is normal whereas the abnormality is represented using the value 1.

Proposed XGBoost-CNN for the elderly patient supervision

A total of 72 elderly people in age 65–95 were taken and monitored throughout the day. The elderly patients mainly suffered from blood pressure and heart disease. The biological and behavioral attributes are retrieved from the smart IoT sensors and biomedical sensors. The XGBoost-CNN architecture is in charge of generating the appropriate health deterioration status of the patient. Every person was monitored for a total of 10 days and this process was done with complete informed consent from the caretakers. The input ECG and pulse data are generated in this paper using a PPG and the details about this dataset are presented in the next section. Data are preprocessed by reducing the noise using a Gaussian type filter. Here in this, the data are sampled where it gets categorized based upon the features. The feature gives the precise range of the predicted signal. These predicted signals give accurate results. Raw data which are collected gets combined and classified using a gradient-based approach. Data that are trained and tested using the Convolutional Neural Network gives the validated results. By comparing the dataset which is clustered with the stored dataset in the cloud which evaluates and returns the data back as an average value.

The data is processed in terms of samples. These samples are featured in cloud storage. The cloud gathers up the information in a structured manner where each data is processed and stored. A large amount of data can be gathered up by the cloud where it connects it to the server through the app. The app stores up the personal information of the user report which analytically identifies the record by evaluating the data in a composed manner. The data is collected by the app and stored on the cloud platform. The cloud is made up of data collected by sensors. If a problem is noticed, the IoT platform collects the data and provides it to the caretaker as graphical data. The patients are constantly monitored 24 h a day, seven days a week. Both senior people and bedridden people can be monitored via our technique. Message and voice calls will be sent as a notification to the caretaker through the app as notifications. The overall architecture of the proposed framework is presented in Fig. 3.

Fig. 3.

Architecture diagram

Experimental results and discussion

The proposed XGBoost-CNN model was built using the Matlab programming language and the experiments are conducted in an 11th GENERATION Intel Core i5-1135G7 processor (4.2 GHz Intel Turbo Boost Technology, 8 MB L3 cache, 4 cores). The number of CNN epochs taken is 1000 and an adam optimizer of 0.01 learning rate is used to enhance the training process. The efficiency of the proposed XGBoost CNN classifier is evaluated with different state-of-art techniques such as Support Vector Machine(SVM), decision tree, and random forest.

Dataset description

The dataset for this work is taken by using the device known as PPG (Photoplethysmography)which is an inexpensive method that can be used for monitoring heart conditions. A non-invasive platform that uses a light source is utilized here and the photodetector is used at the surface of the skin which measures the volumetric variations of circulation in blood. PPG consists of infrared light to calculate the volumetric range of variations of blood changes in the microvascular tissue. To detect the volume of blood changes the microvascular tissue of bed in the skin is used. An Optical based Plethysmograph is used to detect the blood volume changes in the human body. To monitor the perfusion of blood also PPG can be used.

By the emission of light-emitting diode (LED) by the transmission and reflection of the light through the skin, the value can be detected. Each type of time cycle appears to be at the cardiac peak. PPG can able to monitor breathing and another type of circulatory conditions. PPG waveform varies every time according to the location and its manner. Even from the transmission absorption present in the fingerprint and through forehead reflection the data can be evaluated.Noises are the unwanted disturbance in the signal which causes signal variation. To avoid noise, the signals are preprocessed using a Gaussian noise filter.The crucial parameters measured by the PPG device areHeart Rate, Cardiac Arrhythmia, Arterial Oxygen Saturation, Arterial Pressure, Diastolic Blood Pressure Systolic Blood Pressure, Movement Artifacts,and Pulse Dropping. These parameters are attained by using the PPG device by removing the noise using the Gaussian noise filter. The PPG variation analysis is presented in Fig. 4.

Fig. 4.

PPG variation analysis

PPG is pulse oximetry in which arterial oxygen saturation is calculated. It gives the oxygen saturation level and predicts the blood volume if any change. By using arterial blood flow and venous blood flow the oxygen level can be determined Blood pressure is one of the predominant parameters in the human body where it gives the right amount of information for physicians about a patient’s health. Hypotension and hypertension are two different categories in blood pressure. Blood pressure indicates the resistance of the human body. Blood vessels give the accurate movement of blood. The amount depends upon the various function of the heart with its vascular type features. The thick wall of the heart vessel and its elasticity depends. Over 1.4 billion people worldwide suffer from hypertension. When the pressure bounds upper and lower then it is called systolic blood pressure and diastolic blood pressure.

The blood pressure readings can be evaluated by using systole and diastole. The maximum range of the heart while beating is called systolic pressure. The bottom layer where the amount of arteries gets beaten is called diastolic pressure. Blood pressure gets recorded in two different numbers. Both these two parameters are important. If the readings are high, hypertension is present. If things are less then it is due to insufficient blood flow to critical organs. The ventricular type of relaxation is termed diastole. The pressure in the blood during diastole is called diastolic blood pressure. If the blood pressure is low, then it can cause dizziness, organ failure, and lightheadedness. One of the common types of condition which produces systolic hypotension is orthostatic hypotension. The different blood pressure peaks are presented in Fig. 5 and the diastolic and systolic measurements of blood pressure are presented in Table 1. The decline in the health condition of the elderly can be identified by their pulse rate and it is presented in Table 2.

Fig. 5.

Systolic, diastolic, and dicrotic peaks in blood pressure

Table 1.

Diastolic and systolic measurements of blood pressure

Blood pressure	Diastolic (mm hg)	Systolic (mm hg)
Normal	< 80less than 80	< 120less than 120
Elevated upper	< 80less than 80	120–129
Hypertension (stage 1)	80–89	130–139
Sshypertension (stage 2)	90 or higher	140 or higher
Hypertensive cases High blood pressure	> 120 Higher than 120	> 180 Higher than 180

Table 2.

Examining the health deterioration of the elderly based on the pulse rate

Pulse rate	Condition
0–55	Critical
56–100	Normal State
101–200	Normal State
200–250	Critically high

From the above ranges, the experimental way of validating the result is by determining a range to identify the statistical approach of the vital parameters in the human body. The vital signs that are taken as the important features are presented in Table 3.

Table 3.

Vital Parameters present in the dataset

Vital parameters	Low	Medium	High	Abnormality detection
Heart rate	80–136 bpm	90–153 bpm	100–170 bpm	If the heart rate is 80- 136 bpm, then the beat per minute is less than the normal range which indicates it is weak
Irregular heartbeat/Cardiac Arrhythmia	95 bpm	80 bpm	60–65 bpm	If the heartbeat is irregular in condition, then cardiac arrhythmia occurs
Arterial oxygen saturation	< 90%	91–94%	> 95%	If the oxygen level is less than 90% then the saturation level is less
Hemoglobin—Male	13.5 g/Dl	15.5 g/Dl	17.5 g/Dl	If the condition is in a normal state, then the hemoglobin is regular in the state
Hemoglobin—female	12 g/Dl	15.5 g/Dl	12.5 g/Dl	If the condition is in the normal state, then the hemoglobin is regular in the state
Blood flow dynamics	2.80 mm	2.83 mm	2.88 mm	If the blood flow vessel dynamics ahs the range of 2.80 mm then blood flow is less
Skin surface temperature	36.5–37.5 °C	> 37.5 °C	> 40 °C	If the skin temperature has a high heat temperature of more than 40 °C
Arterial pressure	< 80	< 120	> 80	If the blood pressure gets high, then arterial pressure causes hypertension and hypotension
Diastolic blood pressure	< 80	90 or higher	> 120	If the blood pressure ranges greater than 120, then it is Diastolic Blood Pressure
Systolic blood pressure	< 120	140 or higher	> 180	If the blood pressure ranges greater than 180, then it is Diastolic Blood Pressure

The threshold range for the alarm and the voice message to the caretaker obtains the vital parameters range from Table 3. The features are extracted from the statistical data which is later compared with Table 3. Each parameter has a threshold range as depicted in Table 4 and if any critical condition occurs the caretakers are immediately notified.

Table 4.

Threshold range

Rhythm range	Vital parameters
Skin temperature	Temperature ≥ 40 °C
Heart rate(HR)	80–136 bpm (beats per minute) 80 $\le \mathit{H}\mathit{R}\le 136$
Blood pressure	BP ≥ 120 mm Hg
Systolic blood pressure	BP ≥ 180 mm Hg
Diastolic blood pressure	BP ≥ 120 mm Hg
Blood flow volume	V ≤ 2.80 mm
Hemoglobin—male	HB ≤ 13 g/ Dl
Hemoglobin−female	HB ≤ 12 g/ Dl
Arterial oxygen saturation	SpO2 < 90%

Performance metrics

The performance metrics used to verify the efficiency of the proposed XGBoost-CNN are accuracy, precision, recall, F-Score, and execution time. The description of each performance metric is presented below:

Accuracy: It predicts the likelihood of appropriately predicted training samples and it is computed as shown in the below equation

$$A=\frac{True\_pos+True\_neg}{True\_pos+True\_neg+False\_pos+False\_neg}$$ 14

where True_pos, True_neg, False_pos, and False_neg represent the true positive, true negative, false positive, and false negative values respectively. True positive represents the ability of the classifier to correctly predict the abnormal samples as abnormal and true negative represents the capability of the classifier to accurately classify the normal instances as normal. The false-negative indicates the classification of abnormal samples as normal and the false-positive indicates the classification of normal instances as abnormal.

Precision: It is also known as Positive predictive value which computes the ratio of accurately identified instances to the overall abnormal samples.

$$P=\frac{True\_pos}{True\_pos+False\_pos}$$ 15

Recall (R): The proportion of correctly classified instances to the total number of abnormal instances present in the test dataset.

$$R=\frac{True\_pos}{True\_pos+False\_neg}$$ 16

Execution time: The time taken by the classifier to predict the changes in the elder's health condition.

F1-Score: The harmonic mean of precision and recall is called an F-score and it is computed using the below equation.

$$F1\_Score=\frac{2\times P\times R}{P+R}$$ 17

Recognition rate: Recognition rate is the total number of accurately identified instances divided by the total number of instances.

Performance evaluation

Using the k-fold cross-validation, the proposed methodology is evaluated in terms of 5, 10, 15, 20, and 25 folds. For training, the k-onefold is taken and the remaining is allocated for testing. The CNN model has a high number of hidden layers which is utilized to improve the classification accuracy and exhibit strong learning capability. Figures 6 and 7 present the classification accuracy obtained by the XGBoost classifier for a population size of 3000 and 5000. The results show that the increase in population size improves the classifier's classification accuracy. The k-fold cross-validation is applied to the medical dataset and it is partitioned into k equal folds. The cross-validation technique measures accuracy by randomly scattering the input samples into k different folds.

Fig. 6.

Accuracy obtained for a population size of 3000

Fig. 7.

Accuracy obtained for a population size of 5000

The comparison results were obtained by comparing the proposed methodology in terms of accuracy, precision, recognition rate, recall, F1-Score, and execution time as presented in Figs. 7, 8, 9, 10, 11, 12 and 13 in terms of different cross folds. The proposed XGBoost-CNN classifier offers higher accuracy of 98% when compared to the NB, SVM, and DT tree classifiers as illustrated in Fig. 8. The precision of the proposed methodology is 98% which is higher than the other techniques as shown in Fig. 9. The precision value of DT is 92%, SVM is 84%, and RF is 83%.

Fig. 8.

Comparison results in terms of accuracy

Fig. 9.

Comparison results in terms of precision

Fig. 10.

Recognition rate comparison

Fig. 11.

Recall comparison

Fig. 12.

Comparative analysis using F-Score

Fig. 13.

Comparative analysis using execution time

The recognition rate results obtained for comparing the proposed methodology with different classifiers are presented in Fig. 10 concerning different iterations. The proposed methodology gives the recognition rate of 0.93 at the 3500 iterations. The RT, CT, and SVM give a recognition rate of 0.86, 0.82, and 0.78 respectively.

The recall score of the proposed methodology reached 100% in the 20th fold as shown in Fig. 11. It is relatively high when compared to the SVM, DT, and RF classifier Whose recall score is less than 94% in the test dataset. The proposed methodology also attains a higher F1-Score of 94.8% when compared to the other techniques as shown in Fig. 12. The RF and SVM classifier are at the next two places with an F1-score of 90 and 87%. The DT classifier offers the lowest score of 83.4% when compared to the remaining techniques.

The comparative analysis of the proposed method conducted in terms of execution time is presented in Fig. 13. The XGBoost with CNN offers the lowest execution time of 14 ms when compared to the RF, SVM, and DT classifiers on the medical dataset created using PPG. The XGBoost technique used minimizes the time consumption of CNN and yields the results in a low time when compared to the other techniques. The training and development (testing) loss value for both the CNN and XGBoost model is presented in Figs. 14 and 15 which represents that the training of the model was not subjected to overfitting.

Fig. 14.

Loss curve for CNN

Fig. 15.

Loss curve for XGBoost

Discussion

In this paper, we focused on analyzing the vital parameters of the human body by using noninvasive techniques. To avoid deterioration in the health condition of the human body, prior to the intervention by constant monitoring of personal health. In this paper, a deep convolutional neural approach is used where the cardiac dysrhythmia/irregular heart rate is classified with various multilevel parameters. The CNN-based supervised learning model is tested and simulated on a real dataset. Oversampling of the data causes an improper dataset. To overcome the improper dataset, the Xg boost technique is used. Improving life quality by determining health conditions can increase life expectancy. A linear type characteristic gives the statistical graph in the app analysis. The analytical data of the complete vital parameters can be checked by the caretaker. The caretaker can easily check the condition of the patients without any further complications. Caretaker gets message report of the patient’s health condition by statistical data. The main purpose of getting physiological data is to check the vital parameters.

To reduce imbalanced data, use the ensemble method which uses the XGBoost technique along with the easy ensemble method. This can be used in real life and predicts in high sensitivity ensemble learning has more accuracy in classification and ability than single type classifier. A supervised type of learning is used for the regression and classification dataset. The mean error rate computed in terms of taking the reference Peripheral capillary oxygen saturation (SpO2) and blood pressure with a pulse oximeter (BpO2) is presented in Table 5. The error percentage of arterial oxygen estimation based on our observed result is nearly in the range of 0.01–1.11%. The MSE also lies between 0.01 and 0.10% respectively. If the pulse rate is critically high, then the patient’s illness is high in the state. The mean error rate calculated is in the range of 0.1%. The graphical representation of the patient's normal and irregular heart rate is presented in Fig. 16a and b.

Table 5.

Arterial oxygen estimation results in using mean error rate

Bp O2 value	Reference Peripheral capillary oxygen saturation (SpO2)	Mean error rate (%)
95.5	98	0.39
93.6	98	0.19
94.5	98	0.29
96.8	98	0.43

Fig. 16.

Patient's normal and irregular heartbeat signal. a Normal heart rate and b Irregular heart rate (Heart attack)

The results obtained by the different classifiers such as NB, DT, and SVM are satisfactory in terms of accuracy, precision, recall, and F1-score. In terms of execution time, the different classifiers consumed higher time when compared to the proposed methodology. The reason for selecting the XGBoost technique is it is usually ten times faster than other techniques and offers the advantage of parallel processing. The regularization aspect of the XGBoost algorithm prevents the model from overfitting.

Conclusion

This study presents an XGBoost CNN classifier for detecting the vital parameters of elderly patients in a precise manner. An XGBoost technique is used to implement the feature extraction phase. Statistical data are analyzed by the consecutive dataset therefore by using this, caretaker work gets simpler. The report is analyzed and stored in the cloud and then the values are compared with the dataset. The caretaker is alerted if any abnormalities occur. The data imbalance and overfitting issue associated with the CNN is solved with the help of the XGBoost technique. By using this, the accuracy level of the parameters can be attained up to 98%. The efficiency of the proposed work is evaluated by comparing it with different classifiers such as NB, SVM, and DT in terms of precision, accuracy, recognition rate, execution time, etc. The proposed methodology gives a recall value of 100%, F1-Score of 94.8%, precision of 98%, and accuracy of 98% which is relatively higher than the existing approaches such as decision tree, random forest, and Support Vector Machine. The CNN modeling structure can be enhanced because the XGBoost technique is more dependable and optimized. In the future, this technology can be used in real-time and evaluated with a variety of additional features and optimization algorithms.

Funding

Not applicable.

Availability of data and material

Data sharing is not applicable to this article as no new data were created or analyzed in this study.

Code availability

Not applicable.

Declarations

Conflict of interest

The authors declare that they have no conflict of interest.

Consent to participate

Not applicable.

Consent for publication

 Not applicable.

Human and Animal Rights

This article does not contain any studies with human or animal subjects performed by any of the authors.

Informed Consent

Informed consent was obtained from all individual participants included in the study.