An IoT and Fuzzy aware e-Healthcare system using feature optimization tuned T-CNN with high dimensional data

Abstract

Many real-time applications, including some emerging ones, rely on high-dimensional feature datasets. For simplifying the high-dimensional data, the various models are available by using the different feature optimization techniques, clustering and classification techniques. Even though the high-dimensional data is not handled effectively due to the increase in the number of features and the huge volume of data availability. In particular, the high-dimensional medical data needs to be handled effectively to predict diseases quickly. For this purpose, we propose a new Internet of Things and Fuzzy-aware e-healthcare system for predicting various diseases such as heart, diabetes, and cancer diseases effectively. The proposed system uses a newly proposed Intelligent Mahalanobis distance aware Fuzzy Weighted K-Means Clustering Algorithm (IMFWKCA) for grouping the high dimensional data and also applies a newly proposed Moth-Flame Optimization Tuned Temporal Convolutional Neural Network (MFO-TCNN) for predicting the diseases effectively. The experiments have been done by using the UCI Repository Machine Learning datasets and live streaming patient records for evaluating the proposed e-healthcare system and have proved as better than others by achieving better performance in terms of precision, recall, f-measure, and prediction accuracy.

Keywords

Feature optimization clustering e-healthcare system high dimensional data internet of things

1 Introduction

Today, the Internet of things (IoT) is widely applied in various emerging fields, including medical, in the world. More than 25 billion smart devices will be connected by the year 2020. When connected, the huge volume of devices may produce an enormous volume of data that is extracted by using machine learning algorithms. Even though the huge volume of available data is dominating the computational models that are available in the literature and taking more time due to the data movement and the cost difference between memory and processing. For example, the clustering technique that is identified as a well-known unsupervised learning is helpful for extracting the data according to the similarity between the available data. Many solutions are available in the cloud for handling this task and also transmitting the huge volume of data with congestion in the communication paths. These kinds of congestion problems may affect the applications in terms of maintaining privacy preservation and security. To handle these issues smoothly, the feature selection and optimization techniques are incorporated to reduce the data weights.

The role of feature selection is increasing drastically in data analysis due to the rapid growth of data volume in various fields. Data mining, machine learning, and deep learning techniques are widely used for analysing data by applying and taking into account necessary features through dimensionality reduction. For identifying the most useful features, various methods are available such as feature selection, optimization, clustering, and classification. Specifically, the high-dimensional data sets are to be filtered out of the relevant and irrelevant features to reduce the dimensionality. The various feature selection methods are available under the three categories such as filter approach, wrapper approach, and embedded approach [1]. Among them, the filter approach ranks the available features by considering the scores of each attribute. The threshold for making the selection process as relevant and irrelevant features is fixed. All the filter methods are discussed in a survey conducted by Lazar et al. [2]. Next, the wrapper methods [3] are applied for preparing a subset that contains the more relevant features. The subsets are evaluated by calculating scores. Here, the recursive feature elimination process is incorporated like greedy search to identify the relevant features and also eliminate the irrelevant features for performing the feature selection process. The embedded methods include the fitting process that incorporates the Lasso regression [4] and gradient boosting [5] methods.

The clustering method plays a major role in resolving the issues by applying data mining and machine learning techniques. The objective of the clustering method is to form groups of objects into clusters that have high similarity and also neglect dissimilarities. Even with the availability of a large volume of data analysis techniques in the literature, a lot of issues are not resolved on high-dimensional data sets [6] due to the availability of more irrelevant, redundant, and noisy data. The clustering technique is used to categorize and group the relevant features together with the consideration of relevancy scores of features. While handling the different types of data parallelly, the time complexity is also increased and it is also not able to accommodate the data on the available memory spaces. It is necessary to propose a new intelligent clustering method for handling the different dimensionality of data effectively and helpful for making effective decisions. Many clustering algorithms have been proposed by various researchers under categories such as hierarchical, density, and prototype clustering. Even though, these standard clustering algorithms performed well on low dimensional data, they did not achieve enough performance on high dimensional data. For improving the performance on high dimensional data, the deep learning technique is incorporated into the clustering process. This work introduces a new deep clustering technique for resolving the high dimensionality problems and achieving better performance.

The major contributions to this work are as follows:

To propose a new IoT and Fuzzy-aware e-healthcare system for predicting diseases effectively by analyzing high-dimensional medical records.

To propose a new Intelligent Mahalanobis distance based Fuzzy Weighted K-Means Clustering Algorithm (IMFWKCA) algorithm for grouping the relevant features and reducing the dimensionality of the input dataset.

To develop a new Moth-Flame feature optimization tuned Temporal Convolutional Neural Network (MFO-TCNN) technique for making final decisions on patient records with the help of optimized features

To achieve better performance in terms of good precision, recall, f1-score and prediction accuracy with reasonable time taken for the decision making process.

The rest of this research article is organized with four more sections, including Literature Survey, System Architecture, Proposed Model, Result and Discussion, Conclusion, and Future Works. The literature survey section conducts a brief survey highlighting the benefits and drawbacks of clustering, feature selection, deep learning, medical field, and high-dimensional data. The system architecture section describes the workflow of the proposed monitoring system. The proposed model demonstrates the necessary background information and the newly proposed methodology in this work. The experimental results and the explanation are described in the results section. The work is concluded by highlighting the contributions and achievements quantitatively.

2 Related works

Much research work has been done in the direction of handling high-dimensional data by applying various newly proposed clustering and optimization methods [7 –10]. Among them, Thanh et al. [11] developed the density and K-Nearest Neighbor with Kernel Aware Clustering method for handling high-dimensional data. The KNN with a kernel density scheme is helpful for making the possible clusters according to the various densities over the high-dimensional data. Their cluster method was tested with an image dataset and also achieved better performance than other methods. Bouveyron et al. [12] proposed and implemented a new data clustering method that is capable of handling high-dimensional data and also measures the particular subspace and the dimension of intrinsic. Finally, they have conducted experiments with live high-dimensional datasets and proved that they are better than other clustering techniques available in the literature. Jing et al. [13] presented a k-means clustering method to group the high-dimensional data into sub-spaces. The k-means clustering method is used to apply the values for calculating the weight for every dimension in an individual cluster and also apply the values for identifying the suitable subsets. They have achieved better performance than other methods in terms of prediction accuracy. Wu and Gu [14] presented a new clustering scheme to categorize the high-dimensional data. Their method applies supremum and infimum to represent the clusters according to the schema and also assigns the suitable non-sample objects for effective clustering.

Jiang et al. [15] developed a PL-Tree aware clustering method for handling high-dimensional data. Chen et al. [16] proposed a new feature grouping aware K-means clustering method to gather the relevant featured data in a high-dimensional dataset. They have incorporated two different kinds of weight assignment procedures for identifying the relevant features in every group. In addition, they have incorporated a few additional steps for calculating the subspace weights. Finally, they have proved that their clustering method is better than others in terms of prediction and relevancy accuracy.

Wang et al. [17] developed a new soft subspace-aware clustering technique that incorporates the partition index, subspace clustering, and fuzzy clustering. Moreover, they have built a convergence theorem for analyzing the viewpoints. Finally, the experiments have been conducted by using sparse and noisy text data from high dimensional datasets. Muja and Lowe [18] developed a new method to match the binary features by searching multiple clustering trees and also proposed a distributed nearest neighbor matching framework for handling the high-dimensional data.

Radhika et al. [19] developed a new document subspace clustering method for handling high-dimensional data which is capable of handling the high-dimensional data effectively by obtaining good prediction accuracy and clustering process with less processing time. Tao et al. [20] developed a new k-means clustering method that incorporates the intelligence and weights for performing the searching process globally and also finds the centre point for each cluster. The assigned weights are helpful for making effective decisions on high-dimensional data and have also proved better than other methods in terms of prediction accuracy.

Wang et al. [21] developed a new fast adaptive k-means subspace clustering method for providing a flexible clustering process for high-dimensional datasets. Moreover, it is capable of selecting the most useful features for performing an effective clustering process. Finally, they have proved that their method is better than other clustering methods in terms of prediction accuracy. Tapia et al. [22] presented a prediction methodology that was developed for evaluating the success of impatient on high dimensional data and also aimed to learn new combinations between the individual features. They evaluated various feature selection methods and classifiers in order to design the best prediction method with the highest accuracy. In their analysis, they achieved an accuracy of 82% using the best first method. Sui et al. [23] developed a new sparse representation aware DSC method called the evolutionary dynamic sparse subspace clustering approach that consists of two phases, namely static learning and a live clustering process. They have considered the average sparsity concentration index and also compared it with the sparsity concentration index. Finally, they proved that their method is better than other methods in terms of relevancy prediction accuracy and grouping process.

Guan et al. [24] developed a new optimization technique called the history-guided differential evolution method for performing effective grouping and also enhancing the feature selection process on various synthetic datasets that are available as high-dimensional data. Their technique achieved better performance than other methods that are considered to have other sets of features for forming a group. Lei et al. [25] developed a new weighted density aware clustering and analytic hierarchy process to identify a suitable method in the medical field. The hierarchical cluster analysis is incorporated with the newly proposed clustering method for enhancing the clustering performance in terms of a better service provider selection process. All the available methods are not achieving good performance in terms of prediction accuracy on patient records that are available as high dimensional data. To fulfil the current requirement, an IoT-aware e-Healthcare system is proposed in this work.

3 System architecture

The workflow of the proposed e-healthcare system is demonstrated in Fig. 1 that consists of nine important components such as IoT device, Sliding Window, UCI Machine Learning Repository, decision manager, feature optimizer, clustering module, classification module, rule base and temporal agent. The workflow of the proposed e-healthcare system is demonstrated in Fig. 1, which consists of nine important components such as IoT device, Sliding Window, UCI Machine Learning Repository, decision manager, feature optimizer, clustering module, classification module, rule base, and temporal agent.

Fig. 1

Proposed system architecture.

The IoT device is used to collect the necessary live data from the patients in various locations. The collected live streaming data is to be forwarded to the decision manager through a sliding window process, and it applies a feature optimizer to optimize the patient records with the most contributed features. The decision manager analyses the necessary streaming data from the sliding window using the feature optimizer and it incorporates it with the temporal convolutional neural networks for performing hyper tuning and better decision making.

The decision manager applies the clustering algorithm to group the sub datasets that are identified as better subsets with optimized features. Finally, the decision manager decides the patient records whether they are affected by diseases or not. Moreover, it uses the rules that are stored in the rule base for categorizing the disease types and the levels according to the values contained in the important features of the datasets.

4 Proposed work

In this work, a new IoT and Fuzzy-aware e-healthcare system is proposed to predict whether the patient record is affected by diseases such as heart disease, diabetes disease, and cancer disease along with their severity levels effectively. This e-healthcare system works with the newly proposed Intelligent Mahalanobis distance aware Fuzzy Weighted K-Means Clustering Algorithm (IMFWKCA) to group the high dimensional data and a newly developed Moth Flame Optimization Tuned Temporal Convolutional Neural Network (MFO-TCNN) for predicting diseases effectively. In this work, the high-dimensional data is grouped by applying the proposed clustering algorithm according to the differences between the records and reducing the dimensionality as well. Here, the high-dimensional data is subjected to the dimensionality reduction process and grouped into different subgroups. Then, the grouped dataset is forwarded to perform the classification by using the newly proposed feature optimization tuned temporal CNN. Here, the existing MFO algorithm is used for tuning the features with adjusted weights to the Temporal CNN to make a final decision on medical records. This section explains the data clustering process, feature optimization and classification. First, the data clustering is explained by demonstrating the workflow of the proposed IMFWKCA.

4.1 Clustering

In this work, a new enhanced version of the Intelligent Mahalanobis distance-aware Weighted Fuzzy K-Means Clustering Algorithm (IMWFKMCA) is proposed for effectively grouping the relevant data in the high-dimensional dataset. Moreover, an intelligent agent is used for performing the weight assignment process.

4.1.1 K-Means clustering

A group of objects are used in the standard K-means clustering that represents the quantitative value of each input feature as matrix X = (x_hu) where x_hu indicates that the feature value at the specific entity I mean that a group of entities (M), U represents the group of features (L). This method is capable of producing a subgroup SG ={ SG₁, . . . . , SG_k } of H in the K number of non-empty and non-overlapping subsets that represent as SG_k ⊂ I, clusters, the centroid value is pointed out using the variable c_k and the vector is represented with N-dimensional for the feature space $Fk = \bar{1, FK}$ . Moreover, the initial centroid value and the K-means change frequently based on the distance between the centroids and are also identified as different gravity centers. In addition, this algorithm is also used to identify an alternate for minimizing the square error as an alternate that is represented as given in Equation (1). $WT (SG, cp) = \sum_{k = 1}^{k} \sum_{i \in S G_{k}} \sum_{u = 1}^{M} {(x_{hu} - c l_{ku})}^{2}$ (1)

4.1.2 Intelligent agent based k-means clustering method

This subsection describes in detail the intelligent agent-based K-Means clustering method by applying a pattern used in the existing work [26] for identifying the initial cluster (SG) and also finding the centroid value (cp) by minimizing the alternatively represented as given in equation (2). $WT (SG, cp) = \sum_{h \in SG} dst (h, cp) + \sum_{h \in SG} dst (h, 0) .$ (2)

Here, the SG is identified as a subgroup dataset that is removed from the whole dataset and identified as a new subgroup in which the irrelevant records are removed. Finally, the single records are removed, and the remaining centroids are K-means initialized.

4.1.3 Weighted fuzzy clustering

The standard K-Means clustering is enhanced further with the consideration of fuzzy logic and weights. The fuzzy clustering process works based on the membership function in the process of assigning the centroid value. The improvement of the fuzzy clustering is the incorporation of the weight assignment process. The Minkowski Weighted K-Means clustering algorithm was developed by Amorim and Mirkin [26]. They have enhanced equation (1) with the consideration of unknown weights that are to be assigned for the features of input records $u, V = \bar{1, M} 1$ 1: $\begin{matrix} {WT}_{β} (SG, cp, wt) = \sum_{k = 1}^{K} \sum_{h \int m} \sum_{u = 1}^{M} S G_{hk} | w t_{u} x_{hu} - \\ w t_{u} c p_{ku} |^{β} = \sum_{k = 1}^{K} \sum_{h \int S G_{k}} ds t^{β} (x_{h}^{'}, {cp}_{k}^{'}) \end{matrix}$ (3)

Where, the weights are guessed as non-negative value and cumulative value is represented as wt_u ⩾ 0, ∑_u∈MWT_u = 1. The variable β is represent the exponential value and finalized by the supervised method or user.

The right-side expression represented by β power Mahalanobis metric between the points that are rescaled the entity $x_{h}^{'} = (w t_{u} x_{hu})$ and the centroid points ${cp}_{k}^{'} = (w t_{u} c p_{ku})$ in the similar feature. Therefore, the criterion defined in equation (3) leads to the criterion defined in equation (1) by finding the suitable weights for the features. Thus, equation (3) represents the minimization of the Mahalanobis metric. The Weighted K-Means Clustering is a minimization method involving iterations on three different subgroups or clusters.

4.1.4 Proposed Intelligent Mahalanobis distance aware Fuzzy Weighted K-Means Clustering

The fuzzy clustering process is represented by the necessary fuzzy membership function sg_k = (sg_hk) , h ∈ H, in which sg_hk (0 ⩽ sg_ik ⩽ 1) indicates the degree of membership of the entity H for the k number of clusters. Let assumed that the condition ∑_kSG_hk = 1. The extra changes are used for computing the feature weights to every cluster separately. This work is done according to the work [26] for bringing a great flexibility to the method. Thus, the variable wt_uk is used instead of wt_u with wt_uk being the feature weights u (u = 1 . . . M) for the number of clusters as k (k = 1 . . . K). Therefore, the next criterion is finetuned and presented in the equation (4): $\begin{matrix} {WT}_{F β} (SG, cp, wt) = \sum_{k = 1}^{k} \sum_{h \int H} \sum_{u = 1}^{M} {SG}_{hk}^{α} {wt}_{uk}^{β} \\ {| x_{hu} - C_{ku} |}^{β} \end{matrix}$ (4)

The Mahalanobis distance measurement formula is given in equation (5), which is used to calculate the distance between the records that are available in each subgroup. $d (\vec{x}, \vec{y}) = \sqrt{{(\vec{x} - \vec{y})}^{T} S^{- 1} (\vec{x} - \vec{y})}$ (5)

Generally, the centroid values are calculated based on the differences between the records that are available in each cluster or subgroup.

Intelligent Mahalanobis distance aware Fuzzy Weighted K-Means method

The steps of the proposed Mahalanobis Fuzzy Weighted K-means Clustering Algorithm are presented as below:

1. For the given centroid values cp_k and the weights wt_u = (wt_uk), update the cluster based on the formula given in equation (6) that is derived from the equation (4) with the constraints by ∑_kSG_hk = 1: $S G_{hk} = \frac{1}{\sum_{t = 1}^{k} {[\frac{ds t^{β} (x_{h}^{'}, {cp}_{k}^{'})}{ds t^{β} (x_{h}^{'}, {cp}_{t}^{'})}]}^{\frac{1}{α - 1}}}$ (6)

The distance between the points $x_{h}^{'}$ and ${cp}_{k}^{'}$ is computed in β power Mahalanobis distance between the points that are rescaled entities as $x_{h}^{'} = (w t_{uk} x_{hu})$ and the centroid points $x_{h}^{'} = (w t_{uk} c p_{ku})$ : $ds t^{β} (x_{h}^{'}, {cp}_{k}^{'}) = \sum_{u = 1}^{M} {SG}_{hk}^{β} {| x_{hu} - c p_{ku} |}^{β}$ (7)

The positive values are included to every item in the sum for avoiding the division by zero and it is also assumed as α > 1.

2. The fuzzy clusters as sub groups SG_k = (SG_hk) and the necessary weights wt_u = (wt_uk) , change the centroid point value cp_k = (cp_hk) of the every sub group SG_k as its Mahalanobis center point so that, at each u, cp_ku is a value of reducing β value using Mahalanobis distance to the groups objects that are represented as given in equation (8). $dst (cp) = \sum_{h \in H} {SG}_{hk}^{α} {wt}_{uk}^{β} {| x_{hu} - cp |}^{β} \to \min$ (8)

For finding the Mahalanobis center point, the steepest descent method is used for computation process according to the work [26].

3. The subgroups SG_k = (SG_hk) and the center point values cp_k = (cp_hk), change the weights based on the formula given in equation (4) constrained by $\sum_{u = 1}^{M} W T_{uk} = 1$ : $w t_{uk} = \frac{1}{\sum_{t = 1}^{M} {[\frac{\sum_{h \int H} {Sg}_{hk}^{α} {| x_{hu} - c p_{ku} |}^{β}}{\sum_{h \int H} {Sg}_{hk}^{α} {| x_{ht} - c p_{kt} |}^{β}}]}^{\frac{1}{β - 1}}}$ (9)

Here, the positive value is considered for every item in the sum and it also assumed as β ≠ 1 for avoiding the division by 0.

4.2 Feature optimization

Feature optimization is an essential task for enhancing the performance of the prediction process in a short span of time. Generally, the moths are identified as insects that are similar to butterflies, and they have two milestones: as adults and as larvae. Moreover, the larvae are converted into moths. They have evolved to fly at night by applying the moonlight. They have utilized the transverse orientation that is helpful for performing the navigation process. The moths are flying with fixed angle with respect to moon and also be an effective method to travel long distance in a straight path [27]. Suppose the moths are changed by artificial light and also demonstrate their behaviors. The reason for the inefficiency of the transverse orientation process that is used to move in straight line when the light is so far away. When moths see the artificial light, they have tried to maintain a same angle with the light for flying in straight line. Since the artificial light is very closed when it is compared to the moon even though it is maintaining a same angle to the light source that causes a deadly spiral fly path for moths [27]. The existing Moth-flame Optimization Algorithm is explained in the next subsection.

4.2.1 MFO algorithm

The MFO method [28] assumes that the moths are candidate solutions, and the variables are also identified as the positions of the moths. Here, the moths are flying in different dimensions, such as single dimensional, two-dimensional, three-dimensional, and hyper-dimensional by changing their positions. Since the MFOA [28] is a population-aware method, it represents the group of moths in a matrix as below: $MF = [\begin{matrix} m f_{1, 1} & m f_{1, 2} \dots \dots & m f_{1, d} \\ m f_{2, 1} & m f_{2, 2} \dots \dots & m f_{2, d} \\ ⋮ & ⋮ & ⋮ \\ m f_{n, 1} & m f_{n, 2} \dots \dots & m f_{n, d} \end{matrix}]$ (10)

Where the variable n represents the number of moths and the variable d indicates the number of dimensions,

Here, an array is prepared for representing moths for storing the respective fitness values and is presented as a matrix as below: $OMF = [\begin{matrix} OM F_{1} \\ OM F_{2} \\ \begin{matrix} ⋮ \\ OM F_{n} \end{matrix} \end{matrix}]$ (11)

Where the variable n indicates the number of moths, the fitness value is the return value of each moth. The key components of the MFOA are flames and considered as a matrix that is similar to the moth matrix as below: $MF = [\begin{matrix} M F_{1, 1} & M F_{1, 2} \dots \dots & M F_{1, dst} \\ M F_{2, 1} & M F_{2, 2} \dots \dots & M F_{2, dst} \\ ⋮ & ⋮ & ⋮ \\ M F_{n, 1} & m f_{n, 2} \dots \dots & M F_{n, dst} \end{matrix}]$ (12)

Where the variable n indicates the number of moths and the variable d represents the number of dimensions, the equation (12) represents the dimension of the variable MF and OFM arrays are equal. For the moth flames, it is assumed that an array is used to store the respective fitness values as below: $OFM = [\begin{matrix} OF M_{1} \\ OF M_{2} \\ \begin{matrix} ⋮ \\ OF M_{n} \end{matrix} \end{matrix}]$ (13)

Where the variable n indicates the number of moths, the MFO method consists of 3 tuples that approximate the global optimal of the optimization issues and are defined as below: $MFO = (J, Q, S)$ (14)

The variable J indicates the method which creates a random population of moths and their respective fitness values. The mathematical model of this function is as follows: $J : θ \to {M, OM}$ (15)

Here, Q is a function that moves the moths around the search area, and it receives the matrix (M) and also returns the updated value. $Q : M \to M$ (16)

Here, S is a function that returns a true value if the termination criterion is fulfilled, and it returns false when the termination condition is not fulfilled. $S : N \to {true, false}$ (17)

With the variables J, Q, and S, the MFO framework, which is defined as the function J being assigned to M. Then, if the S (M) is false, then the P (M) is assigned to M. Moreover, the function J creates the starting level solutions and also computes the objective values. Any random value is also to be applied in this method that is represented as below:

For I = 1 to n do

For j = 1 to d do

MOTH(i,j)=(UBND(i)-LBND(i))

* RND()+LBND(i);

End

OM = FF(M);

Here, there are two arrays such as UBND and LBND, and the relevant matrices are defined by the upper bound and lower bound of the variables as below: $\begin{matrix} UBND = [UBN D_{1}, UBN D_{2}, \\ UBN D_{3}, . . . ., UBN D_{n - 1}, UBN D_{n}] \end{matrix}$ (18) Where, the variable UBND_i represents the lower bound of the i^th variable. $\begin{matrix} LBND = [LBN D_{1}, LBN D_{2}, \\ LBN D_{3}, . . . ., LBN D_{n - 1}, LBN D_{n}] \end{matrix}$ (19)

Where, the variable LBND_i represents the lower bound of the i^th variable.

After performing the initialization process, the function P is executed iteratively until the function T returns the value as true. The function P is the main method by which the moths in search areas. Now, update the moth’s position according to the flame by applying the equation (12). $M_{i} = S (M_{i}, F_{j})$ (20)

Where, the variable M_i represents the i^th moth, the variable F_j represents j^th flame, and the variable S indicates the spiral method.

In this work, a logarithmic spiral is chosen for updating the moths. Even though any kind of spiral is applied by following the conditions, including the starting point of the spiral is to be started with moth, the final point of the spiral is the flame’s position, and the spiral range fluctuation will not cross the search area. Let us consider the points and define a logarithmic spiral of MFO, which is represented in the equation (13). $S (M_{i}, F_{j}) = D_{i} . e^{bt} . \cos (2 π t) + F_{j}$ (21)

Where, the variable D_i represents the i^th moth distance for the j^th flame, the variable b represents the constant to define the spiral shape, and the variable t represents a random number that is between -1 and 1. Moreover, the value of the variable D is computed by using the formula given in equation (22). $D_{i} = | F_{j} - M_{i} |$ (22)

Where, the variable M_i represent the i^th moth, and the variable F_j represents the j^th moth and the variable D_i represents the distance between the i^th and j^th flames.

Next, the moths’ positions are changed according to the ‘n’ locations in the search area, which degrades the exploitation process of the best solution. This work proposes an adaptive method for the various flames by using the formula given in equation (23). $FLN = RAND (N - l * \frac{N - 1}{T})$ (23)

Where the variable l indicates the number of iterations, the variable N represents the maximum number of flames, and the variable T represents the maximum number of iterations.

Even though the positions of moths are updated according to the best flame in the final steps of iterations. Moreover, it gradually decreases the number of flame balances that are exploiting and exploring the search area. The steps of the function P are as follows:

Apply the equation (22) for changing the flame.

Initialize the FF(M) into OM.

If IT_No.==1 Then

Perform sort operation for M and assigned to F

Perform the sort operation for OM and store the result into OF.

Else

Perform the sorting operation for M_t - 1 and M_t, store the results into F.

End

For i = 1 to n do

For j = 1 to d do

Update the values of r and t

Find the D value by applying the equation (22) according to the respective moth.

Change the M(I,j) value by applying the equations (21) according to the concern moth.

End

Accordingly, the function P is executed until it returns the value true by the function T. After performing the termination process for the P function, it returns the best moth value and it is returned as an optimum.

4.3 Proposed feature optimization tuned temporal CNN

This subsection explains the newly developed, tuned Temporal CNN that is used for predicting diseases effectively by handling the high-dimensional patient records in the form of benchmark datasets and live streaming datasets. In this paper, we propose a new Moth Flame Feature Optimization based Temporal Convolutional Neural Network (MFO-TCNN) for performing disease prediction processes. The proposed model is capable of predicting the diseases and their levels without the intervention of human beings. The proposed MFO-TCNN has a high prediction accuracy due to the MFO’s incorporation of an effective tuning process on the existing T-CNN. The structure of the T-CNN is explained in the next subsection.

4.3.1 T-CNN

The existing Temporal Convolutional Neural Network (T-CNN) [29] works like a standard CNN with the incorporation of the temporal constraints. Generally, the CNN consists of five layers, and each layer is responsible for performing a certain job. Among them, the convolutional layers build the feature maps and the classification layers help finalize the result of CNN. All the layers of CNN are explained one by one in detail in this section. First, the input layer preprocesses the input values of the attributes by eliminating the noisy, redundant, and meaningless data by applying the basic level filter. Second, the convolutional layers confine the features that are extracted or filtered from the input layer by applying the filters, and the feature mapping process is initiated and obtained the results for effective classification. Moreover, the values (neurons) between the two different layers are connected together by considering the necessary weightages for the concerned attributes. The convolutional layer’s output is presented in equation (24). $\begin{matrix} {(R_{r}^{s})}_{P, S} = {(D_{r}^{s})}_{P, S} + \sum_{m = 1}^{A_{1}^{w - 1}} \sum_{n = - A_{1}^{p}}^{A_{1}^{p}} \\ \sum_{o = - A_{2}^{p}}^{A_{2}^{p}} (T_{r, m}^{p}) * {(R_{m}^{p - 1})}_{P + n, S + o} \end{matrix}$ (24) where, * represents the operator of this layer, ${(D_{r}^{s})}_{P, S}$ indicates the finalized feature mapping structure from the R^th layer that is centered at P, S. Now, consider the specific weights for each layer, including P and S, which are also represented as bias values that can be finetuned with the help of MFOA. The variables m, n, and o represent the three different feature maps that are capable of producing the necessary output by every filter in this convolutional layer.

The third layer is a normalization layer, which is available in the convolutional and ReLU layers for enhancing the training attributes of the input dataset. Moreover, it is used to normalize the gradient process and the normalization process effectively by adjusting the training process. The fourth layer is ReLU, which is used to map the layers in the convolutional layer part by applying the non-parametric layer called the ReLU layer and also not inhibiting the weights or bias values. The fifth layer is the Max-Pooling layer that is responsible for performing the feature mapping process and is also generated by applying the down sampling process of the convolutional layer on input data. The sixth layer is a fully connected layer that is used to map the features of convolutional layers and also forward them to the fully connected layer where the final decision is made and the final result is obtained as per the given equation (25). $\begin{matrix} {By}_{r}^{s} = j (R_{r}^{s}) with R_{r}^{s} = {(D_{r}^{s})}_{P, S} + \sum_{m = 1}^{A_{1}^{w - 1}} \sum_{n = - A_{1}^{p}}^{A_{1}^{p}} \\ \sum_{o = - A_{2}^{p}}^{A_{2}^{p}} (T_{r, m}^{p}) * {(R_{m}^{p - 1})}_{P + n, S + o} \end{matrix}$ (25)

The various weightages used in this proposed CNN are identified by using the MFOA and performing the training process. The seventh layer is the softmax layer, which produces an output which is normalized with the help of the softmax activation function for ensuring and enabling the classification tasks. Finally, the classification layer is used for performing classification processes with the help of the activation function on all the inputs of the various disease datasets, including heart, diabetes, and cancer.

4.3.2 Tuning process by MFO

The tuning process is a process of feature selection and optimization over the input datasets. For safeguarding and maintaining human health from various deadly diseases, human health is to be monitored periodically to detect the occurrence of the symptoms. The identification of symptoms is helpful for identifying the various diseases accurately. It is an important task that detects the diseases accurately. If one fails to identify the symptoms or wrongly identifies the symptoms, it may severely affect the human being. To perform this task, the Temporal CNN is applied in this work, which is used to predict and detect diseases. The high prediction accuracy is obtained by using the optimal features along with the necessary weights of the T-CNN using the newly proposed MFO.

In this work, the newly proposed IoT and Fuzzy logic incorporated e-healthcare system is developed to predict which patients are affected by diseases. This system is used to predict heart disease, diabetes disease, and cancer disease and also to know the disease severity level. This e-healthcare system incorporates the newly proposed Intelligent Mahalanobis distance aware Fuzzy Weighted K-Means Clustering Algorithm (IMFWKCA) to group the high dimensional data and a newly developed Moth Flame Optimization Tuned Temporal Convolutional Neural Network (MFO-TCNN) to predict the diseases. Here, the IMFWKCA is used to group the high-dimensional data and also to reduce the dimensionality. Then, the reduced dataset is considered as input for the newly proposed MFO-TCNN to categorize the record as disease affected or not. In addition, the disease severity is also measured by the proposed classifier.

5 Results and discussion

This section provides the detail about the experimental setup and datasets, and evaluation parameters in different subsections and also demonstrated the experimental results and comparative analysis. First, the experimental setup is explained in detail.

5.1 Experimental setup and datasets

The proposed e-healthcare system has been developed and implemented using the Python programming language. The experiments have been done by applying the newly developed software to the IoT devices with Raspberry PI software. The required sensors are fixed in the new e-healthcare system to measure the heartbeat rate, blood pressure and glucose level. All this information has been collected and is also stored as patient records. In addition, the heart, diabetes, and cancer datasets are available in the standard UCI repository for performing the training and testing processes. The IoT devices are equipped with the necessary hardware devices, microcontrollers, and LoRa communication devices to transmit the patient data for the newly developed software. Moreover, the serum cholesterol and glucose levels have been collected with the help of a wearable device, and the electrocardiographic data is also measured by applying the heart monitor board (AD8232). In addition, the Raspberry PI software is used for collecting the highest heart rate, oldpeak, and slop [30].

5.2 Evaluation parameters

The evaluation parameters such as accuracy (ACC), recall (REC), specificity (SPEC), f1-score (FS) value, and Matthew’s correlation coefficient (MCC) are calculated by using the formulae given in the equations (26)–(31) for evaluating the proposed model by considering the True Positive (TP), True Negative (TN), False Positive (FP), and False Negative (FN). $ACC = \frac{TP + TN}{TP + TN + FP + FN}$ (26) $REC = \frac{TP}{TP + FN}$ (27) $PRE = \frac{TP}{TP + FP}$ (28) $SPEC = \frac{TN}{TN + FP}$ (29) $FS = 2 \times \frac{Precision \times Recall}{Precsiion + Recall}$ (30) $\begin{matrix} MCC = \\ \frac{(TP \times TN) - (FP \times FN)}{\sqrt{(TP + FP) (TP + FN) (TN + FP) (TN + FN)}} \end{matrix}$ (31)

5.3 Experimental results

This subsection demonstrates the experimental results that prove the efficiency and effectiveness of the proposed e-healthcare system and also proves it to be better than other healthcare systems that are available in the literature in terms of prediction accuracy.

The prediction analysis over the optimized featured dataset as five different clusters is demonstrated in Fig. 2. This analysis considered the precision, recall, and f1-score values for analyzing the prediction performance.

Fig. 2

Prediction with optimized featured datasets.

The performance significance on the optimized featured dataset is demonstrated by conducting various experiments that consider the different five clusters in terms of the standard evaluation parameters. When compared to other clusters, cluster 5 achieves the best performance in all parameters, while clusters 2 and 3 achieve lower performance.

The comparative analysis between the proposed e-healthcare system, T-CNN, and T-CNN with proposed IMFWKCA is shown in Table 1. Here, five different experiments have been conducted for the standard benchmark dataset and the streaming patient records as an input dataset. Here, heart, diabetes, and cancer disease datasets are considered in both types of datasets for conducting experiments.

Table 1

Comparative analysis

Datasets	Experiments	Prediction accuracy (%)
		T-CNN	T-CNN+IMF-WKCA	MFO tuned T-CNN+IMFWKCA
UCI	E1	93.5	94.5	96.7
Repository	E2	93.8	94.8	97.1
Disease	E3	93.7	94.7	97.2
Datasets	E4	94.2	95.2	97.5
	E5	94.4	95.4	97.3
Streaming	E1	94.55	97.31	99.27
Patient	E2	94.73	97.62	99.59
Health	E3	94.81	97.67	99.79
Records	E4	95.43	98.21	99.46
	E5	95.56	98.22	99.32

The proposed e-healthcare system performed better than the T-CNN and the T-CNN with the proposed IMFWKCA in all the five experiments. Here, the performance of the existing T-CNN is lower than the proposed e-healthcare system and T-CNN with IMFWKCA due to the unavailability of the proposed data clustering and feature optimization algorithm. Moreover, better prediction accuracy is obtained in all the experiments for the streaming dataset than for the standard benchmark dataset.

The training and testing time taken is analyzed and shown in Table 2, which considers two different datasets that are used in this work for evaluating the proposed e-healthcare application. The training time and testing time of two different datasets were considered in this time analysis.

Table 2

Training (TR) and Testing (TST) time analysis

Dataset	Time Taken (sec)
	Proposed e-Healthcare System		Disease Prediction System (Sandhya (et al. 2020)		T-CNN
	TR	TST	TR	TST	TR	TST
UCI Heart	0.52	0.36	0.81	0.38	0.49	0.34
Dataset
Streaming	0.51	0.32	0.56	0.37	0.46	0.31
Dataset

The training and testing time for the e-healthcare application is slightly higher than the existing T-CNN alone and also lower than the existing disease prediction system. The incorporation of the newly proposed clustering and feature optimization algorithms may consume more time than the existing Temporal CNN. However, it achieved better prediction accuracy in all the experiments and clusters that were demonstrated in the previous results.

The prediction accuracy analysis between the domain expert and the proposed e-healthcare system is demonstrated in Fig. 3. The different numbers of records from the various types of diseases such as coronary artery and vascular disease, heart rhythm disorders (arrhythmias), heart failure, structural heart disease, diabetes, and cancer disease affected patients’ records have been considered in this work.

Fig. 3

Prediction accuracy analysis between the domain expert and proposed e-healthcare system.

The newly developed e-healthcare system outperforms the domain expert in prediction accuracy. The prediction accuracy on various patient records goes down and also stabilizes at the same percentage of prediction accuracy when the number of patient records is increased from 1000 records. The reason for the achievement is the use of an effective deep learning algorithm with feature optimization and the clustering method.

Figure 4 shows the comparative analysis between the proposed e-healthcare system and the existing disease prediction systems such as T-CNN with feature selection [29], CNN, FTCM [7], TFMM-PSO [9], and Artificial Neural Network (ANN). Here, the five different experiments, such as Experiment 1 (E1), Experiment 2 (E2), Experiment 3 (E3), Experiment 4 (E4), and Experiment 5 (E5), have been conducted with different sets of patient records with the combination of live streaming datasets and the benchmark dataset for evaluating the proposed e-Healthcare system and the existing disease prediction systems.

Fig. 4

Comparative analysis.

In this comparative analysis, the performance of the proposed e-Healthcare system was obtained with better performance in all the five different experiments, such as E1, E2, E3, E4 and E5. The experiments have been conducted by using various combinations of patient records which are selected randomly from the live streaming dataset and the standard dataset. The reason for the performance improvement is the application of effective feature optimization tuned to Temporal CNN and the proposed Intelligent Mahalanobis distance based Fuzzy Weighted K-Means Clustering Algorithm. Here, the fuzzy logic and temporal constraints are useful for making effective decisions on patient records.

6 Conclusion and future works

In this paper, a new IoT and Fuzzy-aware e-healthcare system has been proposed and implemented for predicting heart disease, diabetes disease, and cancer disease effectively. The proposed system works with the proposed Mahalanobis distance-aware Fuzzy Weighted K-Means Clustering Algorithm (MDFWKCA) to group the high dimensional data and a newly developed Moth Flame Optimization Tuned Temporal Convolutional Neural Network (MFO-TCNN) for predicting the diseases effectively. In this work, the high-dimensional data is grouped by applying the proposed clustering algorithm according to the differences between the records and reducing the dimensionality as well. Here, the high-dimensional data is subjected to the dimensionality reduction process and grouped into different subgroups. Then, the grouped dataset is forwarded to perform the classification by using the newly proposed feature optimization tuned temporal CNN. Here, the existing MFO algorithm is used for tuning the features with adjusted weights to the Temporal CNN to make a final decision on medical records. The proposed IoT-aware e-healthcare system is evaluated by using the standard benchmark datasets and the live streaming patient records that are generated through IoT devices and from various hospital records by conducting various experiments and also achieving better performance in terms of precision, recall, f-measure and prediction accuracy. Future work can be done in this direction with the introduction of a new optimization technique for tuning the deep learning algorithm to improve the disease prediction accuracy.

References

Guyon

and Elisseff

, An introduction to variable and feature selection, Journal of Machine Learning Research 3 (2003), 1289–1305.

Lazar

, Taminau

, Maganck

, Steenhoff

, Coletta

, Molter

, Schaetzen

, Duque

, Bersini

and Nowe

, A survey onfilter techniques for feature selection in gene expression microarray analysis, IEEE/ACM Transactions on Computational Biology and BioInformatics 9(4) (2012), 1106–1119.

Kohavi

and John

G.H.

, Wrappers for feature subset selection, Artificial Intelligence 97(1–2) (1997), 273–324.

Tibshirani

, Regression Shrinkage and Selection Via the LASSO, Journal of Statistical Social Service, B Statistics Methodology 58(1) (1996), 267–288.

Biau

, Cadre

and Rouviere

, Accelerated gradient boosting, Machine Learning 108(6) (2019), 971–992.

Kriegel

, Kroger

and Zimek

, Clustering high-dimensional data, ACM Transactions on Knowledge Discovery from Data 3(1) (2009), 1–58.

Sethukkarasi

, Ganapathy

, Yogesh

and Kannan

, An intelligent neuro fuzzy temporal knowledge representation model for mining temporal patterns, Journal of Intelligent & Fuzzy Systems 26(3) (2014), 1167–1178.

Ayesha

, Hanif

M.K.

and Talib

, Performance enhancement of predictive analytics for health informatics using dimensionality reduction techniques and fusion frameworks, IEEE Access 10 (2022), 753–769.

Ganapathy

, Sethukkarasi

, Yogesh

, Vijayakumar

and Kannan

, An intelligent temporal pattern classification system using fuzzy temporal rules and particle swarm optimization, Sadhana 39(2) (2014), 283–302.

10.

Kanimozhi

, Ganapathy

, Manjula

and Kannan

, An intelligent risk prediction system for breast cancer using fuzzy temporal rules, National Academy Science and Letters 42(3) (2019), 227–232.

11.

Thanh

N.T.

, Wehrens

and Buydens

L.M.C.

, KNN-kernel density-based clustering for high-dimensional multivariate data, Computational Statistics & Data Analysis 51(2) (2006), 513–525.

12.

Bouveyron

, Girard

and Schmid

, High-dimensional data clustering, Computational Statistics & Data Analysis 52(1) (2007), 502–519.

13.

Jing

, Ng

M.K.

and Huang

J.Z.

, An entropy weighting k-means algorithm for subspace clustering of high-dimensional sparse data, IEEE Transactions on Knowledge and Data Engineering 19(8) (2007), 1026–1041.

14.

and Gu

, A cluster description method for high dimensional data clustering with categorical variables, 2010 International Conference on Measuring Technology and Mechatronics Automation (2010), 32–35.

15.

Jiang

, Yu

and Wang

, A high-dimensional data stream clustering algorithm based on damped window and pruning list tree, 2011 4th International Conference on Biomedical Engineering and Informatics (BMEI) (2011), 2036–2040.

16.

Chen

, Ye

, Xu

and Huang

J.Z.

, A feature group weighting method for subspace clustering of high-dimensional data, Pattern Recognition 45(1) (2012), 434–446.

17.

Wang

, Wang

, Chung

and Deng

, Fuzzy partition based soft subspace clustering and its applications in high dimensional data, Information Sciences 246 (2013), 133–154.

18.

Muja

and Lowe

D.G.

, Scalable nearest neighbor algorithms for high dimensional data, IEEE Transactions on Pattern Analysis and Machine Intelligence 36(11) (2014), 2227–2240.

19.

Radhika

K.R.

, Pushpa

C.N.

, Thriveni

and Venugopal

K.R.

, EDSC: Efficient document subspace clustering technique for high-dimensional data, 2016 International Conference on Computational Techniques in Information and Communication Technologies (ICCTICT) (2016), 222–226.

20.

Tao

, Wang

, Gu

, Chen

, Lin

and Lin

, Anovel intelligent clustering approach for high dimensional data in a big data environment, 2017 13th International Conference on Natural Computation, Fuzzy Systems and Knowledge Discovery (ICNC-FSKD) (2017), 1666–1671.

21.

Wang

, Chen

, Yan

, Zeng

and Hong

, Fast adaptive k-means subspace clustering for high-dimensional data, IEEE Access 7 (2019), 42639–42651.

22.

Tapia-Galisteo

, et al., Prediction of cocaine inpatient treatment success using machine learning on high-dimensional heterogeneous data, IEEE Access 8 (2020), 218936–218953.

23.

Sui

, Liu

, Jung

and Li

, Dynamic sparse subspace clustering for evolving high-dimensional data streams, IEEE Transactions on Cybernetics (2020), 1–14.

24.

Guan

, Zhao

, Yin

and Li

, A differential evolution based feature combination selection algorithm for high-dimensional data, Information Sciences 547 (2021), 870–886.

25.

Lei

, Lv

, Yajie

and Xiaohua

, Logistics service provider selection decision making for healthcare industry based on a novel weighted density-based hierarchical clustering, Advanced Engineering Informatics 48(101301) (2021), 1–18.

26.

Amorim

R.C.

and Mirkin

, Minkowski metric feature weighting and anomalous cluster initializing in kmeans clustering, Pattern Recognition 45 (2012), 1061–1075.

27.

Frank

K.D.

, Rich

and Logcore

, Effects of artificial night lighting on moths, Ecological Consequences of Artificial Night Lighting (2006), 305–344.

28.

Mirjalili

, Moth-flame optimization algorithm: A novel nature-inspired heuristic paradigm, Knowledge-Based Systems 89 (2015), 228x–228.

29.

Sandhiya

and Palani

, An effective disease prediction system using incremental feature selection and temporal convolutional neural network, Journal of Ambient Intelligence and Humanized Computing 11 (2020), 5547–5560.

30.

Mohan

, Thirumalai

and Srivastava

, Effective heart disease prediction using hybrid machine learning techniques, IEEE Access 7 (2019), 81542–81554.