A hybrid particle swarm optimization with multi-objective clustering for dermatologic diseases diagnosis

1 Introduction

Most dermatological infections fall under the erythematous-squamous disease (ESD) category. Psoriasis, seborrheic dermatitis, lichen planus, pityriasis rosea, chronic dermatitis, and rubra pilaris are among the six types of ESD. Erythemato-squamous illnesses are crusty inflammatory skin conditions that usually affect the scalp and cheeks. There are six different types of erythema and scale, but they all have had very similar clinical symptoms. Genetic and environmental factors can contribute to dermatological disorders, which tend to manifest themselves at certain life stages, such as late childhood/early adolescence. Skincare differential data analysis is a tough subject since one illness may show early signs of another disease and then develop its own characteristics as the disease progresses through its life cycle.

A huge dataset is mined for hidden, previously unknown, and potentially useful information using a data mining technique. Data mining is an interdisciplinary approach for finding and interpreting important information. On this basis, the following two parts described the data mining techniques utilized for categorizing ESD illnesses. Large amounts of medical information are now available as a result of recent technological advances. These records are full of useful information. As a result, data mining techniques are employed to uncover important patterns. This paper will provide an overview of data analysis and its many methodologies, as well as a review of the medical data mining research. The importance of diverse data mining approaches is highlighted. In medical analytics, classification is a reliable method.

A technique presented in this paper [1] can be used to identify skin disorders. In a two-stage technique for accurate disease diagnosis, computer vision and machine learning are combined with clinically evaluated histological features. When the system was trained and tested on six diseases, it obtained a 95% accuracy rate.

When it comes to preparing effective and efficient data for data mining, data pre-processing is a key component. When it comes to developing complete predictive models, feature selection is crucial since it provides significant features [2]. The use of FSMs such as forward selection and backward elimination to pick the optimal subset of characteristics reduces training time and improves system performance employed in this prediction model.

In this work [3], a predictive framework for ESDs is constructed using a mix of classifiers. It is suggested that the model be broken down into two stages. A simple set-based feature selection approach is used to choose an initial feature subset. In the second phase, an aggregation of three classifiers consisting of MLP, KNN, and SVM is created. Finally, each classifier arrives at a diagnosis result using the majority voting approach. The recommended model achieves 97.78% accuracy in ensemble categorization.

Self-organization maps (SOM) of clinical and pathological findings will be used in the present investigation to explore clusters associated with different illnesses [4]. A SOM is an encompass of neurons that adapts to a dataset’s topology structure, enabling us to view massive datasets and find probable groupings. By constantly shifting its neurons directly to the data points, a SOM understands the form of a dataset. SOM is a size reduction method that uses multidimensional input to reduce the issue. As a consequence, the artificial intelligence program SOM was used to classify the six kinds of erythematous-squamous skin conditions. The classification and regression tree (CART) was used to predict the ESD diagnosis [5]. A CART is a computational intelligence forecasting technique. It illustrates how the quantities of a target attribute can be anticipated using other variables. It is a tree structure with each fork separated into a predictor variable and predictions for the target attribute at the end of each node [6]. The proposed model was 94.84% accurate.

Bayesian network, support vector machine (svm), logic boost, multilayer perceptron (MLP), Random Forest, and J48 data mining classifier were utilized in the current research to predict the illness [7]. They were able to diagnose illnesses with an accuracy of 83.6 and 84.5%, respectively, by using Bayesian Network and Logic Boost methods.

For skin disease detection, a hybrid model combining Case-Based Reasoning and Artificial Neural Networks methods is suggested in ref. [8]. There are nine input factors (attributes) that have a significant impact on the skin diagnosis process in the suggested model. To implement the concept, a user-friendly and interactive computer program has been built.

To identify erythemato-squamous illness, ChiGA (Chi-Square and Genetic Algorithm) is a unique feature selection approach [9]. ChiGA selects the optimum feature subset by combining the benefits of filter and wrapper techniques. Ninety-eight percent of the characteristics are optimal, with 18 of them being particularly accurate.

Dermatology disease processing technique is defined in this study. Each of the three stages of the suggested work paradigm is described in more detail [10]. Statistical parameters are used in the first step of data processing and analysis. To create the predictive cell structure, these factors are monitored with probabilistic measurements in the second step of the process. Bayesian networks are used for illness categorization in the final stage. As a result of its implementation, the method’s outcomes have been very accurate.

Filter technique and Partial Forward Search (F PFS algorithm) is a novel weighted probabilistic strategy for feature selection that determines the best models of support vector machines to detect various skin disorders [11]. Overfitting is reduced and accuracy is improved, especially when using a classifier like a support vector machine.

This paper is structured as follows: In Section 2, we will look at some of the most recent methods for classifying skin disorders. Section 3 explains the shortcomings of the current and the originality of the suggested technique. Section 4 describes the clustered and optimized classifier. Based on the suggested technique, results and discussion are presented in Section 5. Lastly, the paper concludes with a summary, which is expanded in Sections 6 and 7.

2 Literature survey

With the WEKA data mining tool, this article [12] seeks to forecast skin disorders by combining the following algorithms Neural Network, Decision Tree, and Ensemble Method based on Majority Voting. The prediction algorithm was evaluated using a dermatology dataset. Based on accuracy and precision as well as true positive and false positive, the performance evaluation was conducted to determine the effectiveness of the test. It creates a superior skin disease prediction model with significantly greater accuracy and generalization, which avoids both over-fitting and under-fitting.

A combination Dynamic ant colony model with 3 updating stages, wavelets transform, and singular value decomposition incorporating svm was suggested in this article [13]. The suggested method’s intended goal is to reduce a subset of characteristics to provide an acceptable illness diagnosis with the best accuracy, sensitivity, and specificity across a wide variety of disorders. Five benchmark medical datasets of diverse illnesses will be used to assess the proposed approach.

In the categorization of Erythematous-squamous Dataset, Naive Bayes has a 100% success rate [14]. As a consequence of the integration of correlation-based filter methods and different heuristic search methods, the number of characteristics has been decreased from 34 to 19. To locate important class characteristics, the original data are subjected to the correlation-based feature selection (CFS) process. Continuous and discrete problems are filtered using a fast correlation-based filtering technique. To reduce dimension, features are considered to use the relief method.

They [15] represent a new strategy in this paper that combines six distinct data mining categorization techniques. The most important 15 characteristics that will play a key influence in prediction are chosen using a correlation-based feature technique. Then, combining Bagging, AdaBoost, and Gradient Boosting classifier methods, they create an ensemble approach. Gradient boosting is a sort of boosting used in deep learning. It is based on the assumption that whenever the best potential future model is coupled with prior models, the prediction accuracy error is minimal. If a slight modification in a case’s forecast generates no modification in error, the case’s subsequent targeted result is nil [16]. When applied to dermatological datasets, the ensemble approach and feature selection produce better results than individual classifier methods. The ensemble technique seems to be more efficient and precise in predicting skin diseases.

When utilizing data mining algorithms for the prediction of skin disorders, this article [17] explores the effect of feature selection approaches. PCA, GA, and Chi-square were the feature selection methods used to minimize the number of features in a dataset of skin disorders. Feature selection methods did not increase the precision and sensitivities of these methods, according to empirical observations.

Psoriasis – Cronic Dermatitis and Seborreic Dermatitis Pitriasis Rosea were identified in the same cluster, and the diagnoses are confounded as a result of this procedure [4]. Clinical and histological results of the illnesses were grouped by the SOM technique to avoid this misunderstanding [18]. SVM was used to classify clinical and histological results that clustered together. This led to the categorization of psoriasis and chronic dermatitis at 0.89 and 0.93, respectively, and seborreic dermatitis with pitriasis rosea at 0.79 and 0.80.

When applied to dermatological datasets, the ensemble technique and feature selection produce better results than individual classifier approaches [19]. A more accurate and efficient way to predict skin disorders appears to be using the ensemble approach. As a novel method, they [20,21] integrate data mining classification techniques from six different sources.

NB, K-Nearest Neighbor (KNN), SVM, Random Forest (RF), and Multilayer Perception are some of the machine learning algorithms they utilized for skin diseases classification [22,23]. To improve the performance of the trained model, they employed three fundamental ensemble learning: Bagging, Boosting, and Stacking. A model’s utility is measured by its accuracy, precision, sensitivity, and f1-score.

Using Artificial Neural Networks, this research aims to categorize the six types of ESD with great accuracy. Using a backpropagation feed-forward approach, they classified ESDs using the JustNN program [6]. They used the pre-process dataset to train and validate the suggested model. The proposed model has a 98.36% accuracy rate.

Calinski–Harabasz and Davies–Bouldin cluster value indices are used to assist the development of multi-objectives [16]. The clustering is done with the assistance of a k-means cluster and particle optimization to identify the optimal clusters. The multi-objective problem will be solved by PSO to find the optimal cluster.

3 Existing method

In the categorization of Erythematous-squamous Dataset, Naive Bayes has a 100% success rate [14]. As a consequence of the integration of correlation-based filter methods and different heuristic search methods, the number of characteristics has been decreased from 34 to 19. When utilizing data mining algorithms for the prediction of skin disorders, this article explores the effect of feature selection approaches [18]. PCA, GA, and Chi-square were the feature selection methods used to minimize the number of features in a dataset of skin disorders [24]. Feature selection methods did not increase the precision and sensitivities of these methods, according to empirical observations.

The above two algorithms utilized the feature selection approach for skin diseases prediction using supervised classifiers. Due to that, it indicates that feature selection has a lesser role in classification.

Calinski–Harabasz and Davies–Bouldin cluster value indices are used to assist the development of multi-objectives [16]. The clustering is done with the assistance of a k-means cluster and particle optimization to identify the optimal clusters [25]. The multi-objective problem will be solved by PSO to find the optimal cluster. Yet, this paper also suffers from poor classification.

4 Proposed method

With the aid of machine learning methods, an optimized data mining strategy is provided in this work to forecast skin disorders. A hybrid particles swarm and moth flame optimization are used to get an optimal clustered distance value for K-means grouping. The K-means clustering misclassification rates and the neural network classification rate are two cluster value indexes that govern multi-objectives. The multi-objective problem would be solved using hybrid PSO to find the best clusters for cluster analysis [26]. The process in the proposed method is shown in Figure 1.

Figure 1

Proposed method process flowchart.

Figure 1 depicts the fundamental data mining method for predicting skin disorders, which includes the following steps:

1. The UCI machine learning repository website provided the input skin diseases dataset.

2. To balance the dataset and its values, pre-processing processes such as normalization and rows addition are performed.

3. Next, particle swarm optimization is used to identify the best attribute for predicting diseases.

4. A feed-forward neural network is used to verify the specified characteristics.

5. The best distance metric for clustering is then calculated using the optimal attribute and moth flame optimization.

6. After determining the best attribute and distance, the data are clustered using k-means.

7. Finally, accuracy is used to assess the suggested method’s performance.

The working of each block is explained as follows.

4.4 Attribute evaluation using feed-forward neural network

As illustrated in Figure 2, the feed-forward neural network is trained with 70% of the best records from the specified characteristics in this stage. A feedforward network’s purpose is to estimate a mapping function and input to a classification. A feedforward system creates a mapping and trains the values of parameters which create great estimations of the functionality. The data flows from the inputs to the functional getting assessed, then to the intermediate calculations necessary to define it, and finally to the result. That there were no feedback links; therefore, the model’s outputs do not respond to the valuable feedback into it [29].

Figure 2

FFFN process flow.

The optimum characteristics from the previous phase are used as inputs. Each input has its own weights based on the Levenberg–Marquardt function, which are modified in the hidden layer by multiplying with the Gaussian function. Finally, a single hidden layer with 10 hidden neurons is used to connect the input and output classes. The network is trained with the best attribute using this method. The trained network is then tested using 30% of the dataset to assess the particle swarm optimization-selected characteristic. Calculating the accuracy of the testing dataset is used to do the attribute assessment.

After that, using multi-objective moth flame optimization, the best cluster distance is found.

5 Implementation and discussion

The proposed approach was implemented in this study using the MATLAB R2020b program in a Windows 10 environment. Any three-variable optimization issue with two inequity style restrictions can be solved using particle swarm optimization algorithms. In MATLAB scripts which can be added to the Linked data, the programs can simply be modified to include more variables and restrictions. MATLAB can examine larger amounts of data and scale it up to cluster quickly.

The input for the processing in MATLAB is taken in the form of Excel data as shown in Figure 3.

Figure 3

Input dataset format.

Pre-processing is applied to the input dataset, and the values are normalized using Z-score normalization, which transforms the values in the dataset from 0 to 3 to different values as seen in Figure 4.

Figure 4

Pre-process output.

The pre-processed data is then submitted to particle swarm optimization to pick the best attribute, with the associated convergence curve for attribute selection illustrated in Figure 5.

Figure 5

PSO convergence curve.

From Figure 5, it is observed that the optimal attribute is determined after the 20th iteration itself and the corresponding objective function value is 0.47 at the end of the process. From the PSO, it was observed that the scalp involvement and clubbing of the rete edges were important attributes of the classification process.

The records with the best attributes are then put through a neural network to see how well they choose their attributes. Figure 6 depicts the network’s training procedure.

Figure 6

FFNN training.

Then, the FFNN is tested with the testing dataset, then it is evaluated using evaluation metrics and is shown in Figure 7.

Figure 7

Attribute selection evaluation.

Then, the optimal cluster distance for the optimal records is determined using Moth flame optimization and the corresponding convergence curve is shown in Figure 8.

Figure 8

MFO convergence curve.

Finally, from MFO, the optimal distance for K-means clustering is selected as a city-block, and the optimal dataset is classified using an unsupervised algorithm called K-means clustering, and its performance is evaluated in terms of accuracy as shown in Figure 9.

Figure 9

Evaluated in terms of accuracy.

The comparison between the existing and proposed approaches in terms of performance evaluation is listed in Tables 4 and 5.

Table 4 compares the performance of the testing dataset with original and adjusted data for attribute selection using an improved feed-forward neural network. The proposed normalized method increased the accuracy to 100%. It shows that it can categorize all six classes; however, identifying the specific classes is decreased by 6% when compared to prior studies.

The purpose of this article, on the other hand, is to undertake an unsupervised classification using only a few parameters to discover sickness subgroups and give effective treatment. As a result, Table 5 compares the performance of the current and proposed work.

Table 5 demonstrates that when compared to previous work, the suggested increased its unsupervised classification accuracy by 10%. It was accomplished utilizing MFO and a normalized dataset, as well as cluster distance selection. As a result, the suggested hybrid PSO-MFO is ideal for data mining in the categorization of skin disorders.

6 Conclusion

Multi-objective-based Hybrid optimal clustering was used in this work. The findings of their experiments were compared to the existing particle swarm-based clustering approach. Six types of ESDs are included in the dataset, all of which have redness (erythema) as a symptom that causes cell death (squamous). The UCI skin diseases undergo various process steps to determine the optimal attribute and optimal cluster distance for clustering. The research provided comprehensive results for all tests, as well as tables and illustrations for various comparisons. Three distinct performance measures generated from the multiclass confusion matrix were used to make the comparisons. They’re accuracy, sensitivity, and specificity. Based on the results, it was observed that the proposed method improved its performance over the existing work.

• The normalization and data balance improved the prediction process.

• The attribute selection reduces the computational time for processing the testing data.

• The cluster distance selection improved the unsupervised prediction accuracy.

7 Future work

In the future, an ideal clustering technique will be chosen to enhance all of the assessment criteria for the prediction process.