1 Introduction

Healthcare systems around the world are facing huge challenges in responding to the rising trends of chronic diseases, resources constraints, aging population and the growing focus of citizens on healthy living and prevention [1]. From 2012 to 2030 an assessment report declared nearly an economic loss of 3.6 trillion dollars will occur only due to four chronic disorders i.e., cancer, CRDs, CVDs and diabetes. Major Chronic Disease include Coronary Heart Disease [2], Chronic Kidney Disease [3], Parkinson's disease [4], Alzheimer's disease [5], Diabetes and Hypertension [6], Thyroid Disease [7], High Blood Pressure, Asthma, Cardiovascular Disease, Stoke, Peripheral Arterial Disease, Mental Health problems, and Dementia. Usually, Preventing and screening for disease before they start have been identified as the best ways to prevent the rise of chronic disease. However, in primary care most doctors lack the time, resources and tools to prevent chronic diseases. Traditional models of healthcare focus on one disease but prevent chronic diseases we need a compressive model. The easiest way to predict it in early stages is to analyse the already existing big healthcare data [8]. This is where data science and machine learning come in play to assist doctors to predict disease in early stages. Scientifically, we can add more quality to “Skill India and Make in India” By “Make India Healthy” [1]. Chronic diseases that are chief contributors to mortality and economic loss are Breast cancer, Heart disease, Cervical cancer and Diabetes. Diabetes is a chronic disease that occurs due to lack of insulin production in body. In a report published by IDF in 2013 a measure of 382 million people (8.3% of total) was affected by diabetes with 14 million more (184 million women and 198 million men) men than women. Mainly it is of three types: type-1 diabetes, type-2 diabetes and gestational diabetes [6]. Heart disease is also a main cause of mortality in India. Within last 26 years rate of mortality due to cardiac disease increased by 34% from 155.7 to 209.1 per lack population. In 2016 nearly 62.5 million premature mortality reported due to cardiac diseases [2]. Cervical cancer is the major chronic malignant disease and is the fourth most common disease in women’s worldwide [9, 10]. There are several risk factors that develops cervical cancer are smoking, sexual transmitted disease and Human Papilloma virus (HPV) [11]. By identifying all those factors and developing a classification model we can easily predict that the case is malignant or benign [12]. Breast Cancer is one of the malignant tumor that accounts for 25% in women’s globally according to American cancer society. It can be categorized into two types: Malignant (cancerous) and Benign(non-cancerous). Breast cancer is a collection of diseases in which cells present in breast tissue change their shape and divided abnormally, typically resulting in a lump or mass [13]. According to report published by WHO 1.2 million women’s worldwide will be diagnosed breast cancer [14]. Early diagnosis and treatment of such cancerous cells is the only solution to cancer-free environment [15]. Data science is one of the emerging areas of research that incorporates artificial intelligence, machine learning, deep learning, statistics, optimizations and data mining [1]. Data Science approaches already have various applications in early classification of Human activity recognisation, Industrial Process monitoring, Intelligent transportation, Quality monitoring, Medical Diagnosis and others. Healthcare sector produces huge volume of data and Data Science techniques supports to extract hidden knowledge that enable new opportunities and innovations to improve population health by addressing different perspectives [1]: (1) descriptive, to diagnose what happened; (2) diagnostic, to diagnose the reason why it happened (3) predictive, to diagnose what will happen and (4) prescriptive, to detect how we can make it happen [1, 16]. Data analytics technologies provide more effective tools [17] that helps to provide Home care, Lifestyle support, Precision medicine, better treatment of chronic disease by early detection, population health and better treatment of infectious diseases [18]. During last two decades researchers have proposed numerous novel ML techniques for predictive data analysis [1]. These useful techniques have been implemented in various data-intensive research areas like healthcare, biology, astronomy to mine hidden patterns [1]. In this article we review latest high-quality articles from major research databases for computer science like IEEEXplore, ACM Digital Library, GoogleScholar, ScienceDirect and SpringerLink on Breast Cancer, heart disease, cervical cancer, diabetes, and other chronic diseases. Also, we present some limitations of the existing work and their probable solutions that can be extended to other related work.

The Whole structure of this article is described as follows, Sect. 2 discusses Review of latest high-quality work related to major chronic diseases with existing methodologies, datasets and limitations. In Sect. 3 we discuss Material and Methods that are used in this study. Section 4 discusses the results obtained after the implementation of materials and methods listed in Sect. 3. Finally, Sect. 5 ends with a conclusion and future enhancement.

2 Literature review

In the recent few years, several diagnostic techniques involving Data Science, Artificial Intelligence, Machine Learning and Deep Learning has been proposed by researchers for diagnosing diseases. In this study we review few latest articles and novel approaches developed in last few years that improves predictive and diagnostic power of existing healthcare systems and need to be further improved.

Early classification of mitral heart valve and aortic heart valve heart-valve disorder “based on linear discriminant analysis (LDA) and adaptive neuro-fuzzy inference system” [19] using dataset of Doppler heart sound (DHS) signals by ultrasound system proposed by Sengur [19] involves Pre-processing, normalizing and Filtering of DHS. Usage of Wavelet transform, short-time Fourier transform and wavelet entropy on DHS signals to extract waveform patterns along with ANFIS and linear discriminant analysis (LDA) for early classification of abnormal or normal heart valves. Although Healthcare sector is data rich but problem of class imbalance is always there but a new feature selection technique PSO-SVM given by Vijayashree and Sultana [20] for Feature Selection in Heart Disease early Classification. A novel function for selecting an optimal weight and fitness function PSO with SVM for selecting more relevant features. The study compares several feature selection methods such as PSO, gain ratio, CFS, filtered subset, Chi-squared, Consistency subset Relief and Info gain algorithms [20] when PSO-SVM is used for feature selection. SVM based classifiers outperforms other classifiers with accuracy effectively raised by 3.09%. For early and accurate classification of various medical disorders a New era of Hybrid Learning came that combines results of multiple algorithms to generate highly accurate classifiers. For “Early classification of Heart Disease a new Hybrid approach [2]” Using Ensemble Hybrid Machine Learning for assisting medical doctors. Process begins from a pre-processing phase followed by feature selection phase based on DT entropy, classification of modelling performance evaluation, and hyperparameter tuning the results with improved accuracy. Numerous (ML) techniques are used namely, Logistic Regression, Decision Tree, Generalized Linear Model, Support Vector Machine, Gaussian Boosting. This Hybrid learning process involves Decision tree, Random forest and generalized linear model with 13 features extracted based on feature engineering. Basic building block of machine learning process involves data preprocessing, Feature selection, Feature Extraction and selecting best algorithm for classification based on problem. A new intelligent based system for “improved detection of heart disease Based on Random Search Algorithm (RSA) and Optimized Random Forest Model [21]” by JAVEED et al. This involves selection of relevant features and optimized RF with grid search for hyperparameter tuning for heart failure prediction based on Cleveland heart disease dataset. Contribution of Two new experiments: first Random forest and second RSA based RF method. Initially, the dataset is provided to random search algorithm for optimal feature selection. The feature extraction technique RSA extracts 7 most features while the past published work refers 13 features. RSA-RF is efficient that achieves an accuracy of 93.33% on 7 extracted features with improvement in accuracy by 3.3% when compared with other techniques.

Various approaches for Early classification of Time series and diseases have been proposed in recent times. “Early classification of cervical cancer grounded on Support Vector Machine by using various risk factors is [9] proposed by Wen Wu et al. determines relevant 10 risk factors and four target variables Cytology, Hinselmann, Biopsy and Schiller. The experimental study tried to reduce processing time more than other experimental studies by selecting most relevant features by using Principal Component Analysis (PCA) and RFE. Also, over-sampling technique is used with PCA and RFE. Major problem when we work with medical data is class imbalance. To handle this imbalance a new model “using feature reduction technique and SMOTE with random forest for the diagnosis of Cervical Cancer [10]” was designed. Along with random forest and SMOTE two Feature Reduction techniques PCA (principal component analysis) and RFE (Recursive Feature Elimination) is used. Tenfold cross validation is used for training, validation and testing the model. SMOTE-RF shows the rise in accuracy between 1.7 to 3.5% for respective 4 target variables. A new model for Diagnosis of automated cervical cytology and for the realization of squamous epithelial cell automatic detection system [11] was conducted on 500 cervical cells taken from Shenzhen Second People’s Hospital of pathology department. Using neural network model based on faster regions-CNN for classification purpose and cell detection. Model detects five target cells “low grade squamous intraepithelial, endocervical cell, Atypical squamous cells of undetermined significance lesion, metaplastic squamous and high grade squamous intraepithelial lesion [22]. A new approach for the prediction of Cervical cancer Which is a gynaecological cancer [12]. Various algorithms Like “Recursive feature selection (RFE), Boruta algorithm, ctree () and Simulated annealing (SA)” are used to select right features for ML algorithms on UCI dataset. Efficient feature selection reduces features from 36 to 27 and four target variables are combined to make one target feature named ‘cancer’. ML techniques as KNN, C5.0, SVM, RPART and RF are used with tenfold cross validation whereas Random Forest and C5.0 perform significantly well with accuracy 100% and 99% respectively. Now a days Neuro Fuzzy System and various deep learning methods are used for purpose of early classification. Neuro Fuzzy System based diagnosis of cervical cancer using pap smear images [23] obtained from Leica Microsystems website containing 15 cervical cancer pap smear images. It uses clustering algorithm Fuzzy c-means for the segmentation of images. Finally shape theory was used for detection of segmented images under study to detect the abnormality in cells. Features extraction performed on segmented images extract these “nucleus-cytoplasm ratio (NCR), cytoplasm circularity (CC), nucleus area (NA), cytoplasm area (CA), nucleus circularity (NC) and maximum nucleus brightness (MNB)”. The neuro-fuzzy system is trained using Levenberg–Marquardt (LM) back-propagation technique to predict whether the cancer is malignant or benign. Also, for the prediction of Diabetes a Hybrid prediction system was Developed using C4.5 decision tree and classical kmeans clustering algorithm for assisting doctors to diagnose it in early stages efficiently [24]. Data-processing and feature extraction is performed for correct analysis of diabetes on PIMA Indian dataset. Findings of this proposed study achieves an accuracy of 92.38%. New hybrid approach for the diagnosis of diabetes mellitus based on Adaptive neuro fuzzy inference system with decision Tree for achieving higher [25] accuracy. It generates rules in crisp form by using decision tree classifier and fed into ANFIS as an input after application of Gaussian membership method. Following this, the optimization has been carried out using least square estimation, gradient descent approach and tenfold cross validation method. Also Model for the early diagnosis of type-II diabetes using decision tree approach along with particle swarm optimization with java implementation(J48) of decision tree(C45) with optimized parameters [26]. For the identification of optimized parameter set a self-adaptive Inertial weight with PSO is used and fitness function is standardized using J48 algorithm. Risk factors like “mean blood glucose (MBG), fasting plasma glucose (FPG), postprandial plasma glucose (PPG) and glycosylated hemoglobin (A1c)” are consirded for evaluation. Finally, Fisher’s linear discriminant analysis has been used to test the efficiency of the predicted result. Also, a tree-based ML algorithm for diagnosis of diabetes using 8 different ML techniques as base classifiers in 5 different ensembles i.e., DECORATE, random subspace, boosting, bagging, rotation forest [27]. Base classifiers are classification and regression tree, decision tree (C4.5), functional tree, random tree, naïve Bayes tree, reduced error, pruning tree, best-first decision tree and logistic model tree. The exploratory study and performance of base classifiers and different ensembles are thoroughly benchmarked on three different datasets named PIDD, Tabriz dataset, RSMH [28] with AUC as characteristic metric. In Table 1, we listed some important references with source of datasets used, disease specified, algorithm used as well as limitations of study whereas performance measures are specified in italics.

Table 1 Literature review over medical data by various authors
Full size table

The Challenge regarding dataset imbalance can be easily fixed by making use of oversampling and under sampling techniques. Under-sampling need sufficient quantity of data. It reduces the size of an abundant class to convert it into balanced one while Over-sampling handles insufficient quantity of data by increasing the size of rare samples. SMOTE (Synthetic Minority Over-Sampling Technique) is a powerful oversampling approach that is used to handle an imbalanced dataset and is introduced by [10] Chawla et al. Synthetically, it is used to increase the size of rare class by using K-nearest neighbors [32]. Also, to resolve an issue of selecting optimal features, reducing computational time and runtime storage space we can use Particle Swarm Optimization [33] (PSO) that improves predictive power of ML classifiers. PSO is an attractive approach for feature selection which is computationally inexpensive [34] and takes less memory and runtime by taking only few parameters.

3 Materials and methods

In Sect. 3, we are going to confer the materials and methods that have been used for finding results. This section is divided into four subsections, i.e., dataset description, proposed methodology, data preprocessing, Feature Selection Techniques, Algorithms used for comparison and performance metrices.

3.1 Dataset description

The comparative and scientific analysis has been performed on publicly available Wisconsin Breast Cancer (WDBC) datasets acquired from UCI machine learning repository. The dataset was originally given by University of Wisconsin comprises of 569 samples (212-Malignant and 357-benign) with each sample having 32 features. These features present the basic characteristics of the breast mass cell nuclei in the image. Generally, 30 real-valued input features present with id and diagnosis field. Field 2 named “diagnosis” involving 2 classes malignant and benign. By Using every cell nucleus 10 real valued variables are calculated. These variables are: radius, perimeter, area, texture, cancavity, compactness, smoothness, cancave points, symmetry, fractal dimension.

3.2 Proposed methodology

Our proposed methodology for early classification is based on Max Voting Procedure. We used multiple models for predicting each data sample. Final prediction of voting classifier is based on majority vote that we get from the majority of the models [35]. After the data preprocessing we use several feature selection techniques listed in Table 2 and apply various algorithms shown in Table 3. In Table 3 results of best performing classification models are specified in bold. Finally, a best feature selection technique is selected and based on predictions of three top performing classifiers a new voting classifier is proposed. Workflow of the proposed methodology is shown in Fig. 1.

Fig. 1
figure 1

Experiment proposed methodology

Full size image

3.3 Dataset preprocessing

Data preprocessing is a useful step that helps to remove noise, inconsistencies and redundancy to achieve high quality data which improves the performance. During the data pre-processing, all the missing values are filled by the mean of the corresponding feature. Formatting of the given dataset is ensured to be consistent. All the incorrect data types of the features changed to their required datatype. Data normalization is performed on the given dataset to make its range consistent. We used z-score normalization on all the features to restrict the range of values between 3 to − 3. Z-score is calculated by the Eq. (1).

$$Z = \frac{X - \mu }{\sigma }$$
(1)

Generally, some of machine learning algorithms does not handle categorical data. Diagnosis feature in this dataset contains two categories M(malignant) and B(benign) is replaced numeric with M by 1 and N by 0. Also, feature ID is removed from the dataset as it is not necessarily required.

3.4 Feature selection techniques

Feature selection is one of the useful stages in data modelling that helps to discard redundant, irrelevant and noisy data. The proposed study selects relevant features by using following feature selection techniques.

Table 2 Summary of selected features using three feature selection techniques
Full size table

3.4.1 Correlation based feature selection (CFS)

This is one of a useful filter method approach to select features which is faster than wrapper-based approach and is suitably used for Higher-dimensional datasets. CFS technique choose relevant features only on the basics of intrinsic characterstics of the data without implementation of any ML algorithms [36]. Often in a dataset some of the features maybe highly corelated to some other attributes. Such highly corelated features gives redundant information that does not contribute in performance improvement. CFS technique calculates the correlation to other features and excludes the highly corelated (similar) features [37]. Suppose the features p1 and p2 are highly corelated means both carry same type of information. The classification model including both p1 and p2 have same predictive power as dataset containing either p1 or p2. Similarly, the features that are highly corelated with the class label are retained. The most widely used pearson correlation coefficient is expressed by Eq. (2):

$$\rho \left( {X,Y} \right) = \frac{{{\text{cov}} \left( {X,Y} \right)}}{{\sigma_{X} \sigma_{Y} }}$$
(2)

3.4.2 Sequential feature selection (SFS)

It is a feature selection technique that reduces n-dimensional feature space to m-dimensional feature space with (m < n) based upon a greedy search approach. It selects or discards features automatically one at a time on the basics of classifier performance until a new feature subset with best performance is obtained [38]. This wrapper approach is slower than correlation because it selects an optimal subset that increases the performance of classifier.

3.4.3 Information Gain (IG) attribute evaluation

IG is a useful feature selection technique that selects features by building a decision tree by using test attribute at every node of DT. This particular approach was introduced by J. R. Quinlan. [39] Consider node N that represent the partition D records from dataset. Finding IG helps in selecting an attribute for splitting at node N. The attribute having maximum IG is carefully chosen for splitting. Higher the IG of an attribute minimizes the information required for classification of objects in partitions. Such an approach helps to improve the performance of decision tree-based algorithms. IG required for classifying objects in partition D is calculate by Eq. (3). Where \(p_{i }\) represents a probability of an object D belonging to class \(c_{i }\).

$$Info\left( D \right) = - \mathop \sum \limits_{i = 1}^{m} p_{i } \log \left( {p_{i } } \right)$$
(3)

3.5 Algorithms used for study

A brief description of the algorithms selected for scientific study is given under:

3.5.1 Logistic regression

It is one of a useful supervised learning algorithm that can solve regression as well as classification problem. This statistical model is used to predict binary values (zero or one). It is widely used in healthcare sector and gives definite output values. Disadvantage of linear regression is capability of predicting only continuous variables but in case of categorical feature logistic regression is found useful [40]. This classification technique is based on the sigmoid or logistic function \(\left( {\frac{1}{{1 + {\varvec{e}}^{{ - {\varvec{t}}}} }}} \right)\). The logistic regression is represented by Eq. (4).

$$LR = \frac{1}{{1 + e^{{ - \left( {Y = \beta_{0} + \beta_{1} X_{1} + \beta_{2} X_{2} + \ldots + \beta_{n} X_{n } } \right)}} }}$$
(4)

3.5.2 K-nearest neighbor

KNN is a non-parametric method of classification and can be used to solve both regression as well as classification problem. This algorithm responds to an input vector where the units are located near each other. It works only on the basis of the stored trained database without construction of any general model so called a lazy learner. It categorises the new input vector on the basis of majority vote of k nearest neighbours irrespective of their labels assigned. In case we have M training vectors, the KNN technique computes K nearest neighbors to the test data, then categorise new input by taking majority vote of classes among k nearest neighbors. It is a distance-based function so scaling and normalization of features is useful in improving performance of K nearest neighbor classifier. The Euclidean distance is calculated by the Eq. (5).

$${\text{Euclidean }} = \sqrt {\mathop \sum \limits_{i = 1}^{m} \left( {x_{i} - y_{i} } \right)^{2} }$$
(5)

3.5.3 Decision tree

It is a supervised machine learning classification algorithm that is constructed for training samples of dataset D on the basics of high entropy values. Construction of tree is very fast and simple by using recursive top-down DAC (divide and conquer) approach. Also, irrelevant samples on D are removed using tree pruning. Entropy is represented by Eq. (6).

$${\text{Entropy}} = - \mathop \sum \limits_{j = 1}^{m} p_{ij} \log_{2} p_{ij}$$
(6)

3.5.4 Random forest

It is a supervised ML learning classification ensemble technique that is based on many decision trees. This ensemble model constructs various decision trees that are incorporated to get the better performance. Mainly bagging or bootstrap aggregating is applied for tree learning. Consider a given data, X = {\(x_{1} ,x_{2} , \ldots ,x_{n}\)} be input vectors and Y = {\(y_{1} ,y_{2} , \ldots ,y_{n}\)} be a response variable with bagging b repeated from 1 to B. Finally, by averaging the predictions of all individual decision tree a new unseen sample \({x}^{^{\prime}}\) are created as Eq. (7):

$$j = \frac{1}{B}\mathop \sum \limits_{b = 1}^{B} fb(x^{\prime})$$
(7)

The standard deviation is used for measuring the uncertainty of their predictions on these trees specified by Eq. (8)

$$\sigma = \sqrt {\frac{{\mathop \sum \nolimits_{b = 1}^{B} \left( {fb\left( {x^{\prime}} \right) - \hat{f}} \right)^{2} }}{B - 1} }$$
(8)

3.5.5 Support vector machine

It is a supervised ML technique that has been effectively used for classifying linear as well not linear problems. Because of its complexity, even it is highly accurate in higher dimensional spaces for classification as well as outlier detection. Let the dataset D having training samples data = {\(x_{i} ,y_{i}\)}| for all i = 1, 2…, n where \(y_{i} and x_{i}\) € \(R^{n}\) be a target item and an ith vector [41]. It represents an optimal hyperplane by \(f\left( x \right) = w^{T} x + b\) where w and b are dimensional coefficient and offset vector.

3.5.6 Neural networks

It is a supervised ML classification algorithm that is inspired by the functioning of human brain. It has three main components i.e., one input layer (\(x_{i} )\), hidden layers and one output layer (\(y_{i} )\). The number of hidden layers present in any ANN is atleast one [40]. The strength of neural network is dependent upon weight associated with it. Activation function plays a vital role in NN for giving final result by adding of a bias value [42]. Non-linearity in NN is achieved by this function and is represented by Eq. (9).

$$f\left( {b + \mathop \sum \limits_{i = 1}^{n} x_{i} u_{i} } \right)$$
(9)

3.6 Performance measures

These classification evaluation measures used are accuracy, Precision, Recall, specificity, F-measure [43]. These metrics are calculated by using the elements of confusion matrix that represents an information about predicted and actual values. The performance metrics are depicted by following Eqs ((10, 11, 12, 13, 14):

$${\text{Accuracy}} \frac{TP + TN}{{TP + TN + FP + FN}}$$
(10)
$${\text{Recall }} \frac{TP}{{TP + FN}}$$
(11)
$${\text{Precision }} \frac{TP}{{TP + FP}}$$
(12)
$$F1 {\text{Score}} \frac{{2 \times {\text{Precision}} \times {\text{Recall}}}}{{ {\text{Precision}} + {\text{Recall}}}}$$
(13)
$${\text{Specificity}} \frac{TN}{{FP + TN}}$$
(14)

4 Results and discussion

In this section, we are going to discuss the prediction results of the breast cancer classification before and after using feature selection techniques using Logistic Regression (LR), Decision Tree (DT), Support vector machine (SVM), Artificial neural network (ANN), Adaboost and XGBoost. Based on these results best feature selection technique is decided and a voting classifier is proposed.

Experimental study implements several state-of-the-art classifiers on the features listed in Table 2 using corresponding feature selection technique. Table 3 compares the accuracy, precision, recall, F-measure of various classifiers by considering all the features along with features using CFS, SFS and Information Gain. Findings shows that Neural network achieved a higher accuracy with measure 98.83% and naive bayes achieved minimum accuracy of 95.90% using all the features. But due to the presence of some redundant features it takes more CPU time and memory for computation. Results shown in Table 3 insights that there is slightly increase in performance of some models but the overall CPU time and memory required is comparatively less than using all the features. Firstly, Correlation based feature selection (CFS) approach helps to select only 11 features having high correlation value with the class(diagnosis) and least correlation value with other features. Features such as perimeter, area and radius show almost linear patterns that insights the presence of multicollinearity between these variables. Similarly, the relationship between compactness, concave_points and concavity also display the presence of multicollinearity. Absence of some of these features may not affect the performance of the model. We select features with filter value 0.65 with the feature diagnosis representing class labels. Features such as ‘radius_worst’, ‘perimeter_worst’ and ‘area_worst’ having correlation value of 0.78, 0.78 and 0.73 with the diagnosis are retained. Further, ‘concave points_mean’, ‘concavity_mean’, ‘concavity_worst’ and ‘concave points_worst’ features having 0.78, 0.7, 0.66 and 0.79 are also retained. ‘texture_mean’, ‘symmetry_se’ and ‘smoothness_worst’ features are least correlated with all other features present are also retained. Machine learning classifiers are developed by considering these 11 features selected using a CFS technique and accuracy is computed for each modelling technique. Table 3 compares the accuracy, precision, recall, F-measure on features selected using CFS. The peak accuracy is achieved by support vector machine with measure 99.11%.

Table 3 Performance comparison of classifiers using different feature subsets
Full size table

Secondly, based on sequential SFS approach used to select or discard one feature at a time until an optimal feature subset with best performance is achieved. Based on SFS we select 7 features named 'concavity_mean', ‘texture_se’, ‘concave points_se’, ‘radius_worst’, ‘perimeter_worst’, ‘texture_worst’ and ‘smoothness_worst’ listed in Table 2. Logistic Regression outperforms all other ML classifiers with accuracy measure of 98.83%. Finally, an optimal feature set selection based on highest Information Gain we consider 14 features that are listed in Table 2. Performance measures for different ML classifiers based on IG are computed shown in Table 3. Based on results shown in Table 3 we have seen that the performance order of various models. Overall individual performance comparison of various ML Classifiers by using four feature subsets (All features, CFS, SFS, Info_Gain) is depicted in “Fig. 2”. It depicts that ANN surpassed all other predictors with least error rate of 1.17% by using all features. The performance order based on CFS SVM > LR > NN > RF > KNN > DT. Overall Highest performance is achieved by SVM Classifier followed by Logistic Regression and neural network classifier as shown in Fig. 2 in case of CFS features. Finally, we make a voting based classifier by using SVM, LR and ANN that predicts results with 99.41% accuracy. In Table 4 we Compared results of our proposed model with various existing models where bold text signifies the accuracy and error of proposed classification model.

Fig. 2
figure 2

Overall performance comparison of various classifier

Full size image
Table 4 Comparison of proposed model with various models
Full size table

5 Conclusion and future scope

In the current era, large volume of data is generated by using various sensors and machines in the healthcare domain named as Big Data. Early classification and analysis of this data is a challenge to predict diseases in early stages and is a hot topic of research. Machine learning and Deep learning evolution helps to do predictive analysis of such high-volume data. In our work we choose well established ML algorithms such as KNN, LR, DT, ANN, SVM, RF, Adaboost etc. based on various feature selection techniques such as CFS, SFS and Information Gain. Features selected using CFS achieves higher accuracy as comparative to other techniques. Finally, we build a voting classifier by combining results of best three models SVM, LR and ANN to classify the new test samples. Results shows that voting classifier can predict results with 99.41% accuracy.

In future, machine learning methodologies can be integrated with neuro-fuzzy and deep learning approaches for efficient diagnosis. A new hybrid models can be proposed by hybridization of Machine learning, CNN/Auto-encoder, neuro-fuzzy systems and swarm optimizations to integrate Optimum feature selection, Feature extraction and Classification for medical imaging data. Further, there are some innovative evolutionary optimization algorithms like Genetic Programming, Simulated annealing, Gradient Descent, Stochastic Optimization, Swarm optimization, SMOTE, Cukcoo search algorithm, ant colony optimization, Hunting search algorithm, Firefly algorithm, lyapunov exponents, Glow worm algorithm, Bat algorithm and wavelet transformations can be combined with Machine learning, Neuro-fuzzy systems, Deep learning and Artificial intelligence methodologies to develop hybrid models.