A Framework for Blended Sub Feature Engineering for Chronic Disease Prediction Using in-Memory Computing

Chronic illnesses are widely recognized as the most severe and deadly diseases in humans. The rising prevalence of chronic illnesses with high death rates poses a considerable danger and burden to healthcare systems globally. Chronic illnesses are more common in males than females, especially in middle- and old-age populations, while youngsters with similar health conditions exist. When educated on appropriate data, machine learning algorithms can be excellent in identifying illnesses [1]. Heart disease datasets are freely available for model comparison. With the advent of machine learning and artificial intelligence, academics may now construct the greatest prediction models by employing the massive databases that are now available. The latest evidence on heart-related issues in adults and children has emphasized the importance of lowering chronic disease mortality. Because the clinical datasets supplied are inconsistent and redundant, proper pre-processing is essential. It is critical to choose the key traits that may be employed as risk factors in prediction models. Various supervised models using feature selection methods such as AdaBoost (AB), Decision Tree (DT), Gradient Boosting (GB), Stochastic Gradient Descent (SGD), Lasso Regression (Lassos), and Random Forest (RF) are used in this work, along with classifiers. The findings are compared to previous research [1].

Feature extraction applications pose new challenges in the selection of streaming features [2]. The feature extraction applications have several characteristics, including a) characteristics are evaluated consecutively with a set number of occurrences; and b) the trained model does not exist in advance. In a text mining assignment for spam filtering, for example, additional features (e.g., words) are dynamically created and must therefore be exploited to filter out the spam instead of waiting for every characteristic to be collected. Traditional methodologies, which have not been developed for streaming information applications, cannot be employed in this situation since they demand that the whole extracted feature set be known beforehand to evaluate the effective attributes effectively and scientifically [2, 3]. Parkinson's disease is a widespread neurological disorder. One of the early warning indicators is a speech or voice issue. Acoustics and voice communication system technologies can evaluate and measure the effects of Parkinson's disease on the voice. The current research proposes a unique feature extraction framework based on two different stages of the subset of features approach for diagnosing voice loss in Parkinson's patients. The PCA and ECFS techniques are computed individually during the initial level of selection, and the selected features from each method are treated as separate lists, namely the ECFS chosen features sublist and the PCA picked features sublist, in the first set [4, 5]. Traditional feature selection techniques presume that prior to learning, all data instances and attributes are known. However, we are much more likely to encounter streams of data, feature streams, or both in many real-world applications. Feature streams are features that appear sequentially over time while the number of training samples remains constant. Existing streaming feature selection algorithms concentrate on removing unneeded and redundant features and selecting the most relevant features, but they ignore feature interaction. A quality might have a poor connection with the goal idea on its own, but when paired with other characteristics, it may have a high correlation with the goal concept [6, 7]. Feature selection is a crucial part of good learning algorithms, and it also helps to explain machine-driven judgments. Algorithms such as decision trees and the Least Absolute Shrinkage and Selection Operator (LASSO) may be used to choose features during training. These embedded techniques, however, are limited to a subset of machine learning models. Wrapper-based approaches, which are often computationally expensive, can extract features from machine learning models. Many randomized strategies are now being developed to improve their efficiency [8]. Because rolling bearings are one of the most important components of rotating machinery, this study proposes a balanced multidimensional feature fusion-based feature selection and fusion technique. To begin, features are taken from many domains to construct the initial high-dimensional feature collection. Given the huge number of erroneous and redundant features in the initial feature set, a feature selection strategy was developed that incorporated support vector machines (SVM), single feature analysis, correlation coefficients, and principal component analysis. weighted load analysis [8]. Multi-label feature selection has increased in popularity in deep learning, statistic processing, and related domains, and it is currently widely used for a broad range of tasks ranging from music identification to text mining, picture annotation, and so on. Traditional multi-label attribute selection procedures, on the other hand, employ a cumulative accumulation strategy to create solutions, which has the disadvantage of overvaluing the redundancy of some candidate features. As a data preparation strategy, feature selection has been proven to be successful and effective in preparing data (especially high-dimensional data) for various machine learning and data mining concerns. The aims of feature engineering include producing simpler and more transparent models, increasing data mining performance, and giving clean, intelligible data. The growing amount of big data has presented some important challenges and opportunities for feature selection. In this research, we give a complete and systematic assessment of current advances in attribute selection studies. Motivated by current difficulties and opportunities in the big data era, we investigate feature selection research from a data perspective and examine sample feature selection strategies for traditional data, structured data, heterogeneous data, and streams [9]. Because of the vast number of different, highly changeable inputs in today's decision-making system, feature selection is unavoidable. It is vital to pick the suitable feature set to prevent computational strain and educate algorithms to function correctly. If only a few elements are chosen, it is possible that adequate results may not be obtained; if many characteristics are selected, productivity may decrease. To improve accuracy, more salient points can be included [10].

Recursive feature selection and the six ensemble feature selection methods were used to create this system. There have previously been several studies on various sorts of feature selection algorithms based on classifiers. We picked three of the most often used techniques (DT, RF, and LASSO) as well as three less commonly used approaches (SGD, GB, and ADA). A previous study has discovered that existing feature algorithms have a high level of expected accuracy when compared to other existing approaches. In addition, a limited number of tests have revealed that the ensemble work of wrapper-based feature selection may perform rather well with extremely high accuracy. Except for DT and kNN, none of the research endeavors came close to our proposed techniques as a basic classifier. As a result, all of the prior approaches have been studied further in this work, utilizing ensemble techniques to make the proposed model more efficient. Although the literature review shows that hypotheses put forward yielded good prediction accuracy, it was not high enough in comparison to our investigation.

5.1 The purpose of feature selection

The goal of feature selection in ML is to determine the optimal collection of characteristics that will allow one to develop meaningful and constructive models of the subject under investigation. Machine learning feature selection approaches are classified as follows:

5.1.1 Supervised

These approaches are utilized for labeled data and to categorize the important characteristics for boosting the effectiveness of supervised models such as classification and regression.

5.1.2 Unsupervised

These approaches are applied to unlabeled data.

7.1 Recursive feature elimination (RFE)

It is a popular feature selection method. RFE is popular because it is easy to set up and use, and it is effective at identifying features (columns) in a training dataset that are relevant for predicting the target variable. When using RFE, there are two critical configuration variables to consider: the number of features to select and the technique used to aid in selecting features. Both model parameters can be explored, although their importance in the method's effectiveness cannot be overstated [12, 13].

7.2 RFE with random forest

Random Forest is a well-known machine learning technique for supervised learning. It may be used in machine learning for both classification and regression problems. It is based on the concept of ensemble learning, which is the process of combining several classifiers to solve a complex problem and improve the model's performance [14, 15].

7.3 RFE with decision tree

A decision tree is a non-parametric supervised learning approach that may be used for both classification and regression. It has a tree structure with a root node, branches, internal nodes, and leaf nodes. A decision tree begins with a root node that has no incoming branches. Outgoing branches from the root node feed into internal nodes, also known as decision nodes. Based on the specified features, both node types perform assessments to yield homogeneity subsets known as leaf nodes or terminal nodes [16,17].

7.4 RFE with Lasso

"LASSO" is an acronym that stands for "Least Absolute Shrinkage and Selection Operator." This model employs shrinkage. Shrinkage simply means that the data points are realigned by imposing a penalty that causes the coefficients to be reduced to zero if they are not significant. It makes use of the L1 regularization penalty method. This type of regression is most suited for models with a high degree of multi-collinearity or when we need to automate model selection aspects, such as parameter removal or feature selection [18, 19].

Lasso loss functions can be represented mathematically.

$\sum_{i=1}^n\left(y_i-\sum_j x_{i j} \beta_j\right)^2+\alpha \sum_{j=1}^m\left|\beta_j\right|$

7.5 Reasons for selecting above methods

The above methods deal with large datasets and reduce the chances of overfitting the data. Recursive feature elimination with the above methods provides output effectively.