Bioinformatics and Medical Applications. Группа авторов
Читать онлайн книгу.
Healthcare and biomedicine are increasingly using big data technologies for research and development. Mammoth amount of clinical data have been generated and collected at an unparalleled scale and speed. Electronic health records (EHR) store large amounts of patient data. The quality of healthcare can be greatly improved by employing big data applications to identify trends and discover knowledge. Details generated in the hospitals fall in the following categories.
• Clinical data: Doctor’s notes, prescription data, medical imaging reports, laboratory, pharmacy, and insurance related data.
• Patient data: EHRs related to patient admission details, diagnosis, and treatment.
• Machine generated/sensor data: Data obtained from monitoring critical symptoms, emergency care data, web-based media posts, news feeds, and medical journal articles.
The pharmaceutical companies, for example, can effectively utilize this data to identify new potential drug candidates and predictive data modeling can substantially decrease the expenses on drug discovery and improve the decision-making process in healthcare. Predictive modeling helps in producing a faster and more targeted research with respect to drugs and medical devices.
AI depends on calculations that can gain from information without depending on rule-based programming while big data is the type of data that can be supplied to analytical systems so that a machine learning model could learn or, in other words, improve the accuracy of its predictions. Machine learning algorithms is classified in three sorts, particularly supervised, unsupervised, and reinforcement learning.
Perhaps, the most famous procedure in information mining is clustering which is the method of identifying similar groups of data. The groups are created in a manner wherein entities in one group are more similar to each other than to those belonging to the other groups. Although it is an unsupervised machine learning technique, such collections can be used as features in supervised AI model.
Coronary illness, the primary reason behind morbidness and fatality globally, was responsible for more deaths annually compared to any other cause [1]. Fortunately, cardiovascular failures are exceptionally preventable and straightforward way of life alterations alongside early treatment incredibly improves the prognosis. It is, nonetheless, hard to recognize high-risk patients because of the presence of different factors that add to the danger of coronary illness like diabetes, hypertension, and elevated cholesterol. This is where information mining and AI have acted the hero by creating screening devices. These devices are helpful on account of their predominance in pattern recognition and classification when contrasted with other conventional statistical methodologies.
For exploring this with the assistance of machine learning algorithms, we gathered a dataset of vascular heart disease from Kaggle [3]. It consists of three categories of input features, namely, objective consisting of real statistics, examination comprising of results of clinical assessment, and subjective handling patient related information.
Based on this information, we applied various machine learning algorithms and analyzed the accuracy achieved by each of the methods. For this report, we have used Naive Bayes, Decision Tree, Random Forest, and various combinations of using these algorithms in order to further improve the accuracy. Numerous scientists have just utilized this dataset for their examination and delivered their individual outcomes. The target of gathering and applying methods on this dataset is to improve the precision of our model. For this reason, we gave different algorithms a shot on this dataset and successfully improved the accuracy of our model.
We suggested using the ensemble method [2] which is the process of solving a particular computer intelligence problem by strategically combining multiple models, such as classifiers or experts. Additionally, we have take the wrongly classified records by all the methods and tried to understand the reason for wrong classification and modify it mathematically in order to give accurate results and improve model performance continuously.
1.1.1 Scope and Motivation
Exploring different classification and integration algorithms to perceive teams in an exceedingly real-world health record data stored electronically having high dimension capacity and find algorithms that detect clusters within reasonable computation time and ability to scale with increasing data size/features while giving the highest possible accuracy. Diagnosis is a challenging process that, as of today, involves many human-to-human interactions. A machine would increase the speed of giving a diagnosis and lead to a more rapid treatment decision and would be able to detect rare events easier than humans.
1.2 Literature Review
Over the years, many strategies have been used regarding data processing and model variability in the field of cardiovascular diagnostics. Authors in [4] show that splitting the data into 70:30 ratio using for tutoring and examination purpose and 10-fold cross proofing putting logistic regression into operation improved the accuracy of the UCI dataset to 87%.
Authors in [5] have used ensemble classification techniques using multiple classifiers followed by score level ensemble for improving the prediction accuracy. They pointed out that maximum voting produces the highest level of development. This functionality is enhanced by using feature selection.
Hybrid approach has been proposed in [6] by consolidating Random Forest along with Linear method leading to a precision of around 90%. In [7], Vertical Hoeffding Decision Tree (VHDT) was used accuracy of 85.43% using 10-fold cross-validation.
Authors in [8] outline a multi-faceted voting system that can anticipate the conceivable presence of coronary illness in humans. It employs four classifiers which are SGD, KNN, Random Forest, and Logistic Regression and joins them in a consolidated way where group formation is performed by a large vote of the species making 90% accuracy.
The strategy utilized in [9] finds these features by way of correlation which can help enhanced prediction results. UCI coronary illness dataset is used to evaluate the result with [6]. Their proposed model accomplished precision of 86.94% which outflanks Hoeffding tree technique which reported accuracy of 85.43%.
Different classifiers, mainly, Decision Tree, NB, MLP, KNN, SCRL, RBF, and SVM have been utilized in [10]. Moreover, integrated methods of bagging, boosting, and stacking have been applied to the database. The results of the examination demonstrate that the SVM strategy utilizing the boosting procedure outflanks the other previously mentioned techniques.
It was exhibited in [11] after various analyses that, if we increase the feature space of RF algorithm while using forecasts and probability of a tuple to belong to a particular class from Naive Bayes model, then we could increase the precision achieved in identifying the categories, by and large.
Studies in [12] suggested that Naive Bayes gives best result when combined with Random Forest. Also, when KNN is combined with RF or RF+NB, the errors remain same suggesting that it is the dominating method.
Authors in [13] compared the precision of various models in classification of coronary disease taking Kaggle dataset of 70,000 records as input. The algorithms used were Random Forest, Naive Bayes, Logistic Regression, and KNN among whom Random Forest was the winner with an accuracy of 73%.
Creators in [14] have fused the results of the AI examination applied on different informational collections focusing on the CAD illness. Common features are compared and extracted from different datasets, and advanced concepts such as fast decision trees and pruned C4.5 tree are administered on it resulting in higher classification accuracy.
Ensemble Optimization is applied in [15] wherein fuzzy logic is used for extraction of features, Genetic Algorithm for reducing them and Neural Network for classifying them. The results have been tested on a sample of size 30 and accuracy achieved is 99.97%
Based on the detailed research discussed above, we analyze by comparing different strategies suggested by different authors in their respective papers. This helps us to quickly understand where we stand presently with respect to these techniques