Ph.D in computer Engineering, Assistant Professor of Computer Engineering, Department of Computer Engineering Dept., Yazd University, Yazd, Iran.
Abstract: (5802 Views)
Introduction: Combination of artificial intelligence and data mining has been resulted to considerable progress in the prevention and diagnosis of diseases. Complex models have been proposed for the diagnosis of acute leukemia from genetic information, but significant results have not been achieved. This study aimed to predict the type of blood cancer by examining a wide range of parametric and non-parametric methods and to increase the generalization of learning by extracting fewer essential features.
Methods: This descriptive and analytical study used Leukemia1 dataset from the Vanderbilt University of USA. This dataset contains a set of bone marrow and blood samples of patients having leukemia used for classification based on three subgroups of leukemia, namely ALL B-cell, ALL T-cell and AML. Parametric classification including linear algorithms, Naïve Bayes, Euclidean distance, nearest average, template matching as well as non-parametric classification using basic estimator algorithms, kernel, k-nearest neighbors and k-nearest neighbors based on the kernel has been used.
Results: Considering all features, the best method was nearest mean prediction method achieving the accuracy of 92.86%. By applying the PCA feature reduction method, too, the best result was related to the nearest mean algorithm and by average number of features of 6.8, the accuracy became 96%. Finally, using data-balancing methods and quadratic algorithm resulted in the average number of features and the accuracy of 5.41 and 98.59% respectively.
Conclusion: The results show the effectiveness of essential features extraction in improving the accuracy of Bayes-based models and its preference over the existing complex models.
Type of Study:
Original Article |
Subject:
Data Mining Received: 2017/11/16 | Accepted: 2018/05/7