High-Performance Multiclass Classification Framework Using Cloud Computing Architecture

Abstract

Multiclass classification methods that combine a genetic algorithm with support vector machine have been shown to have superior performance in digital biomedical data analyses. However, they usually require a lot of computing resources. Cloud computing has recently played a more important role in this area, as it provides large computing power, storage space, and network bandwidth. The present study proposes a high-performance multiclass classification framework that adopts cloud computing technology to speed up analysis and enhance performance. The MapReduce parallel programming paradigm supported in the Apache Hadoop environment is adopted. A benchmark mRNA dataset of 14 tumor types was used to evaluate the classification framework. Accurate tumor classification based on gene expression profiles (GEP) has great promise for early diagnosis and clinical prognosis of cancers. However, GEP samples are usually very highly dimensional and statistically differentially expressed genes are not always tumor-related. In the experiment, when the number of servers increased from 1 to 10, the training time of the classifier was significantly reduced from 22.06 to 2.33 days. The classification accuracy reached to 94 %, a significant improvement over previous studies (78 and 90 %). Among 16,063 expressed genes, several cancer-related genes were identified after further investigation was made. The experimental results demonstrate the great effectiveness and efficiency of the proposed framework in terms of multiclass biomedical data classification.