Efficient Support Vector Machine Toward Medical Data Processing

Abstract

This work explores the efficient and practical scheme of medical data analysis through machine learning algorithms. The support vector machine (SVM) mechanism is specifically employed for building an artificial intelligence (AI) assistant diagnosis systems. Considering the practical demands on clinical diagnosis, the plain SVM algorithm is hardly used since the poor number of classes (typically, two classes) and explosion of samples. Therefore, a sample domain description technology is developed to realize a one-class SVM for flexibly expending the number of classes. Furthermore, a constantly online learning strategy is proposed to implement high-performance classification/diagnosis with greatly reduced database. For proof-of-concept, several medical databases are employed for diagnosis test. From the test results, the diagnosis correct rate is improved with compact database; and the scale of database is reduced while the similar correct rate is achieved by plain SVM algorithm. Maintaining the best accuracy, the proposed online learning SVM reduces the numbers of active samples (support vectors) to $23.4 \%$, $54.6 \%$, $70.9 \%$ of the plain for diagnosing the breast cancer, diabetes, and liver disorders, respectively, where the best accuracy is superior or similar to state-of-the-arts.