b iframe width height src
“Breast cancer” refers to a malignant tumor resulted from the erratic growth of Dexmedetomidine that originate in the breast tissue (Akay, 2009). A recent survey of US population investigated that there were 1,685,210 new cancer cases and 595,690 cancer deaths in 2016, among which breast cancer (BC) is the leading cancer death cause for women aged 20 to 59 years (Siegel, Miller, & Je-mal, 2016). Due to the high annual increasing rate, approximately 252,710 new BC cases will be diagnosed in 2017, affecting 0.15% of US women population (DeSantis, Ma, Goding Sauer, Newman, & Jemal, 2017; United Nations, Department of Economic & Social Af-fairs, Population Division, 2017). Based on a recent study, a woman
∗ Corresponding author. E-mail addresses: [email protected] (H. Lu), [email protected] (H.
Wang), [email protected] (S.W. Yoon).
living in the US has a 12.4% lifetime risk of being diagnosed with BC (DeSantis et al., 2017). As one of the most common cancer types among US women, it is necessary to develop an effective and ef-ficient method in BC diagnosis and prognosis cases from its early stage.
Effective risk analysis for BC patients is a critical factor for clin-icians to accomplish the central issues in BC treatment, such as implementing effective treatment strategies and fostering informed decisions by patients (Pace & Keating, 2014). BC survivability pre-diction, as a major clinical risk prognosis problem, significantly fa-cilitates the decision making of clinicians by grouping BC patients into different risk levels. Hence, research for BC survivability pre-diction has been conducted during recent years where data-driven BC survivability prediction has been extensively explored. The mo-tivations of this research include improved patient data manage-ment systems, continuous updates of machine learning technolo-gies, and statistical evidence that certain patient clinical data and
microarray data can indicate the chance of survival. The clinical data include patient history, body indexes, etc. The microarray data are extracted from a selection of probes on tumor samples, which contain reverse transcribed mRNA to represent the gene expres-sion levels of tumor samples (Gevaert, Smet, Timmerman, Moreau,
& Moor, 2006). Both the clinical and microarray data are sequen-tially collected. Given that genetic information is associated with microarray data, the datasets are generally large.
Prevailing data-driven classification techniques that have been intensively evaluated for BC research include the Support Vec-tor Machine (SVM) (Zheng, Yoon, & Lam, 2014), Artificial Neural Networks (ANN) (Park et al., 2013), ensemble learning techniques (Kourou, Exarchos, Exarchos, Karamouzis, & Fotiadis, 2015), etc. The state-of-art techniques are mainly o ine models that require one-pass training of the entire dataset. When incoming patient data are incrementally added to clinical systems, single pass training be-comes unfeasible; e.g., large scale microarray data are included as indicators. In addition, with the progress in diagnostic technolo-gies, the instances and dimensions of data related to tumor gene expression are expanding, and it’s necessary to address computa-tional complexities and memory costs. As a result, adaptive ma-chine learning models that are effective in processing data streams are highly desired. Online learning, as a typical sequential learn-ing technique, has been sparsely studied for BC diagnosis or prog-nosis. Among the limited online classification models for BC re-search, limited research has focused on improving the adaptiveness and stability of online learning models. It is necessary to tackle the challenges to implement practical online learning models.