A dynamic gradient boosting machine using genetic optimizer for practical breast cancer prognosis. (February 2019)

• Record Type: Journal Article
• Authors: Lu, Hongya; Wang, Haifeng; Yoon, Sang Won
• Is Part Of: Expert systems with applications. Volume 116(2019)
• Page Start: 340
• Page End: 350
• Publication Date: 2019-02
• Subjects: Breast cancer prognosis -- Online learning -- Gradient boosting -- Genetic algorithm -- Adaptive linear regression; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2018.08.040
• Languages: English
• ISSNs: 0957-4174

Abstract

Highlights: Devising an adaptive gradient boosting model using online weak learners. Enhancing the online boosting performance using a genetic optimizer. Dynamic breast cancer prognosis using the proposed technique. Comprehensive evaluation on state-of-art online learning techniques. Validation of the proposed technique on benchmark datasets. Abstract: This research proposes a novel genetic algorithm-based online gradient boosting (GAOGB) model for incremental breast cancer (BC) prognosis. The development of clinical information collection technologies has brought in increasingly large amounts of stream data for BC research. Traditional batch learning models have shown limitations in: (1) real-time prognosis accuracy from losing the information of incremental changes of a patient's pathological condition by time; (2) high redundancy due to the time required to retrain models every time new data are received. Online boosting is an efficient technique for learning from data streams. However, difficulties in parameter assignment and the lack of adaptiveness for batch learning base learners can degrade the performances of typical online boosting algorithms. The main objective of this research is to propose an incremental learning model for BC survivability prediction. To render a boosting algorithm with superiority in global optimal parameters, the genetic algorithm (GA) is integrated to an online gradient boosting scenario at the parameter selection phase, enabling real-time optimization. To enhance adaptiveness, an adaptive linear regressor is adopted as the base learner with minimal computational efforts, and updated in symphony with the online boosting model. The proposed GAOGB model is comprehensively evaluated on the U.S. National Cancer Institute's Surveillance, Epidemiology, and End Results (SEER) program breast cancer dataset in terms of accuracy, area under the curve (AUC), sensitivity, specificity, retraining time, and variation at each iteration. Experimental results show that the proposed GAOGB model achieves statistically outstanding online learning effectiveness. With a highest 28% improvement on testing accuracy over its base learners, outperforming current state-of-art online learning methods, and approximating batch learning boosting algorithms, the GAOGB algorithm validates the impact of parameter, adaptiveness and convergence in devising practical online learning algorithms. The proposed GAOGB model demonstrates potential for practical incremental breast cancer prognosis, promising a combination of training effectiveness and efficiency.