Addressing Optimization Challenges in Health Survey Research Using Multi-Objective Constrained Binary Particle Swarm Optimization

• Activity Number: 53 - Applications of Data Linkage and Machine Learning Techniques
• Type: Contributed
• Date/Time: Monday, August 3, 2020 : 10:00 AM to 2:00 PM
• Sponsor: Survey Research Methods Section
• Author(s): Di Xiong* and Honghu Liu
• Companies: UCLA SPH and UCLA
• Keywords: Feature Selection; Partial AUC; Crowding Distance; Multi-objective optimization; Synthetic Minority Over-sampling Technique (SMOTE)

Abstract

Effective screening surveys can assist in detecting early diseases among high-risk children who need treatment intervention. However, it is challenging to optimize the survey protocol for low incidence diseases with large health surveys. This paper proposes a Multi-objective Constrained Binary Particle Swarm Optimization (MCBPSO) method to identify effective and optimal survey items for disease detection. The algorithm balances on dual objectives, minimizing feature redundancy and maximizing partial AUC (Area under the ROC curve) with a constraint sensitivity at 0.8 for training data. Meanwhile, it realizes the variability by controlling velocity and the best performance of the swarm using mutation and resetting operators. Multiple machine learning algorithms were ensembled by a Super Learner to improve prediction performance. The proposed algorithm is applied to a recent oral health survey of children with 192 self-reported items. MCBPSO-based feature selection algorithms can be effectively applied to detect diseases with a low incidence rate. The cost-effective screening toolkit developed can be used in oral health screening for large school-age children in the future.