Identifying Feature Pattern for Weighted Imbalance Data: A Feature Selection Study for Thoracolumbar Spine Fractures in Crash Injury Research

Abstract

In motor vehicle crash study, spine injury investigation has a greater impact due to the serious physical, mental and financial consequences. Even though spine fracture deteriorates the quality of life significantly, to the best of our knowledge, there is no study that searched for the exhaustive thoracolumbar spine fracture(TL-fx) feature space to discover potential feature pattern in the motivation of illustrating the increasing risk phenomenon as a function of vehicle model year. This study investigates National Automotive Sampling System Crashworthiness (NASS-CDS) database, year 2000 to 2015. Each year, approximately 4000 to 6000(weighted) occupants are diagnosed with one or multiple TL-fx in road crashes. Even though the TL-fx data support is less than 1.6%, a two-fold feature selection model in a combination of random forest and lift measure based Apriori algorithm generates insightful association rules yielding prominent feature patterns and promotes further investigation to build causal model for the TL-fx study.