Machine Learning for Injuries Cause of Death Assignment: A New Method for the Global Burden of Disease Study

Abstract

Globally, injuries were responsible for 8% of deaths in 2017 and have been a neglected source of burden, especially in many low income countries.1,2 Effective public health interventions and decision making rely on accurate estimates of injuries burden; however, inconsistent and unreliable coding of injuries deaths has complicated this task over time. In particular, a large portion of injuries deaths are coded to Exposure to unspecified factor (International Classification of Diseases (ICD) 10: X59) and Unspecified event, undetermined intent (ICD 10: Y34), when there is insufficient information regarding the circumstances of an injuries death.3 These garbage-coded deaths, or deaths assigned to ICD codes that are insufficiently specific or for which death is impossible, have a deleterious effect on cause-specific public health interventions. The Global Burden of Disease Study (GBD) has developed an overall algorithm for redistributing garbage codes to a predefined cause list in order to attribute these deaths to more informative causes.4 This process involves grouping similar ICD codes into “packages” and defining a list of “target causes” for each package to then be redistributed onto by age, sex, location, and year. Current redistribution methods are based on either statistical models, literature review, or expert opinion; however, a growing field of interest in the GBD is the use of multiple cause of death (MCoD) data to inform garbage code redistribution.1 In GBD 2019, a novel regression method was introduced using MCoD data to redistribute X59 and Y34 deaths, however it relied largely on an algebra-based preliminary proportional redistribution method prior to modeling.5 This analysis seeks to improve upon this preliminary method using machine learning.