Circumstances associated with injury-related deaths and changes in causes of injury-related deaths over three decades in rural northeast South Africa: Insights from text mining of verbal autopsy narratives from Agincourt

Abstract

Injuries remain an underprioritized public health issue despite contributing to approximately 5.3 million deaths annually and leaving millions with temporary or permanent disabilities. Low-to-middle-income countries bear the greatest burden from injuries. However, due to the lack of comprehensive medical certification of deaths occurring outside health facilities, most injury-related deaths may remain unreported. Verbal autopsy (VA) serves as a crucial method for assigning causes of death in such settings, offering valuable insights into the sequence of events leading to death. However, VA narratives consist of highly unstructured text, making analysis a challenge. This retrospective longitudinal study employed natural language processing (NLP) techniques, including n-gram analysis, Latent Dirichlet Allocation (LDA), and BERTopic, to extract meaningful patterns from VA narratives. Data were drawn from the Agincourt Health and Demographic Surveillance System, covering deaths recorded between January 1, 1993, and December 31, 2023. The text-mining workflow involved five key steps: data acquisition, pre-processing, feature extraction, topic modeling, and knowledge discovery. Results revealed that road traffic accidents, homicide, and suicide were the predominant causes of injury-related deaths over the 30-year period (1993-2023) while drowning and fall-related fatalities remained consistently low. Road accidents primarily involved vehicle collisions and pedestrian fatalities, with most victims dying at the scene and only a few accessing medical care. Homicide was largely driven by gun and knife violence, often influenced by alcohol, particularly in stabbing-related incidents. Suicide cases were primarily due to hanging, with a smaller proportion involving poison ingestion. Additionally, falls were more prevalent among the elderly, while drowning incidents disproportionately affected younger individuals. This study provides critical insights into the circumstances surrounding injury-related deaths, the possible risk factors contributing to injuries, and the most prevalent types of injuries by analyzing unstructured textual data from verbal autopsy narratives. The findings underscore the effectiveness of automated text-mining techniques in extracting meaningful patterns and enhancing injury surveillance. These results can be used to strengthen injury prevention efforts through targeted, information-driven interventions aimed at reducing the overall injury burden.