Chu, Y., & Clark, S.. (2025, April). Automating cause of death classification from verbal autopsy: using pretrained language models and multi-modal learning [Poster presentation]. Population Association of America Annual Meeting, Washington DC.

In this study, we explored the performance of current state-of-the-art pretrained language models in biomedical/clinical domain on cause of death (COD) classification with verbal autopsy (VA) narratives. We also proposed multi-modal learning frameworks to combine knowledge from structured and unstructured data for COD classification using VA.   
Timely and accurate estimates of COD are essential for informing health policy and monitoring population health. In countries lacking civil registration and vital statistics systems, VA serves as a key source of COD data. VAs collect information from next-of-kin or caregivers of the deceased about details on the signs, symptoms and circumstances preceding the death, through both unstructured narratives and structured questions. The narratives play a crucial role in conventional physician-coded COD assignments. However, current automated COD classification algorithms rely solely on structured data, overlooking the rich information embedded in the VA narratives.