Original title: Automating narrative analysis to enhance computer-coded verbal autopsies: a scoping review
Mahadia Tunga, Sangyal Dorjee, Isaac Lyatuu, Erin Nichols +2 · September 2026
When people die outside hospitals, health workers interview family members to understand what happened. This interview is called a verbal autopsy. Usually, computer programs figure out the cause of death by reading simple multiple-choice questions. These programs mostly ignore the open-ended written stories also collected during the interviews. These unstructured stories are tricky for computers to process. Researchers reviewed studies that use artificial intelligence to read these death narratives. These tools turn written text into computer-readable numbers and patterns, spotting clues like timelines and context. The researchers found that reading the text alone often works worse than using multiple-choice questions. However, mixing both the written stories and the structured questions improves predictions compared to using just one type of data alone. Still, accuracy varies widely between studies because researchers used different methods and data, showing that researchers need open, updated datasets to compare models fairly.
You are trying to guess what illness killed someone by reading a medical form. The form has checkboxes for symptoms, but also a blank space where the family wrote a free-hand account. You have always ignored that account because it is messy and different each time. Now imagine you start reading those accounts and picking up clues from the story itself—it fills in missing context. Here, the form is the verbal autopsy, the checkboxes are the structured questions, and the blank space is the narrative. This paper reviews how computers can be taught to read that blank space, and finds that combining it with the checkboxes gives a better guess than either alone.
Where a doctor isn't available to certify death, computers already assign causes using checklist answers. But families also tell a story about what happened—and that story holds context computers skip. Adding those narratives could sharpen mortality data in the places that need it most.
Performance varies substantially across studies due to differences in datasets, preprocessing, feature representation, and evaluation procedures—so no single best method is identified. Narrative-only models often underperform those using only structured questions. The review does not propose new algorithms or run new experiments, and it highlights the need for open, up-to-date datasets for fair comparison.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Automating narrative analysis to enhance computer-coded verbal autopsies: a scoping review — Mahadia Tunga, Sangyal Dorjee, Isaac Lyatuu et al. — Population Health Metrics — doi:10.1186/s12963-026-00507-z
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper.