Verbal autopsy through integrated language analysis (vital)

The National Institute of Child Health and Human Development (NICHD) awarded The Johns Hopkins University a $434,068 Project Grant effective September 15, 2025 through August 31, 2027, under the Child Health and Human Development Extramural Research program (CFDA 93.865). The VITAL (Verbal Autopsy Through Integrated Language Analysis) project addresses critical gaps in global mortality surveillance by developing innovative natural language processing (NLP) and artificial intelligence (AI) tools to ascertain cause of death from verbal autopsy narratives. The project encompasses three primary objectives: assembling the most extensive and heterogeneous collection of verbal autopsies paired with gold-standard, medically certified causes of death to enable machine learning model training; evaluating both classical and cutting-edge NLP algorithms for cause-of-death determination while conducting detailed error analysis to identify potential demographic and contextual bias; and applying advanced statistical methods to correct for algorithmic biases and improve the accuracy of epidemiological estimates regarding mortality patterns and trends across populations.

The University of Washington serves as a sub-grantee on this award, contributing to the streamlining of verbal autopsy processes through AI-enhanced natural language processing and model development. The sub-grantee's efforts focus on developing AI models for cause-of-death prediction, evaluating the integration of AI-estimated causes into epidemiological analyses, and piloting the refined methodology in Bangladesh. These coordinated efforts directly support improved mortality monitoring and evidence-based public health policy development in regions where medically certified death records are unavailable, ultimately addressing significant gaps in global health surveillance infrastructure.

Abstract

Understanding trends and patterns in cause of death is essential for designing and implementing effective public health policy. In large parts of the world, most deaths happen outside of a healthcare setting and there is no medically certified record of the cause, stymying efforts to improve health in such settings and also creating downstream effects for health systems globally. Verbal autopsies (VAS) consist of interviews with a person familiar with the circumstances surrounding a recent death and involve both a structured interview and a free-text narrative. The narrative is currently underutilized or wholly ignored in most settings. Our proposal leverages innovative natural language processing (NLP) tools to ascertain cause of death from VA narratives. Our first aim produces the most extensive and most heterogeneous collection of VAS with gold standard (medically certified) causes, which is essential for training machine learning and artificial intelligence (AI) models. Aim 2 critically evaluates both classical and cutting-edge NLP models to ascertain the cause of death. Along with evaluating the accuracy in ascertaining cause of death, we perform a detailed error analysis to understand potential sources of bias across demographic or other contextual factors that arise from the combination of currently available NLP algorithms and available data to train these models. Ascertaining cause of death is incredibly challenging, and even the most advanced algorithm will make mistakes. Ignoring this reality leads to inaccurate and biased estimates of patterns and trends in cause of death (e.g., the fraction of deaths due to malaria from urban areas compared to rural areas or whether the burden of cardiovascular disease has changed from one time to another). Aim 3 leverages cutting-edge statistical tools and the datasets assembled in aim 1 to correct for these biases. We also perform detailed sample size calculations to understand the number of VAS required to detect changes in patterns and trends. 1