Sixty million people die each year, most with no known cause of death (CoD). According to a 2012 report by the United Nations Department of Economic and Social Affairs, Population Division, two-thirds of the deaths each year have no recorded CoD. The highest ratios of the unknown-to-known cause of death are in developing countries where many people die at home. Verbal autopsies (VA) with structured surveys with family members of the deceased are the most common way to determine CoD.
Nonmedical personnel administers the surveys and two or more physicians review them to determine the CoD. Verbal autopsies also include free-text narratives from family members who are asked to describe the events and circumstances around the time of death.
Researchers from the University of Toronto’s Department of Computer Science, the Vector Institute for Artificial Intelligence, and others have published several studies that compared artificial intelligence machine learning classifiers with human doctor VA coding. The long-term goal is AI tools that could reduce the time and costs of physician coding. To date, studies published in 2014 and 2015 did not show high correlations between human and AI conclusions. A study published earlier this month finally reported marked success. The study used data from India’s Million Death Study (MDS), a government CoD program to determine the estimated leading CoDs.
The researchers did not find comparable results to doctor coding from machine language classifiers with individual CoD classifications. The AI classification engine did, however, perform well with population-level CoD distributions. The machine learning engine and the doctor CoDs agreed 96% of the time with 15 CoD categories and 91% with 48 additional categories. The machine also showed it could score well with free-text narratives, not just completed surveys which opens avenues for future study of words-based relationships with specific CoDs.
There is no plan at this point to implement machine learning CoD coding. The Toronto researchers suggest a path for eventual implementation could include physician review of machine CoD coding, a change that would significantly reduce the time required from human doctors.
