In regions of the world lacking resources to conduct comprehensive autopsies, health officials often establish probable cause of death by interviewing people who knew the decedent and then filing written reports. The process is known as a “verbal autopsy.”
Now researchers at the University of Toronto have shown AI can automatically draw conclusions from the reports.
The advance could represent an important step forward for the developing world, where many deaths occur in homes rather than hospitals, obscuring information that might inform public health planning and resource allocation.
Working with data from India’s Million Death Study, the researchers designed their method to analyze language after manual preprocessing of free-text reports and correcting of spelling mistakes.
The technique uses these narratives as inputs for four different machine learning classifiers—naïve Bayes, random forest, support vector machines and a neural network.
Led by senior study author Graeme Hirst, PhD, a professor of computational linguistics, the team found that, for adult deaths—the largest group of deaths in their dataset—their best AI model hit 90% agreement with causes of death as assigned by physician experts.
“No current method for automatically determining cause of death for verbal-autopsy records has sufficient accuracy to be a replacement for human doctors,” the authors concluded. However, their method’s performance “demonstrates that narrative-based machine learning methods are a promising option.”
Hirst and colleagues note that similar methods of text-based machine learning “could be applied to other tasks in the healthcare domain, such as automatic diagnosis or treatment recommendations based on hospital records.”
The study was published July 9 in BMC Medical Informatics and Decision Making and is available in full for free.