Prospects for automated diagnosis of verbal autopsies

Verbal autopsy is a method for assessing probable causes of death from lay reporting of signs, symptoms and circumstances by family members or caregivers of a deceased person. Several methods of automated diagnoses of causes of death from standardized verbal autopsy questionnaires have been developed recently (Inter-VA, Tariff, Random Forest and King-Lu). Their performances have been assessed in a series of papers in BMC Medicine. Overall, and despite high specificity, the current strategies of automated computer diagnoses lead to relatively low sensitivity and positive predictive values, even for causes which are expected to be easily assessed by interview. Some methods have even abnormally low sensitivity for selected diseases of public health importance and could probably be improved. Ways to improve the current strategies are proposed: more detailed questionnaires; using more information on disease duration; stratifying for large groups of causes of death by age, sex and main category; using clusters of signs and symptoms rather than quantitative scores or ranking; separating indeterminate causes; imputing unknown cause with appropriate methods.