Classifying verbal autopsy records by cause of death using neutral networks and temporal reasoning

Dr. Graeme Hirst
Department of Computer Science, University of Toronto

It is important for public health planning and resource allocation for
authorities to have statistics on the varying causes of death in each
region. But in developing countries, where people are more likely to die
at home than in a hospital and where there are insufficient resources
for physical autopsies, the cause of a person’s death is frequently
never formally established by a physician. To mitigate this problem, the
family of the deceased may be interviewed about the circumstances of
death, resulting in a so called “verbal autopsy”, which will be
subsequently coded by physicians with their judgement as to the cause of
death. We aim to develop automated text-analysis methods assist in this.

Current automated methods primarily use structured data from the verbal
autopsies to assign a cause-of-death category, but the results have been
poor. We present a neural-net-based classification method based on
textual features to automatically classify cause-of-death categories
from free-text verbal autopsy narratives alone. Features used, in
addition to lexical cues, include events, temporal sequences, and
symptom words. We are presently porting the system from English to


Time: Wednesday, 3rd of July 2019, 6:30 p.m. sharp

Location: Oesterreichisches Forschungsinstitut
fuer Artificial Intelligence, OFAI
Freyung 6, Stiege 6, 1010 Wien

Univ.-Prof. Ing. Dr. Robert Trappl