Identifying the Markers of Suicidal Ideation among Twitter Users in Sub-Saharan Africa.

Emmanuel Olamijuwon, University of St Andrews
Elton Mukonda, University of Cape Town
Ronald Musizvingoza, United Nations University
Garikayi G. B. Chemhaka, University of the Witwatersrand and University of Swaziland
Caroline Kiarie, University of Kwazulu Natal in Durban
Vissého Adjiwanou, Université Du Québec à Montréal

Although preventable, about one million people die from suicide every year. The monitoring and surveillance of suicidal ideation could help reduce its prevalence. However, this is lacking in Africa. We attempt to identify the markers of suicidal ideation and assess whether a machine learning technique can be used to correctly classify suicidal tweets. We collected 6,025 tweets with keywords such as “kill myself”, “end my life” and many others posted in May 2019. Human annotators were asked to classify the downloaded tweets as either suicidal or safe to ignore. We trained the random forest model on a rebalanced 80% of the data and validated the model on the remaining 20%. The results suggest that the model could accurately classify suicide ideation in tweets. The overall accuracy of the random forest was 82%. The model was correctly classified 71% of suicide ideation tweets and 92% of safe to ignore tweets.