Precision identification of high-risk phenotypes and progression pathways in severe malaria without requiring longitudinal data

Mutual information approach to identify features predicting mortality. At each level (horizontal axis), patient data are greedily split into two subsets according to the remaining feature that most strongly predicts mortality. The algorithm stops when no feature is statistically significantly associated with death. The figure shows a tree generated by this algorithm: cerebral malaria (CM), respiratory distress (RD), splenomegaly (SP), abnormal posturing (PO) and transfusion (TF) are selected as informative features. Nodes are shown as pie charts representing the composition of WHO classifications in each cluster. Solid (dashed) edges indicate that the feature was present (absent) and their width is proportional to the number of patients. The vertical axis corresponds to the mortality log odds ratio compared with the average mortality. Partition 8 has infinite log odds ratio (LOR) because all patients surviv