
Researchers have recently used artificial intelligence (AI) to better predict outcomes in patients with severe malaria. In their study, quantitative analysis showed features that predicted not only the magnitude of each patient’s outcome, but the dynamic pathways of malaria progression as well. These pathways ultimately allowed the team to find each patient’s individual risk of mortality. This work was recently published in npj Digital Medicine.
Malaria’s Threat to Public Health
Malaria presents as a considerable threat to public health, with 3.3 billion people living in regions where the disease is transmitted through mosquitoes. 445,000 deaths were reported from an estimated 216 million cases of malaria in 2016, with most of these fatalities occurring in sub-Saharan Africa. Identification of severe malaria is intended to target children at the highest risk of death, such as those with cerebral malaria or respiratory distress. It is particularly challenging to identify these individuals, however, and current patient outcomes are failing to improve.
With prevalent data regarding malaria patient outcomes and advances in technologies like machine learning AI, this UK research team saw an opportunity to better understand the dynamics malaria. The scientists used mutual information (MI) to learn clinical factors that predict patient outcomes and an algorithm that learned the dynamic pathways of the disease progression. The authors note that this dual approach is not specific to malaria and that it can be applied to the study of any disease progression.
Background of the Study
The dataset was comprised of 2,915 children aged 4 months to 15 years who meet the WHO criteria for severe malaria. These patients were admitted to the Royal Victoria Teaching Hospital (RVTH) in Banjul, The Gambia between 1997 and 2009. From this set, 2,904 children were used with available outcome data, with 387 of these patients dying (case-fatality rate: 13.3%). The researchers divided the children into 3 different categories based on features of disease severity: respiratory distress (40.2%), cerebral malaria (36.5%), and severe anemia (22.7%).
The team then used their algorithm to reveal the features most strongly correlated with increased risks of mortality. They found that cerebral malaria was the strongest predictor of mortality in the patient cohort. After splitting the cohort into cerebral vs. noncerebral malaria cases, they found respiratory distress to be the second strongest predictor. Further analysis showed abnormal posturing, absence of transfusion and lack of splenomegaly to be tightly correlated to mortality as well.
The researchers also surveyed 11 clinical practitioners in the malaria field to compare to their algorithm’s predictions. These predictions were found to be significantly correlated to the computational ones from the dataset. These survey results correlated better with the inferred results of living patients, rather than those who died.
This dual usage of mutual information and machine learning algorithms allowed these researchers to identify the most informative clinical features that correlate with death in children with severe malaria, as well as the sequence of appearance of these features. This was made possible through the analysis of many patient samples together and putting them all in one framework for disease progression.
“The value of this powerful approach is clear: we can simultaneously learn the dynamic pathways of disease progression, identify key predictors of clinical outcome, and use this unprecedented elucidation of disease dynamics to facilitate novel and clinically informative classification of the clinical risk associated with individual patients,” the authors concluded.