AI advances malaria prediction outcomes

AI in healthcare is advancing many areas of care and diagnosis, but it is also helping predict the clinical risk and severity of malaria.

Researchers from Imperial College London applied machine learning and model-based inference tools to detect patterns in clinical features association with severe malaria in a sample of nearly 3,000 Gambian children who were admitted to the hospital with malaria. Their findings were published in npj Digital Medicine.

Severe malaria claims more than 400,000 lives annually, with most deaths being African children, according to researchers. However, “the sequence of events leading to SM is poorly understood,” and determinants of risk––and death––from malaria are rarely captured in clinical studies, according to study author Iain G. Johnston, associate professor at the University of Bergen, and colleagues.

The researchers used mutual information (MI) to learn clinical factors that could predict patient outcomes. They also used a HyperTraPS (hypercubic transition path sampling) algorithm to create probabilistic pathways of disease progression. Clinical features of the severity of the disease were used to classify patients into three categories––respiratory distress, cerebral malaria and severe anaemia––and associate the features with death.

Children with cerebral malaria had the highest risk of death and the absence of it reduced the odds of death “significantly,” researchers wrote. Respiratory distress increased the odds of death for those with and without CM. Blood transfusion, which was the next informative prognostic feature in patients with both CM and RD, seemed to reduce mortality.

Researchers also found that patients with an enlarged spleen had a better clinical outcome, as the organ may have been attempting to clear the infection from the blood.

Beyond identifying the features that predicted death, they looked at the sequence of events that led to death in a data-driven approach validated by 11 experienced clinicians, most of whom said they were uncertain about disease progression.

“This novel approach captures not just a snapshot of individual risk factors but the full probabilistic information about the learned pathways of disease progression, allowing the histories of previous patients to inform the clinical analysis of new patients,” Johnstone et al. wrote. “This approach, validated with a test dataset, aligns with the goals of precision medicine and makes full use of available biomedical data; we anticipate that it may also find use in numerous other diseases and clinical contexts."