Machine learning could help reduce hospitalizations by nearly 30% during pandemic, study finds

A new study highlights a promising approach using machine learning to more efficiently allocate medical treatments during a pandemic or any time there is a shortage of therapeutic products.

The results, published in JAMA Health Forumfound a significant reduction in predicted hospitalizations when using machine learning to help distribute medications using the COVID-19 pandemic to test the model. The model was found to reduce hospitalizations by about 27% relative to actual, observed care.

“During the pandemic, the health care system was at a breaking point, and many health care facilities relied on first-come, first-served or a patient’s medical history to determine who received treatment,” said study lead author Adit Ginde, MD, professor of emergency medicine at the University of Colorado Anschutz Medical Campus.

“However, these methods often fail to account for the complex interactions that can occur in patients taking medications to determine expected clinical efficacy and may overlook patients who would benefit most from treatment. We show that machine learning in these scenarios is a way to use real-time, real-world evidence to inform public health decision-making,” adds Ginde.

In the study, the researchers showed that using machine learning to examine how individual patients benefit differently from treatment can provide physicians, health systems, and public health officials with more accurate, real-time information than traditional allocation score models. Mengli Xiao, Ph.D., assistant professor of biostatistics and computer science, developed the machine learning-based mAb allocation system.

“Existing allocation methods mainly target patients who have a high risk profile for hospitalization without treatment. They may overlook patients who benefit most from treatments. We developed a monoclonal antibody allocation point system based on heterogeneity estimates of treatment effects from machine learning. Our allocation prioritizes patient characteristics associated with important causal treatment effects, seeking to optimize overall treatment benefits when resources are limited,” Xiao said.

Specifically, the researchers investigated the effectiveness of adding a novel policy learning tree (PLT)-based method to optimize the allocation of COVID-19 neutralizing monoclonal antibodies (mAbs) during periods of resource constraints.

The PLT approach was designed to decide which treatments to allocate to individuals in a way that maximizes the overall benefits to the population (ensuring that those most at risk of hospitalization receive treatments, particularly when treatments are in short supply). This is done by taking into account how different factors affect the effectiveness of treatment.

The researchers compared the machine learning approach with real-world decisions and a standard scoring system used during the pandemic. They found that the PLT-based model demonstrated a significant reduction in predicted hospitalizations compared to observed scoring. This improvement also outperformed the monoclonal antibody screening score, which looks at antibodies for diagnostic purposes.

“Using an innovative approach like machine learning goes beyond crises like the COVID-19 pandemic and shows that we can provide personalized public health decisions even when resources are limited in any scenario. To do this, however, it is important that robust, real-time data platforms, like the one we developed for this project, are implemented to provide data-driven decisions,” adds Ginde, head of the Colorado Clinical and Translational Sciences Institute at CU Anschutz.

The paper in JAMA Health Forum This will be the 15th publication from a project called Monoclonal Antibody (mAB) Colorado. The project has focused on doing the most good for the most people, using real-world evidence to make data-driven decisions during the COVID-19 pandemic.

The researchers hope this paper will encourage public health entities, policymakers, and disaster management agencies to consider methods such as machine learning for implementation in the event of a future public health crisis.