Machine learning could have reduced hospitalizations by more than a quarter during the COVID-19 pandemic by determining how best to allocate scarce medication, a study suggests.
The findings, in the journal JAMA Health Forum, could have implications for other public health crises in which limited resources are available.
The AI-based system allocated points by estimating how much COVID-19 patients gained from treatment with neutralizing monoclonal antibodies (mAbs) based on their individual characteristics.
It then prioritized those who gained the most, thereby optimizing population health benefits in a scenario of therapeutic scarcity.
“Using an innovative approach like machine learning expands beyond crises like the COVID-19 pandemic and shows we can provide personalized public health decisions even when resources are limited in any scenario,” said senior researcher Adit Ginde, MD, a professor of emergency medicine at the University of Colorado School of Medicine.
“To do so, though, it’s important that robust, real-time data platforms, like what we developed for this project, are implemented to provide data-driven decisions.”
Existing methods to allocate treatment primarily target patients who are at high risk of hospital admission if they do not receive this.
But this could overlook patients who could benefit most and not optimize overall treatment benefits when resources are in short supply.
To investigate further, the research team retrospectively examined the impact of using a machine-learning method based on novel Policy Learning Trees (PLTs) to optimize the distribution of COVID-19 neutralizing mAbs.
The system was designed to assign resources to patients in a way that maximized the overall benefits for the population, making sure those at the highest risk of hospitalization were certain to receive treatment.
The mAb allocation point system considered how different factors affect the effectiveness of the treatment and the results were compared with real-world decisions and a standard point allocation system that was used in the pandemic.
Applying the PLT-based allocation method to electronic health record data from more than 15,000 patients with COVID-19 within a large health care system resulted in significantly fewer expected hospitalizations than actually occurred.
Among 9,542 patients in a cohort used to train the AI-based system using electronic record data from 1 October to 11 December 2021, 40.5% received mAbs. Among 6,248 eligible patients in a testing cohort using health record data from June to 1 October 2021, 21.3% received mAbs.
Allocating treatment based on the PLT led to an estimated absolute 1.6% reduction in overall expected hospitalization compared with observed treatment. The expected overall hospitalization rate was estimated to be 6.0%.
This translated into 63 patients who would need to be treated under the PLT-based allocation in order to prevent on hospitalization.
The system was even superior in terms of 28-day hospitalization to the Monoclonal Antibody Screening Score, which assesses antibodies for diagnosis.
“During the pandemic, the healthcare system was at a breaking point and many health care facilities relied on a first-come, first-serve or a patient’s health history to implement who received treatments,” said Ginde.
“However, these methods often don’t address the complex interactions that can occur in patients when taking medications to determine expected clinical effectiveness and may overlook patients who would benefit the 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.”
