Improving COVID-19 Treatment Allocation With Machine Learning
The COVID-19 pandemic presented many challenges to the health care system, including the effective distribution of limited treatments and resources. When monoclonal antibodies (mAbs) emerged as a treatment that could reduce disease progression, it became challenging to allocate and administer these treatments due to scarcity. To prevent similar issues in future situations, a retrospective cohort study published in JAMA Health Forum aimed to determine whether a machine learning-based allocation method, specifically using policy learning trees (PLTs), could improve the allocation of scarce COVID-19 treatments and reduce hospitalization rates.
The study used data from 15 000 patients with COVID-19. The data set was divided into training and testing cohorts to evaluate the PLT’s effectiveness. Using electronic health record (EHR) data, the study included data from October 2021 to December 2021 for the training cohort and data from June 2021 to October 2021 for the testing cohort. It analyzed patient demographics, comorbidities, and vaccination status.
Researchers developed a point-allocation scoring system that used PLT to analyze the data and create a plan to allocate limited mAb treatment and maximize health benefits, which would reduce hospitalization through optimized resource distribution. Policy learning (PL) methods were applied to create decision trees that optimized hospitalization outcomes by identifying ideal treatment candidates within the training cohort.
This method showed an estimated 1.6% reduction in hospitalization compared with observed treatment allocation. Additionally, the PLT-based point system demonstrated a greater reduction in hospitalization rates than the traditional Monoclonal Antibody Screening Score (MASS). The PLT-based system also showed adaptability, allowing for real-time updates in resource allocation. Limitations of the study include generalizability, as the data was specific to a Colorado health care system and may not be directly generalizable to other populations, and the possibility that the effectiveness of the PLT system could be impacted by unobserved confounders not captured in the EHR data. However, researchers are optimistic that the PLT-based allocation model could improve therapeutic distribution efficiency and health outcomes during times of scarcity.
“Using electronic health record data to show that machine learning methods, namely policy learning trees, can improve the allocation of scarce therapeutics; therefore, policy learning tree-based allocation should be considered in potential future episodes of therapeutic scarcity, including pandemics,” researchers stated.
Reference
Xiao M, Molina KC, Aggarwal NR, et al. A machine learning method for allocating scarce COVID-19 monoclonal antibodies. JAMA Health Forum. 2024;5(9):e242884. doi:10.1001/jamahealthforum.2024.2884
