Department of Biostatistics and Informatics, Colorado School of Public Health, University of Colorado Anschutz Medical Campus, Aurora.
Department of Emergency Medicine, University of Colorado School of Medicine, Aurora.
JAMA Health Forum. 2024 Sep 6;5(9):e242884. doi: 10.1001/jamahealthforum.2024.2884.
During the COVID-19 pandemic, the effective distribution of limited treatments became a crucial policy goal. Yet, limited research exists using electronic health record data and machine learning techniques, such as policy learning trees (PLTs), to optimize the distribution of scarce therapeutics.
To evaluate whether a machine learning PLT-based method of scarce resource allocation optimizes the treatment benefit of COVID-19 neutralizing monoclonal antibodies (mAbs) during periods of resource constraint.
DESIGN, SETTING, AND PARTICIPANTS: This retrospective cohort study used electronic health record data from October 1, 2021, to December 11, 2021, for the training cohort and data from June 1, 2021, to October 1, 2021, for the testing cohort. The cohorts included patients who had positive test results for SARS-CoV-2 and qualified for COVID-19 mAb therapy based on the US Food and Drug Administration's emergency use authorization criteria, ascertained from the patient electronic health record. Only some of the qualifying candidates received treatment with mAbs. Data were analyzed between from January 2023 to May 2024.
The primary outcome was overall expected hospitalization, assessed as the potential reduction in overall expected hospitalization if the PLT-based allocation system was used. This was compared to observed allocation using risk differences.
Among 9542 eligible patients in the training cohort (5418 female [56.8%]; age distribution: 18-44 years, 4151 [43.5%]; 45-64 years, 3146 [33.0%]; and ≥65 years, 2245 [23.5%]), a total of 3862 (40.5%) received mAbs. Among 6248 eligible patients in the testing cohort (3416 female [54.7%]; age distribution: 18-44 years, 2827 [45.2%]; 45-64 years, 1927 [30.8%]; and ≥65 years, 1494 [23.9%]), a total of 1329 (21.3%) received mAbs. Treatment allocation using the trained PLT model led to an estimated 1.6% reduction (95% CI, -2.0% to -1.2%) in overall expected hospitalization compared to observed treatment allocation in the testing cohort. The visual assessment showed that the PLT-based point system had a larger reduction in 28-day hospitalization compared with the Monoclonal Antibody Screening Score (maximum overall hospitalization difference, -1.0% [95% CI, -1.3% to -0.7%]) in the testing cohort.
This retrospective cohort study proposes and tests a PLT method, which can be linked to a electronic health record data platform to improve real-time allocation of scarce treatments. Use of this PLT-based allocation method would have likely resulted in fewer hospitalizations across a population than were observed in usual care, with greater expected reductions than a commonly used point system.
在 COVID-19 大流行期间，有效分配有限的治疗方法成为一个关键的政策目标。然而，利用电子健康记录数据和机器学习技术（如政策学习树 [PLT]）来优化稀缺治疗药物的分配，这方面的研究还很有限。
评估基于机器学习的 PLT 稀缺资源分配方法是否可以优化 COVID-19 中和单克隆抗体（mAb）治疗期间稀缺治疗资源的治疗效益。
设计、设置和参与者：这是一项回顾性队列研究，使用了 2021 年 10 月 1 日至 12 月 11 日的电子健康记录数据进行训练队列分析，以及 2021 年 6 月 1 日至 10 月 1 日的数据进行测试队列分析。队列纳入了 SARS-CoV-2 检测结果呈阳性且符合 COVID-19 mAb 治疗条件的患者，该条件基于美国食品和药物管理局的紧急使用授权标准，从患者的电子健康记录中确定。只有部分符合条件的候选患者接受了 mAb 治疗。数据分析于 2023 年 1 月至 2024 年 5 月进行。
主要结局是总体预期住院率，评估如果使用基于 PLT 的分配系统，总体预期住院率的潜在降低情况。将这一结果与观察到的使用风险差异的分配情况进行比较。
在训练队列的 9542 名合格患者中（5418 名女性 [56.8%]；年龄分布：18-44 岁 4151 人 [43.5%]；45-64 岁 3146 人 [33.0%]；≥65 岁 2245 人 [23.5%]），共有 3862 人（40.5%）接受了 mAb 治疗。在 6248 名合格的测试队列患者中（3416 名女性 [54.7%]；年龄分布：18-44 岁 2827 人 [45.2%]；45-64 岁 1927 人 [30.8%]；≥65 岁 1494 人 [23.9%]），共有 1329 人（21.3%）接受了 mAb 治疗。与测试队列中的观察性治疗分配相比，使用经过训练的 PLT 模型进行治疗分配，估计可降低 1.6%的总体预期住院率（95%CI，-2.0% 至 -1.2%）。直观评估显示，与常用的单克隆抗体筛选评分（最大总体住院差异为-1.0% [95%CI，-1.3% 至 -0.7%]）相比，PLT 评分系统在 28 天住院方面的降低幅度更大。
本回顾性队列研究提出并测试了一种 PLT 方法，该方法可以与电子健康记录数据平台相连接，以改进稀缺治疗方法的实时分配。与常规护理相比，这种基于 PLT 的分配方法可能会导致更多的人群住院率降低，并且预期的降低幅度会大于常用的评分系统。
