Dai MF, Li SY, Zhang JF, Wang BY, Zhou L, Yu F, Xu H, Ge WH. Warfarin anticoagulation management during the COVID-19 pandemic: The role of internet clinic and machine learning.
Front Pharmacol 2022;
13:933156. [PMID:
36225580 PMCID:
PMC9549053 DOI:
10.3389/fphar.2022.933156]
[Citation(s) in RCA: 0] [Impact Index Per Article: 0] [Reference Citation Analysis] [Abstract] [Track Full Text] [Download PDF] [Figures] [Journal Information] [Subscribe] [Scholar Register] [Received: 04/30/2022] [Accepted: 08/30/2022] [Indexed: 11/22/2022] Open
Abstract
Background: Patients who received warfarin require constant monitoring by hospital staff. However, social distancing and stay-at-home orders, which were universally adopted strategies to avoid the spread of COVID-19, led to unprecedented challenges. This study aimed to optimize warfarin treatment during the COVID-19 pandemic by determining the role of the Internet clinic and developing a machine learning (ML) model to predict anticoagulation quality.
Methods: This retrospective study enrolled patients who received warfarin treatment in the hospital anticoagulation clinic (HAC) and “Internet + Anticoagulation clinic” (IAC) of the Nanjing Drum Tower Hospital between January 2020 and September 2021. The primary outcome was the anticoagulation quality of patients, which was evaluated by both the time in therapeutic range (TTR) and international normalized ratio (INR) variability. Anticoagulation quality and incidence of adverse events were compared between HAC and IAC. Furthermore, five ML algorithms were used to develop the anticoagulation quality prediction model, and the SHAP method was introduced to rank the feature importance.
Results: Totally, 241 patients were included, comprising 145 patients in the HAC group and 96 patients in the IAC group. In the HAC group and IAC group, 73.1 and 69.8% (p = 0.576) of patients achieved good anticoagulation quality, with the average TTR being 79.9 ± 20.0% and 80.6 ± 21.1%, respectively. There was no significant difference in the incidence of adverse events between the two groups. Evaluating the five ML models using the test set, the accuracy of the XGBoost model was 0.767, and the area under the receiver operating characteristic curve was 0.808, which showed the best performance. The results of the SHAP method revealed that age, education, hypertension, aspirin, and amiodarone were the top five important features associated with poor anticoagulation quality.
Conclusion: The IAC contributed to a novel management method for patients who received warfarin during the COVID-19 pandemic, as effective as HAC and with a low risk of virus transmission. The XGBoost model could accurately select patients at a high risk of poor anticoagulation quality, who could benefit from active intervention.
Collapse