gms | German Medical Science

Introduction: The COVID-19 pandemic has sparked a tsunami of research on various facets of the disease. The number of published studies grew at an unprecedented rate, leading to information overload, requiring solutions to reduce screening work while maintaining accuracy.

Hospitalized patients with COVID-19 appeared to be at increased risk for experiencing thromboembolic complications compared with other hospitalized patients. As a foundation to develop living recommendations for the use of anticoagulation in COVID-19 patients, we need up-to-date reliable estimates for the baseline risks of patient-important outcomes, e.g., mortality, VTE, or major bleeding.

Methods: We conducted a systematic review (registered in PROSPERO as CRD42020204021) to establish the baseline risk of various thromboembolic complications in patients with COVID-19. We then rerun the search strategy every month as part of the living process and used the Laser AI system to automatically deduplicate the references and train a machine learning classifier based on the inclusion and exclusion decisions from the original review. The model was used to calculate the inclusion likelihood of the references from subsequent updates to create a prioritized screening list. We also retrained the model periodically with decisions made by screeners as part of the living process.

Results: The initial review was completed on July 19th, 2020, and involved screening 14,879 references. During title and abstract screening, 1771 citations with 1769 studies were included. 77 studies passed full-text screening.

Between August 2020 and July 2021, a total of 158,726 new citations were returned by the search strategy. Automatic deduplication reduced the number of references to screen to 73,422. The machine learning model was able to further reduce the number of references needed to screen to achieve 95% sensitivity to 37,739 (24% of the total number). The screening of references during the update was prioritized to first evaluate the most relevant references.

The living review is being used to develop American Society of Hematology living guidelines on the use of anticoagulation for thromboprophylaxis in patients with COVID-19, with continuous updates as needed. We plan to present the most up-to-date figures during the conference.

Conclusions: The rapid increase of research output may force reviewers to make trade-offs between the currency and accuracy of systematic reviews. We believe that our application of machine learning enables keeping the review process systematic and accurate while making the screening work better manageable and up-to-date. Our estimates continue to be an important component for providing healthcare practitioners with the best possible guidance.

Keywords: living systematic review, COVID-19, thromboprophylaxis, machine learning