OBJECTIVES: Continued morbidity following COVID-19 (long COVID (LC)) is defined by the World Health Organization (WHO) as conditions lasting > 2 months and identified ~ 3 months after COVID-19 onset. Predictors for LC following COVID-19 infection are not known.

METHODS: Patients with COVID-19 (first date = index) from April 1, 2020 onwards, with ≥ 6 months of continuous enrollment pre- and post-index, in IBM® MarketScan® Commercial and Medicare Supplemental databases, were identified and stratified by severity (mild, moderate and severe/critical). Only severe cases were used herein. Variables included demographics, comorbidities (Elixhauser index (EI) and all 31 Elixhauser disease domains), and distinct COVID-19 signs and symptoms during index disease. Duration of disease was defined as follows: from 5 days before positive test to last related visit/prescription, with a maximum gap of 35 days between visits/prescription. The primary target for prediction was LC, defined as duration > 5 months. Different machine learning algorithms (>50) were evaluated to develop prediction models (DataRobot Inc.).

RESULTS: 19,776 severe COVID-19 patients were included (56.5% male, average age (standard deviation (SD)): 54.5 (14.9), average EI (SD): 1.9 (2.1), LC: 2,716 patients (13.7%). The 2 most accurate models included ENET Blender (ENET) and Nystroem Kernel Support Vector Machine classifier (NK) (both: AUC = 0.7639). (NK: F1 Score: 0.4254, sensitivity: 0.5, specificity: 0.8716; ENET: F1 Score: 0.4282, sensitivity: 0.4924, specificity: 0.8782). In both cases the 3 most important features included two COVID-19 symptoms (thrombosis and chest pain) and the Elixhauser comorbidity indicator (EI). In both models, age was less important than thrombosis or chest pain at time of index.

CONCLUSION: Predictive models may help identify patients at increased risk for LC using claims data. Our analysis suggests that thrombosis, chest pain and EI at index in patients with severe COVID-19 may be a risk factor for LC.