Data source

The present study used the TriNetX COVID-19 Network, an international collaboration of health research platforms that compiles de-identified patient data from electronic health records (EHRs). These records encompass a wide variety of patient information, including demographic details, medical diagnoses, procedures, medication records, laboratory results, genomic data, and types of healthcare organization visits. Over 120 healthcare organizations (HCOs) worldwide, predominantly academic health centers, have contributed data from their main hospitals, affiliated institutions, and outpatient clinics to TriNetX. For this specific analysis, we utilized the COVID-19 network, encompassing data from more than 114 million patients from 87 HCOs. TriNetX offers integrated tools for patient-level data analysis and delivers aggregated results to the researchers. Detailed information on the database can be accessed online [24]. Written informed consent was not required because TriNetX contains anonymized data. The Institutional Review Board of the Chi Mei Medical Center approved the study protocol (no. 11202–002).

Patient selection

In the patient selection process, the TriNetX database was used, which contains 86 HCOs as of July 4, 2023. The initial patient pool consisted of individuals who had visited these HCOs at least twice between March 1, 2020, and January 1, 2023. Patients who tested positive for SARS-CoV-2 or were diagnosed with COVID-19 between January 1, 2022, and January 1, 2023, and those who were prescribed antiviral agents and were initially hospitalized were identified from this pool. The prescribed antiviral medications included molnupiravir, remdesivir, and nirmatrelvir plus ritonavir. The selection process was identical for all patients diagnosed with influenza within the same timeframe. The analysis was restricted to patients aged ≥ 18.

Subsequently, several exclusion criteria were used. For the COVID-19 group, patients who were also diagnosed with influenza and those with long-term COVID-related symptoms one year before the index date were excluded. Similarly, for the influenza group, patients who tested positive for SARS-CoV-2 or were diagnosed with COVID-19 and those with long-term COVID-related symptoms one year before the index date were excluded.

Finally, patient selection involved propensity score matching on a 1:1 basis for age at index, race, sex, adverse socioeconomic determinants of health, and comorbid medical conditions. This resulted in two comparable groups for this study: a COVID-19 and an influenza group (Tables S1 and S2).

Covariates

We considered 47 variables to adjust for imbalances in baseline characteristics between the COVID-19 and influenza groups. The list included both confirmed and suspected risk factors for COVID-19 and more severe cases of the illness, such as demographics (eg, age, sex, and ethnicity), adverse socioeconomic determinants of health (including "problems related to education and literacy," "problems related to employment and unemployment," and "problems related to housing and economic circumstances," as defined by ICD-10), and comorbidities (such as obesity, hyperlipidemia, hypertension, diabetes mellitus, chronic kidney disease, asthma, chronic lower respiratory diseases, ischemic heart disease, neoplasm, chronic liver diseases, stroke, dementia, rheumatoid arthritis, lupus, psoriasis, human immunodeficiency virus infection, mood disorders, and psychotic disorders).

Outcome measurement

The primary outcome of this study was a composite outcome consisting of 12 clinical features of post-COVID conditions, observed 90–180 days after the index event. These features include chest/throat pain, abnormal breathing, abdominal symptoms, fatigue/malaise, anxiety/depression, pain, headache, cognitive dysfunction, myalgia, loss of taste or smell, sleep disturbances, coughing, and palpitations [25,26,27].

We investigated the secondary outcomes of all-cause hospitalization, all-cause emergency department (ED) visits, and deaths during the follow-up period. Table S3 provides additional details regarding the diagnostic, visiting, and procedural codes used to define these outcomes [21, 28, 29].

Statistical analysis

We used the built-in propensity score-matching (PSM) function of the TriNetX platform to ensure a 1:1 match between the participants in the COVID-19 and influenza groups. This was achieved using a nearest-neighbor greedy matching algorithm with a caliper width of 0.1 pooled standard deviation. Standard differences were then computed to assess the balance between groups, with differences in absolute values < 0.1, indicating a good match between groups [30].

Subsequently, we performed Kaplan–Meier analysis, followed by log-rank tests and calculation of hazard ratio (HR) with 95% confidence intervals (CI) to compare the two groups. Statistical significance was set at p < 0.05. The HR was used to describe the relative risk of post-COVID conditions based on a comparison of time-to-event rates calculated using a proportional hazard model, which is a built-in function in TriNetX.

For subgroup analysis, we compared the primary and secondary outcomes between the two groups, stratified by age (18–64 and ≥ 65 years), sex, vaccine status (unvaccinated, 1 or 2 doses of vaccine, boosted), and race (Caucasian and non-Caucasian). The vaccine type used was Pfizer with CPT code 91,307 (0051A, 0052A, 0071A, 0072A), Janssen 91,303 (0031A), Novavax (0041A, 0042A), and Moderna 91,301 (0011A, 0012A, 0013A, 0111A).