Diagnosis-wide analysis of COVID-19 complications: an exposure-crossover study

Data source

This study used de-identified United States medical claims from HealthVerity’s Marketplace data set. These data are claims from nationally representative health plans that encompass all major payer types (commercial, Medicaid and Medicare). The data include open claims, which are sourced in near real time from practice management systems, billing systems and claims clearinghouses, as well as closed claims, which are sourced from insurance providers and payers. Data from open claims can be captured within days of a health care encounter, while data from closed claims have longer lag times. Thus, open claims provide more recent data, and closed claims encompass a more complete view of a patient’s interactions with the health care system. These data sources are further described in Appendix 1, Methods, available at www.cmaj.ca/lookup/doi/10.1503/cmaj.201686/tab-related-content. Claims were available with dates of service from Nov. 1, 2019, through May 30, 2020.

Study population

We included all patients who had had at least 1 medical encounter related to COVID-19, defined as any claim with an ICD-10-CM diagnosis code of U07.1 (COVID-19, virus identified) or B97.29 (other coronavirus as the cause of diseases classified elsewhere) that occurred within the period from Mar. 1, 2020, to Apr. 30, 2020. Apr. 30 was the last eligibility date to allow for a month of possible follow-up for all patients. Codes were selected based on US Centers for Disease Control and Prevention (CDC) coding guidance for confirmed COVID-19 cases.10^,11 The first eligible claim within the study period was the index date. Patients with a code for U07.1, B97.29 or B97.21 (SARS-associated coronavirus) occurring before index, and those with a missing year of birth, were excluded. Patients with a code for Z03.818 occurring between 7 days before index and 30 days after index were also excluded, because this code indicates that SARS-CoV-2 exposure has been ruled out.10 To ensure that all analyzed patients were observable before their COVID-19 diagnosis, patients were excluded if they did not have any medical claim between 120 days before index and 30 days before index. We also assessed subgroups of patients treated only on an outpatient basis, patients treated as inpatients, and patients who were admitted to the intensive care unit (ICU) (Appendix 1, Methods).

Study design

We used an exposure-crossover design12 to identify diagnoses whose odds increased after the onset of COVID-19. This design minimizes unmeasured confounding by comparing a patient’s condition upon COVID-19 diagnosis (the “hazard period”) with their own condition at a previous time (the “baseline period”). Any increased odds of a condition are likely attributable to COVID-19 sequelae or treatment.

To capture activity related to COVID-19 that occurred just before the patient received a diagnosis of the disease (Appendix 1, eFigure 2), we defined the hazard period as the 7 days before the index date through the 30 days after index. We defined the baseline period as the 120 days before the index date through 30 days before index. We chose this 90-day window as it was long enough to capture most chronic conditions while ensuring equivalent capture of baseline status for all patients. We considered the 23 days between the baseline and hazard period as a period of unknown COVID-19 disease state (Appendix 1, eFigure 2) and excluded that person-time from the analysis. We also performed sensitivity analyses to assess the effect of the duration of the baseline period by conducting an analysis where the baseline period was identical in length to the hazard period (days 68 through 30 before index; 38 days total). The study design is illustrated in Appendix 1, eFigure 3.

Statistical analysis

To identify ICD-10-CM codes associated with COVID-19, we paired observations of each patient in the baseline and hazard periods and computed matched odds ratios (ORs) to estimate the strength of association of the code with COVID-19, as well as McNemar’s test p values.13 We used exact tests if the number of discordant pairs was less than 25. ICD-10-CM codes are constructed as a 3-character header indicating system classification and disease and can be followed by up to 4 digits offering greater specificity.14^,15 For our primary analysis, we analyzed all 1724 ICD-10-CM diagnosis codes present in the data set, aggregated at the header level (e.g., “R43 – Disturbances of smell and taste”). For any header code from the primary analysis found to be significantly associated, we also analyzed it at the individual code level (e.g., “R43.0 – Anosmia”). For our secondary analysis, we analyzed all 64 931 individual ICD-10-CM codes. We applied Bonferroni-corrected significance thresholds, defined as a nominal type 1 error of 0.05 divided by the number of analyzed codes (2.9E-05 and 7.7E-07 for primary and secondary analyses, respectively).16 We excluded codes used to identify cases (B97 and U07). A candidate COVID-19 complication was any code that increased in odds with COVID-19 at a Bonferroni-corrected level of significance.

We also calculated the absolute risk of becoming newly diagnosed with each ICD-10-CM code upon having a COVID-19 diagnosis. This was the percentage of patients with the code in the hazard period, calculated among patients who did not have the code in the baseline period (Appendix 1, eFigure 3).