Increased vaccine sensitivity of an emerging SARS-CoV-2 variant
Study setting
We undertook this retrospective observational study within the KPSC healthcare system. As a comprehensive, integrated care organization, KPSC delivers healthcare across telehealth, outpatient, emergency department, and inpatient settings for ~4.7 million members enrolled through employer-provided, government-sponsored, and pre-paid coverage schemes. EHRs across all clinical settings, together with laboratory, pharmacy, and immunization data, provide a complete view of care delivered by KPSC. These observations are augmented by insurance claims for out-of-network diagnoses, prescriptions, and procedures, enabling near-complete capture of healthcare interactions for KPSC members. The KPSC Institutional Review Board reviewed and provided ethical approval for the study.
During the study period, roughly 15% of all outpatient cases with SARS-CoV-2 infection confirmed by molecular testing received nirmatrelvir-ritonavir, although this population accounted for only ~25% of all nirmatrelvir-ritonavir prescribing (as the majority of patients received treatment on the basis of clinical or at-home antigen test results). Although also available during the study period, molnupiravir was rarely used (<0.1% of all cases), as it was reserved for patients unable to receive nirmatrelvir-ritonavir due to potential drug-drug interactions48.
Eligibility criteria
We included cases who: (1) received a positive molecular test result in any outpatient setting between December 1, 2022, and February 23, 2023; (2) had specimens processed using the ThermoFisher TaqPath COVID-19 Combo Kit (as described below); (3) had not received any prior SARS-CoV-2 test result in a clinical setting or prior COVID-19 diagnosis within 90 days before their index test; (4) had been enrolled in KPSC health plans for ≥1 year prior to their index test, allowing for up to a 45-day lapse in membership to account for potential delays in re-enrollment; and (5) were not hospitalized at the time of their index test, and had not been hospitalized at any point in the preceding 7 days. Excluding individuals without ≥1 year of membership in KSPC health plans (N = 3749 out of 35,488 otherwise eligible cases) enabled us to ensure COVID-19 vaccine doses, infection history, comorbid conditions, and healthcare utilization were captured accurately for the analytic sample. Restricting analyses to outpatient cases was expected to provide several design advantages. First, this strategy helped to ensure cases infected with each lineage were similar to each other in terms of healthcare-seeking behavior34,49. Second, initiating follow-up from the point of outpatient testing helped to ensure cases were ascertained at similar stages of their clinical course, facilitating unbiased comparisons of subsequent progression to severe disease. Outpatient-diagnosed cases at KPSC were automatically enrolled in a home-based symptom monitoring program with standardized criteria for emergency department referral and inpatient admission as a measure to preserve hospital capacity throughout the study period50. Thus, hospital admission was considered to represent an internally consistent measure of disease severity within the sample followed from the point of outpatient testing10, whereas cases first intercepted in hospital settings may have had greater variation in clinical status at the point of testing; in the event that infections were acquired in the hospital, these cases may also have differed with respect to clinical status, SARS-CoV-2 exposure, and time to treatment initiation in comparison to those acquiring infection in the community. This approach also helped our study to avoid the inclusion of incidental SARS-CoV-2 detections among patients who were tested at the point of inpatient admission for causes unrelated to COVID-19. Last, whereas the TaqPath COVID-19 Combo Kit was the primary assay used at regional reference laboratories for outpatient testing, cases diagnosed in hospital settings may have had tests processed in-house using other assays, without readout enabling lineage determination.
Lineage calling
The TaqPath COVID-19 Combo Kit assay included probes for the spike (S), nucleocapsid (N), and orf1a/b genes. Cases with cycle threshold (cT) values below 37 for ≥2 probes were considered positive for SARS-CoV-2. We interpreted S-gene target failure (SGTF), defined as cT ≥ 37 for the S-gene but cT < 37 for the N and orf1a/b genes, as a proxy for infection with BA.4/BA.5 sublineages, whereas S-gene detection provided a proxy for infection with XBB/XBB.1.5 lineages, consistent with our validation data (Results) and prior US studies30,31.
Exposures
We characterized the following attributes of included cases using data from their EHRs: age (defined in 10-year age bands), biological sex; race/ethnicity (white, black, Hispanic of any race, Asian, Pacific Islander, and other/mixed/unknown race, as self-reported by individuals and recorded in their medical record); neighborhood socioeconomic status, measured as the median household income within their Census block (<$30,000, $30,000–59,999, $60,000–89,999, $90,000–119,999, $120,000–149,999, and ≥$150,000 per year); cigarette smoking status (current, former, or never smoker); body mass index (BMI; categorized as underweight, normal weight, overweight, obese, or severely obese); Charlson comorbidity index (0, 1–2, 3–5, and ≥6); prior-year emergency department visits and inpatient admissions (each defined as 0, 1, 2, or ≥3 events); prior-year outpatient visits (0–4, 5–9, 10–14, 15–19, 20–29, or ≥30 events); documented prior SARS-CoV-2 infection; and history of COVID-19 vaccination (receipt of 0, 1, 2, 3, 4, or ≥5 doses, according to manufacturer, type, and time from receipt of each dose to the date of the index test). While our analyses include recipients of BNT162b2, mRNA-1273, Ad.26.COV2.S, and NVX-COV2373, we do not distinguish protection associated with receipt of mRNA or non-mRNA vaccine series, as non-mRNA vaccine doses accounted for only 2.0% of all vaccine doses received within the study population (1741/86,076). For individuals with multiple prior COVID-19 diagnoses or positive SARS-CoV-2 test results, we considered these infections to be distinct if they were not preceded by any other COVID-19 diagnosis or positive test result within 90 days. For data anonymization, index test dates were jittered by random addition of –1, 0, or 1.
Outcomes
We followed outpatient-diagnosed cases from the point of their index test through death, disenrollment, or March 5, 2023, the date of the database cut (providing ≥10 days of follow-up for individuals who did not die or disenroll). The primary endpoint was hospital admission for any cause within 30 days after the index test. Additional endpoints monitored over 60 days after the index test included ICU admission, initiation of mechanical ventilation, and death.
Multiple imputation of missing data
To accommodate missing data on cases’ BMI (n = 4799; 15.1%), cigarette smoking (n = 4229; 13.3%), and neighborhood household income (n = 6476; 20.4%), we generated 10 pseudo-datasets completed by sampling from the conditional distribution of these variables, given all other observed characteristics of cases, via multiple imputation. We conducted complete-case statistical analyses across each of the 10 pseudo-datasets and pooled results across these analyses according to Rubin’s rules51.
Logistic regression analysis
Within this analytic sample of cases testing positive for SARS-CoV-2 infection, potential outcomes were binary (infection with XBB/XBB.1.5 or non-XBB/XBB.1.5 lineages). We estimated adjusted odds ratios of prior vaccination and prior documented infection among cases with XBB/XBB.1.5 and non-XBB/XBB.1.5 cases via logistic regression. Models used in primary analyses controlled for variables as categorized in Table 1, as motivated by a directed acyclic graph (Supplementary Fig. S1), and included fixed intercepts for the calendar week of testing. As a sensitivity analysis, we also fit models defining calendar time continuously via polynomial transformations of individuals’ calendar date of testing. Models including a fourth-order polynomial function were found to provide optimal penalized fit to data on the basis of minimizing the Bayesian Information Criterion (Supplementary Table S3).
For all analyses, we report unadjusted associations of each exposure with XBB/XBB.1.5 or non-XBB/XBB.1.5 infection, associations accounting only for testing week (“time-adjusted” odds ratios), and associations accounting for all confounders (“adjusted” odds ratios). Models using alternative adjustment strategies for vaccination and infection (Supplementary Fig. S1) provided results similar to those of the primary analysis (Supplementary Tables S2, S8, S9, S11 and S12).
Survival analysis
For analyses addressing the association of prior vaccination, infection, and infecting lineage with cases’ risk of hospital admission or ICU admission, we fit Cox proportional hazards models to data on cases’ times to each of these events or censoring (at study end date, end of follow-up at 30 or 60 days, or disenrollment, whichever occurred earliest). A survival analysis framework was motivated by the fact that XBB/XBB.1.5 infections accounted for an increasing share of all diagnosed cases over time, and thus had higher likelihood of censoring within <30 days or <60 days in comparison to non-XBB/XBB.1.5 cases. Models defined strata according to cases’ calendar week of testing to control for potential changes in testing and healthcare-seeking practices. We verified the proportional hazards assumption by testing for non-zero slopes of the Schoenfeld residuals52.
Software
We conducted analyses using R (version 4.0.3; R Foundation for Statistical Computing, Vienna, Austria). We used the survival53 package (version 3.5-3) for time-to-event analyses, and the Amelia II package54 (version 1.81.1) for multiple imputation.
Reporting summary
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.
