Effect Modifiers of the Association of High-Flow Nasal Cannula and Bronchiolitis Length of Stay Free

We obtained data for this multicenter retrospective cohort study from a combination of the local hospital electronic health records (EHRs) at the 7 participating institutions and the Pediatric Health Information System (PHIS; Children’s Hospital Association, Lenexa, KS). PHIS is an administrative and billing database that contains inpatient, emergency department, ambulatory surgery, and observation encounter-level data from 49 not-for-profit, pediatric hospitals in the United States. Data quality and reliability are assured through a joint effort between the Children’s Hospital Association and participating hospitals. Data are deidentified at the time of submission, and data are subjected to multiple reliability and validity checks before inclusion in the PHIS database. For this study, only data from the participating hospitals were included. These hospitals were chosen to reflect variability in hospital size and geographic region, including each of the 4 US census regions.

We included inpatient and observation patient encounters identified in PHIS between September 1, 2018 and March 31, 2020 from patients who were <2 years old at the time of admission if they had a principal diagnosis of acute bronchiolitis (International Classification of Diseases, Tenth Revision codes J21.0, J21.1, J21.8, or J21.9).²⁸  Repeat encounters for the same patient were treated as independent. By using EHR data, patients were excluded if their PHIS record could not be paired with a hospital encounter or they did not have a full initial set of vital signs in room air without respiratory support (Fig 1).

The primary outcome of interest was hospital LOS, measured from the local EHR as the difference in time (in hours) between the patient’s initial registration (including emergency department) and hospital administrative discharge. We chose to use registration time rather than admission time to remove any possible effect of differences in the admission and transfer process from the emergency department to the inpatient unit.

For each patient, the usage of HFNC, initial measured weight, respiratory rate, heart rate, oxygen saturation, and level of respiratory support were obtained from the local EHR at each participating hospital. Tachypnea and tachycardia for age were defined as the 95th percentile by age group.²⁹  Demographic data, including sex, age in months, race, ethnicity, primary payer type, median household income (HHI) quartile by each patient’s residential zip code, child opportunity index (COI) category³⁰  by residential zip code, the hospitalization resource intensity scores for kids (H-RISK),³¹  and ICU utilization were obtained from PHIS. The number and types of complex chronic conditions (CCC)³²  were also obtained from PHIS.

Hospital protocols during the study period were collected and summarized (Supplemental Table 4). HFNC utilization was summarized by hospital by using point estimate and 95% confidence interval (CI), both before and after adjustment for age group, race/ethnicity, weight, payor type, median HHI quartile, COI category, tachypnea for age, tachycardia for age, room air saturation <90%, H-RISK, and ICU utilization (Fig 2). HFNC usage, demographics, baseline clinical characteristics, and clinical outcomes were summarized for each hospital (Supplemental Table 5).

Demographics, clinical characteristics, and outcomes were summarized by using frequencies and percentages for categorical variables, median and interquartile range (IQR) for nonnormal continuous variables; mean and standard errors (SE) were summarized for normally distributed continuous variables.

Multivariable Poisson regression was performed to assess the association of covariates with LOS. All covariates, including admission hospital, were treated as fixed effects. The model included main effects (Table 2), as well as an interaction term for HFNC usage with each of the other covariates (Table 3, Supplemental Table 6). Hospital LOS was summarized using the longest LOS (Hospital B) as a reference value. An interaction term between the hospitals and ICU category was added to account for differences in hospital ICU referral patterns. Main effects were reported with unadjusted and adjusted LOS for demographic and clinical covariates (Table 2) and summarized with mean ratios (MR) and 95% CI. The interaction term is summarized by demographic and clinical covariates using the MR of LOS between those who did and did not receive HFNC and its 95% CI. Interaction results are reported in Table 3 for covariates associated with LOS and for all covariates in Supplemental Table 6. The MR can be interpreted as the predicted ratio between 2 means; for example, an MR of 1.50 would mean that, on average, patients in a group have 1.5 times longer (50% longer) LOS than those in the reference group. In the case of the interaction term, the mean ratio is the expected ratio of LOS between patients receiving HFNC than those who did not receive HFNC. If the mean ratio for a given group were 2, for example, then patients receiving HFNC would have twice as high LOS on average than those who did not receive HFNC during their admission.

All analyses were performed by using SAS v 9.4 (SAS Institute, Cary, North Carolina) and P values <.05 were considered statistically significant. This study was approved by the local institutional review boards at each of the participating hospitals.

We reviewed 9524 patient encounters from 7 hospitals. We included 8060 patient encounters in the analysis and excluded the remaining 1464 because of absent initial vital signs (Fig 1). Overall, 27.0% of patients received HFNC during their admission. There were differences in age, weight, race, ethnicity, median HHI quartile, and COI by HFNC status and no differences in sex or primary payer type (Table 1). Tachypnea (49.3% vs 27.9%, P < .001), tachycardia (45.3% vs 36.4%, P < .001), and room air desaturation <90% (8.5% vs 3.4%, P < .001) were more common in patients who received HFNC than those who did not. ICU services (40.2% vs 3.2%, P < .001) were more common in patients who received HFNC than those who did not, but the pattern of day 0 ICU versus transfer after day 0 was the same in both groups.

HFNC usage at individual hospitals ranged from 10.1% to 47.9% [median 21.9%; IQR 15.7%, 30.0%]. After adjusting for potential confounders, the variation in HFNC use ranged from 7.2% to 56.7% (median 16.8%; IQR 1.1%, 24.8%; Fig 2).

Hospital protocols during the study period are summarized in Supplemental Table 4. Hospital-level demographics and clinical covariates are summarized in Supplemental Table 5. Four of the hospitals had protocol changes during the study period (Supplemental Table 4). Five hospitals allowed HFNC outside of the ICU throughout the study period, 1 started to allow it during the study period, and 1 allowed HFNC to be delivered in the ICU only. The hospital that started to allow HFNC outside of the ICU during the study had the third lowest HFNC usage and the hospital with HFNC only in the ICU had the lowest.

HFNC was significantly associated with increased unadjusted LOS (Table 2); patients who received HFNC had a mean LOS of 74.9 hours (95% CI 72.7–77.2), whereas those who did not had a mean LOS of 37.1 hours (95% CI 36.4–37.7). This difference was smaller after adjustment; patients who received HFNC had a mean adjusted LOS of 57.8 hours (95% CI 56.2–59.5), 42% longer (95% CI 37%–47%) than those who did not (mean adjusted LOS 40.8 hours, 95% CI 40.2–41.5).

There was significant variation in the LOS among hospitals, both before and after adjustment for other covariates. Age group and weight were significantly associated with LOS, with age >12 months and increased weight both being associated with shorter LOS. Patients with tachypnea for age at presentation had increased LOS, as did patients who had desaturation in room air on presentation. ICU services were associated with increased LOS both for patients initially admitted to the ICU on hospital day 0 and for those transferred to the ICU after day 0.

The effect of HFNC on LOS differed among hospitals (P < .001), with the estimated increase in LOS ranging from 32% to 139%. The effect of HFNC on LOS differed among age groups (P <.001), with the largest effect being seen in the youngest children (1–6 months MR 2.13, 95% CI 1.15–3.96; 7–12 months MR 1.69, 95% CI 0.91–3.14; 13–23 months MR 1.65, 95% CI 0.89–3.05). The effect of HFNC on LOS differed between patients who did (MR 1.64, 95% CI 0.88–3.07) and did not (MR 1.99, 95% CI 1.08–3.69) have initial room air saturation <90% (P = .003). The effect of HFNC on LOS also differed (P = .002) between patients receiving no ICU services (MR 2.25, 95% CI 1.22–4.15), those who were admitted to the ICU on hospital day 0 (MR 1.68, 95% CI 0.90–3.15), and those who were transferred to the ICU after hospital day 0 (MR 1.56, 95% CI 0.83–2.96). Mean ratio of length of stay for patients receiving HFNC versus patients not receiving HFNC is summarized for covariates with significant main effects in Table 3 and for all covariates in Supplemental Table 6.

In this study, we demonstrated that the association of HFNC usage with LOS may differ between subsets of patients. Although there were significant interactions between HFNC and a demographic or clinical factor, the less severe subset of patients (eg, no initial desaturation <90% in room air, no ICU services) had a greater increase in LOS related to HFNC usage than the more severe subset. Additionally, we demonstrated significant variation in HFNC policy and usage among hospitals, consistent with previous research,^11,19  with the added finding that the association of HFNC use with LOS varied by hospital, indicating the need for local evaluation and quality improvement efforts.

HFNC had significant interaction with multiple individual demographic or clinical factors when predicting LOS, specifically patients aged >6 months, with initial desaturation <90% in room air, or who received ICU services during their admission as compared with their opposing subsets. Knowing that the effect of HFNC is modified by other factors, it may be possible to identify groups of patients who are at risk for prolonged admission if they receive HFNC.

The finding that HFNC usage is associated with the largest increase in LOS among the least severe patients is consistent both with previous bronchiolitis research and with general principles of health care value. The best evidence that exists for the effect of HFNC is in the ICU setting to prevent positive pressure ventilation.^4,8  The evidence for its benefit in less severe bronchiolitis, on the other hand, is more equivocal.^{15,24,33–35}  In general, the risk-benefit ratio of a given intervention increases as the severity of an illness decreases because iatrogenesis is relatively fixed,³⁶  but the potential benefit decreases.

This study also reveals significant variations in HFNC policies and utilization between hospitals. Additionally, we found that the association of HFNC use with longer LOS varied by hospital. Much of the variability in HFNC policy utilization occurred among patients admitted outside of the ICU. Thus, there is potentially significant unwarranted variation in HFNC practice that may increase the LOS for children admitted with less severe bronchiolitis. Local quality improvements at the hospital level may help reduce HFNC in patients for whom there may not be a clinical benefit, such as those admitted to the regular inpatient floor.³⁷ 

This study has multiple limitations. First, there is the possibility of residual confounding to explain the association of HFNC use with LOS. Although we adjusted for multiple demographic and clinical factors in our analysis, including markers of severity of illness at presentation, patients who received HFNC may have been sicker, thereby resulting in longer LOS. Second, the data were collected from a combination of EHR extraction and the PHIS database, a process that is restricted by the limits of administrative data and EHRs. Third, the hospitals were chosen in a nonrandom fashion and may not be representative of PHIS-participating hospitals, let alone pediatric hospitals. Fourth, there was high variation in policies among hospitals and within hospitals over the course of the study. This included variability in the maximum flow allowable outside the ICU, and we were not able to assess the association of HFNC use with LOS in a standardized manner. Fifth, the study was retrospective based on data available at the time of presentation. Therefore, results are only an association, and no causative inference can be made. Sixth, factors that occurred after admission were not measured, and it may be beneficial to look at the progression of clinical signs and symptoms over time. Seventh, because of limitations in the granularity of available PHIS data, we could not differentiate patients admitted directly to the ICU from those transferred to the ICU from the inpatient unit during the first calendar day. Eighth, we were only able to assess HFNC at any point in hospitalization as a binary variable rather than measure the duration or flow of HFNC to ensure effective, weight-based dosing. This could potentially introduce bias, particularly in the measurement of LOS. Lastly, our study period ended in March 2020 at the start of the coronavirus disease 2019 pandemic, and thus, bronchiolitis care may have changed throughout and after the pandemic.

We identified multiple factors that may serve as effect modifiers for the relationship between HFNC usage and LOS. Specifically, HFNC was associated with a significant increase in LOS for patients without initial desaturation <90% in room air or ICU services. There is significant variability in HFNC protocols, utilization patterns, and the association of HFNC use with LOS between hospitals. Future prospective studies should investigate the effect of HFNC usage on LOS in admitted patients, particularly those without desaturations and not severe enough to require ICU services, as these are a subgroup for which HFNC is potentially low value.

The authors would like to acknowledge Britney Byars, Veena Myers, Mary Verner, the Division of Biomedical Informatics at Cincinnati Children’s Hospital Medical Center, the Yale Joint Data Analytics Team, and the other data analysts at each site for their assistance in obtaining local hospital EHR data.