Outcomes in Patients with Acute Hypoxemic Respiratory Failure Secondary to COVID-19 Treated with Noninvasive Respiratory Support versus Invasive Mechanical Ventilation

Abstract

Purpose

The goal of this study was to compare noninvasive respiratory support to invasive mechanical ventilation as the initial respiratory support in COVID-19 patients with acute hypoxemic respiratory failure.

Methods

All patients admitted to a large healthcare network with acute hypoxemic respiratory failure associated with COVID-19 and requiring respiratory support were eligible for inclusion. We compared patients treated initially with noninvasive respiratory support (noninvasive positive pressure ventilation by facemask or high flow nasal oxygen) with patients treated initially with invasive mechanical ventilation. The primary outcome was time-to-in-hospital death analyzed using an inverse probability of treatment weighted Cox model adjusted for potential confounders. Secondary outcomes included unweighted and weighted assessments of mortality, lengths-of-stay (intensive care unit and hospital) and time-to-intubation.

Results

Over the study period, 2354 patients met inclusion criteria. Nearly half (47%) received invasive mechanical ventilation first and 53% received initial noninvasive respiratory support. There was an overall 38% in-hospital mortality (37% for invasive mechanical ventilation and 39% for noninvasive respiratory support). Initial noninvasive respiratory support was associated with an increased hazard of death compared to initial invasive mechanical ventilation (HR: 1.61, p < 0.0001, 95% CI: 1.33 – 1.94). However, patients on initial noninvasive respiratory support also experienced an increased hazard of leaving the hospital sooner, but the hazard ratio waned with time (HR: 0.97, p < 0.0001, 95% CI: 0.96 – 0.98).

Conclusion

These data show that the COVID-19 patients with acute hypoxemic respiratory failure initially treated with noninvasive respiratory support had an increased hazard of in-hospital death.

Introduction

The optimal management strategy for patients with acute hypoxemic respiratory failure due to SARS-CoV-2 infection has undergone particular interest and scrutiny. Early, intense discussion occurred in the published literature and on social media. High failure rates with noninvasive positive pressure ventilation (NIPPV) during the SARS-CoV outbreak in 2003,¹ concerns over aerosol transmission with noninvasive support strategies,^2–8 potential novelty of respiratory physiology with COVID-19,^9–18 impending ventilator shortages,^19–21 and the high mortality initially reported with invasive mechanical ventilation (IMV)^22,23 all factored in the discussion over timing of intubation and the utility of noninvasive strategies.

The current literature on noninvasive respiratory support (NIRS) — NIPPV or high flow nasal oxygen (HFNO) — for patients with COVID-19 report disparate outcomes. Importantly, these studies compare noninvasive strategies to each other, or to conventional oxygen therapy, but not to invasive mechanical ventilation. The goal of this study was to explore the optimal initial treatment for COVID-19 patients admitted with acute hypoxemic respiratory failure by comparing invasive versus noninvasive strategies.

Methods

Study Design and Setting

This study is a retrospective cohort observational study using clinical data. Clinical data were obtained from the Banner Health Network clinical data warehouse, covering 26 hospitals across six states in the western United States. Data were extracted for all adult patients (≥18 years) admitted for respiratory failure associated with COVID-19 between January 1, 2020 and January 7, 2021. Data consist of de-identified structured data generated from the Cerner (Cerner Corporation, North Kansas City, MO, USA) electronic health record. This work adheres to the STROBE reporting guidelines and was approved by the University of Arizona (IRB #1907780973) and Banner Health Institutional Review Boards (IRB #483-20-0018).

Study Participants and Treatment Assignment

We generated seven cohorts using a phenotyping algorithm²⁴ based on the sequence of therapies received: (1) IMV only, (2) NIPPV only, (3) HFNO only, (4) NIPPV requiring subsequent IMV, (5) HFNO requiring subsequent IMV, (6) IMV extubated to NIPPV, (7) IMV extubated to HFNO, and (8) evidence of all three treatments but unclear treatment ordering. Those treated with any noninvasive respiratory support modality first were compared to those treated with invasive mechanical ventilation first. Subsequent analyses separated NIPPV and HFNO and included pairwise comparisons of initial IMV, NIPPV, and HFNO.

We estimated the propensity to be given each modality using generalized boosted models and used inverse probability of treatment weighting in the models to account for non-random treatment assignment.²⁵ The variables for propensity score estimation included age, body mass index, sex, ethnicity (non-Hispanic, Hispanic), race (white, non-white), respiratory rate and S_PO₂/F_IO₂ ratio most immediately prior to first treatment, hospital size by American Hospital Association category (small [< 100 beds], medium [100–499 beds], large [> 500 beds]), and either hours from hospital admission to first treatment or hours from ICU admission to first treatment, transformed via the Box-Cox method with negatives.²⁶ These variables were additionally included in Cox models to further improve balance between treatment groups.

Outcomes and Data Analysis

The primary outcome is time-to-in-hospital death. We fit a Cox model from first treatment initiation to death with censoring occurring at hospital discharge. However, there are multiple potential pathways for a patient (1. death, 2. discharge alive [from hospital or ICU], 3. remain in the hospital, and 4. intubation [if on NIRS]). We modeled discharge alive and intubation with three additional hazard models. First, we fit a cause-specific hazard model for time-to-hospital discharge alive with death as a competing risk where time zero was set as time of first treatment initiation. For this outcome, we performed a planned sensitivity analysis with time zero set as time of hospital entrance. Second, we fit a cause-specific hazard model for time-to-ICU discharge alive (from ICU admission) with death as a competing risk. Only visits with an ICU stay were included in this analysis; thus, results are conditional on having been admitted to the ICU and by nature excluded ICU-level of care patients who remained outside of the ICU. Third, we conducted a competing risks analysis for time-to-intubation with death as a competing risk and censoring occurring at hospital discharge. Only visits with known time to event (96.3% of possible visits) and a noninvasive initial treatment were included. The competing risks analysis used a modification of Gray’s test that incorporates inverse probability of treatment weighting.²⁷ Statistical significance was judged via the median p-value approach.²⁸ Planned sensitivity analyses were performed to assess preprocessing and analysis decisions made for the competing risks (see online supplement for details).

We analyzed unweighted outcomes of mortality, intubation rate, duration of mechanical ventilation in patients who failed initial NIRS treatment, length-of-hospital-stay, and ICU-free days using Fisher’s Exact and Kruskal-Wallis rank sum tests where appropriate. For all outcomes, the predictor of interest was first treatment (noninvasive respiratory support or invasive mechanical ventilation). We assessed if the proportional hazard assumption was violated in the Cox models by including an interaction of time by first treatment. If the interaction was statistically significant at α = 0.05, we report the interaction model; otherwise, we report the proportional hazards model.

Data preprocessing is described in detail in the online supplement. As a retrospective study based on routinely collected EHR data, it is expected that the extent of observed data will be different for different clinical variables.^29–31 Missing data were handled by using multiple imputation by chained equations (MICE).^32,33 For each non-competing risks analysis, we created 50 imputed data sets using all variables in the propensity score construction, the Nelson-Aalen estimate of the available time-to-event data, the time-to-event itself, and event information (e.g., death, hospital discharge alive, ICU discharge alive). Age, body mass index, SpO₂/FiO₂, the transformed time-from-hospital admission (or ICU admission) to first treatment, respiratory rate, time-to-event, and the Nelson-Aalen estimate of the time-to-event were imputed via predictive mean matching. Sex, ethnicity, race, the outcome event, hospital size, and first treatment were imputed with logistic regression (two-category variables) or multinomial log-linear models via neural networks (> two category variables) as appropriate. Raw time-to-event was imputed but not used to predict other variables in the MICE algorithm. Instead, for the Cox models and some competing risk models, temporal information was used in the prediction of other missing values via the Nelson-Aalen estimate. For the competing risk models using the median p-value inference approach, no outcome information was allowed to predict other variables in the MICE algorithm. Since each imputed data set had different values for variables used in the propensity score algorithm, we estimated propensity scores for each imputed data set separately. For the Cox models, the propensity scores from a specific imputed data set were used to do inverse probability of treatment weighting with the analysis of that data set,³⁴ and results were combined using Rubin’s Rules. For the competing risks analyses, the propensity scores from each imputed data set were handled differently depending on the inference approach. See the supplementary information for further details. All data preprocessing and statistical analyses were done using R version 4.0.4³⁵ and included the following packages: twang³⁴, survival^36,37, survminer,³⁸ MICE³², xtable,³⁹ and tidyverse.⁴⁰

Results

During the study period, there were 2354 COVID-19 patients that met criteria for inclusion. There were 326 that were not classified by the phenotyping algorithm but received all three treatments. Of the 2028 patients that were reliably classified, invasive mechanical ventilation was used as the first therapy in 947 (47%) patients and noninvasive respiratory support in 1081 (53%), Figure 1, Table 1. Of those on noninvasive respiratory support, 811 (75%) received NIPPV first and 270 (25%) received HFNO first. There was imbalance in NIRS use among hospitals of different sizes with larger hospitals disproportionately using NIPPV first (large hospitals 81% [433/536], medium hospitals 70% [346/495], small hospitals 64% [32/50]) Table 1. For the 326 unclassifiable records, the first treatment was imputed using multiple imputation. Since treatment assignments may have varied between imputed data sets, demographics for this group are reported separately in the online supplement.

Figure 1:

STROBE Statement

326 participants had evidence of high-flow nasal oxygen, noninvasive positive pressure ventilation, and invasive mechanical ventilation but were unable to be reliably classified by the phenotyping algorithm because of nonsensical time stamps for timing of therapies. See text and online supplement for further description and demographics of this group. “Additional exclusions” includes 4 participants that were excluded because of nonsensical hospital discharge time stamps, and an additional group of patients excluded only from the model assessing time-to-ICU exit.

Table 1:

Demographics

Measure	Invasive Mechanical Ventilation	Noninvasive Positive Pressure Ventilation	High Flow Nasal Oxygen	Total
N (%)	947 (47%)	811 (40%)	270 (13%)	2028
Sex
Male	555 (59%)	465 (57%)	159 (59%)	1179 (58%)
Female	392 (41%)	346 (43%)	111 (41%)	849 (42%)
Age	60 (49 – 69)	65 (52 – 75)	70 (56 – 80)	63 (51 – 73)
BMI	31.43 (26.57 – 37.74)	32.37 (26.9 – 39.6)	30.29 (25.64 – 35.73)	31.64 (26.58 – 37.96)
Ethnicity
Not Hispanic or Latino	537 (59% of 917)	489 (61% of 800)	154 (58% of 266)	1180 (60% of 1983)
Hispanic or Latino	380 (41% of 917)	311 (39% of 800)	112 (42% of 266)	803 (40% of 1983)
Race
White	652 (71% of 914)	636 (80% of 799)	215 (81% of 267)	1503 (76% of 1980)
Black or African American	43 (5% of 914)	45 (6% of 799)	10 (4% of 267)	98 (5% of 1980)
Asian/Native
Hawaiian/Other Pacific Islander	11 (1% of 914)	18 (2% of 799)	6 (2% of 267)	35 (2% of 1980)
American Indian or Alaska
Native	136 (15% of 914)	52 (7% of 799)	25 (9% of 267)	213 (11% of 1980)
Other	72 (8% of 914)	48 (6% of 799)	11 (4% of 267)	131 (7% of 1980)
Hospital Sizea
small	21 (2%)	32 (4%)	18 (7%)	71 (4%)
medium	329 (35%)	346 (43%)	149 (55%)	824 (41%)
large	597 (63%)	433 (53%)	103 (38%)	1133 (56%)
APACHE IVa Score on Admission	67 (51 – 90)	56 (44.25 – 74)	60 (45.25 – 76.75)	63 (48 – 83)
Vital Signs on Treatment Assignment
HR	96 (78 – 112)	83 (70 – 97)	85 (72 – 101.75)	89 (74 – 105)
SBP	120 (103 – 142)	126 (111.5 – 142)	122 (110 – 136)	123 (108 – 141)
DBP	68 (59 – 80)	72 (64 – 81)	71 (63 – 79)	71 (61 – 80)
SpO2	96 (93 – 99)	94 (91 – 96)	93 (89 – 95)	95 (91 – 98)
Oxygen Flow Rate (L)	15 (15 – 50)	15 (8 – 40)	40 (27.25 – 40)	15 (15 – 40)
FiO2b	100 (60 – 100)	70 (50 – 100)	100 (66 – 100)	90 (60 – 100)
SpO2:FiO2	100 (95 – 152.08)	130.3 (94 – 198)	96.44 (90.16 – 141.3)	104.44 (94 – 166.67)
Temperature (C)	37 (36.76 – 37)	37 (37 – 37)	37 (36.6 – 37)	37 (36.8 – 37)
Respiratory Rate	22 (18 – 26)	24 (20 – 30)	22 (19 – 27)	22 (18 – 28)
Comorbiditiesc
Diabetes	513 (61% of 846)	466 (64% of 733)	151 (61% of 246)	1130 (62% of 1825)
Chronic Kidney Disease	226 (27% of 846)	168 (23% of 733)	55 (22% of 246)	449 (25% of 1825)
Heart Failure	203 (24% of 846)	179 (24% of 733)	45 (18% of 246)	427 (23% of 1825)
Hypertension	663 (78% of 846)	576 (79% of 733)	191 (78% of 246)	1430 (78% of 1825)
Chronic Liver Disease	182 (22% of 846)	96 (13% of 733)	36 (15% of 246)	314 (17% of 1825)
Neoplasm or Immunosuppression	70 (8% of 846)	49 (7% of 733)	14 (6% of 246)	133 (7% of 1825)
COPD	216 (26% of 846)	252 (34% of 733)	55 (22% of 246)	523 (29% of 1825)
Labs on Admission
PaO2 (Worst Value)	70 (58 – 90)	65 (56 – 82)	64 (56 – 79.75)	68 (57 – 86)
PaO2:FiO2 (Worst Value)	82.5 (62.9 – 133.5)	86.44 (61 – 136.74)	74 (59 – 109.62)	83 (62 – 132)
White Blood Cell Count	9.6 (6.6 – 13.65)	8 (5.8 – 11.4)	8.2 (5.73 – 11.1)	8.7 (6.2 – 12.3)
Lactate	1.75 (1.2 – 2.8)	1.6 (1.2 – 2.2)	1.6 (1.2 – 2.2)	1.6 (1.2 – 2.4)
pH	7.37 (7.28 – 7.43)	7.41 (7.36 – 7.46)	7.43 (7.38 – 7.46)	7.4 (7.32 – 7.45)
PaCO2	38 (33 – 47)	35.1 (31 – 42)	34 (28 – 39)	37 (31.4 – 44)
HCO3	21 (18 – 24)	22 (20 – 24)	22 (20 – 24)	22 (19 – 24)
BNP	619 (177 – 3034)	559 (194 – 2154)	427.5 (145.75 – 2009.25)	562.5 (180 – 2413.5)
Creatinine	1.05 (0.76 – 1.56)	0.99 (0.74 – 1.4)	0.95 (0.74 – 1.31)	1 (0.75 – 1.45)
Therapies
Therapeutic Anticoagulationd	726 (77%)	651 (80%)	225 (83%)	1602 (79%)
Hydroxychloroquine	146 (15%)	21 (3%)	15 (6%)	182 (9%)
Remdesivir	229 (24%)	525 (65%)	152 (56%)	906 (45%)
Corticosteroidse	779 (82%)	755 (93%)	244 (90%)	1778 (88%)
Tocilizumab	1 (0%)	0 (0%)	0 (0%)	1 (0%)
Continuous Paralysis	297 (31%)	--	--	406 (20%)
Vasopressors
Before Treatment	263 (28%)	68 (8%)	31 (11%)	362 (18%)
At or After Treatment	745 (79%)	266 (33%)	89 (33%)	1100 (54%)
Location at Treatment Assignment
Emergency Department	218 (23% of 940)	73 (9% of 809)	13 (5% of 270)	304 (15% of 2019)
ICU	640 (68% of 940)	305 (38% of 809)	100 (37% of 270)	1045 (52% of 2019)
Non-ICU ward	34 (4% of 940)	120 (15% of 809)	43 (16% of 270)	197 (10% of 2019)
Stepdown	48 (5% of 940)	311 (38% of 809)	114 (42% of 270)	473 (23% of 2019)

^ahospital size categorized by American Hospital Association category (small [< 100 beds], medium [100–499 beds], large [> 500 beds]).

^bFiO2 determined by documented FiO2, if documented, or by FiO₂ = 100(0.21 + oxygen flow [L/min⁻¹] × 0.03 if a flow rate was documented.

^ccomorbidities determined by documented ICD codes

^dTherapeutic anticoagulation included: continuous infusions of an anticoagulant (e.g., heparin, bivalirudin), or therapeutic weight-based dosing of subcutaneous injections, oral direct anticoagulants, therapeutic dosing anti-platelet agents, and oral warfarin.

^eCorticosteroids included systemic oral or intravenous hydrocortisone, prednisone, dexamethasone, and methylprednisolone.

All-cause in-hospital mortality was 38%. For those on IMV first, mortality was 37%. For those treated with noninvasive respiratory support first and never required intubation, mortality was 29%, but rose to 60% for those that required intubation. There was also an imbalance in mortality rates between NIPPV and HFNO, Table 2. Intubation rates for those treated with HFNO and NIPPV were 32% (87/270) and 33% (268/811), respectively.

Table 2:

Unmatched outcomes

Measure	Invasive Mechanical Ventilation	Noninvasive Positive Pressure Ventilation	High Flow Nasal Oxygen	Total	P-Value
All-cause in-ICU mortality
Total	253 (28% of 891)	138 (32% of 426)	75 (57% of 131)	466 (32% of 1448)	<0.001
Failure	--	99 (41% of 244)	47 (64% of 74)	146 (46% of 318)	0.001
Success	--	39 (21% of 182)	28 (49% of 57)	67 (28% of 239)	<0.001
All-cause hospital mortality
Total	348 (37% of 947)	281 (35% of 811)	140 (52% of 270)	769 (38% of 2028)	<0.001
Failure	--	145 (54% of 268)	67 (77% of 87)	212 (60% of 355)	<0.001
Success	--	136 (25% of 543)	73 (40% of 183)	209 (29% of 726)	<0.001
Failure (intubation rate)	--	268 (33% of 811)	87 (32% of 270)	355 (33% of 1081)	0.001
Days to Intubation	--	0.24 (0.13 – 1.18)	0.14 (0.06 – 0.56)	0.21 (0.11 – 1.05)	<0.001
Duration of mechanical ventilation (days)	--	8.08 (3.6 – 13.98)	8.36 (1.4 – 17.37)	8.09 (3.34 – 14.91)	0.516
ICU length of stay (days)
Total	8.04 (3.49 – 16.75)	6.2 (2.25 – 12.7)	7.15 (2.59 – 13.66)	7.48 (3 – 15.2)	<0.001
Failure	--	10.77 (5.02 – 18.03)	12.26 (3.5 – 19.36)	10.81 (4.65 – 18.73)	0.916
Success	--	2.68 (1.19 – 5.5)	3.99 (2.3 – 7.61)	2.91 (1.33 – 6.04)	0.014
ICU free days (among those who were in the ICU)
Total	5.5 (0.56 – 12.59)	6.69 (2.03 – 11.96)	3.59 (0.3 – 9.41)	5.62 (0.99 – 12.22)	0.003
Failure	--	6.97 (1.69 – 12.71)	4.72 (0.74 – 11.81)	6.72 (1.43 – 12.43)	0.115
Success	--	5.6 (2.4 – 10.61)	2.27 (0.14 – 6.53)	5.03 (1.67 – 9.91)	<0.001
Hospital length of stay (days)
Total	17.59 (9.47 – 28.23)	12.32 (6.97 – 21.06)	11.75 (7.22 – 18.98)	14.46 (7.95 – 24.24)	<0.001
Failure	--	20.17 (13.79 – 29.27)	17.61 (11.44 – 27.77)	19.85 (13.22 – 29.02)	0.157
Success	--	9.8 (5.84 – 15.22)	9.86 (5.93 – 14.8)	9.82 (5.85 – 15.03)	0.655

Estimates are n (% of column n with available data) for categorical characteristics and median (interquartile range) for continuous characteristics. Inferences for categorical variables are the result of Fisher Exact Tests, with p-values computed via Monte Carlo simulation when necessary for computational efficiency. Inferences for continuous variables come from Kruskal-Wallis rank sum test of group differences.

NIRS vs IMV

Initial noninvasive respiratory support was associated with an increased hazard of in-hospital death compared to initial invasive mechanical ventilation (HR: 1.61, p < 0.0001, 95% CI: 1.33 – 1.94), and there was no significant interaction of treatment and time with this association (HR: 1.00, p = 0.6038, 95% CI: 0.98 – 1.01) Figure 2. However, being on initial noninvasive respiratory support was also associated with an early increased hazard of leaving the hospital alive, but this hazard ratio decreased over time due to an interaction between time and treatment. This eventually resulted in the hazard ratio reversing such that initial noninvasive respiratory support was later associated with a decreased hazard of leaving the hospital alive (HR: 0.97, p < 0.0001, 95% CI: 0.96 – 0.98) FIGURE 2. This was consistent with the sensitivity analysis where time zero is hospital admission (online supplement). There was also a similar pattern in the difference between initial NIRS and initial IMV in the time-to-ICU discharge alive: early on, initial noninvasive respiratory support was associated with an increased hazard of discharge alive from the ICU, but this hazard ratio decreased over time because of a statistically significant interaction between treatment and time (HR: 0.98, p = 0.0321, 95% CI: 0.97 – 1.00), resulting in no significant differences between non-invasive and invasive initial support at later time points..

Figure 2:

Kaplan-Meier curves for NIRS vs IMV for death (left) and discharge alive (right)

HFNO vs NIPPV vs IMV

Both HFNO (HR: 2.37, p < 0.0001, 95% CI: 1.85 – 3.03) and NIPPV (HR: 1.44, p = 0.0003, 95% CI: 1.18 – 1.74) were associated with increased hazards of in-hospital mortality compared to invasive mechanical ventilation. HFNO had an increased hazard of in-hospital death compared to NIPPV (HR: 1.65, p = 0.0001, 95% CI: 1.29 – 2.10). The interactions of time and both NIPPV and HFNO were not statistically significant and were excluded from the final model (both p > 0.05).

For time-to-hospital discharge alive, there were significant interactions between treatment and time for both NIRS modalities (time by HFNO HR: 0.96, p = 0.0011, 95% CI: 0.94 – 0.98 and time by NIPPV HR: 0.98, p = 0.0231, 95% CI: 0.97 – 1.00). Similar to the models that grouped NIPPV and HFNO, these models showed increased hazards of discharge alive for both HFNO and NIPPV compared to invasive mechanical ventilation that decreased over time but at different rates (HFNO faster, NIPPV slower), FIGURE 3. This resulted in no statistically significant differences between NIPPV and invasive mechanical ventilation at later time points but a decreased hazard of discharge alive for HFNO compared to invasive mechanical ventilation.

Figure 3:

Estimated Hazard Ratios for Time-to-Discharge Alive HFNO vs. IMV (left) and NIPPV vs. IMV (right) at various points in time. Pointwise 95% confidence intervals included.

Patients who started on HFNO first and were intubated were more likely to be intubated sooner than those who started on NIPPV (p < 0.001). First treatment with both IMV and NIPPV were associated with a higher hazard of ICU discharge alive than first treatment with HFNO (IMV HR: 1.50, p = 0.0226, 95% CI: 1.06 – 2.13; NIPPV HR: 1.84, p = 0.0014, 95% CI: 1.27 – 2.68), and first treatment with NIPPV was associated with a higher hazard of leaving the ICU alive than first treatment with IMV (HR: 1.23, p = 0.0341, 95% CI: 1.02 – 1.48). There was no significant interaction between time and treatment for time-to-ICU discharge alive (time by HFNO HR: 0.99, p = 0.4065, 95% CI: 0.97 – 1.01; time by NIPPV HR: 0.99, p = 0.3047, 95% CI: 0.97 – 1.01), and these interactions were left out of the final model. All competing risk sensitivity analyses showed the same pattern and were also statistically significant (all p < 0.001, see online supplement). In the sensitivity analysis of time-to-hospital discharge alive, the interaction of time with HFNO was statistically significant (p = 0.0055) but the interaction of time with NIPPV was not (p = 0.0848). Despite this, the relationship between therapies with the time by treatment interaction was consistent with the model where time zero was set to first treatment initiation.

Discussion

Our results show that initial noninvasive respiratory support presents a double-edged sword for COVID-19 patients with acute respiratory failure. Patients supported first by NIRS experience a greater hazard of in-hospital death compared to those intubated first. However, those same patients also experience a greater hazard of being discharged alive up to about 15–35 days, after which that hazard decreases either to the point of reversing direction (for NIRS in general and HFNO specifically) or no longer being statistically significant (for NIPPV). Essentially, patients initially treated noninvasively have higher hazards of both in-hospital death and discharge alive up to about a month. After that, their hazard of in-hospital death remains elevated but their hazard of discharge alive decreases. Moreover, those started on HFNO first experience faster intubation compared to those started on NIPPV first.

These results offer several important contributions. Studies utilizing NIRS for patients with COVID-19 compare one or both noninvasive respiratory support strategies either to each other or to conventional oxygen therapy and have reported mixed findings. Some report improved outcomes with high flow nasal oxygen compared to conventional oxygen therapy,^41–43 or to noninvasive positive pressure ventilation,⁴⁴ while others report improved outcomes with noninvasive positive pressure ventilation compared to conventional oxygen therapy,⁴⁵ or to high flow nasal oxygen.⁴⁶ Yet some studies report no difference in outcomes compared to conventional oxygen therapy,^43,45,47 or to each other.^46–50 Those studies utilize intubation directly as the primary outcome alone or as part of a composite outcome with mortality.^{41–48,51,52} We found that invasive mechanical ventilation as a comparator intervention may be the optimal method of respiratory support for COVID-19 patients that need more than conventional oxygen.

Additionally, most studies are limited to patients admitted to the ICU, which carries inherent selection bias and confounding potentially reflected as higher intubation rates of around 50%. We included all patients regardless of location as many patients were cared for outside of the physical ICU location or the ICU service from pandemic-strained resources.

We found that the benefit of leaving the hospital alive has a relationship with time. There are several potential explanations for this. NIRS cohorts are classified by a phenotyping algorithm based on the first therapy received but could have crossed over between NIRS modalities at any point with or without intubation. Crossover in those patients requires failing NIRS twice before intubation. Second, intubation could have been delayed for various reasons at the cost of worsening acute lung injury that may have benefitted from lung protective ventilation. Gershengorn found similar results in that the overall 46% failure rate on HFNO evolved from a decreased odds of failure for those on HFNO for a short time to an increased odds of failure for those on HFNO for longer.⁵² Similar findings were reported with NIPPV, which reduced mortality only in patients with short hospital stays, but was associated with higher mortality in those hospitalized longer than seven days.⁵³

There are several issues to consider when generalizing and interpreting data on noninvasive respiratory support and invasive mechanical ventilation in COVID-19 patients. Early in the pandemic, early intubation was often recommended to avoid aerosol exposure to healthcare staff and based on the high failure rates of NIPPV in the SARS epidemic. This was followed by a concerningly high early mortality reported in patients on mechanical ventilation²² and debate over the clinical syndrome seen in COVID-19 that contributed to variability in mechanical ventilation (e.g., high tidal volumes or alternative ventilator modes) or an avoidance of intubation altogether in some patients.^54–59 Simultaneously, pharmacological adjuncts to standard of care have changed over time, especially during the duration of this study⁶⁰, including convalescent plasma ^61–63, corticosteroids,^64–68 interleukin ^69–76 and janus kinase ^77–79 inhibitors, antivirals ^80–83, hydroxychloroquine ^80,84,85, and anticoagulation strategies.^86–88 Lastly, patient surges⁸⁹ and surge capacity almost certainly contributed to patient outcomes as more patients were managed outside of traditional ICUs, or in expansion ICUs, and facilities were faced with impending ventilator shortages, staffing ratio changes, and an increase in traveling staff. These results are, however, hypothesis generating for non-COVID-19 acute respiratory distress syndrome patients where NIPPV has been shown in some studies to be associated with higher mortality than IMV.⁹⁰

There are also important limitations to our study. Our data are limited to the COVID-19 patients in 2020 and the evolution of COVID-19 management potentially limits generalizability. Additionally, results are based on the first assigned therapy and patients who crossed over between NIRS and any imbalance between those crossovers could confound the findings. There were also some necessary assumptions based on the nature of the observational dataset from the electronic health record. For instance, we only considered a given patient’s first COVID-19 hospital visit where they required respiratory support. If two visits that appeared to belong to the same person were less than 24 hours apart, we assumed a hospital transfer occurred and combined these records into the same visit. We attempted to control for confounding, including by indication, by weighting our analysis by the inverse probability for treatment assignment, further adjusting for potential confounders in the Cox models, and by a competing risks analysis where death and intubation were competing risks. Additionally, some patients had evidence of all three treatments but no clear treatment ordering for which we performed sensitivity analyses to assess how dependent our results are on our specific inclusion, exclusion, or imputation decisions. Lastly, clinical care was not protocolized and there were likely important differences between flow rates with HFNO, airway pressures with NIPPV (of which only facemask NIPPV was available), and mechanical ventilator settings that could confound these results. Despite these limitations, the size of our dataset and statistical analyses still provide valuable results.

We found that patients intubated early without a trial of noninvasive respiratory support had a lower hazard of in-hospital death, but also a lower hazard of hospital discharge alive up to about a month. Clinical focus should shift to not delaying intubation in patients in whom noninvasive respiratory support has not reduced work of breathing or hypoxemia. Meanwhile, studies are needed to identify optimal patients for each noninvasive respiratory support modality and early prediction and identification of failure of that modality.