Routine low-dose continuous or nocturnal oxygen for people with acute stroke: three-arm Stroke Oxygen Supplementation RCT

Prevalence of hypoxia and effects on outcome

Mild hypoxia is common in stroke patients and may have significant adverse effects on the ischaemic brain. 7 Hypoxia with oxygen saturations falling below 92% has been observed in 24% of continuously monitored stroke patients within the first 24 hours of symptom onset. 8 While healthy adults with normal cerebral circulation can compensate for mild hypoxia by an increase in cerebral blood flow,9 this is not possible in the already ischaemic brain after stroke. 10–13 Hypoxaemia with oxygen saturations falling below 90% in the first few hours after hospital admission is associated with a doubling of mortality14 and a trebling of the rate of institutionalisation. 15 Patients on a stroke unit are more likely to receive oxygen than patients on a non-specialised general ward16 and less likely to be hypoxic. 17 An observational study of processes of care has shown that hypoxia increases the risk of an adverse outcome fivefold if only some of the hypoxic episodes are treated with oxygen, but has no adverse effect on outcome if all episodes of hypoxia are treated with oxygen. 15 Prophylactic oxygen treatment could prevent hypoxia and secondary neurological deterioration.

Potential adverse effects of oxygen treatment

However, oxygen treatment is not without side effects. 18 It impedes early mobilisation and could pose an infection risk. There is evidence from animal models and in vitro studies that oxygen encourages the formation of toxic free radicals, leading to further damage to the ischaemic brain,19–22 especially during reperfusion. Marked changes in adenosine triphosphate and related energy metabolites develop quickly in response to acute ischaemia and tissue hypoxia. These alterations are only partially reversed on reperfusion despite improved oxygen delivery. Ischaemia-induced decreases in the mitochondrial capacity for respiration result in reduced oxygen consumption and further increase free radical generation during reperfusion. 23 Oxidative stress has also been implicated in the activation of cell signalling pathways, which leads to apoptosis and neuronal cell death. 24,25 While much research points towards adverse effects of hyperoxia in the ischaemic brain, there is also evidence to support the notion that therapy-induced eubaric hyperoxia may be neuroprotective. 19,20,26,27 Routine oxygen supplementation (ROS) for acute myocardial infarction has been abandoned after a clinical trial showed no benefit and potential harm. 28

Randomised controlled trials of oxygen treatment after acute stroke

Evidence from randomised controlled trials of oxygen supplementation after acute stroke is conflicting and insufficient to guide clinical practice. A quasi-randomised study of oxygen supplementation for acute stroke by Rønning and Guldvog29 has shown that routine oxygen treatment in unselected stroke patients does not reduce morbidity and mortality. Subgroup analyses suggested that patients with severe strokes were more likely to benefit than those with mild strokes, but the study size was too small to define patients who are likely to derive benefit with certainty. 29 A very small (n = 16) study of high-flow oxygen treatment after acute stroke showed that cerebral blood volume and blood flow within ischaemic regions improved with hyperoxia. By 24 hours, magnetic resonance imaging of the brain showed reperfusion in 50% of hyperoxia-treated patients compared with 17% of control patients (p = 0.06) but no long-term clinical benefit at 3 months. 30 In the Stroke Oxygen Pilot Study (ISRCTN 12362720), the flow rate of oxygen was lower (2 or 3 l/minute dependent on baseline oxygen saturation) and treatment was continued for longer (72 hours). Neurological recovery at 1 week was better in the oxygen group than in the control group. 31 While there was no difference in the mRS at 6 months, there was a trend for better outcome with oxygen after adjustment for differences in baseline stroke severity and prognostic factors. 32 In contrast to the earlier study by Rønning and Guldvog,29 oxygen was as effective in mild as in severe strokes. These results are promising, but need to be confirmed in a larger study.

Recommendations for oxygen treatment in national and international clinical guidelines

Clinical guidelines for oxygen supplementation after stroke are not based on evidence from randomised clinical trials, and have changed over time without obvious reason. The European Stroke Organization (2008)33 stated that ROS to all stroke patients had not been shown to be effective, but that adequate oxygenation was important, and that oxygenation can be improved by giving oxygen at a rate of > 2 l/minute (no target saturation or supporting evidence given). In 2003, the American Stroke Association Guideline34 recommended keeping the oxygen saturation at or above 95%. The 2005 update25 of the guideline made no change to the recommendations. In 2007, the advice was revised to say that oxygen saturation should be maintained at or above 92%,35 but in 2013 the Association reverted to recommending maintenance of an oxygen saturation at or above 95%. 36 In the UK, the National Clinical Guideline for the management of people with stroke recommended keeping oxygen saturation within normal limits in 2005 and in 200837,38 and in 201239 it specified that this means maintaining an oxygen saturation ≥ 95%. None of the recommendations is based on evidence from controlled clinical trials. A survey of British stroke physicians showed that there is uncertainty among physicians treating patients with stroke about which treatment approach to take, and when to give oxygen. 40

Rationale for the fixed dose oxygen regimen used in the Stroke Oxygen Study

A fixed dosage scheme was chosen to keep the design of the study as simple as possible, so that any recommendations resulting from the study outcome can be carried out in day-to-day clinical practice.

Rønning and Guldvog29 have shown that giving oxygen at a rate of 3 l/minute to all stroke patients during the first 24 hours after hospital admission does not improve overall outcome. They did not report baseline oxygen saturation, or changes in saturation on treatment. It is therefore possible that some patients were undertreated, and that others achieved too high oxygen levels, leading to an increase in free radical generation in the ischaemic penumbra. There were no other data from clinical studies to inform recommendations for the dose of oxygen to give for routine supplementation at the time the study was designed. The European Stroke Organization suggested a dose of 2–4 l/minute and the American Stroke Association Guideline recommended keeping the oxygen saturation at or above 95%,34,41,42 but these recommendations were not based on evidence from controlled clinical trials. In the absence of data to the contrary, it was reasonable to assume that treatment should restore oxygen saturation to the normal range.

Normal oxygen saturation in adults is 95–98.5%,43 in healthy older individuals it is reported to be lower, at 95% [standard deviation (SD) 2.5%]. 44 Oxygen saturation in stroke patients who are normoxic at recruitment is about 1% lower than that of age-matched community control patients. 45 We have conducted a dose titration study for oxygen after acute stroke and found that 2 l/minute oxygen by nasal cannula increases oxygen saturation by 2% and 3 l/minute by 3%. 46 We also found that oxygen masks were less likely to be tolerated than nasal cannulae, leading to poorer treatment compliance with the former. For this study, we therefore decided to give oxygen by nasal cannula. A dosage regimen of 3 l/minute in individuals with a baseline oxygen saturation of ≤ 93% and 2 l/minute in individuals with a baseline saturation > 93% was therefore considered likely to prevent hypoxia without increasing oxygen saturation beyond the upper limit of the normal range.

Rationale for giving routine oxygen supplementation at night only

Recruitment

Patients were recruited by research nurses and clinicians from 136 hospitals (secondary care) across England, Northern Ireland and Wales (see Appendix 1 for a list of recruiting centres). All recruiting hospitals had acute stroke units, were able to carry out the required observations four times a day and had a stroke-trained principal investigator. Screening, baseline, randomisation and 1-week follow-up data were collected and entered online into the trial database by the research staff based in the local hospital.

Inclusion criteria

• Adult patients (aged ≥ 18 years) with clinical diagnosis of acute stroke.

• Within 24 hours of hospital admission.

• Within 48 hours of stroke onset.

Exclusion criteria

• Definite indication or contraindication to oxygen treatment at a rate of 2–3 l/minute.

• Stroke not the main clinical problem or patient has another serious life-threatening illness likely to lead to death within the next 12 months.

Consent

Fully informed consent was sought from all research participants. Research nurses or clinicians explained the study to potential participants, who were also given a patient information sheet explaining the trial. Owing to the acute nature of stroke and the intervention being tested, there was not a 24-hour consideration period between receiving the information and taking informed consent. In patients who were unable to give fully informed consent, assent was sought from either the patient’s next of kin or from an independent physician. Fully informed consent was obtained in patients who regained capacity during the first week following randomisation.

Baseline assessment

This was done by the research team randomising the patient, and was either entered online for patients randomised via the web or sent to the trial centre by fax for patients randomised via telephone. The initial assessment included baseline demographics, date and time of the event, whether or not the patient had been given oxygen in the ambulance or in the emergency department, and how much, comorbidities [chronic obstructive pulmonary disease, other chronic lung problems, heart failure (congestive cardiac failure), ischaemic heart disease, and atrial fibrillation], the Glasgow Coma Scale (GCS) score,53 score on the Six Simple Variables (SSV) outcome predictor tool,54 the National Institutes of Health Stroke Scale (NIHSS) score,55,56 the type of consent (by patient or legal representative) and the date and time of randomisation.

Week 1 assessment

The week 1 assessment was performed by a member of the local research team trained in the assessment tools at 7 days (± 1 day) after enrolment. In patients who were discharged before the end of 1 week, or who could not be followed at 7 days, the assessment was conducted at discharge. Data were entered online or sent to the trial centre via fax. The week 1 assessment included neurological function (NIHSS), vital status, adverse events, whether or not the patient was prescribed oxygen for clinical indications during the first 72 hours, information on compliance with the treatment, details of other treatments (antibiotics during week 1, thrombolysis, sedatives or antipsychotics), physiological variables (highest heart rate, highest systolic and diastolic blood pressure, highest and lowest oxygen saturation during the first 72 hours, and the highest temperature during week 1), the result of the computerised tomography (CT) or magnetic resonance imaging head scan (cerebral infarct/primary intracerebral haemorrhage/subdural haemorrhage/brain tumour/head scan not performed/other), the final diagnosis [ischaemic stroke/transient ischaemic attack (TIA)/primary intracerebral haemorrhage/cerebrovascular accident (CVA) without CT confirmation of aetiology/other], and the Trial of Org 10172 in Acute Stroke Treatment (TOAST) classification of stroke aetiology. 57 Compliance was assessed by asking whether or not oxygen was prescribed on the drug chart, whether or not it was signed, and whether or not it was stopped before the end of 72 hours. After an amendment of the case report form [version 1, amendment 3 (30 August 2009)], additional details were recorded for the final 4143 patients. These included a more in-depth assessment of compliance with a record of oxygen saturation at 06:00, 12:00 and 00:00, whether or not oxygen treatment was in place at 06:00 and 00:00; whether or not the patient had been enrolled into another study, and which, the pre-stroke modified Rankin Scale (mRS) and the European Quality of Life-5 Dimensions, three levels (EQ-5D-3L), whether or not this was reported by the patient or a relative, discharge destination (if discharged), and whether or not a new brain haemorrhage was identified on a second CT of the head (if conducted).

Three-, 6- and 12-month assessments

The 3-, 6- and 12-month follow-up assessments were performed centrally by the Stroke Oxygen Study team. Following a call to the participant’s general practitioner to confirm that the participant was alive and the contact details were the same as those on the trial database, postal questionnaires were sent to all participants at 3, 6 and 12 months post randomisation. The questionnaires contained the discharge date, the mRS,58 the Barthel Index (BI),59 the EQ-5D-3L and the EQ-VAS,60–62 the Nottingham Extended Activities of Daily Living scale (NEADL),63 questions regarding current abode, whether or not the patient had been readmitted to hospital since the stroke, patient-reported outcome measures (sleep, speech and memory), and a question on whether or not they remembered which treatment they were randomised to. If the questionnaire was not returned within a few weeks, a data assistant at the trial co-ordinating centre telephoned the patient and completed the questionnaire over the telephone with either the patient or a relative/carer. See Table 1 for a summary of all patient assessments and timings.

Primary outcome

The primary outcome measure was disability assessed by the mRS at 90 days post randomisation. 58,64 The mRS is an ordinal scale that ranges from 0 for a patient with no disability to 5 for extreme disability. Death was included in the scale as a score of 6.

Secondary outcome measures at 1 week

The number of patients with neurological improvement (≥ 4-point decrease from baseline in the NIHSS score or a value of 0 at day 7),55,56 NIHSS score, mortality, the lowest and highest oxygen saturation during the 72-hour treatment period and the number of patients whose oxygen saturation fell below 90% were all secondary outcome measures.

Secondary outcome measures at 3, 6 and 12 months

Mortality, the number of patients alive and independent (mRS score of ≤ 2), the number of patients living at home, the BI of activities of daily living, quality of life EQ-5D-3L and European Quality of Life Visual Analogue Scale (EQ-VAS) and the NEADL index were all secondary outcome measures for 3-, 6- and 12-month time points.

Data at 6 and 12 months

The longer-term follow-up data at 6 and 12 months were analysed at each time point using the same methods as those listed previously. In addition, analyses were performed across 3-, 6- and 12-month time points using a longitudinal repeated measures analysis, in this case linear mixed models. 67

The treatment effect was initially assumed to be constant over time; further analyses were carried out to investigate the effects of including time and a treatment-by-time interaction in the models.

Mortality was analysed using log-rank methods (unadjusted analysis) with Kaplan–Meier plots. The adjusted analysis used Cox regression methods, including the covariates listed above. In the covariates, the prognostic index for 30-day survival replaced that for independence at 6 months. The proportional hazards assumption associated with the Cox regression was tested via Schoenfeld residuals. Hazard ratios and 95% CIs are reported for both the unadjusted and adjusted analyses.

Planned subgroup analyses

These were performed in respect of the primary outcome measure only, based on a risk-stratification approach. 68 The subgroups comprise:

• NIHSS score at baseline as indicator of stroke severity (0–4, 5–9, 10–14, 15–20 and > 20)

• baseline % oxygen (O₂) saturation (< 94%, 94–94.9%, 95–97% and > 97%)

• treatment with O₂ prior to randomisation (yes/no)

• time in hours since onset of stroke (< 4 hours, 4 to < 7 hours, 7 to < 13 hours, 13 to < 24 hours and ≥ 24 hours)

• final diagnosis (haemorrhage, infarct, TIA and other)

• TOAST classification of infarct aetiology

• GCS score (motor plus eye score:< 10 and 10)

• age (< 50 years, 50–80 years and > 80 years)

• history of chronic obstructive airway disease or asthma (yes/no)

• history of heart failure (yes/no)

• thrombolysed (yes/no)

• baseline SSV risk score for independence at 6 months (≤ 0.1, > 0.1 to 0.35, > 0.35 to 0.7 and > 0.7).

These subgroup effects were analysed by means of an interaction term;69 however, pairwise hypothesis tests between the levels of the subgroup factor were not performed owing to the likely low level of statistical power. Subgroup-specific estimates are reported descriptively with 99% CIs and displayed graphically on a forest plot.

Overview

A within-trial economic evaluation was conducted alongside the clinical trial in order to estimate the cost-effectiveness and cost–utility of ROS compared with no oxygen supplementation (no ROS) after stroke, over 12 months’ follow-up. Further analysis compared all three trial arms: (1) no ROS, (2) nocturnal oxygen for three nights, and (3) 72 hours’ continuous oxygen. The base-case economic evaluation adopted an NHS/Personal Social Services perspective. The cost-effectiveness analysis calculated the cost per additional day of home time gained, using information on length of stay in hospital and discharge destination. The cost–utility analysis (CUA) used quality-adjusted life-years (QALYs) as the benefit measure, in which QALYs take into account the survival and quality of life of an individual. The reporting of this analysis follows the Consolidated Health Economic Evaluation Reporting Standards (CHEERS). 70

Health outcomes

Information on length of stay in hospital due to stroke, discharge destination and any readmissions to hospital were used to calculate the number of home time days per patient over a 12-month period. Home time has been previously used as an outcome measure in other stroke trials. 71,72 When data on actual discharge location were not available, the response to a question regarding place of residence at 3 months was used as a proxy for the discharge location. Discharge to institutional care (nursing home or residential care) was not counted as home time. The actual length of stay for readmissions was not available for all patients in the trial; therefore, a sample of 100 readmissions was analysed to determine the mean length of stay for a readmission. This value was attached to all patients who had a readmission, and used in the calculation of home time gained. Although the intention was to calculate home time gained in the first 90 days, as there were poor data on the date of readmission, home time gained was calculated over the full 12 months.

The EQ-5D-3L questionnaire73 was completed at 3, 6 and 12 months, and when a patient had died during the 12 months, the date of death was noted and a value of zero (equivalent to dead) was assumed from the date of death. EQ-5D-3L data were not collected at baseline, as it was not possible for a patient admitted for a stroke to fill in a health-related quality of life questionnaire. Therefore, we used a previously published method74 and assumed that the EQ-5D-3L score for all patients at baseline was zero, and the change in quality of life between baseline and 3 months was linear. An alternative method for baseline EQ-5D-3L score was used in a sensitivity analysis by assuming that the baseline quality of life was equal to the EQ-5D-3L value at 3 months. Quality of life estimates were derived from EQ-5D-3L responses provided by patients at each time point by applying the standard UK tariff values. 75 These estimates were then used to calculate total QALYs over 12 months for every individual in the study, using the area under the curve approach.

Resource use and costs

The costs included in the analysis related to oxygen administration as prescribed by the trial, additional oxygen required for other clinical reasons, length of stay in hospital, readmission to hospital and long-term care on discharge from hospital. All costs in the analysis were in UK pounds (£), based on a price year of 2013–14. Unit costs were obtained from published standard sources of costs for NHS procedures,76 staff costs77 and previously published research (Table 2). Health and Community Health Services pay and price indices were used to inflate costs, when appropriate. 77 Unit costs are listed in Table 2.

For the purposes of estimating the cost of oxygen administration as prescribed in the trial, it was assumed that, once a patient was allocated to a treatment arm, costs of treatment were incurred. The cost of oxygen administration was adjusted for those patients for whom continuous oxygen was prescribed for clinical reasons outside the trial treatment, as described in Table 3. Information on resources required for oxygen administration, including any additional staff time and equipment required for each treatment group, was collected prospectively during the trial using a short questionnaire filled in by a representative in 24 participating hospitals (see Appendix 3). The cost of each resource type was calculated in order to determine the cost of oxygen supplementation per trial arm (Table 4). All institutions included in the trial were assumed to have the same expertise and to have followed similar protocols in the management of patients.

Patients in the trial had a stroke, TIA or a stroke mimic. For a stroke, a NHS reference cost for a CVA was assumed, as the majority of strokes were ischaemic. As there were five different categories for CVA, taking into account the number of complications and comorbidities, the median was calculated for the cost of a non-elective long stay, cost of an excess bed-day and average length of stay. These values were then adjusted for the length of stay in hospital for each patient, either adding or subtracting a bed-day cost for length of stay under or over the median length of stay. For patients who had a TIA or stroke mimic, the NHS reference cost for a TIA was assumed, using the same methodology as for CVA to adjust for length of stay. As full data on the details and length of stay of any non-elective readmissions during the 12 months were not available, the overall average cost of a non-elective admission (for all categories) was used.

Patient-level resource use on long-term care beyond the initial hospital admission was unavailable; however, the responses to the mRS provided information on the level of dependence. Patients were categorised as independent (mRS score of 0–2) or dependent (mRS score of 3–5) using data from the 3-month questionnaire. The annual cost of independent and dependent stroke after discharge from hospital was obtained from Sandercock et al. 79 and updated to 2013/14 costs. This annual cost was then adjusted to take into account the time not in hospital over the 12-month period.

For the purpose of the analysis, the costs of acute stay in hospital and long-term care were not included in the cost-effectiveness analysis to avoid double counting, as the measure of outcome was home time gained.

Analysis

The EQ-5D-3L and mRS data were not available for all randomised patients, and therefore, multiple imputation (MI) was used. MI is a statistical technique that retains overall population variability and the relationship between observations, and is considered useful when > 10% of data are missing. As > 10% of the data were missing within the trial, these were treated as missing at random and estimated using the Markov chain Monte Carlo MI method. Imputation of missing EQ-5D-3L scores used methods proposed by Simons et al. ,80 in which the whole index score was imputed. The percentage of missing EQ-5D-3L data at 12 months was 18%; therefore, 25 simulated, complete versions of the data set were produced using Stata 12.1 software (StataCorp LP, College Station, TX, USA). The results of each of the simulated data sets were combined to produce estimates and CIs to incorporate missing data uncertainty. MI was carried out at all time points. Appendix 4 contains further detail on the imputation methods.

As the majority of cost and outcome data are usually skewed, normal parametric methods are not appropriate for calculation of the differences in means. Bootstrapping is a non-parametric approach that can be used to compare arithmetic means without making any assumptions regarding the sampling distribution. In this analysis, 3000 bootstrapping replications were undertaken in order to calculate the 95% CIs around the differences in mean costs and outcomes.

The incremental cost-effectiveness analysis was carried out at 12 months, based on the outcome of home time gained, and the incremental cost-effectiveness ratio (ICER) was expressed in terms of the cost per 1 additional day of home time gained at 12 months. A CUA was also carried out at 12 months and the ICERs were expressed as cost per additional QALY gained. The presentation of results in QALYs allows comparison of the results with other available published studies. The analysis was conducted according to the intention-to-treat principle, in line with the main trial analysis, and discounting was not applied as the duration of follow-up was only 1 year. First, an analysis of ROS compared with no ROS was conducted by including all patients in the nocturnal and continuous ROS arms in the overall ROS comparator. A subsequent analysis considered all three trial arms, using the principle of dominance: that is, if one of the trial arms was shown to be both more costly and less effective than at least one of the alternative interventions, then that option would be seen as dominated and excluded from remainder of the analysis.

A range of one-way deterministic sensitivity analyses were carried out to explore the robustness of the base-case cost–utility results for ROS compared with no ROS, and to assess the uncertainty associated with input parameters. The following sensitivity analyses were undertaken:

• changing the costs of oxygen treatment (increasing/reducing the costs by 20%)

• changing the acute inpatient costs of stroke (reducing/increasing the costs by 20%)

• changing the costs of stroke care after discharge (reducing/increasing the costs by 20%)

• assuming that quality of life changes between base-case and 3 months took place immediately, therefore allocating the 3 month EQ-5D-3L value to baseline.

In addition, a probabilistic sensitivity analysis of the base-case analysis (cost-effectiveness analysis and CUA) was carried out to enable the simultaneous exploration of uncertainty in the cost and outcome data, using 5000 simulations. The results of the probabilistic sensitivity analysis are presented using cost-effectiveness planes and cost-effectiveness acceptability curves (CEACs). The CEAC graphically represents the probability that an intervention is cost-effective at different ICER thresholds.

All analyses were carried out using Stata version 12.1 and Microsoft Excel^® 2007 (Microsoft Corporation, Redmond, WA, USA).

Sensitivity analyses for the primary outcome

Sensitivity analyses (Table 9) show very similar results for the complete case analysis and the analysis using MI for missing values, both confirming that oxygen treatment does not improve the primary outcome. The best- and worst-case imputations indicate the plausible maximum bounds of any potential bias from missing data (under a missing not at random assumption), showing improvement in outcome for oxygen for the best-case analysis and worse outcomes with oxygen for the worst-case analysis, with similar effect sizes in both directions. The adherers-only analyses showed a minor difference between the combined oxygen groups and control, with slightly lower odds for a good outcome with oxygen for the comparison of the combined oxygen groups with control, which was not statistically significant.

Oxygenation

Results for the highest oxygen saturation, the lowest oxygen saturation, the number of patients who had desaturations below 90% and the number of patients who needed additional oxygen over and above the trial prescription during the first 72 hours are shown in Table 10. The highest and lowest oxygen saturations recorded during the treatment period increased significantly (p < 0.001), by 0.8% and 0.9% respectively, in the continuous oxygen group when compared with the control group. This was also seen in the nocturnal oxygen group with highest and lowest oxygen saturations increased by 0.5% and 0.4%, respectively (p < 0.001). Severe hypoxia was recorded in a small number of patients (143, 2%), but was significantly less common in the combined oxygen group than in the control group. Significantly more patients in the combined oxygen group than in the control group required oxygen in addition to the trial intervention (OR 1.36, 99% CI 1.07 to 1.73; p = 0.0008). This was also not statistically different when comparing continuous with nocturnal oxygen (OR 1.23, 99% CI 0.96 to 1.59; p = 0.03), as significance was defined as < 0.01 for secondary outcomes.

Neurological recovery at 1 week

Data for neurological recovery and mortality at 1 week are shown in Table 11. The median (IQR) NIHSS score at week 1 was 2 (1–6) in all three treatment groups. There was no difference in the number of patients who improved by 4 or more NIHSS points between baseline and week 1. Mortality by day 7 was very low in all three treatment groups with 50 (1.9%), 35 (1.3%) and 45 (1.7%) deaths in the continuous oxygen, nocturnal oxygen and control groups, respectively.

Explanatory analysis at 1 week

Exploratory analyses (see Table 11) did not show evidence of increased stress levels (higher heart rates, higher blood pressure, or need for sedation) in oxygen-treated patients compared with control patients. There was also no evidence that oxygen treatment was associated with more infections, with no differences in the highest temperature or the need for antibiotics.

Survival

Mortality at 90 days (Figure 6) was similar in the oxygen (both groups combined) and control groups (hazard ratio 0.97, 99% CI 0.78 to 1.21; p = 0.8) and in the groups recieving continuous oxygen or oxygen at night only (hazard ratio 1.15, 99% CI 0.90–1.48; p = 0.1).

FIGURE 6.

Kaplan–Meier survival graph comparing (a) oxygen (combined) with control (no ROS) at 90 days; and (b) continuous vs. nocturnal oxygen. Oxygen compared with no oxygen: unadjusted hazard ratio for a worse outcome – 0.97 (99% CI 0.78 to 1.21; p = 0.8). Continuous compared with nocturnal oxygen: unadjusted hazard ratio for a worse outcome –1.15 (99% CI 0.90 to 1.48; p = 0.1).

Survival was the same throughout the 365 days of follow-up for patients treated with continuous oxygen and with nocturnal oxygen, and those in the control group (Figure 7).

FIGURE 7.

Kaplan–Meier survival graph comparing (a) oxygen (combined) with control (no ROS) at 365 days; and (b) continuous compared with nocturnal oxygen.

Functional outcomes

Figure 8 shows the mean mRS score at 3, 6 and 12 months. There was no change in the level of disability over the 3-, 6- and 12-month follow-up assessment points in any of the three groups. The proportion of patients in each mRS score category is shown in Figure 9 for the combined oxygen group compared with the control group and in Figure 10 for the continuous oxygen group compared with the nocturnal oxygen group. Oxygen has no effect on the level of disability at any time point, whether given continuously or at night only.

FIGURE 8.

Modified Rankin Scale score at 3, 6 and 12 months.

FIGURE 9.

Functional outcome of the combined (continuous and nocturnal) oxygen group compared with the control group at (a) 3; (b) 6; and (c) 12 months post randomisation.

FIGURE 10.

Functional outcome of the continuous oxygen group compared with the nocturnal only oxygen group at (a) 3; (b) 6; and (c) 12 months post randomisation.

The number of patients who are alive and independent was similar in all three treatment groups and did not change with time (52%, 53% and 53%, respectively, at each of the three time points). Performance of activities of daily living (BI), ability to conduct extended activities of daily living (NEADL), and quality of life (EQ-5D-3L) were no different between the combined oxygen groups or between continuous and nocturnal oxygen at 90, 180 and at 365 days and were similar at each of the three time points (Table 13).

Outcomes considered important by stroke survivors

Outcomes considered important by stroke survivors during the pre-study focus group meetings (sleep, speech and memory) are shown in Table 13. There was also no difference in these outcomes between the continuous oxygen, nocturnal oxygen and control groups. At 90 days, 87% had no significant speech problems and 65% considered their sleep as good as before the stroke, but only 44% reported their memory as being as good as before the stroke. The data were similar at 180 and 365 days.

Length of hospital stay and readmissions

Data on length of stay were available for 2435 (91%), 2401 (90%) and 2446 (92%) of participants in the continuous oxygen, nocturnal oxygen and control groups. The mean (SD) length of stay in hospital was 18.6 days (50.9 days), 18.6 days (52.0 days) and 18.0 days (56.4 days), respectively (Table 14). The rate of readmissions (Table 15) increased with time, but was similar for all three treatment groups (14%, 17% and 20% at 3, 6 and 12 months, respectively).

Place of abode

Details of place of abode throughout the follow-up period are shown in Table 15. The proportion of patients living in their own homes, with relatives, in residential nursing, or continuing care homes was similar in all three groups. The proportion of participants living in their own homes or with family members was 76%, 74% and 68% at 3, 6 and 12 months. Only a few patients were residing in institutions (care home, nursing home or continuing NHS care) at each of the three time points (7%, 7% and 6%, respectively).

Base-case analysis

Comparison of routine oxygen supplementation with no routine oxygen supplementation

Table 19 presents the mean outcomes per patient in terms of hospital stay, home time, EQ-5D-3L scores and QALYs. Both mean hospital stay and home time were marginally higher in the ROS group, but by less than half a day. There were no differences in mean EQ-5D-3L scores at 3, 6 and 12 months and the difference in QALYs was very small, at 0.0004 QALYs, and not significant. Table 20 presents disaggregated mean (SD) health-care costs per patient for trial oxygen and additional oxygen, acute stay costs, readmissions and long-term care, as well as total health-care costs. As expected, oxygen treatment costs were higher in the ROS arm, with all other costs slightly higher for ROS, and £206 (95% CI –£283 to £695) higher overall, compared with no ROS.

TABLE 19 - Mean outcomes per patient by treatment group over 12 months

Variable	Intervention
	No ROS (N = 2629)	ROS (N = 5269)
Mean hospital stay (days) (95% CI)	18.4 (17.7 to 19)	18.7 (17.9 to 19.5)
Mean home time (days) (95% CI)	313.0 (310.4 to 315.7)	313.4 (310.2 to 136.5)
EQ-5D-3L scores [mean (SD)]
Baseline	0	0
3 months	0.48 (0.38)	0.48 (0.37)
6 months	0.48 (0.37)	0.48 (0.37)
12 months	0.46 (0.37)	0.46 (0.37)
Total QALYs over 12 months	0.4133 (0.31)	0.4137 (0.30)
Difference in QALYs (95% CI)	0.0004 (–0.0139 to 0.0148)

TABLE 20 - Mean (95% CI) costs (£) per patient by treatment group over 12 months

Cost category	Intervention
	No ROS (N = 2629)	ROS (N = 5629)
Trial prescribed oxygen	0.65 (0.42 to 0.94)	72.89 (72.62 to 73.15)
Additional oxygen	5.52 (4.76 to 6.37)	9.40 (8.44 to 10.32)
Total oxygen treatment	6.18 (5.40 to 7.05)	82.29 (81.30 to 83.38)
Acute care costs	5130 (4856 to 5406)	5196 (5008 to 5418)
Readmission	978 (929 to 1032)	986 (951 to 1022)
Stroke care after discharge	6587 (6310 to 6865)	6643 (6447 to 6835)
Total cost	12,702 (12,309 to 13,115)	12,908 (12,619 to 13,183)
Mean difference (95% CI)	206 (–283 to 695)

The cost-effectiveness analysis presented in Table 21 shows that it costs an additional £71 to gain a day of home time. The CUA gave an ICER of £463,338 per QALY gained (Table 22). The cost-effectiveness plane in Figure 11 demonstrates the uncertainty around both costs and QALY differences, with points in all four quadrants. The corresponding CEAC in Figure 12 shows a 27% probability of ROS being cost-effective compared with no ROS if society was willing to pay up to £20,000 per additional QALY, also suggesting ROS is not cost-effective.

TABLE 21 - Cost-effectiveness analysis

Intervention	Mean costs (£)	Cost difference (£)	Mean days of home time	Outcome difference (days)	ICER (£/day of home time gained)
No ROS	6.2	–	312.28	–	–
ROS	82.3	76.1	313.36	1.08	70.74

TABLE 22 - Cost–utility analysis

Intervention	Mean costs (£)	Cost difference (£)	Mean QALYs	QALY difference	ICER (£/QALY gained)
No ROS	12,702	–	0.4133	–	–
ROS	12,908	206	0.4137	0.0004	463,338

FIGURE 11.

Cost-effectiveness plane for ROS compared with no ROS.

FIGURE 12.

Cost-effectiveness acceptability curve for ROS compared with no ROS.

Comparison of continuous routine oxygen supplementation and nocturnal routine oxygen supplementation with no routine oxygen supplementation

Table 23 presents the mean outcomes per patient in terms of hospital stay, home time, EQ-5D-3L scores and QALYs for all three trial arms. Mean hospital stay was the greatest in the nocturnal ROS group. Home time was lowest in the nocturnal ROS group and highest for continuous ROS, by almost 3 days compared with no ROS. There was almost no difference in mean EQ-5D-3L scores at 3, 6 and 12 months. Overall QALYs for nocturnal ROS were slightly lower (mean difference –0.0023 QALYs) than for no ROS, and slightly higher than for continuous ROS (mean difference 0.0032 QALYs), but in both cases differences were very small and non-significant. Table 24 shows the disaggregated mean (SD) health-care costs per patient for both trial and additional oxygen, acute stay costs, readmissions and long-term care, as well as total health-care costs. Oxygen treatment costs were highest in the continuous ROS arm, and there were small differences in other health-care costs between trial arms. Total health care costs were higher in both ROS arms and highest in the continuous ROS arm, £246 (95% CI –£322 to £814) higher overall compared with no ROS.

TABLE 23 - Mean outcomes per patient by treatment group over 12 months

Variable	Intervention
	No ROS (N = 2629)	Nocturnal ROS (N = 2633)	Continuous ROS (N = 2636)
Mean hospital stay (days) (95% CI)	18.4 (17.7 to 19.0)	19.0 (17.9 to 20.1)	18.5 (17.4 to 19.6)
Mean home time (days) (95% CI)	313.0 (310.4 to 315.7)	310.9 (306.6 to 315.6)	315.9 (311.4 to 320.2)
EQ-5D-3L scores, mean (SD)
Baseline	0	0	0
3 months	0.48 (0.38)	0.48 (0.37)	0.48 (0.37)
6 months	0.48 (0.37)	0.47 (0.37)	0.48 (0.36)
12 months	0.46 (0.37)	0.46 (0.37)	0.46 (0.36)
Mean (SD) total QALYs over 12 months	0.4133 (0.31)	0.4109 (0.31)	0.4164 (0.30)
Difference in QALYs (95% CI) (vs. no ROS)	–	–0.0023 (–0.018 to 0.014)]	0.0032 (–0.0132 to 0.0019)

TABLE 24 - Mean (95% CI) costs (£) per participant by treatment group over 12 months

Cost category	Intervention
	No ROS (N = 2629)	Nocturnal ROS (N = 2633)	Continuous ROS (N = 2636)
Trial prescribed oxygen	0.65 (0.42 to 0.94)	64.32 (64.12 to 64.50)	81.45 (81.21 to 81.67)
Additional oxygen	5.52 (4.76 to 6.37)	3.83 (3.27 to 4.46)	14.97 (13.32 to 16.73)
Total oxygen treatment	6.18 (5.40 to 7.05)	68.15 (67.50 to 68.79)	96.42 (94.61 to 98.26)
Acute care costs	5130 (4855 to 5404)	5125 (4862 to 5406)	5265 (4990 to 5583)
Readmission	978 (929 to 1032)	976 (924 to 1030)	997 (948 to 1052)
Stroke care after discharge	6587 (6310 to 5406)	6697 (6420 to 6984)	6588 (6305 to 6855)
Total cost	12,702 (12,309 to 13,115)	12,867 (12,471 to 13,286)	12,948 (12,523 to 13,347)
Mean difference (95% CI) (vs. no ROS)	–	165 (–393 to 725)	246 (–322 to 814)

The cost-effectiveness analysis presented in Table 25 shows that nocturnal ROS is dominated by no ROS. An intervention is dominated if it is more costly but less effective. Here, the outcomes for nocturnal ROS are worse, with fewer days of home time but at a higher cost than no ROS. Continuous ROS costs £25 per extra day of home time gained compared with no ROS. In the CUA, nocturnal ROS was again dominated because of lower QALYs and higher costs. Continuous ROS had an ICER of £76,997 per QALY gained, well above standard NICE willingness-to-pay thresholds of £20,000 to £30,000 per QALY (Table 26). The cost-effectiveness plane in Figure 13 displays a high level of uncertainty in both cost and QALY differences, with points in all four quadrants, and the corresponding CEAC (Figure 14) shows a 31% probability of continuous ROS being cost-effective compared with no ROS at a £20,000 per additional QALY threshold, suggesting that it is not a cost-effective intervention.

TABLE 25 - Cost-effectiveness analysis

Intervention	Mean costs (£)	Cost difference (£)	Mean days of home time	Outcome difference (days)	ICER (£/day of home time gained)
No ROS	6.2	–	312.28	–	–
Nocturnal ROS	68.2	62.0	310.87	–1.41	Dominated
Continuous ROS	96.4	90.2	315.87	3.59	25.13

TABLE 26 - Cost–utility analysis

Intervention	Mean costs (£)	Cost difference (£)	Mean QALYs	QALY difference	ICER (£/QALY gained)
No ROS	12,702	–	0.4133	–	–
Nocturnal ROS	12,867	165	0.4109	–0.0023	Dominated
Continuous ROS	12,948	246	0.4164	0.0032	76,997

FIGURE 13.

Cost-effectiveness plane for continuous ROS compared with no ROS.

FIGURE 14.

Cost-effectiveness acceptability curve for continuous ROS compared with no ROS.

Sensitivity analysis

Deterministic sensitivity analysis was undertaken for the CUA for ROS versus no ROS. This was to assess (1) the impact on results of changing the health care cost inputs and (2) an alternative approach to valuing baseline quality of life. Three main types of cost (oxygen, acute care, long-term care) were individually changed by increasing and then decreasing the base-case value by 20% (Table 27). The difference in costs between ROS and no ROS changed very little across the board and all ICERs were over £400,000 per QALY gained. The differences in ICERs do appear to be large; however, this is a result of the sensitivity of the ICER to small changes in cost, owing to the extremely small difference in QALYs.

A further analysis was undertaken to look at the impact of applying a different method for valuing the baseline quality of life of patients (Table 28). The base-case assumed an EQ-5D-3L score of zero and a linear increase in quality of life to 3 months. The alternative analysis used the EQ-5D-3L value at 3 months for the baseline value. This increased the total mean QALYs over 12 months for both ROS and no ROS and the difference in QALYs also increased to 0.0008, but this is still a very small difference. The ICER decreased to £257,500 per QALY gained, still considerably higher than the cost per QALY threshold of £20,000 to £30,000 per QALY.

Mismatch of baseline clinical and demographic characteristics

The large sample size (n = 8003) makes mismatch of baseline clinical characteristics highly unlikely. Further safeguards against mismatch were built into the protocol. Randomisation by minimisation including key prognostic factors (age, sex, living alone, normal verbal component of the GCS, ability to lift both arms, the ability to walk, routine oxygen treatment during ambulance transfer and baseline oxygen saturation) ensured equal distribution of these factors across the three treatment groups. Effectiveness of minimisation was monitored continuously during the course of the study via online minimisation reports. The results in Table 6 confirm that there were no differences in demographics and baseline clinical characteristics between the three groups. The negative results of this study cannot be explained by a mismatch in baseline clinical characteristics.

The patients recruited to the study were not representative of acute stroke patients

To ensure that patients in SO₂S were representative of acute stroke patients, inclusion criteria both for participating centres and for patient enrolment were wide. This is the largest and highest recruiting acute stroke study in the UK. In total 136 centres, including both acute and hyperacute stroke units, participated in the study. This means that over 50% of stroke services in the UK enrolled patients into SO₂S. Stroke patients enrolled into clinical studies tend to have less severe neurological deficits than the overall stroke population unless the inclusion criteria specify that only severe patients can be included. Key reasons for this are both ethical constraints (it would not be ethical to include moribund patients who have nothing to gain from the trial intervention) and problems with obtaining informed consent in a patient group who have suffered a brain injury and are therefore often unable to make a fully informed decision. In spite of these unavoidable constraints, stroke severity in SO₂S was representative of the stroke population in the UK. Participants in SO₂S had a median NIHSS score of 5. While this seems low, it reflects patients admitted to acute stroke units. Data from 23,199 patients in the UK Sentinel Stroke National Audit Programme show a median NIHSS score of 4 for stroke patients in the UK. 82 The median NIHSS score of 127,950 patients with acute ischaemic stroke in the US Get with the Guidelines Register was 5,83 as in SO₂S. A median NIHSS score of 5 was also recorded at baseline in a recently published Dutch study of antibiotic prophylaxis after stroke,84 which had similarly wide inclusion criteria to those of SO₂S. This is a study of routine prophylactic oxygen supplementation in acute stroke. Patients with clinically significant hypoxia (e.g. oxygen saturations below 90%) who thus had indications for oxygen supplementation were not included. Our results therefore do not apply to clearly hypoxic patients. However, 717 of the participants had mild hypoxia (oxygen saturation below 95%) at the time of enrolment. The lack of benefit was the same in this subgroup as in patients who had normal or high-normal oxygen saturations at enrolment. While this study was not powered to examine this question, the results do not support the hypothesis that patients with lower baseline oxygen saturations are more likely to benefit from oxygen treatment. While we have not conducted formal screening logs for this study, discussions with participating centres about barriers to recruitment identified the requirement to recruit within 24 hours of hospital admission and the need for informed consent as the main barriers to enrolment, while clinical indications for oxygen treatment were rarely a problem. Apart from the protocol-driven exclusion of hypoxic patients, the participants in this study were as representative of the general stroke patient population as it is possible to achieve in a clinical study.

The amount of oxygen given may have been insufficient

The dose of oxygen given in SO₂S was based on the results of a dose-ranging study and aimed to keep oxygen saturation within the normal range, avoiding both hyperoxia and hypoxia. The dose of oxygen given might have been too low. Studies of short-duration high-flow oxygen do not support this possibility but were too small to identify benefit or exclude harm. 30,85 The largest (n = 85) study of short-burst high-flow oxygen was terminated early because of excess mortality in the actively treated group,86 but it cannot be excluded that this difference was a result of baseline imbalance in the study groups rather than a true effect of oxygen. All three studies tested oxygen for 12 hours or less. It remains unknown whether or not higher-flow oxygen given during the whole hyperacute phase of stroke, as in SO₂S, could be effective, but, given the results so far, this seems unlikely. Indeed, there is increasing concern that high-dose oxygen treatment could be harmful after stroke and in other forms of neurological injury. Hyperoxia (PaO₂ > 39.99 kPa) was independently associated with mortality in a large retrospective cohort study of ventilated stroke patients. 87 Hyperoxia has also been related to delayed cerebral ischaemia in patients with subarachnoid haemorrhage (PaO₂ > 23.06 kPa),88 to higher mortality and worse functional outcomes in traumatic brain injury (PaO₂ > 26.66 kPa),89 and to increased in-hospital mortality following resuscitation from cardiac arrest (PaO₂ > 39.99 kPa). 90 It is therefore highly unlikely that a higher dose of oxygen would have been more effective.

Oxygen treatment may have been ineffective because of poor compliance

Oxygen treatment can be effective only when it is given as prescribed. It is relatively easy to ensure that tablets or injections are given and to check compliance, as all drug treatments are recorded in the drug chart. Oxygen treatment is also prescribed on the drug chart, but given continuously, so that it is not possible to record compliance at all time points. Acute stroke patients are often confused and restless, and intentional or accidental removal of the nasal cannulae or oxygen mask is common. Minor benefits from oxygen treatment might therefore have been masked by poor compliance. Compliance with oxygen treatment was checked by two methods. Investigators were asked to document whether or not trial oxygen was prescribed. This was done as instructed in > 99% patients in the oxygen groups. They were also required to document whether or not the oxygen was actually given. The latter was recorded unreliably in the drug charts. Oxygen is normally prescribed only once as a continuous treatment, and nurses were not used to signing for ongoing oxygen treatment as they would for doses of medicines. To encourage more reliable recording of administration, we instructed centres to circle time points for signatures in the drug chart and put marks onto each field/time point requiring a signature (four times daily). In addition, we amended the protocol to require spot checks of compliance with oxygen treatment at different time points (00:00 and 06:00). This applied to the final 4144 patients. The results showed that oxygen was administered as prescribed in 65–83% of participants. These spot checks also showed that 3% of the control patients had oxygen in place at midnight and 6 a.m. Prescription of oxygen for clinical indications was permitted in all three treatment groups, and it was therefore appropriate that a small proportion of patients in the control group were found to have oxygen in place. While the main analysis of the study was intention to treat, a sensitivity analysis was conducted to examine the effects of non-compliance. This confirmed the main results that oxygen treatment, whether given continuously or at night-time only, did not improve outcome. The OR of the comparison of the combined oxygen groups with the control group was below 1, suggesting a possibility of harm, rather than benefit. However, this effect was very small and not significant statistically. It does nevertheless make poor compliance a less likely reason for ineffectiveness. Furthermore, oxygen treatment increased oxygen saturation in both the continuous and the night-time only treatment groups. Despite suboptimal compliance, we found highly significant increases of 0.8% and 0.9% in the highest and lowest oxygen saturations, respectively, in the oxygen groups compared with the control group during the intervention. While these changes in saturation are small, because of the S-shape of the haemoglobin dissociation curve, even small changes in oxygen saturation can reflect significant differences in the amount of oxygen dissolved in the blood. The high mean baseline oxygen saturation (96.6%) placed many participants on the flat part of the curve, where saturation changes little with increasing blood oxygen concentrations. A dose-ranging study of oxygen supplementation in stroke patients showed a 2.2% increase in oxygen saturation with 2 l/minute and a 2.9% increase with 3 l/minute. 31 In this study patients were observed continuously so that non-compliance could be excluded. However, the baseline oxygen saturation was lower at 95% and this will also have contributed to the relatively larger response. SO₂S was a pragmatic study and aimed to reflect clinical practice rather than strictly supervised experimental conditions. Better compliance would require continuous observation of patients, and this is impracticable outside an intensive care unit.

Few participants in the Stroke Oxygen Study were recorded as being hypoxic

When planning the study we may have overestimated the number of patients who develop hypoxia. If hypoxia is uncommon, the effect of prophylactic treatment would be diluted. In SO₂S very few (1.8%) participants in either group were recorded as having oxygen desaturations below 90%, with no differences between groups. In the Stroke Oxygen Pilot Study, 20% of patients in the oxygen group and 33% in the control group spent > 5 minutes with an oxygen saturation below 90% at night, but only 4% and 5%, respectively, spent > 60 minutes with an oxygen saturation below 90%. 31 Age, baseline oxygen saturation and NIHSS scores were similar in both studies. In the pilot study, oxygen desaturation was recorded continuously by pulse oximetry, while in SO₂S pulse oximetry readings were taken from intermittent measurements recorded in the patients’ notes. It is likely that short episodes of hypoxia were missed in SO₂S. This fits with evidence from a meta-analysis of continuous compared with intermittent monitoring of stroke patients, which also shows that intermittent monitoring detects fewer episodes of hypoxia than continuous monitoring. 91 A study of patients recovering from non-cardiac surgery showed that intermittent monitoring by ward nurses in a non-intensive care setting missed 90% of hypoxaemic episodes in which saturation was < 90% for at least 1 hour. 92 It is reasonable to assume that intermittent monitoring on stroke units is no more effective than that.

Neither the Stroke Oxygen Pilot Study nor SO₂S found a significant difference in severe desaturations between the treatment and control groups, which may indicate that low-dose oxygen supplementation is not sufficient to prevent severe desaturations. 31 Intermittent spot checks of oxygen saturation are likely to miss a significant number of transient episodes of hypoxia. This study was not designed to monitor oxygen saturation. It is unlikely that the Stroke Oxygen Study population has had fewer episodes of hypoxia than other stroke populations monitored continuously. It is therefore very unlikely that the lack of effect can be explained by rarity of hypoxic episodes.

Oxygen may have been started too late to have an effect

In SO₂S oxygen treatment was started at a median of 20 hours 43 minutes after symptom onset. This is well after the hyperacute phase of the stroke. It is therefore not possible to determine from the results of this study whether or not earlier oxygen treatment might have been more effective. However, other studies of oxygen treatment after stroke have enrolled patients earlier (Rønning and Guldvog29 within 24 hours, Singhal et al. 30 and Singhal 201086 within 8 hours, and Chiu et al. 93 within 12 hours), but also showed no beneficial effect.

Oxygen might still be beneficial, if given to the right subgroup of patients

Subgroup analyses did not identify any class of patient that was more likely to benefit from oxygen treatment. This included patients enrolled soon after onset of stroke, those with lower baseline oxygen saturation and patients with more severe strokes, the groups that were expected to be most likely to benefit from oxygen. However, patients who clearly needed oxygen because of dyspnoea or hypoxia were not enrolled into the study. The results of SO₂S can therefore not be extrapolated to this subgroup.

Incomplete blinding may have introduced bias

Patients tend to expect benefit from active treatment rather than from control. For oxygen, this may apply more than for unknown new interventions, as many patients know that the brain needs oxygen and that oxygen cannot get to the brain when the blood supply is blocked. Giving more thus makes sense intuitively. Knowing whether or not they have had oxygen could thus have affected their assessment of recovery. Just under 50% of participants remembered their treatment allocation correctly at 3 months, reducing with time to 35% at 12 months. Bias towards effectiveness of oxygen treatment cannot be excluded with an open design. Bias towards assuming a worse outcome with oxygen treatment is very unlikely. SO₂S did not show any effect of active treatment on any of the outcomes. While strong positive bias of the patients could have masked an adverse effect, this is unlikely. In addition, the lack of effect on functional outcomes was mirrored by a lack of effect on mortality, and the latter is not subject to bias.