Supplemental oxygen in Queen Elizabeth Central Hospital Malawi: a prospective cohort study of patients admitted to medical wards [version 2; peer review: 2 not approved]

Background: Oxygen is designated an essential drug by the World Health Organisation, and reduces mortality in hypoxic patients. In low-resource settings the provision of oxygen seldom meets its demand. This study explores predictors and observed time-course of hypoxaemia in order to help inform needs assessments for oxygen in hospitals in lowand middle-income countries. Methods: A prospective cohort study of adults with hypoxaemia admitted to medical wards of a teaching hospital in Malawi between February and March 2020. Vital signs and oxygen therapy were recorded daily. We analysed outcomes (death, discharge from hospital or ongoing inpatient care at 14 days after admission) using KaplanMeier and Cox regression time-to-event analysis. Results: 33 patients were recruited with median age 45 years (IQR 3361). 13 (39%) were female. Median pre-treatment oxygen saturations were 84% (IQR 76-87%). Oxygen delivery devices were often shared with other patients (n=10, 33%) and the flow rate was often unknown (n=14, 47%), mostly because of broken equipment (n=8, 57%). Median duration of oxygen therapy was 3 days (IQR 1-7). Death occurred in 16 (49%). Hazard ratios for short oxygen therapy were reduced in patients who had a chest radiograph performed (HR 0.08, 95% CI 0.02–0.30), in ex-smokers (HR 0.01, 95% CI 0.00-0.22) and in never smokers (HR 0.03, 95% CI 0.00 – 0.78). Conclusions: Delivering oxygen therapy in lower-middle income countries is challenging; broken equipment and shared delivery devices prevented titration of flow rates. Patients were relatively young and at a high risk of death. Patients with a chest radiograph Open Peer Review


Amendments from Version 1
We would like to thank the peer reviewers for the comments on our manuscript and hope that the changes made reflect some of their feedback.
In particular we have now highlighted clearly the limitations to our study, most importantly the fact it is single centred and has a short period. Whilst we recognize that our results may not be fully generalizable, the results can be used to generate hypothesis and thus have suggested possible relevant future studies.
We have also clarified how 33 participants were included in the study, and that 3 patients or their families declined consent to participation. We have explained how we believe that we have accurately identified all hypoxaemic patients during the study period due to the AETC triage system as well as the daily recordings of oxygen saturations. Previous studies in a regional hospital in Malawi (Queen Elizabeth Central Hospital, Blantyre; QECH) showed that less than one-third of patients that required supplemental oxygen received it, highlighting the disparity between supply and demand (Evans et al., 2012). This prospective cohort study explored the observed time course of hypoxaemia and predictors of supplemental oxygen therapy requirements in medical patients in Blantyre to help inform needs assessment for oxygen at the hospital level.

Setting and study design
We conducted a prospective cohort study of adults with hypoxaemia at the QECH in Blantyre, Malawi, between 18 th February 2020 and 20 th March 2020. The hospital has 200 medical beds, providing healthcare to approximately 1 million people in a context of high community prevalence of human immunodeficiency virus (HIV, 10.6%) and incidence of tuberculosis (159 cases per 100,000 population) ('Government of Malawi', 2015; World Health Organization, 2015a; World Health Organization, 2017).
Adults, aged 18 years and above, were included if they were hypoxaemic, were treated with supplemental oxygen and were admitted under the medical teams. Hypoxaemia was defined as peripheral oxygen saturations of less than 90% by finger pulse oximetry. Hypoxaemic patients were identified at triage in the emergency department, or from daily pulse oximetry assessment within 48 hours of admission to the medical wards. Written, informed consent was gained before any data collection was commenced. Those unwilling to consent to participation and those not receiving supplemental oxygen were excluded.

Data collection
Baseline demographics, admission details including vital signs and oxygen delivery method, diagnoses and co-morbidities were documented on admission. Diagnoses made by the attending physician and the haemoglobin level were extracted from medical notes. Results of chest radiograph performed at any point during the admission were recorded. Vital signs and details of supplemental oxygen delivery were recorded daily for participants until outcomes were determined (i.e. death, discharge, or ongoing inpatient care after 14 days). Oxygen saturations were measured with oxygen therapy present and then with temporary removal of supplemental oxygen, providing it was clinically safe to do so. Data were collected using an Open Data Kit platform on tablet devices (www.kobotoolbox.org).

Statistical analysis
A sample size of 62 participants was calculated to power the study to estimate the duration of oxygen therapy with a precision 6 hours with a 95% confidence level, and a standard deviation of 24 hours. The primary outcome was the duration of oxygen therapy in days. Secondary outcomes were the availability and delivery of supplemental oxygen, length of hospital stay, factors associated with the duration of supplemental oxygen provision and clinical outcome.
Data were presented as means, standard deviations, medians, ranges or percentages, based on the type and distribution of data. Those experiencing other events (death, discharge) before the end of the study contributed to the proportional hazard until that event. The Kaplan-Meier method was used to assess time to event data and probabilities. Time to event in univariable analysis was estimated by log-rank test. Cox regression models were used for uni-and multi-variable analysis and to calculate hazard ratios. Competing risk of events were calculated using estimated cumulative incidence (Putter et al., 2007). Conditional survival analysis was used to predict chances of requiring oxygen therapy on given days (Zabor et al., 2013). Statistical significance was defined as p<0.05, and 95% confidence intervals given where appropriate. Data were analysed using R (R Core Team, 2020).

Ethical considerations
Ethical approval was gained from the College of Medicine Research Ethics Committee, Blantyre, Malawi, (reference number P.05/19/2693) and the Liverpool School of Tropical Medicine Research Ethics Committee, United Kingdom (reference number 19-088). All participants provided written informed consent prior to participating in the study.

Baseline results
Between 18 th February 2020 and 11 th March 2020 525 patients were admitted under the medical team at QECH. 36 patients were identified by triage in the emergency department and on the medical wards as meeting inclusion criteria, and 33 participants were recruited into the study (3 patients or their families declined participation). Study recruitment was stopped on 11 th March due to safety concerns resulting from the evolving Coronavirus-19 pandemic, but all already enrolled participants were followed up until one of the outcomes was achieved. Last day of study follow up was 20th March 2021.
Thirteen study participants were women (39.4%). The median age was 45 years (interquartile range (IQR) 33-61 years, Table 1). One third of participants were HIV positive, two were currently receiving treatment for tuberculosis (TB) and six
Of a total of 167 cumulative follow up days, oxygen was given for 107 days (64%, Table 2). During this time, oxygen was usually supplied from cylinders (83%) and through nasal cannulae (83%). For most days one oxygen source was shared between at least two patients (71% of days). The median of peripheral saturations whilst receiving supplemental oxygen was 93% (IQR 88-97%) compared to 91% (IQR 85-96%) without supplemental oxygen. The median of oxygen saturations was 99% when supplemental oxygen was no longer required.

Duration of oxygen therapy
The use of oxygen declined by two-thirds between study commencement and day five of follow up, and then approximately halved from days 5 to 10 and from days 10 to 15 ( Figure 2). Two participants continued to receive supplemental oxygen up to day 15. The duration of therapy was associated with smoking status (p=0.03), the availability of a CXR (p<0.01), and the initial delivery device (p<0.01, Figure 3). In the univariable Cox regression analysis (Table 3), obtaining a chest radiograph had a reduced hazard ratio of still requiring oxygen on day 15 (HR 0.04, 95% CI 0.01-0.33, p< 0.01), i.e. an increased hazard of prolonged therapy. Requiring a reservoir mask on admission had an increased hazard of requiring supplemental oxygen at the end of the study (HR 14.5, 95% CI 2.28-92.7,   Unknown 11 (12 %) p <0.01). CXR, being an ex-smoker, or a never smoker remained significantly associated in multivariable analysis (respective HR 0.08, 95%CI 0.02-0.30, p<0.01, HR 0.01, 95%CI 0.00-0.22, p<0.01, and HR 0.03, 95%CI 0.00-0.78, p=0.03).
There was a statistically non-significant trend to an association with age (HR 0.95,95% CI 0.89-1.00, p=0.05). The likelihood of receiving supplemental oxygen at the end of the study increased with prolonged oxygen therapy (Table 4).

Discussion
Despite oxygen therapy resolving hypoxaemia in most participants, mortality was almost 50% and three quarters of these deaths occurred in the first three days of admission, indicating that hypoxaemia was a marker of severe underlying pathology. It recapitulates previous observations where hypoxaemia was associated with increased 30-day mortality in adult patients with pneumonia (Aston et al., 2019).
In the absence of piped oxygen, it is recommended that oxygen concentrators are used when power supply is reliable (such as in QECH) as they are comparatively cheap and provide consistent oxygen (Duke et al., 2010a; World Health Organization, 2015b). In our study, which immediately predates the introduction of an on-site oxygen plant, oxygen cylinders were preferentially used. Oxygen cylinders are used preferentially across Africa, as they are cheaper to purchase and easier to use than concentrators ( , the most common diagnoses in our cohort were pneumonia, TB and heart failure (14(42%), 5(15%) and 5(15%) respectively). Pneumonia, sepsis and heart failure were the commonest causes of death (6, 3 and 3 cases respectively). Hence hyperoxaemia is unlikely to have affected mortality.
The median duration of supplemental oxygen therapy was three days, which was influenced by the large number of early  having a chest radiograph and c) initial device used to deliver oxygen. Censored data appears as crosses, grey area: 95% confidence interval. Patients who were investigated with a CXR received oxygen for longer than those who had no radiography performed. A CXR may have allowed the identification of reversible pathology and thus justified the medical decision to continue oxygen therapy. However, sicker patients with a higher oxygen requirement and poorer prognosis may have been less likely to have had a CXR initially, creating confounding bias.
Oxygen therapy duration was also statistically significantly associated with the initial delivery interface that was used. However, three out of the four participants who were using a reservoir mask died, two within two days of admission, indicating sicker patients may have needed a reservoir mask. HIV status was not associated with oxygen therapy duration.

Limitations
We recognize that this study was single centered with a relatively small sample size, and thus may only provide a 'snapshot' of the situation. It is liable to be affected by random error as well as selection bias. It was conducted over a short study period that only represents one of the seasonal variations that are known to affect the number of admissions and severity of illness, and thus in turn the need for oxygen. It was also not possible to longitudinally follow flow rates for individual patients due to frequent switches between delivery devices as well as broken equipment.
During our recruitment period 535 patients were admitted to medicine, 36 of whom we identified as being hypoxaemic (7% admissions). We are confident that we have included the majority of patients who were hypoxaemic during the study period due to the thorough admission triage process as well as close collaboration with staff to identify any newly hypoxaemic patients on the wards via daily saturation readings. Whilst the results may not be wholly generalizable, they can be used to shape further hypothesis. For example, future work may explore the association of sharing oxygen devices and complications such as nosocomial infections or acute desaturating events. There could also be further exploration into the cost of oxygen supply methods which could better inform formal health economic analysis.

Conclusions
It has previously been demonstrated that programmatically introducing oxygen in paediatric populations is beneficial (Duke et al., 2008), however data from adult practice are few. Our study provides some patient-level data which describe periods of treatment, and outcomes which could help generate further hypothesis to better inform oxygen demand.  duration, length of oxygen therapy simply reflects provider variation? I would strongly suggest removing all of these regression analyses and K-M curves. The authors seem to make causal conclusions here that are not substantiated by the results.

Data availability
The Results state that 525 patients were admitted under the medical team during the study period. The paper's quality is significantly diminished by the fact that the authors do not include a CONSORT diagram showing the reasons why only 33 people were recruited for this study. How many were hypoxic but didn't receive oxygen? How many were both hypoxic and received oxygen, but declined consent? 2.
Minor Comments: The abstract states that the study took place January -March 2020, but the methods state it was 18 Feb -20 March (actually 18 Feb -11 March per the results). Please correct whichever sections are incorrect.

1.
Why is no percentage included for the patients receiving ARV therapy (Table 1)? 2.

If applicable, is the statistical analysis and its interpretation appropriate? No
Many thanks for the review of the manuscript. Your comments have been reviewed and hopefully you can see the reflection of this in our most recent submission.
Of note, we perhaps did not clearly highlight the limitations to our study, and thus have added a paragraph about this. We agree this paper is likely hypothesis generating work and have suggested possible relevant future work.
Whilst we have not included a consort diagram, we hopefully have explained in the discussion that of the 525 patients admitted under the medical team 36 patients were identified as being hypoxaemic and 3 patients/their families declined consent, leading to 33 participants in the study. We believe that both the triage system in AETC as well as the daily recording of oxygen saturations allowed us to highlight all the hypoxaemic patients during the study period.
We have corrected the study dates in the abstract and have included the percentage for patients receiving ARV therapy.

Kathryn Maitland
Faculty of Medicine, Imperial College London, London, UK The introduction starts with a bold statement 'Oxygen reduces morbidity and mortality in hypoxaemic patients and then goes on to state 'failure of oxygen supply to tissues precipitates organ failure and death'. The two references supporting these are a 'before and after' observational study conducted in Papua New Guinea (Duke et al.) in early 2000's in children with hypoxaemia implying that better use of oxygen (although not actually recorded rate/volume of oxygen used) resulted in better outcomes -has numerous potential biases to draw on as definitive evidence. The second reference is to a case report (Goldhill) of an elderly gentleman admitted with severe pneumonia who received oxygen, escalated to CPAP and the ICU care but died 24 days into admission. Perhaps the authors may be more speculative or provide better references to support this statement.
The introduction also includes a study conducted in the same hospital (Evans et al. 2016) in 144 adults including 14 eligible to receive oxygen therapy. Some of the same issues that the current paper discusses were brought out. I noticed that the short period of the study from 18th February 2020 and 20th March 2020 and refer the authors (some being co-authors of the Evans paper) to the limitations highlighted in that study. 'as a 24-hour cross-sectional study it only provides a brief snapshot and is open to random error as well as information and selection bias. Malawi essentially has three seasons: hot and wet, hot and temperate. As a result, there are seasonal variations in the number of admissions as well as the severity of illness with a consequent impact on the need for oxygen'.
My concern is about generalisabilty. I understand that the study halted as a result of COVID -why was it not possible restart it once recruitment was allowed? There needs to be some more reflection on this limited window of the study. I note that median O 2 sats were 84% on admission and that on oxygen (all received oxygen therapy, despite many having to share the supply) median O 2 sats were normal 98% (95 -99). The conclusion that sharing oxygen resource needs to be tempered by this fact. The authors 'show' that having a chest X-ray reduced the risk of dying. This is reported as 'CXR on admission'. Does that mean CXR taken AT admission time (to hospital) or if a CXR was ever taken during admission was included in this variable. This is rather problematic both ways. First the patients who got a CXR AT admission were probably more likely to less critically sick hence less likely to die. If this parameter includes CXR at any time then since most deaths occurred <72 hours of admission and likely to be to critical sick to send off for an X-ray it is mostly that these deaths did not have a CXR i.e. only only means those who were able to survive long enough to get a chest X-ray had a better outcome. Unless the authors have an alternative explanation?
The authors have reported overall use of oxygen used in the study. Is it possible that the authors could report median (and interquartile range) of oxygen use per patient in those with known flow rates (although assumptions may have to made that this flow rate may not have been constant over time)? In addition, it would be very good to provide, based on the cost of a bottle of oxygen, how much oxygen therapy cost -other variable costs to include would be mask, tubing etc. There is very little published on the 'costs' of oxygen therapy.
When discussing the issues of difficulties in providing sustainable sources of oxygen there are some reference they may want to include 1 . These include a survey of oxygen supply and associated infrastructure was performed at 231 health centres and hospitals in twelve African countries and some more discussion based on this paper for the challenges of using Oxygen concentrators, the WHO preferred option, when electricity supplies as reported in this survey at most centres was very unreliable. Our research group has also reported on this 2 .
At the beginning of the discussion the authors note that despite correction of oxygen saturations there was a high mortality. It would be good to see some speculation over whether hypoxaemia was on the causal pathway to mortality? Oxygen therapy can have risks and benefits -perhaps these could be weighed up 3, 4 ?
Malawi. We recognise that whilst our results may not be fully generalizable, the results can be used to generate hypothesis, and we have made suggestions of future relevant work.
In the discussion we have a paragraph related to the presence of a CXR and length of oxygen therapy received ('Patients who were investigated with a CXR received oxygen for longer than those who had no radiography performed. A CXR may…). We agree that there is potential for confounding bias in our findings and have hopefully explained this.
Whilst we would have liked to explore the longitudinal relationship of individual patients oxygen use, this was not possible due to broken equipment and frequent switching of patients oxygen delivery devices. We have referenced the Belle et al. study of oxygen supply and associated infrastructure to highlight difficulties in providing sustainable sources of oxygen.

Competing Interests:
No competing interests were disclosed.