Use of Proton-Pump Inhibitor Is Not Associated with Adverse Clinical Outcomes in COVID-19 Patients: A Territory-Wide Cohort Study

Background . Evidence regarding the use of proton-pump inhibitors (PPIs) in COVID-19 patients remains elusive. Aim. To examine the potential e ﬀ ects of PPI use on the clinical outcomes of COVID-19 patients in a territory-wide cohort. Methods . A retrospective cohort study was performed using data from a territory-wide database in Hong Kong. Patients diagnosed with COVID-19 from 23 January 2020 to 1 January 2021 were identi ﬁ ed by virological results. The primary endpoint was a composite of intensive care unit admission, use of invasive mechanical ventilation, and/or death. PPI users were identi ﬁ ed by PPI use within 12 months prior to their diagnosis of COVID-19. Results . We identi ﬁ ed 8,675 COVID-19 patients (mean age 46 years, 49% male, 97.6% of all reported cases in Hong Kong), of which 579 (6.7%) patients had used PPI. PPI users were found to be older, more likely to have comorbidities, concomitant medications and unfavourable laboratory parameters than nonusers. Of the 8,675 COVID-19 patients, 500 (5.8%) developed the primary endpoint. After propensity score (PS) balancing for patients ’ demographics, comorbidities, laboratory parameters, and use of medications, PPI use was not found to be associated with the development of primary endpoint in PS weighting (weighted hazard ratio (HR) 1.10, 95% con ﬁ dence interval (CI) 0.82 – 1.46, P = 0 : 529 ), and PS matching analysis (weighted HR 0.79, 95% CI 0.56 – 1.13, P = 0 : 198 ). Consistent nonassociation was observed after multivariable adjustment (adjusted HR 0.84, 95% CI 0.67 – 1.06, P = 0 : 142 ), and in subgroups of current and past PPI users. Conclusion . PPI use is not found to be associated with adverse clinical outcomes in COVID-19 patients. The result remains robust after PS weighting, PS matching, and multivariable adjustment.


Introduction
COVID-19, caused by severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), has infected over 238 million people and caused over 4.8 million deaths worldwide as of 13 October 2021 [1]. COVID-19 is a heterogeneous disease with a case-fatality ratio that varies substantially among different patient populations. Identified risk factors for adverse clinical outcomes include advanced age, preexisting cardiovascular disease, diabetes mellitus, chronic kidney disease, and liver injury [2][3][4][5]. In addition, various prediction models on the risk of hospital admission, adverse clinical outcomes, and mortality have been developed and published [6][7][8].
Proton-pump inhibitor (PPI) is an acid suppression therapy commonly used worldwide to treat gastrooesophageal reflux disease and peptic ulcers. As gastric acid can inhibit swallowed infectious microorganisms and prevent them from entering the intestine, PPI may alter its users' susceptibility to enteric pathogens [9]. Indeed, it was observed in an American online survey that the use of PPI increases the risks of contracting COVID-19 among community-dwelling people [10], whereas a separate Korean nationwide study suggested that PPI use does not increase users' susceptibility to SARS-CoV-2 infection. This Korean study, however, suggested that PPI use is correlated with worse clinical outcomes of COVID-19 [11]. Moreover, PPI treatment may even be a risk factor for the development of secondary infections among patients with an existing SARS-CoV-2 infection [12]. In contrast, the use of famotidine, a histamine-2 receptor antagonist (H2RA), is reported to be associated with a lower risk of clinical deterioration in COVID-19 patients [13]. In a case series, famotidine use is also correlated with improved patient-reported outcomes on symptoms among nonhospitalised COVID-19 patients [14]. Nonetheless, the association between famotidine use and better clinical outcomes for COVID-19 patients was not observed in a similar territory-wide study conducted in Hong Kong, after adjusting for patients' concomitant medications and laboratory parameters [15]. The contradictory findings in the aforementioned studies reflect the betweenstudy heterogeneity and different sources of bias that had driven the effect estimates. In particular, most previous studies on the association between PPI use and severe clinical outcomes of COVID-19 involved a small sample size and did not adjust for important confounding factors, as shown in a meta-analysis [16]. Given the rapidly growing number of COVID-19 cases and the widespread use of PPI globally, this study is aimed at examine the impact of PPI use on clinical outcomes of COVID-19 using robust methodology to identify and adjust for different sources of confounders.

Study Design and Data Source.
A territory-wide retrospective cohort study was conducted using data from the Clinical Data Analysis and Reporting System (CDARS) under the management of the Hospital Authority, Hong Kong [17]. CDARS is an electronic healthcare database that covers patients' demographic, death, diagnoses, procedures, drug prescription and dispensing history, and laboratory results of all public hospitals and clinics in Hong Kong [18]. The Hospital Authority is the sole public healthcare provider in Hong Kong and accounts for over 90% of all healthcare services provided to the Hong Kong population. All suspected and confirmed cases of COVID-19 are reported to the Department of Health, and all were hospitalised under the care of the Hospital Authority. SARS-CoV-2 reverse transcription polymerase chain reaction tests were performed on symptomatic patients presenting to outpatient clinics and hospitals, as well as on asymptomatic close contacts of infected patients and inbound travellers. All data are anonymised in CDARS to ensure confidentiality. Territorywide epidemiological studies of various infectious diseases were previously conducted using CDARS [3,[19][20][21]. The International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) coding was used in CDARS. The use of ICD-9-CM codes in CDARS to identify medical conditions has been found to be 99% accurate when referenced to clinical, laboratory, imaging, and endoscopy results from the electronic medical records [22].

Subjects.
Consecutive laboratory-confirmed COVID-19 patients between 23 January 2020 and 1 January 2021 were identified by virological results (Supplementary Table 1). The baseline date was defined as the date of diagnosis of COVID-19 by virological results. Patients were followed from the baseline date to the earliest of the following: (i) discharge from hospital, (ii) the last follow-up date (i.e., 1 January 2021), (iii) admission to the intensive care unit (ICU), (iv) use of invasive mechanical ventilation (IMV), or (v) death. PPI users were defined as patients who had used PPI within 12 months before baseline date (i.e., the diagnosis of COVID-19) to prevent immortal time bias introduced when treatment status is determined by a prescription issued or received at some point during follow-up of their hospitalisation [23]. The study protocol was approved by the Joint Chinese University of Hong Kong-New Territories East Cluster Clinical Research Ethics Committee (reference number: 2020.074); informed consent was waived due to the study's retrospective nature and the use of anonymised clinical data.

Clinical Evaluation.
All COVID-19 patients in the study were admitted to medical wards or ICU with isolation facilities. Initial investigations included a complete blood count (with a differential count), clotting profile (prothrombin time, activated partial-thromboplastin time, international normalised ratio), and serum biochemical measurements (electrolytes, renal and liver biochemistries, C-reactive protein and lactate dehydrogenase, glucose, and procalcitonin). These laboratory assessments and chest radiography were performed regularly as clinically indicated. A reverse transcription polymerase chain reaction (RT-PCR) assay was used to detect a conserved region in the E gene of SARS-CoV and SARS-CoV-2 as well as other bat-associated SARS-related viruses (Sarbecovirus) as screening [24]. All positive samples were sent out to the Public Health 2.5. Definitions. The primary endpoint was a composite endpoint of ICU admission, use of IMV, and/or death. The secondary endpoints were ICU admission, use of IMV, and death, respectively. The use of PPI, H2RAs, aspirin, and nonsteroidal anti-inflammatory drugs (NSAIDs) were defined as use within 12 months before the baseline date (i.e., the diagnosis of COVID-19). Among PPI users, the cumulative days of the use of PPI within 12 months before the diagnosis of COVID-19 were categorised into <30 days, 30-89 days, 90-179 days, and ≥180 days. In a subgroup analysis, current PPI users were defined as patients who used PPIs within 1 month prior to their diagnosis of COVID-19; past PPI users were defined as patients who used PPIs 1 to 12 months prior to their diagnosis of COVID-19 [11]. On sensitivity analysis, short-term new NSAID users were defined as patients who began using NASID within 1 month prior to their COVID-19 diagnosis. New users of PPI were defined as patients who began using PPI within 12 months prior to their COVID-19 diagnosis, without any exposure to PPI between 12 and 36 months prior to their COVID-19 diagnosis. Details on definitions of comorbidities are described in the Supplementary methods (available here).
2.6. Statistical Analysis. Data were analysed using Statistical Product and Service Solutions (SPSS) version 25.0 (SPSS, Inc., Chicago, Illinois), SAS (9.4; SAS Institute Inc., Cary, NC), and R software (4.0.2; R Foundation for Statistical Computing, Vienna, Austria). Continuous variables were expressed in mean ± standard deviation or median (interquartile range), as appropriate. Categorical variables were presented as numbers (percentage). Qualitative and quantitative differences between subgroups were compared by the Chi-square test or Fisher's exact tests for categorical parameters and Student's t-test or Mann-Whitney test for continuous parameters, as appropriate.
Differences in baseline clinical characteristics were observed between PPI users and nonusers (Tables 1 and 2). Propensity score (PS), the conditional probability of receiving PPI, was estimated to control for 23 confounders and reduce selection bias (Table 2) [25,26]. PS weighting and 1 : 3 PS matching were used to balance patients' baseline clinical characteristics. The balance of baseline clinical characteristics between PPI users and nonusers was assessed by absolute standardised mean difference (ASMD), where an ASMD of below 0.2 indicated a good balance [27,28]. Before estimating PS, missing data were imputed by multiple imputation. Details on PS estimation and multiple imputation are described in the Supplementary methods (available here).
Hazard ratios and adjusted hazard ratios (aHRs) with 95% confidence interval (CI) of PPI use on the primary endpoint were estimated by Cox proportional hazards regression. Weighted Cox proportional hazards regression was used in PS weighting and matching analysis. Details of modelling are described in the Supplementary methods (available here). All statistical tests were two-sided. Statistical significance was taken as P < 0:05. Subgroup analyses on current and past PPI users were performed. As patients who received short-term NSAID for early pneumonia symptoms may start PPI simultaneously, sensitivity analysis was performed after excluding short-term new NSAID users to minimise protopathic bias, i.e., reverse causation bias [11,29]. Another sensitivity analysis was performed on new PPI users who began using PPI within 12 months prior to the diagnosis of COVID-19. In addition, patients who used H2RA within 12 months prior to the diagnosis of COVID-19 were analysed as active control on the risk of adverse clinical outcomes of COVID-19 in a sensitivity analysis after excluding all PPI users.

Results
3.1. Demographic Characteristics. We identified 8,675 COVID-19 patients between 23 January 2020 and 1 January 2021 which represented 97.6% of all patients who reported to the Department of Health during the study period. Among these patients, their mean age was 45:8 ± 19:9 years; 48.5% were male; 579 (6.7%) patients had used PPI before their diagnosis of COVID-19 (516 pantoprazole, 46 lansoprazole, 15 esomeprazole, and 2 dexlansoprazole) ( Table 1). Compared to PPI nonusers, PPI users were older, more likely to have diabetes mellitus, malignant tumours, and cardiovascular, digestive, nervous system, respiratory, and kidney diseases. PPI users had worse renal and liver functions, and higher C-reactive protein and LDH; they also had higher neutrophil counts, and lower lymphocyte and platelet counts compared to PPI nonusers. More PPI users received H2RAs, NSAIDs, aspirin, corticosteroids, antibiotics, antifungals, and antiviral treatment for COVID-19 as compared to PPI nonusers (Table 1).  Table 2 shows the result in 1 of the 20 imputed data sets; consistent patterns were obtained across other imputed data sets.

PPI Use and Clinical Outcomes after PS Matching. PS matching led to greater similarity in distributions of the 23 clinical characteristics between PPI users and non-users
and reduced all ASMDs to <0.2 (Table 2 and Supplementary Figure 1C). Among 579 PPI users, 395 (68.2%) were matched to at least 1 PPI nonuser; 54.2%, 18.5%, and 27.3% were matched to 3, 2, and 1 PPI nonuser, respectively. Consistent patterns were also observed across other imputed data sets. Compared to PPI users who were matched to PPI nonusers, PPI users who were not matched were older, more likely to have co-morbidities, had worse liver and renal function, and higher C-reactive protein and LDH, had higher neutrophil counts, and had lower lymphocyte and platelet counts (Supplementary Table 3).

PPI Use and Clinical
Outcomes before PS Balancing. The development of adverse clinical outcomes was more common in PPI users relative to nonusers (Table 1). On univariate analysis, the use of PPI was found to be associated with a higher risk of adverse clinical outcomes of COVID-19 (HR 3.36, 95% CI 2.71-4.18, P < 0:001). Moreover, the use of H2RAs, NSAIDs, aspirin, age, gender, preexisting comorbidities, and baseline laboratory parameters were found to be associated with adverse clinical outcomes (Table 4). However, after adjusting for patients' age, gender, comorbidities, and baseline laboratory parameters, the use of PPI (aHR 0.84, 95% CI 0.67-1.06, P = 0:142) was not found to be associated with adverse clinical outcomes of COVID-19. Other factors, such as advanced age, male gender, preexisting circulatory system disease, diabetes mellitus, respiratory disease, chronic kidney disease, elevated levels of alanine aminotransferase, LDH, C-reactive protein, and respiratory rate, and lower albumin and platelet counts were found to be associated with a heightened risk of adverse clinical outcomes on multivariable analysis (Table 4). Subgroup analyses on current and past PPI users showed comparable results (Supplementary Table 7).
3.6. Sensitivity Analysis. After excluding short-term new NSAID users, 566 PPI users and 8,018 nonusers were included in a sensitivity analysis. PS weighting and matching led to greater similarity in distributions of the 23 clinical characteristics between PPI users and non-users and reduced all ASMDs to <0.2. The result was comparable to the main analysis (Supplementary Table 8 & Table 3). After excluding the prevalent PPI users, 269 PPI new users and 8,096 PPI nonusers were included in another sensitivity analysis (Supplementary Table 9). New use of PPI was found not to be associated with adverse clinical outcomes of COVID-19 (Supplementary Table 10). In the sensitivity analysis on H2RA users as an active control, the use of H2RA was also found not to be associated with adverse clinical outcomes of COVID-19 (Supplementary  Tables 10-11).

Discussion
In this study, the use of PPI in COVID-19 patients and its relationship with adverse clinical outcomes were examined in a territory-wide cohort in Hong Kong. Based on the data collected, PPI use was found not to be associated with adverse clinical outcomes including admission to ICU, use of IMV, and death. Furthermore, the result remains robust after PS weighting, PS matching, and multivariable adjustment.
Gastrointestinal symptoms including vomiting, diarrhoea, or nausea have been reported in COVID-19 patients [16]. Studies have shown that angiotensin-converting enzyme-2, the SARS-CoV-2 host receptor, is expressed in gastrointestinal epithelial cells and may potentially cause gastrointestinal infection [30]. This supposition is further    [11]. The study found that current PPI use within 30 days prior to the onset of COVID-19 was associated with a 79% increase in the risk of developing severe clinical outcomes of COVID-19, whereas the same association was not seen in patients with past use of PPI. Following Lee et al.'s study, a metaanalysis published as a letter in an academic journal also showed that current or regular use of PPI was associated with severe clinical outcomes of COVID-19. However, it is important to note that there was evidence of substantial between-study heterogeneity that impaired the validity of the results [29]. Additionally, most of the studies in the meta-analysis involved a relatively small population of PPI users and did not adjust for confounders; therefore, the effect estimates were more susceptible to confounding and selection biases [33].
Interestingly, a recent PS-matched territory-wide study conducted in Hong Kong by Zhou et al. concluded that PPI use was associated with worse clinical outcomes of COVID-19 [34]. However, the authors did not explicitly define "PPI users" in their study. As the authors also included medications used after COVID-19 infection in their PS matching analysis, one may interpret that the definition of "PPI users" contemplates those who used PPI at the time of or after their COVID-19 diagnosis and during hospitalisation. The inclusion of patients' clinical data after contracting COVID-19 is also reflected by the unexpectedly high prevalence of prior comorbidities of respiratory diseases (98%) and gastrointestinal diseases (97%) in their data. It is plausible that in certain cases, COVID-19-induced respiratory and gastrointestinal symptoms were inadvertently considered as preexisting comorbidities. As PPI may be prescribed to critically ill patients requiring intensive care 9 GastroHep for stress ulcer prophylaxis, the association between PPI use and severe clinical outcomes of COVID-19 may be inevitably influenced by protopathic bias, or reverse causation bias, if patients who used PPI after their diagnosis of COVID-19 were included as PPI users in the studies. Reverse causation bias has been raised previously as a source of overestimated association in studies on the use of PPI and the risk of pneumonia [29].
As outlined in the sections above, our findings were different from Lee et al.'s and Zhou et al.'s studies. One possible explanation lies in the substantial clinical characteristic differences between PPI users and nonusers. For instance, whereas Lee et al.'s study captured less comprehensive data at the patient level, our study incorporated more comprehensive patient-level data such as patients' comorbidities, laboratory parameters, and concomitant medications. The inclusion of more complete patient parameters enabled more precise adjustments for confounders on the adverse clinical outcomes of COVID-19 through PS weighting, PS matching, and multivariable analysis. In our study, COVID-19 patients who used PPI were indeed found to be at greater risk of developing adverse clinical outcomes on univariate analysis, probably due to the fact that those patients are prone to more risk factors (namely older age and more comorbidities) for adverse clinical outcomes at the time of their COVID-19 diagnosis (Table 1). Therefore, a fairer comparison between PPI users and nonusers on clinical outcomes could only be drawn after balancing for these confounding factors using stringent statistical approaches, namely, PS weighting, PS matching, and multivariable adjustment. Moreover, PPI users were defined in our study as COVID-19 patients who used PPI before their COVID-19 diagnosis to minimise the reverse causation bias. Sensitivity analysis that excluded patients who started recent and short-term use of NSAID as a possible treatment for early pneumonia symptoms also showed comparable results to that of the main analysis.
The strength of our study includes a territory-wide cohort that covers 97.6% of all COVID-19 patients in Hong Kong with detailed patient-level clinical data. Notwithstanding, our study has a number of limitations. Firstly, missing data on laboratory parameters might lead to biases as in other observational studies. These biases, however, can partially be compensated by our extensive cohort size. Missing data were uncommon for routine laboratory parameters that are checked as part of our clinical practice. However, less routine laboratory parameters, such as the international normalised ratio, may not be checked for each patient due to minor variations in clinical practice in different hospitals. Multiple imputation with 20 imputed data sets was used to reduce the possible selection bias due to missing data [35]. Secondly, COVID-19 patients who used and did not use PPI might have been different in terms of the baseline clinical characteristics (e.g., age and gender) such that our study might be subjected to confounding as in other observational studies. We were not able to accurately identify patients with diseases associated with PPI use included gastroesophageal reflux disease, Helicobacter pylori infection, Barrett's oesophagus, achalasia, and stricture. Barrett's oesophagus is uncommon in Hong Kong [36]. The prevalence of Helicobacter pylori infection is over 50% in Eastern Asia including Hong Kong [37]. The information on body mass index was not available in most of the patients, while the information on the presence of radiographic chest infiltrates was not available. Due to the difference in clinical characteristics, some older PPI users (mean age: 74 years) with more comorbidities were not able to be matched with PPI nonusers in the PS matching analysis (Supplementary Table 3). This may limit the generalizability of the result to these older patients with comorbidities. Thus, in addition to PS matching, we also applied PS weighting and multivariable adjustment on important clinical characteristics aiming to include all PPI users in the cohort. Thirdly, PPI may be purchased by patients without a prescription. Thus, the use of PPI prior to hospitalisation might go unreported in certain patients. Fourth, ascertainment bias may affect the reliability of the study due to inaccurate entry of certain diagnosis codes for comorbidities, namely diabetes mellitus and cardiovascular disease. Nonetheless, every endeavour was made to minimise such bias by including laboratory and medication data for certain diagnoses (diabetes mellitus and hypertension). The use of ICD-9-CM codes in CDARS to identify medical conditions has also been found to be 99% accurate when referenced to clinical, laboratory, imaging, and endoscopy results from the electronic medical records [22].
In conclusion, PPI use was found not to be associated with adverse clinical outcomes in a territory-wide cohort of COVID-19 patients. The results remained robust in PS weighting, PS matching, and multivariable adjustment analysis. Despite the ongoing pandemic with millions of COVID-19-related casualties, this study's findings do not favour withholding PPI use.

Data Availability
The data that support the findings of this study are available from the Hospital Authority, Hong Kong. Restrictions apply 10 GastroHep to the availability of these data, which were used under license for this study. Data are available with the permission of the Hospital Authority, Hong Kong.