Prior fluid and electrolyte imbalance is associated with COVID-19 mortality.

Background: The COVID-19 pandemic represents a major public health threat. Risk of death from the infection is associated with age and pre-existing comorbidities such as diabetes, dementia, cancer, and impairment of immunological, hepatic or renal function. It remains incompletely understood why some patients survive the disease, while others do not. As such, we sought to identify novel prognostic factors for COVID-19 mortality.
Methods: We performed an unbiased, observational retrospective analysis of real world data. Our multivariable and univariable analyses make use of U.S. electronic health records from 122,250 COVID-19 patients in the early stages of the pandemic.
Results: Here we show that a priori diagnoses of fluid, pH and electrolyte imbalance during the year preceding the infection are associated with an increased risk of death independently of age and prior renal comorbidities. Conclusions: We propose that future interventional studies should investigate whether the risk of death can be alleviated by diligent and personalized management of the fluid and electrolyte balance of at-risk individuals during and before COVID-19.

Introduction

In 2020, the world ground to a halt owing to the COVID-19 pandemic that still continues on its course in large parts of the world. Achieving durable universal sterilizing immunity through population-penetrating and transmission-halting vaccinations still remains a remote prospect: as of now, only 39.9% and 1.8% of the entire world population and that in low income countries, respectively, have started their COVID-19 immunization regimes[1,2]. Moreover, vaccine hesitancy percentages of up to 40% are reported in some large countries such as the U.S., Italy, and Russia[3]. As it seems possible that numerous COVID-19-naive individuals are yet to contract the virus, we anticipate that improvements in fortifying the health of risk group individuals in anticipation of a potential infection, as well as improving the outcomes of patients succumbing to the severe form of COVID-19, may translate to large savings in loss of life.

The SARS-CoV-2 virus that causes COVID-19 enters human cells via the ACE2 receptor. The infection first occurs in upper airways and at later stages may proceed to the lung, gastrointestinal tract, kidney, heart, or brain[4]. ACE2 and its antagonistic homolog ACE are core enzymes of the renin–angiotensin–aldosterone system (RAAS), which regulates electrolyte homeostasis, blood pressure, and cardiovascular health[5], as well as restores balance upon volume disturbance of extracellular fluid[6,7]. The antagonistic effects of ACE and ACE2 are largely achieved by increases or decreases of the amount of circulating Angiotensin II, respectively. Angiotensin II is a potent secretagogue of aldosterone, an adrenal cortex hormone that enhances renal reabsorption of sodium and water, excretion of potassium, and the maintenance of acid–base balance[8,9].

The underlying mechanisms of infection and viral spread are not fully understood, and despite the advances in prevention and treatment of severe COVID-19, an unmet need remains to better understand the clinical course and the risk factors of severe disease and death. Here, we present an agnostic and data-driven analysis of real world data from U.S. electronic health records (EHR) of 122,250 COVID-19 patients to identify a priori factors associated with death during a COVID-19 infection. Our unbiased analyses reveal pre-existing aberrations of fluid, pH or electrolyte levels as risk factors for COVID-19 mortality. We suggest that balancing electrolyte homeostasis in COVID-19 patients offers opportunities for better care and/or prevention of the severe disease.

Methods

Study design and participants

We extracted 122,250 COVID-19 cases collected by Optum® with a diagnosis date between 20 February and 1 July 2020 (Supplementary Data 1). The Optum® de-identified COVID-19 EHR dataset contains patient-level medical and administrative records from hospitals, emergency departments, outpatient centers, and laboratories from across the United States. Mortality information is derived from combining data from the Social Security Death Master File, hospital reports on patient deaths, and third party obituary sources. Data de-identification is performed in compliance with the HIPAA Expert Method and managed according to Optum® customer data use agreements. The COVID-19 EHR dataset sources clinical information from hospital networks that provide data meeting Optum’s internal data quality criteria. We confirmed COVID-19 diagnosis either by documented ICD-10 codes (Supplementary Information: Data Preparation) or via positive PCR test result (Supplementary Data 2). Survival time was computed as the number of days between the date of COVID-19 diagnosis and last documented clinical activity (vitals, labs, medication, encounter, collected until 13 July 2020) or documented death.

We analyzed variables that were observed at least a month before the infection. We selected such a conservative buffer time period, since some patients might have had an undiagnosed COVID-19 infection for days before an opportunity to get tested. We also wanted to exclude any physiological changes potentially incurred by the virus during the incubation period, which may be up to 14 days[10]. We used the median value captured between 1 and 12 months before the initial COVID-19 diagnosis for vitals and laboratory measurements, and the entire past medical history (mean length ~5.4 years, Supplementary Data 1) for variables with a long-term effect, such as diagnoses of chronic indications. Prior inoculations were handled analogously.

We investigated all disease entities that are a part of the Charlson comorbidity index[11], the AHRQ[12], or the former Elixhauser definition[13]. A detailed description of assignment of ICD codes to disease entities can be found in Supplementary Data 3.

After quality control (Supplementary Information: Quality control, transformation, and handling of missing data), 249 variables were available for primary univariable analysis. For multivariable analysis, patients with an exaggerated proportion of missing data were removed, leaving 55,757 patients for analysis. Variables available for less than 10,000 patients were mean-imputed, reducing the overall missing rate from 22.1% to a remaining missingness rate of 12.4% in total. The remaining missing values were imputed using the missForest R-package[14].

Association analysis and model development

We originally set out to identify prognostic biomarkers that could identify patients at risk of COVID-19 mortality already before the onset of the disease. As primary analysis, we performed time-to-event analysis using Cox regression[15]. Univariable analysis was conducted using age, sex, ethnicity, race, insurance status, and US region/division as covariate parameters for adjustment. We applied a Bonferroni-correction with the number of variables (m = 249) to account for multiple testing and required a significance level of α = 0.05/m = 2 × 10−4. The univariable associations were calculated for the entire patient cohort, as well as separately for the age groups <50, 50–70, 70–80, and >80 years (Supplementary Data 4).

In order to allow comparison of hazard ratios (HRs) between different variables, we report the 2-standard-deviations hazard ratio “HR2SD”. It is computed as HR2SD = HR{2 × SD}, where SD is the standard deviation of the respective variable.

As secondary analysis, multivariable modeling was performed. We pursued two approaches in parallel. First, we performed a backward selection procedure on the Cox regression model of all eligible variables. We iteratively removed the variable with least impact on model performance until all remaining parameters were significant at α1 = 0.05/249 = 2 × 10−4 (Bonferroni-correction). By construction, the procedure controls the family-wise error rate at α = 0.05. In parallel, we derived a regularized Lasso model[16]. We fitted a L1 (Lasso) regularized Cox-Proportional Hazards Model using glmnet version 3.02[16], with the concordance index (C-index)[17] as the performance measure. The regularization parameter λ was optimized using ten-fold cross-validation. We selected λ such that we extracted the most regularized model with a C-index within one standard error of the best performing model.

More details on model assumption checking, calibrations and performance measures can be found in Supplementary Information: Supplementary Methods and Supplementary Figs. S1 and S2.

Ethical framework under which this study was conducted

Use of the Optum EHR data for research purposes has been determined by the New England Institutional Review Board (IRB) to not constitute research involving human subjects. This study has also been exempted from further IRB oversight in Switzerland and Germany by Kantonale Ethikkommission Kanton Zürich and Ethik-Kommission der Bayerischen Landesärztekammer, respectively.

The data licensed by Optum® to support the study consists of only data de-identified in compliance with 45 CFR 164.514(a)-(c). The data has identifying information removed and is not coded in such a way that the data could be linked back to the subjects from whom it was originally collected.

The resulting research with this data would utilize data that did not include Human Subjects, as there is no interaction or intervention with living individuals, and neither can the provider of the data nor the recipient link the data with identifiable individuals, as defined in HHS regulation 45 CFR 46.102(f).

Our research involving the data licensed by Optum® and described above, does not require an IRB review, as analyses with the data would not meet the definition of “research involving human subjects”.

Table 1

Combined multivariable model of a priori prognostic factors for COVID-19 mortality (n = 55,757), complemented with univariable results and potential clinical associations.

Variable	Multivariable model, P	HR 2 SD	LCL 2 SD	UCL 2 SD	Univariable model, P	HR 2 SD	LCL 2 SD	UCL 2 SD	N	Possible clinical associations
Age	2.54E-191	6.95	6.11	7.91	0.00E+00	17.8	16.3	19.5	122,250	complex
Male gender	7.31E-41	1.82	1.66	1.98	2.39E-100	1.76	1.67	1.86	122,250	complex
Albumin	1.56E-18	0.731	0.682	0.784	5.10E-112	0.471	0.442	0.503	46,191	renal
Red cell distribution width	7.29E-17	1.34	1.25	1.44	4.86E-82	1.84	1.73	1.96	50,295	hematological insult
Insurance status: uninsured	1.14E-16	1.26	1.19	1.33	2.69E-17	1.21	1.16	1.26	122,250	NA
Blood urea nitrogen	5.21E-14	1.33	1.23	1.43	2.58E-67	1.9	1.77	2.04	53,846	renal
Respiratory rate	6.43E-13	1.46	1.32	1.62	5.00E-43	2.21	1.97	2.47	51,169	hypoxia
Fluid, pH and electrolyte imbalance (FPEI)	9.67E-12	1.3	1.21	1.41	3.97E-76	1.51	1.45	1.58	122,250	homeostasis
African-American race	1.30E-11	1.29	1.2	1.38	3.27E-14	1.22	1.16	1.29	122,250	complex
Dementia	1.23E-08	1.15	1.09	1.2	8.84E-22	1.17	1.13	1.21	122,250	neurology
Hemoglobin A1C	1.44E-08	1.23	1.15	1.33	4.28E-15	1.48	1.35	1.64	27,169	diabetes
Metastatic carcinoma	1.49E-08	1.15	1.09	1.2	5.02E-11	1.11	1.08	1.15	122,250	cancer
Insurance status: medicare	2.96E-08	1.23	1.14	1.32	4.14E-10	1.16	1.11	1.22	122,250	NA
Oxygen saturation	1.41E-07	0.838	0.785	0.895	7.99E-10	0.818	0.767	0.872	64,154	hypoxia
Height	2.09E-07	0.79	0.723	0.864	6.89E-04	0.832	0.748	0.925	61,694	pulmonary disadvantage
Moderate or severe liver disease	1.46E-06	1.11	1.07	1.16	7.77E-27	1.17	1.14	1.21	122,250	hepatic
Triglycerides	8.04E-06	1.21	1.11	1.31	4.78E-09	1.38	1.24	1.53	31,566	diabetes
Carbon dioxide total	8.70E-06	0.867	0.814	0.923	5.43E-15	0.763	0.713	0.817	51,259	homeostasis
Congestive heart failure	1.08E-05	1.14	1.08	1.21	1.97E-58	1.34	1.29	1.38	122,250	cardiovascular
Boostrix DTP vaccine	1.77E-05	0.843	0.78	0.912	8.91E-15	0.784	0.737	0.833	122,250	cross-reactive immunity

Table 2

Mortality rates of patient cohorts with or without a prior history of fluid, pH and electrolyte imbalance (FPEI), and/or renal comorbidities.

	Neither FPEI nor renal history	Both FPEI and renal	Renal only	FPEI only	Total
Survivors	91903 (79.5%)	6967 (6.0%)	3603 (3.1%)	13059 (11.3%)	115532
Non-survivors	3117 (46.4%)	1745 (26.0%)	546 (8.1%)	1310 (19.5%)	6718
Mortality	3.4%	25.0%	15.2%	10.0%	5.8%