Hyponatraemia in Emergency Medical Admissions—Outcomes and Costs

Healthcare systems in the developed world are struggling with the demand of emergency room presentations; the study of the factors driving such demand is of fundamental importance. From a database of all emergency medical admissions (66,933 episodes in 36,271 patients) to St James’ Hospital, Dublin, Ireland, over 12 years (2002 to 2013) we have explored the impact of hyponatraemia on outcomes (30 days in-hospital mortality, length of stay (LOS) and costs). Identified variables, including Acute Illness Severity, Charlson Co-Morbidity and Chronic Disabling Disease that proved predictive univariately were entered into a multivariable logistic regression model to predict the bivariate of 30 days in-hospital survival. A zero truncated Poisson regression model assessed LOS and episode costs and the incidence rate ratios were calculated. Hyponatraemia was present in 22.7% of episodes and 20.3% of patients. The 30 days in-hospital mortality rate for hyponatraemic patients was higher (15.9% vs. 6.9% p < 0.001) and the LOS longer (6.3 (95% CI 2.9, 12.2) vs. 4.0 (95% CI 1.5, 8.2) p < 0.001). Both parameters worsened with the severity of the initial sodium level. Hospital costs increased non-linearly with the severity of initial hyponatraemia. Hyponatraemia remained an independent predictor of 30 days in-hospital mortality, length of stay and costs in the multi-variable model.


Data Collection
For audit purposes we employed an anonymous patient database assembling core information about each clinical episode from elements contained on the patient administration system, the national hospital in-patient enquiry (HIPE) scheme, the patient electronic record, the emergency room and laboratory systems. HIPE is a national database of coded discharge summaries from acute public hospitals in Ireland [30]. Ireland used the International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) for both diagnosis and procedure coding from 1990 to 2005 and ICD-10-CM since then.
Data held on the database includes the unique hospital number, admitting consultant, date of birth, gender, area of residence, principal and up to nine additional secondary diagnoses, principal and up to nine additional secondary procedures, and admission and discharge dates. Additional information cross-linked and automatically uploaded to the database includes physiological, haematological and biochemical parameters. Data was related to all emergency general medical patients admitted to SJH in the twelve years between 2002 and 2013.
Each emergency medical patient was referred to the team of the "on-call" Acute Medicine Consultant-on-take for a 24 hours period-most ~90% remained under the care of the admitting consultant for the duration of their admission. Approximately 9.9% of our patients stay >30 days with a median LOS of 54.8 days (IQR 38.8,97.2). Consequently the LOS data represents a highly skewed distribution. Although the clinical episode is complete for the majority by day 30, some patients remain for social reasons related to the lack of long-term care facilities. We have therefore chosen a truncated end-point (at the 30-day endpoint) for analysis, to avoid these additional confounders.
We assessed the ability of known predictors-Acute Illness Severity [13,31], Charlson Co-Morbidity Index [32], Manchester Triage Category [33] and Chronic Disabling Score [34] to predicts outcomes (30 days in-hospital mortality and Length of Stay) and Episode Costs. Derangement of haemodynamic and physiological admission parameters has been utilised to derive an Acute Illness Severity Score that predicts clinical outcomes [13,31,35]. From modelling laboratory data collected at time of hospital admission we developed a predictive algorithm based on serum sodium, potassium, urea, albumin, red cell distribution width, white blood cell count and troponin level. The underlying principle is that deviation beyond the bounderies of "normal homeostasis" is an estimate of risk, although the relationship is non-linear and differs for each variable, it is possible to calculate an "aggregrate" risk score from the admission biochemistry profile [13]. Six groups were originally defined with a 30 days mortality risk increasing in an exponential fashion.
The Charlson Co-morbidity index provides an evaluation of Co-morbidity [32]. Co-morbidity is the presence of one or more additional disorders (or diseases) co-occurring with a primary disease or disorder. The Charlson Co-morbidity index predicts the ten-year mortality outcomes for patients who may have a range of Co-morbid conditions, such as heart disease, AIDS, or cancer (a total of 22 conditions). Each condition is assigned a score of 1, 2, 3, or 6, depending on the mortality; scores are then summed into three classifying groups (Groups 0, 1 and 2).
We recently described a chronic disability score, derived from counts of discharge ICD9/ICD10 codes, that strongly correlated with mortality and length of stay [34].
We have data that permitted the cost of each clinical episode for emergency patients admitted between 2008-2012 to be calculated. The Republic of Ireland has proposed to introduce a Money Follows the Patient system, where a case based funding model with Diagnosed Related Groups (DRG's), compares hospital costs, quality and efficiency. The calculation of costs per case is adjusted by reference to the relative cost weight of each DRG. The hospital costing of the price of an episode of care encompasses all costs appropriately associated with the delivery of that care including pay costs, non-pay costs and costs of diagnostics, medical services, theatres, laboratories, wards and overhead allocations as appropriate.
The hospital uses a number of standard accounting costing methodologies. The predominant approaches used in this exercise were Activity Based Costing and Absorption Costing [36,37]. Both methods are used in parallel to cost individual patient episodes of care by directly linking cost to patient clinical data (e.g., laboratory and radiology tests, inpatient bed days). The accuracy of the costing is greatly enhanced because the hospital has utilized a robust devolved accounting and budgetary framework since 2004. The financial data is validated by externally audited annual Financial Statements; in addition strong relationships between costing and clinical risk profile/outcomes data would suggest that the financial calculations provide a realistic view of the costs of care provision.

Statistical Methods
Descriptive statistics were calculated for background demographic data, including means/standard deviations (SD), medians/interquartile ranges (IQR), or percentages. Comparisons between categorical variables and mortality were made using chi-squared tests. The 30-day in-hospital survival outcome bivariate variable was assessed by fitting a logistic regression model for variables that were univariately predictive. Combining the significant predictors gave an AUROC of 0.87 (95% CI: 0.86, 0.87) to predict an in-hospital death by day 30. We used margins to estimate and interpret adjusted predictions for sub-groups, while controlling for other variables, using computations of average marginal effects [38]. Margins are statistics calculated from predictions of a previously fitted model at fixed values of some covariates and averaging or otherwise over the remaining covariates.
For the LOS count data, we employed a truncated Poisson regression model, including some categorical variables (e.g., disabling score groups) in the model as a series of indicator variables. The dependent variable LOS is a positive integer; it cannot have zero value. The data are truncated because there are no observations on individuals who stayed for zero days; the predictor variables were therefore regressed against LOS using the zero-truncated Poisson model. We used robust standard errors for the parameter estimates, as recommended by Cameron and Trivedi [39]. The Poisson regression coefficients are the log of the rate ratio: the rates at which events occur are the incidence rates. Thus with the Truncated Poisson regression model, we can interpret the coefficients in terms of incidence rate ratios (IRR).
As hospital costs typically have considerable heteroscedasticity, we examined the impact of the predictor variable (hyponatraemia grouping), using quantile regression; this method models the relationship between the hospital costs and the conditional quantiles (25%, median, 75%) of a predictor variable. Thus, traditional least-squares regression, requiring both normality and equal variance, does not perform well for these types of data [40]. Quantile regression, as a method, can be used to model the effects of covariates on the conditional quantiles of a response variable for such datasets [41]. The approach is robust, making no distributional assumption about the error term in a model. It is also robust to extreme points in the response space (outliers); confidence intervals for the estimated parameters are based on inversion of a rank test [42]. Quantile analysis concentrates on the dependent variable and its distribution; it is particularly appropriate where one might anticipate marked differences in the dependent variable at different quantiles of the predictor variable.
Hyponatraemia and hospital Episode Cost (Figure 3, Table 3).  Hyponatraemia at time of hospital admission was a significant predictor of in-hospital costs. Quantile regression demonstrated ( Figure 3) the non-linear relationship between the predictor variable (level of admission hyponatraemia) and total hospital episode cost. The standard Ordinary Least Squares (OLS) regression model-€550 (95% CI: 439, €661) over-estimated the costs at lower and underestimated at upper cost quantiles; these at Q25 point of the duration of hyponatraemia distribution were €194 (95% CI: €146, €422), at the median of €403 (95% CI: €330, €477) but increased at the Q75 point to €638 (95% CI: €502, €774). The estimates of cost increase per unit change in quantile demonstrated that hyponatraemia was only second to chronic disabling disease as a predictor variable for hospital episode costs (Table 3).

Discussion
The current study demonstrates that hyponatraemia is an independent predictor of mortality, hospital LOS, and costs in unselected general medical admissions. Our study adds to the previous literature by demonstrating in our cohort that the effect of hyponatraemia is not fully explained by illness severity or co-morbidity [27]. Furthermore our data demonstrates a biologic gradient, strengthening the support for an independent effect of hyponatraemia. Our study has shown hyponatraemia to be an extremely powerful predictor of the cost of hospital admission. Of the analysed variables only disabling disease was a more significant predictor of costs in our model.
There are significant potential clinical implications to these findings. As a minimum, the data suggest that hyponatraemia should arouse concern in clinicians that affected patients are at risk for adverse outcomes. Patients at higher risk of adverse outcomes have the potential for a greater benefit from the dedication of a physician's limited time resources than those at low risk of events. This further justifies the role of serum sodium as a component of risk prediction scores and tools. There may be benefits to the active modification of the hyponatraemic state in these patients; however this will require further study to determine. This is of particular interest as unlike outcome predictors such as co-morbidity and disabling disease score, hyponatraemia is amenable to both preventive and corrective interventions.
The impact of serum sodium on outcomes has been evaluated in a number of previous studies. Serum sodium has been shown to be an important prognostic factor in a range of diagnostic groups including heart failure, pulmonary hypertension and subarachnoid haemorrhage [22,[24][25][26]. The influence of sodium on outcomes has also been shown for intensive care unit admissions and amongst general medical admissions [23,43]. We have previously reported five year data on our patient cohort, with differential influences of acute illness severity on the strength of the association of serum sodium with mortality demonstrated for hyponatraemic and hypernatraemic patients [2]. Evaluation of the impact of hyponatraemia on outcomes by case-control methodology has shown consistent results [44,45].
Our study has several strengths. The use of a large dataset has allowed us to demonstrate that hyponatraemia is not an epiphenomenon, low serum sodium actively impacts on mortality rather than being a reflection of other risk predictors. We have included all unselected medical admissions including those sickest of the sick who are admitted to ICU or HDU. Our study therefore reflects real world clinical practice enhancing its relevance to practicing clinicians. The demonstration of a dose-response curve further enhances confidence in the veracity of the results.
Like any study ours also has limitations. We have demonstrated an association between hyponatraemia and outcomes. While we have controlled for a multitude of other risk predictors this does not neccessarily imply causation. Residual unmeasured confounders may be present which modify the effects demonstrated here. Our study was based on a retrospective cohort, we are dependant on the accurate coding and recording of data over the course of the study. In addition our study was based in a single centre, the results will require verification in other centres to establish external validity.

Conclusions
We have demonstrated an independent relationship between hyponatraemia and the key outcome measures of 30 days in-hospital mortality, length of stay, and cost of hospital admission.