Although residual confounding is a widely acknowledged limitation of observational data, it is particularly problematic in the critical illness literature. The complexity of underlying disease processes in the ICU is often difficult to delineate clinically and even more so to quantify in large datasets (1–3). Since these mechanisms might affect both the exposure and outcome of interest, they remain a source of potential confounding.
As an example, hyponatremia has been widely associated with an increased risk of mortality, thought largely due to water excess, cerebral edema, and downstream neurologic effects (4–7). However, a multitude of disparate disease mechanisms can lead to hyponatremia, including disorders of volume overload, volume depletion, thirst, and vasopressin release (8,9). Standard binary diagnostic codes often lack the sensitivity to identify these diseases or to quantify their severity accurately. Since these diseases are likely to be associated with both hyponatremia and mortality, they are potential sources of bias due to unaccounted confounding.
To better understand how underlying pathophysiologic determinants influence the association of hyponatremia and mortality, we examined physician choice of diuretic (Fig. 1). Potent loop diuretics are used for the most sodium-avid states, such as heart failure, cirrhosis, and renal disease, which influence both the risk of hyponatremia and mortality. The challenge of residual confounding similarly exists for hyponatremic patients without concurrent diuretic use, since the mechanism of hyponatremia could represent untreated cardiac, liver disease, or kidney disease, or disorders of vasopressin, such as can be seen in malignancy or lung disease. Thiazide diuretics are unique, since by inducing mild volume depletion with simultaneous preservation of renal concentration mechanisms, they can directly cause water retention, which usually improves with drug withdrawal (10). Since thiazides are primarily prescribed for hypertension, and since hypertension does not typically lead to hyponatremia otherwise, thiazide-associated hyponatremia provides an opportunity to reevaluate the question of hyponatremia and mortality in the absence of significant confounding from underlying disease. Put more formally, we hypothesize that the association of hyponatremia with mortality differs by diuretic use (i.e., effect modification) due to differences in the degree of confounding by underlying diseases and their severity across strata.
To address this hypothesis, we used a large inception cohort of critically ill medical and surgical patients admitted to a single tertiary medical center for whom pre-admission medications were available. We evaluated whether the associations between hyponatremia and 90-day mortality differed among thiazide and loop diuretic users and nondiuretic users.
We used the Multiparameter Intelligent Monitoring in Intensive Care (MIMIC)–II database, a joint venture managed by the Laboratory for Computational Physiology at Massachusetts Institute of Technology (MIT) and the Department of Medicine at the Beth Israel Deaconess Medical Center (BIDMC) (11). MIMIC-II contains data from 23,455 critical care admissions between 2001 and 2008 at BIDMC, a 700-bed urban academic medical center with 77 adult ICU beds. The database contains high temporal resolution data from clinical systems, including laboratory results, electronic documentation, and bedside monitor trends and waveforms. Use of the MIMIC II database has been approved by the Institutional Review Boards of BIDMC and MIT. We developed and validated a natural language processing algorithm that searched for pre-admission home medications in the discharge summary and then processed the medications to find individual entries of diuretics (12). A total of 17,896 patients had an identifiable medication section. Of these, 15,078 had a documented serum sodium and glucose at admission. Age, gender, and race were used to impute missing admission vital signs for 746 individuals, calculated means were used for 257 with missing white count, hematocrit, or creatinine measurements, and 1,417 hypernatremic patients (serum sodium > 145 mEq/L) were excluded, leaving a final cohort of 13,661 unique first ICU admissions.
The primary outcome was death within the critical illness hospitalization or within 90 days or discharge, as determined by the Social Security Death Index.
Admission serum sodium concentrations, obtained within 12 hours of admission to the ICU, were the primary exposure. Sodium concentrations were corrected for elevated serum glucose (corrected sodium = measured sodium + 0.024 × [serum glucose–100]) (13). Sodium concentrations were examined as a binary threshold according to our hospital definition of hyponatremia (≤ 133 mEq/L) and continuously.
Demographic information included age and gender. Histories of heart failure, hypertension, liver disease, and kidney disease were categorized by the Elixhauser comorbidity index. Admission systolic blood pressure, heart rate, WBC count, hematocrit, and creatinine were taken from the first available laboratory data within twelve hours of ICU admission.
Diuretics were classified into the following categories; loop diuretics (furosemide, torsemide, bumetanide, and ethacrynic acid); thiazide diuretics (hydrochlorothiazide, chlorthalidone, chlorothiazide, indapamide, and metolazone); and other diuretics (acetazolamide, spironolactone, eplerenone, amiloride, and triamterene). We categorized 101 patients prescribed both a thiazide and a loop diuretic as loop diuretic users. A sensitivity analysis excluding these 101 individuals did not significantly alter the results.
We present baseline characteristics, stratified by hyponatremia, among thiazide, loop, and diuretic-native patients. We first used logistic regression to examine the adjusted association between hyponatremia and mortality using the covariates above, including separate indicator variables for each diuretic type. To explore whether the association of hyponatremia and mortality differed between thiazide and loop diuretic users, we tested individual multiplicative interaction terms between thiazide and loop diuretics users with serum sodium less than or equal to 133 mEq/L, considering diuretic-naïve patients as the reference. We then repeated our models stratified by diuretic use. We examined the associations of sodium as a continuous variable similarly.
Of 13,661 critically ill patients, 8% (n = 1,110) were hyponatremic upon ICU admission. The overall frequency of mortality was 17.3% (n = 2,161), and 34.1% (n = 382) among hyponatremic patients.
Admission medications included thiazide diuretics in 9% (n = 1,188) and loop diuretics in 18% of patients (n = 2,498). Hyponatremia was observed in 9% of thiazide users (n = 110) and 10% of loop diuretic users (n = 254), compared with 7% of diuretic-naïve patients (n = 722). The rates of mortality were 15% for thiazide users (n = 178), 27% for loop diuretic users (n = 663), and 17% for diuretic-naïve patients (n = 1,657).
Table 1 illustrates the characteristics of patients according to hyponatremia and their admission diuretic use. Hyponatremic and normonatremic thiazide users tended to have similar blood pressure, urine output, and IV fluid administration. In contrast, hyponatremic loop and diuretic-naïve patients tended to have lower blood pressures and urine outputs and received more IV fluid than loop and diuretic naïve patients with normal serum sodium concentrations. Among diuretic-naïve patients, hyponatremia also tended to be associated with a history of malignancy and weight loss.
The adjusted odds of mortality associated with admission hyponatremia was 2.31 (95% CI, 2.00–2.67; p < 0.001), but this estimate differed according to outpatient diuretic type (multiplicative interaction terms between thiazide and loop diuretics with serum sodium ≤ 133 mEq/L: ß = –0.22; 95% CI, –0.38 to –0.09; p = 0.002 and ß = 0.07; 95% CI, –0.01 to 1.15; p = 0.08, respectively). In adjusted analysis (Fig. 2; Supplemental Table 1, Supplemental Digital Content 1, https://links.lww.com/CCX/A130), hyponatremia was associated with a three-fold higher risk of mortality among loop diuretic users and two-fold higher risk among diuretic-naïve patients, but it was not associated with an increased risk among thiazide diuretic users (odds ratio [OR], 0.87; 95% CI, 0.47–1.51; p = 0.63).
The association of hyponatremia severity with mortality similarly differed by according to diuretic use (multiplicative interaction terms between thiazide and loop diuretics with serum sodium defined continuously: ß = 0.21; 95% CI, –0.002 to 0.04; p = 0.02 and ß = –0.001; 95% CI, –0.02 to 0.001; p = 0.11, respectively). A one mEq/L increment in serum sodium was associated with 8% (OR, 0.92; 95% CI, 0.90–0.94%; p < 0.001) lower odds of mortality in loop diuretic users and 5% (OR, 0.95; 95% CI, 0.93–0.96; p < 0.001) lower odds among diuretic-naïve patients, but it was not associated with mortality risk among thiazide users (OR, 0.99; 95% CI, 0.95–1.02; p = 0.45).
The context in which hyponatremia occurs modifies the association between hyponatremia and outcomes. Assuming that there is generally less underlying morbidity in thiazide users, and therefore lower risk of confounding due to unaccounted illness severity, the absence of an association between hyponatremia and mortality in thiazide users undermines a causal role for water excess (14–16). In contrast, the stronger effect of hyponatremia on mortality in loop diuretic users, where there is likely more unaccounted pathophysiology, suggests that other mechanisms, such as osmolar independent activated water retention, are pathogenically explanative.
The modifying effect of diuretic use does not eliminate the possibility of a causal effect of hyponatremia on mortality in all scenarios, but suggests that any potential link should have a plausible and identifiable biologic pathway. Hyponatremia induced alteration in neurologic function, such as might occur with cerebral edema, could potentially be explanatory, and clearly, the water excess in symptomatic hyponatremia must be carefully and judiciously ameliorated. But in the absence of such explanatory mechanisms, the observed effect might simply reflect underlying confounding. Given that osmolar intendent water retention, such as sensed volume mediated activation of vasopressin in heart failure, is often difficult to measure or quantify, yet is simultaneously closely linked with mortality, confounding seems likely.
Our analysis questions whether asymptomatic hyponatremia, barring other pathogenic signatures of water intoxication, needs to be aggressively treated. Hyponatremia remains a common cause of hospitalization and ICU admission, associated with downstream risks and costs, and whether more conservative treatment in asymptomatic patients might improve overall patient care needs to be determined.
Our analysis has important limitations. We were not able to directly measure whether confounding was less among the thiazide users than other diuretic groups, and the observation of an effect modification does not rule out the possibility that hyponatremia might contribute to mortality with more severe disease (such as loop diuretics). In addition, we did not know the cause of hyponatremia, and since we lacked longitudinal sodium data, the absence of an effect of hyponatremia on mortality in thiazide users might simply reflect the rapid correction of sodium concentrations with drug cessation.
In a large population of critical care patients, hyponatremia was associated with higher mortality only among loop diuretic users and nondiuretic users, but not among thiazide users. Because the association of hyponatremia and mortality is likely to be least confounded in the latter group, our results suggest that the high mortality rates observed with hyponatremia are in part due to the underlying pathophysiology that causes water retention, rather than water excess per se. Our analysis highlights how difficult it is to account for the complex determinants of many exposures in critical illness and cautions against over-interpretation of observational data.
1. Ekbia H, Mattioli M, Kouper I, et al. Big data, bigger dilemmas: A critical review. J Assoc Inf Sci Technol. 2015; 66:1523–1545
2. Celi L. Big data in the intensive care unit. Am J Respir Crit Care. 2013; 187:1157–1160
3. Rush B, Stone DJ, Celi LA. From big data to artificial intelligence: Harnessing data routinely collected in the process of care. Crit Care Med. 2018; 46:345–346
4. Zilberberg MD, Exuzides A, Spalding J, et al. Epidemiology, clinical and economic outcomes of admission hyponatremia among hospitalized patients. Curr Med Res Opin. 2008; 24:1601–1608
5. Funk GC, Lindner G, Druml W, et al. Incidence and prognosis of dysnatremias present on ICU admission. Intensive Care Med. 2010; 36:304–311
6. Upadhyay A, Jaber BL, Madias NE. Incidence and prevalence of hyponatremia. Am J Med. 2006; 119:S30–S35
7. Cumming K, Hoyle GE, Hutchison JD, et al. Prevalence, incidence and etiology of hyponatremia in elderly patients with fragility fractures. PLoS One. 2014; 9:e88272
8. Danziger J, Zeidel ML. Osmotic homeostasis. Clin J Am Soc Nephrol. 2015; 10:852–862
9. Danziger J, Hoenig MP. The role of the kidney in disorders of volume: Core curriculum 2016 Am J Kidney Dis. 2016; 68:808–816
10. Schrier RW. Water and sodium retention in edematous disorders: Role of vasopressin and aldosterone. Am J Med. 2006; 119:S47–S53
11. Saeed M, Villarroel M, Reisner AT, et al. Multiparameter Intelligent Monitoring in Intensive Care II (MIMIC- II): A public-access intensive care unit database. Crit Care. 2011; 39:952–960
12. Danziger J, William JH, Scott DJ, et al. Proton-pump inhibitor use is associated with low serum magnesium concentrations. Kidney Int. 2013; 83:692–699
13. Hillier TA, Abbott RD, Barrett EJ. Hyponatremia: Evaluating the correction factor for hyperglycemia. Am J Med. 1999; 106:399–403
14. Hoorn EJ, Zietse R. Hyponatremia and mortality: Moving beyond associations. Am J Kidney Dis. 2013; 62:139–149
15. Hoorn EJ, Zietse R. Hyponatremia and mortality: How innocent is the bystander? Clin J Am Soc Nephrol. 2011; 6:951–953
16. Sterns RH. Disorders of plasma sodium–causes, consequences, and correction. N Engl J Med. 2015; 372:55–65