Are we correcting sodium too slowly — and is it killing patients?

by Dr. J. Shriki

Dustin G. Mark, Mubarika Alavi, Joshua R. Nugent, et al.Sodium Correction Rates and Associated Outcomes Among Patients With Severe Hyponatremia: A Retrospective Cohort Study. Ann Intern Med.2026;179:330-339. [Epub 27 January 2026]. doi:10.7326/ANNALS-25-03676

A deep dive into a landmark Annals paper on hyponatremia correction rates, the methods behind it, and what it means for critical care practice.

Good day, Eh!, and Welcome to my first blog post on the Meducation website reboot. I wanted to kick things off with a paper that I think perfectly embodies what this site is about. It challenges entrenched dogma — specifically the US approach to sodium correction, which, as we'll see, the Europeans have quietly been doing better for years. It uses rigorous, thoughtful methodology to squeeze real signal out of a dataset that, while large in clinical terms, is modest by the standards of modern big data. And it covers a condition we deal with constantly in the cardiac ICU.

The paper is Sodium Correction Rates and Associated Outcomes Among Patients With Severe Hyponatremia by Mark et al., published in Annals of Internal Medicine in January 2026 — fresh off the press. Let's get into it.

Background: the dogma we inherited

Severe hyponatremia — defined as serum sodium at or below 120 mEq/L — affects roughly 1% of hospitalized patients and carries substantial morbidity and mortality. For decades the clinical response has been governed by one overriding fear: osmotic demyelination syndrome (ODS), previously called central pontine myelinolysis. The mental image is vivid: correct sodium too fast, osmotic stress tears apart the myelin sheath, and you've traded a fixable electrolyte problem for a permanent neurologic catastrophe.

That fear has been productive in some ways — it prompted systematic thinking about correction rates. But it's also driven guidelines toward targets that may be dangerously conservative. The American guidelines in particular have recommended correction as slow as 4–6 mEq/L per 24 hours. The problem is a growing body of evidence suggesting that slow correction isn't neutral — it's independently associated with higher mortality. Patients who are persistently hyponatremic are suffering ongoing hyponatremic encephalopathy, aspiration events, and prolonged critical illness. We've been so afraid of the rare catastrophe of ODS (incidence 0.2–0.5%) that we may have been causing net harm to the majority.

The core tension: ODS is devastating but rare. Persistent hyponatremia is common and also deadly. Guidelines optimized to prevent the former may be increasing the latter — and this paper is the largest observational evidence yet that this is exactly what's happening.

The study: what they did

This is a retrospective cohort study from Kaiser Permanente Northern California — 21 community hospitals, 16 years of data (2008–2023). The cohort is adults admitted from the ED with sodium at or below 120 mEq/L, with the first qualifying admission per patient used. After exclusions for dilutional hyponatremia (glucose ≥400) and lack of continuous health plan coverage, they ended up with just under 14,000 patients for the 90-day outcome analysis.

The primary outcome was a composite of 90-day death or delayed neurologic events — importantly, the neurologic endpoint was deliberately broad, capturing demyelinating disease, paralytic syndromes, epilepsy, and coma/altered consciousness occurring between day 3 and day 90. This was smart design: ODS alone is underdiagnosed and heavily subject to confirmation bias (you're more likely to call a radiographic lesion ODS if you know the patient overcorrected). Broadening the net reduces that bias.

The methods: why this is worth taking seriously

Let’s take a second to remember how to read primary literature. Never Read the intro, discussion or conclusion! First read the title, then maybe the abstract, then always START with the methods section. Then read the results section, more importantly look at the tables and numbers not the words. Now the important part… GENERATE YOUR OWN CONCLUSIONS!!! Now you are ready to read the intro and conclusions sections. When you do that you will start to see the methods section of this paper is where the paper earns its credibility. They used targeted maximum likelihood estimation (TMLE) as their primary estimator — a doubly robust, semiparametric approach that I think is genuinely the right tool for this kind of causal inference problem in observational data. Let me explain what that means in practice.

TMLE — THE PRIMARY ESTIMATOR

TMLE uses two models: an outcome model predicting the probability of death or neurologic events given treatment and covariates, and a propensity model predicting the probability of each correction rate category given covariates. In other words it makes two different regression equations as a way to compare internal validity of the paper. The key is the targeting step — it uses the propensity scores to compute a correction term that updates the outcome model specifically to reduce bias for the estimand you care about (here, the standardized risk difference). The final estimate comes from marginal standardization over updated predictions — effectively asking what the outcome rate would be if the entire population received slow versus fast correction, holding covariates fixed.

The doubly robust property means you only need one of the two models to be correctly specified for the estimate to be consistent. That's a meaningful protection in complex real-world data where you can't be certain either model is perfect.

SENSITIVITY ESTIMATORS — AIPW AND IPWRA

Augmented Inverse Probability Weighting (AIPW)

Starts with an IPW estimator (reweighted observed outcomes using propensity scores) and adds an augmentation term derived from outcome model residuals that corrects for propensity model imperfection. Doubly robust — consistent if either model is correct.

Inverse Probability Weighted Regression Adjustment (IPWRA)

Fits weighted outcome regression models within a propensity-score reweighted pseudo-population, then marginalizes. Also doubly robust. Slightly less efficient than TMLE in finite samples but conceptually complementary.

The fact that all three estimators agreed is important. It means the result isn't an artifact of any single modeling choice. That convergence across independent methodological pathways substantially increases confidence in the direction of the finding.

E-VALUES — QUANTIFYING RESIDUAL CONFOUNDING

The E-value answers a specific question: how strong would an unmeasured confounder need to be — simultaneously on both the exposure-confounder and confounder-outcome axes — to fully explain away the observed association?

Weak: 1.0–1.5

Modest: 1.5–2.0

This paper: 2.3–2.7

Strong: 3.0–4.0

Almost impossible: >4.0

An E-value of 2.3–2.7 is in the moderate range — not dismissible but not airtight. The honest concern is that the most plausible unmeasured confounder here — the pathophysiology driving hyponatremia — may operate near exactly this magnitude. A patient with volume-depleted thiazide hyponatremia or beer potomania will correct fast naturally and has low baseline mortality. The LAPS and Elixhauser adjustments capture chronic comorbidity and admission physiology but can't fully disentangle why sodium is low or why it's rising. It's plausible residual confounding in the 2–3x range exists, which is precisely why this finding is suggestive but not causal proof.

The results: what they found

Relative to slow correction, both medium and fast rates were associated with meaningfully lower adjusted risk for the primary composite outcome.

• Fast versus slow had a standardized risk difference of −9.0 percentage points (95% CI −11.1 to −6.9).

• Medium versus slow was −5.6 percentage points (−7.1 to −4.0).

• These are clinically large numbers in a population with an 18% 90-day mortality rate.

The heterogeneity analysis was one of the more interesting features. They stratified by predicted risk quartile and found that absolute risk differences became more negative (i.e., benefit of faster correction grew larger) as baseline risk increased — but risk ratios remained roughly constant. This is what you'd expect from a real treatment effect rather than a confounding artifact: higher-risk patients have more to gain absolutely from the same relative benefit.

The exploratory U-shape finding: When correction rate was modeled as a continuous variable, predicted risk reached a nadir around 15–20 mEq/L per 24 hours before rising again at higher rates. This aligns with animal data showing ODS risk at rates exceeding 20–25 mEq/L. The U-shape is plausible biology — but the right tail had sparse data and this should be treated as hypothesis-generating only.

The ODS-specific signal was also notable for what it didn't show: only 19 coded cases of demyelinating disease in nearly 14,000 patients. Coma and altered consciousness accounted for 60% of delayed neurologic events — and faster correction was associated with fewer of those events, not more. This is the opposite of what ODS-centric guidelines predict.

Where European guidelines are already ahead

The 2014 ESE/ERA-EDTA European guidelines target approximately 10 mEq/L per 24 hours with a ceiling of 10–12 — which maps almost exactly to the "medium" correction category in this paper. The medium group had significantly better outcomes than the slow group and, critically, did not have worse neurologic outcomes than the slow group. So in a sense, this paper is large-scale retrospective validation that European practice is on the right track, and that the American insistence on 4–6 mEq/L targets for most patients is probably causing net harm. The Europeans have been operating in the medium zone while Americans have been trained to fear it.

The honest limitations

• Residual confounding. The pathophysiology of hyponatremia is largely unmeasured. Self-correcting causes inflate the apparent fast-correction benefit. E-values of 2.3–2.7 are moderate, not overwhelming.

• ODS-risk subgroup power. The subgroup analyses in high-risk patients (cirrhosis, malnutrition, hypokalemia) were not powered to detect harm from faster correction in that specific population. The absence of a signal is not the same as absence of risk.

• ICD-10 outcome ascertainment. Neurologic events captured by diagnostic codes will miss cases and may misclassify others. The authors acknowledge neurologic events missing at random — reasonable but not verifiable.

• No preadmission sodium. Chronicity of hyponatremia — a major ODS risk factor — is largely uncaptured. Acute-on-chronic versus purely chronic patterns likely behave very differently.

• Observational design. Causation cannot be inferred. The RCT this field needs — randomizing patients to different correction rate targets with appropriate ODS-risk stratification — remains undone.

Bottom line

CLINICAL TAKEAWAY

This paper is not practice-changing on its own. It is observational, residual confounding is a legitimate concern, and the ODS-high-risk subgroup remains genuinely uncertain territory.

What it does do is add substantial weight to a growing body of evidence — including a 2025 JAMA Internal Medicine meta-analysis showing 6–13% absolute mortality increase with slow correction — that the 4–6 mEq/L target is probably causing net harm in most patients. The medium range of 8–12 mEq/L, already endorsed by European guidelines, now has both biologic plausibility and the largest observational cohort yet behind it.

The practical translation: don't panic about medium correction rates, treat hyponatremic encephalopathy aggressively when it's present, and reserve the most conservative targets for the truly high-risk ODS patients — while acknowledging that even that recommendation rests on weak prospective evidence. And recognize that desmopressin for rescue overcorrection remains a critical tool in your back pocket.

What the field ultimately needs is a properly powered RCT with ODS-risk stratification. Until then, this paper is the best data we have — and it points clearly toward loosening the reins.