Clinical background
Acute decompensated heart failure is a condition in which in-hospital mortality varies widely between patients, from a few per cent to over twenty. Assessing risk on arrival is central to several decisions: the level of care (monitored ward, intermediate care or intensive care), the intensity of diuretic and vasodilator therapy, and when and whether inotropic support should be started. Without a structured instrument, clinical assessments of risk tend to vary widely between assessors and to overestimate mortality, particularly in patients with a normal or raised blood pressure on arrival.
The ADHERE algorithm was developed to provide a simple bedside assessment based on three variables already available on arrival: blood urea nitrogen (BUN), systolic blood pressure and serum creatinine. The aim was to identify both patients with very low mortality, in whom less intensive management may be appropriate, and patients with high mortality, in whom early aggressive therapy and a higher level of care may be warranted.
Applying the ADHERE algorithm
The ADHERE algorithm is a classification tree (CART, classification and regression tree) based on three variables collected on arrival at hospital:
- Blood urea nitrogen (BUN) in mg/dL
- Systolic blood pressure in mmHg
- Serum creatinine in mg/dL
The tree can be expressed as a sequential decision rule:
The derivation cohort consisted of 33,046 hospitalisations from the ADHERE registry (Acute Decompensated Heart Failure National Registry), which collected data from 263 hospitals in the USA between October 2001 and February 2003 [1]. The mean age was 72.5 years and 52% were women. All patients had a primary diagnosis of acute decompensated heart failure. In-hospital mortality was 4.2% in the derivation cohort. Of 39 variables analysed, BUN was identified as the single strongest predictor, followed by systolic blood pressure and serum creatinine. The model was validated prospectively in 32,229 subsequent hospitalisations (March to July 2003), in which mortality was 4.0% and the risk stratification retained its ability to separate the groups [1].
Interpretation in practice
The tree divides patients into three risk groups with in-hospital mortality ranging from 2.1% to 21.9% in the derivation cohort [1,2]. The odds ratio for mortality between the high- and low-risk groups was 12.9 (95% CI 10.4 to 15.9) [1].
| Risk group | Criteria | In-hospital mortality |
|---|---|---|
| Low | BUN <43 mg/dL and SBP ≥115 mmHg | approx. 2% |
| Intermediate | All other combinations apart from low and high | approx. 5.7 to 6.4% |
| High | BUN ≥43 mg/dL, SBP <115 mmHg and creatinine ≥2.75 mg/dL | approx. 21.9% |
Low risk means that the patient can be managed on a general ward with standard diuretic treatment and monitoring, without an immediate indication for intensive care on the basis of mortality risk alone. It is the commonest group and includes the majority of patients.
Intermediate risk requires clinical judgement beyond the score. This group includes patients with impaired renal function but a preserved blood pressure, as well as patients with a low blood pressure but a normal BUN. The level of care should be determined by the overall picture: respiratory failure, oxygenation, the need for intravenous vasodilator treatment and the response to initial therapy.
High risk implies a mortality of approximately one in five patients during the admission. These patients should be considered for intensive or intermediate care, early invasive haemodynamic monitoring and a structured search for the underlying causes of the decompensation. The risk of cardiogenic shock and renal failure is substantial.
Validation and performance
In the prospective internal validation within the ADHERE registry, the algorithm retained its discrimination, with results in the validation cohort similar to those in the derivation cohort [1].
In an external validation by Lagu et al. (2016), seven mortality prediction models for acute decompensated heart failure were tested in 13,163 patients from 62 American hospitals (the HealthFacts database, 2010 to 2012) [3]. The median age was 74 years, half were women and 27% were Black. In-hospital mortality was 4.3%. The ADHERE algorithm showed a c-statistic of 0.68 to 0.70, comparable to the other clinical models (EFFECT and two GWTG-HF variants) but poorer than models based on administrative and electronic health record data (LAPS2 and Premier+, c-statistic 0.76 to 0.81) [3]. The prediction range of the ADHERE algorithm in this population was 1.2 to 17.4%, somewhat narrower than in the original cohort [3].
In a more recent study by Palmer et al. (2025) of 5,602 patients admitted to cardiac intensive care in the MIMIC-IV database, the AUC of the ADHERE algorithm was 0.63 (95% CI 0.57 to 0.70), lower than both GWTG-HF (AUC 0.66) and a machine learning model (AUC 0.72) [4]. This population was, however, selected to intensive care patients with a higher mortality (6.2%), which does not correspond to the general hospital population for which the ADHERE algorithm was developed.
A subanalysis within the ADHERE Emergency Module (ADHERE-EM) showed that the CART model stratified men and women equally well, with no significant difference in mortality between the sexes within each risk group [5].
Limitations
The ADHERE algorithm applies to acute decompensated heart failure and has not been validated for chronic stable heart failure, cardiogenic shock or high-output heart failure. It was developed for in-hospital mortality and says nothing about long-term prognosis or the risk of readmission.
The tree uses only three variables and therefore ignores other important predictors such as sodium, heart rate, respiratory rate, troponin, NT-proBNP, oxygen saturation and mechanical ventilation. This is a deliberate simplification to ensure that the tool can be used immediately on arrival, but it means that patients with a low classification may nonetheless be at high risk because of factors the tree does not capture.
The thresholds are based on American patient data from the early 2000s. Management of acute heart failure has since developed markedly, with wider use of ACE inhibitors, beta blockers, mineralocorticoid receptor antagonists and SGLT2 inhibitors, which may have altered mortality in the different risk groups. The absolute mortality figures should therefore be read as historical reference values rather than as current estimates.
A common error is to use BUN values in SI units (mmol/L) instead of mg/dL. The threshold of 43 mg/dL corresponds to approximately 15.4 mmol/L. Likewise, the creatinine threshold of 2.75 mg/dL corresponds to approximately 243 µmol/L. Entering values in the wrong unit produces a wholly incorrect classification.
The algorithm should not be used as the sole basis for deciding the level of care. It gives a crude risk stratification that must be complemented by clinical judgement, particularly in patients at intermediate risk, where the tree does not distinguish between different underlying causes of the decompensation.
References
- Fonarow GC, Adams KF Jr, Abraham WT et al. Risk stratification for in-hospital mortality in acutely decompensated heart failure: classification and regression tree analysis. JAMA 2005. PMID: 15687312
- Adams KF Jr, Uddin N, Patterson JH. Clinical predictors of in-hospital mortality in acutely decompensated heart failure: piecing together the outcome puzzle. Congest Heart Fail 2008. PMID: 18550923
- Lagu T, Pekow PS, Shieh MS et al. Validation and comparison of seven mortality prediction models for hospitalized patients with acute decompensated heart failure. Circ Heart Fail 2016. PMID: 27514749
- Palmer DQ, Gismondi RA, Gemal P et al. Mortality prediction in heart failure patients: machine learning versus Get With The Guidelines-Heart Failure (GWTG-HF) and Acute Decompensated Heart Failure National Registry (ADHERE). Cureus 2025. PMID: 40895687
- Diercks DB, Fonarow GC, Kirk JD et al. Risk stratification in women enrolled in the Acute Decompensated Heart Failure National Registry Emergency Module (ADHERE-EM). Acad Emerg Med 2008. PMID: 18275445
