Development and internal validation of a rapid bedside score for predicting in-hospital mortality in acute heart failure

Abstract

BACKGROUND

Early risk stratification for in-hospital mortality in acute heart failure (AHF) is a challenging topic, particularly at the time of presentation, when rapid clinical decisions are required. Many existing risk scores rely on variables that are not immediately available or are impractical for bedside use.

AIM

To develop and internally validate Fast Assessment Score for triage in heart failure (FAST-HF), a rapid bedside risk score for predicting in-hospital mortality in patients with AHF.

METHODS

A total of 123 patients with AHF were retrospectively analysed. The primary outcome was in-hospital mortality. A multivariable logistic regression model included admission parameters-NT-proBNP, arterial lactate, serum sodium, systolic blood pressure and clinical congestion summarized by a congestion index. The multivariable model was converted into a simplified point-based bedside score using a Framingham-type scaling approach. Model performance was assessed using discrimination, calibration, decision curve analysis (DCA), Youden-based risk stratification and internal validation by bootstrapping with comparative area under the curve (AUC) analysis.

RESULTS

The multivariable model demonstrated excellent discrimination for in-hospital mortality (AUC = 0.889), which was preserved after simplification into the FAST-HF bedside score (AUC = 0.867). Calibration remained acceptable, with similar Brier scores for both models. A Youden-derived cut-off identified a high-risk subgroup (≥ 15 points) with 38.6% in-hospital mortality, while no deaths occurred in the low-risk group (≤ 8 points). DCA demonstrated consistent net clinical benefit across clinically relevant threshold probabilities.

CONCLUSION

FAST-HF provides a simple, rapid bedside approach for early in-hospital mortality prediction in AHF, with good internal performance and potential clinical utility. External validation in independent cohorts is warranted.

Key Words: Acute heart failure; In-hospital mortality; Bedside risk score; Prognostic model; Emergency department; Risk stratification

Core Tip: This single-centre retrospective study developed and internally validated the Fast Assessment Score for triage in heart failure (FAST-HF) score, a practical prognostic tool based on readily available clinical and biological parameters, including systolic blood pressure, congestion status, NT-proBNP, lactate and serum sodium. FAST-HF showed good discriminatory performance for predicting in-hospital mortality and may help identify high-risk patients early during admission. The score is intended as an exploratory bedside instrument that requires external validation before routine clinical implementation.

INTRODUCTION

Heart failure (HF) constitutes a major public health burden worldwide, affecting millions of individuals, with a growing incidence of rehospitalization and mortality[1,2]. HF accounts for 3%-8% of total healthcare expenditures in developed countries, imposing a substantial burden on healthcare systems[3]. Despite advances in prevention and treatment that have improved survival, the overall prevalence of HF remains high and is expected to increase further[4,5]. By 2030, the number of individuals affected is expected to rise substantially, with associated healthcare costs projected to at least triple[3,6]. Hospitalization expenditures account for approximately two thirds of total HF-related costs and are frequently driven by patients’ comorbidities rather than by the form of presentation[7-9]. Patients with the most severe forms of acute HF (AHF) frequently require complex management in intensive cardiac care units or intensive care units, resulting in even higher hospitalization-related expenditures[10,11].

Over the past two decades, HF has become a major focus of academic research. From early HF registries to contemporary studies, numerous clinical, biological, electrocardiographic and echocardiographic parameters have been identified for risk stratification[9]. Based on these predictors, various prognostic models and scoring systems have been developed to estimate short- and long-term outcomes, with their prognostic relevance varying according to individual patient profiles and comorbidity burden[12].

While some risk scores focus on chronic HF[13], the majority target patients with AHF[14-17]. These models are primarily based on clinical and biological parameters, with only more recent scores incorporating biomarkers or echocardiographic data. Several scores using easily obtained data in the emergency department (ED) (e.g. patient history, clinical and paraclinical parameters) have been proposed to aid early decision making[18]. Examples include the Ottawa Heart Failure Risk Scale for identifying patients at risk of serious adverse events within 14 days, the MEESSI-AHF risk score to predict 30-day mortality, the Get with the Guidelines Heart Failure Risk Score (GWTG-HF) for in-hospital mortality and the Emergency Heart Failure Mortality Risk Grade (EHMRG) which predicts 7-day mortality[15,19-21].

Clinical congestion represents a key pathophysiological hallmark of AHF and has been incorporated into multiple congestion-based scores and indices[22]. However, congestion-focused approaches alone may incompletely capture early mortality risk, which is also influenced by systemic hypoperfusion, neurohormonal activation and hemodynamic instability[23-25].

Despite the availability of these scores, they differ substantially in their intended clinical use, complexity and outcome of interest. Some tools were developed to support ED discharge decisions or short-term outcomes beyond hospitalization, while others require a relatively large number of variables that may not be immediately available at presentation.

The present study sought to develop and internally validate a rapid bedside score, the Fast Assessment Score for triage in HF (FAST-HF), based on easily obtainable parameters for predicting in-hospital mortality among patients with AHF in the ED.

MATERIALS AND METHODS

Study population

This single-centre retrospective study included 123 consecutive patients diagnosed with AHF and admitted to the Department of Cardiology of the Clinical Emergency Hospital “Sfântul Pantelimon” between January and December 2024. All patients underwent comprehensive clinical, laboratory and transthoracic echocardiographic evaluation at admission. The diagnosis of AHF was established according to the current European Society of Cardiology guidelines, integrating typical clinical symptoms and signs of congestion and/or hypoperfusion, elevated NT-proBNP levels and echocardiographic evidence of functional or structural cardiac abnormalities[26]. This study was conducted in a hospital with a level II intensive cardiac care unit, capable of managing patients with AHF requiring advanced monitoring, vasoactive therapy and non-invasive ventilatory support. The primary outcome was all-cause in-hospital mortality[27].

Patients with missing data regarding congestion status, arterial blood gas analysis, NT-proBNP levels or other laboratory parameters were excluded from this study. Patients with amyloidosis, hypertrophic cardiomyopathy, myocarditis, Takotsubo cardiomyopathy or acute coronary syndrome were also excluded from this study to avoid confounding aetiologies of AHF and to maintain a homogeneous pathophysiological substrate. The reason for these exclusions was to avoid AHF of acute ischemic origin or secondary forms that have distinct pathophysiological mechanisms and prognostic profiles from AHF of non-ischemic or chronic ischemic origin. Infiltrative cardiomyopathies, such as amyloidosis and hypertrophic cardiomyopathy, have different structural and hemodynamic patterns than conventional HF regardless of ejection fraction. In acute coronary syndromes, acute myocardial ischemia is the main cause of the onset of AHF, while in myocarditis and Takotsubo cardiomyopathy transient ventricular dysfunction results from inflammatory or stress-induced myocardial injury.

This study was approved by the Hospital’s Ethics Committee (approval No. 77). Given the retrospective design and the use of anonymized data, the requirement for written informed consent was waived in accordance with institutional regulations. All study procedures were performed in accordance with the Declaration of Helsinki for research involving human subjects.

Data collection

Clinical data were retrospectively collected from patients’ charts. Clinical examination, laboratory tests and echocardiography were performed in the ED. Physicians collected information regarding cardiovascular risk factors (hypertension, obesity, smoking status, diabetes and dyslipidemia), as well as prior medical history (prior diagnosis of HF, valvular heart disease, cardiomyopathies, ischemic heart disease, atrial fibrillation, chronic kidney disease or chronic obstructive pulmonary disease). A comprehensive physical examination was performed on each patient, including heart rate, blood pressure, signs of congestion, oxygen saturation and prior treatment. Systolic blood pressure (SBP) was measured noninvasively at presentation using a calibrated device.

Upon admission, within minutes of arrival in the ED, all patients underwent arterial blood gas analysis, performed on a dedicated analyzer (RAPIDPoint 500, Siemens Healthineers, Erlangen, Germany). This method enables rapid, accurate assessment of metabolic and electrolyte parameters for real-time hemodynamic evaluation. NT-proBNP levels were determined using a rapid immunoassay kit, with results available within 15 minutes of venous blood sample collection. The laboratory tests were processed at the hospital’s laboratory for all patients and were performed in accordance with standard hospital protocols. Arterial blood gas analysis and NT-proBNP followed manufacturer-recommended calibration and quality standards.

Transthoracic echocardiography was performed in the ED according to institutional protocol, using dedicated ultrasound equipment (GE Logiq S7 XDclear 2.0 ultrasound system), as part of the standard diagnostic assessment. Collected echocardiographic data included left ventricular ejection fraction (LVEF), systolic and diastolic function of the left ventricle (LV), LV filling pressures and signs of pulmonary hypertension and congestion. However, these parameters were not included in the final model to preserve the purely clinical and biological nature of the FAST-HF score.

Definition of the bedside score

The newly developed risk score was designed to predict the risk of in-hospital mortality in patients admitted with AHF using simple and easily obtainable clinical and biological parameters that are available within the first minutes of ED admission. Five independent parameters were selected based on their independent prognostic value and their immediate availability in the ED: NT-proBNP, lactate, serum sodium, SBP, congestion index. The signs of congestion (jugular vein distension, pulmonary rales, pleural effusion and peripheral oedema) were later integrated into a congestion index, representing the sum of positive findings for each patient (0-4).

Statistical analysis

All statistical analyses were performed using IBM SPSS Statistics, version 26.0. The baseline characteristics of patients were presented as mean ± SD for continuous variables and as absolute n (%) for categorical variables. Patients were divided into two groups: Survivors and non-survivors. Differences between the two groups were compared using Student’s t-test or Mann-Whitney U test for continuous variables and the χ² test or Fisher’s exact test for categorical variables. Statistical significance was defined as a two-sided P value < 0.05.

To identify patients at risk for in-hospital mortality, we developed a clinical and biological predictive score, combining variables that were statistically significant in univariate analyses with established prognostic parameters reported in the literature. The FAST-HF score was developed using a two-step modelling approach. For statistical robustness and clinical applicability, we derived a continuous multivariable logistic regression model, which will be referred to as the “FAST-HF multivariable model”, followed by its transformation into a simplified, point-based score (the “FAST-HF bedside model”).

The variables comprising the FAST-HF score are easily and routinely obtainable parameters in any ED-NT-proBNP (pg/mL), arterial lactate (mmol/L), serum sodium (mmol/L), SBP (mmHg) and a congestion index, defined as the sum of peripheral oedema, jugular vein distension, pulmonary rales and pleural effusion, ranging from 0 to 4. All parameters were obtained in the first hour of hospital admission, thus ensuring that the score reflects early clinical presentation, before any therapeutic intervention.

All continuous variables were evaluated for distributional properties. Lactate, sodium and SBP were tested using restricted cubic splines and locally weighted scatterplot smoothing (LOWESS) to verify linearity, with lactate and sodium retained as continuous variables. In these analyses, mortality rose steeply when SBP decreased below approximately 100 mmHg, while differences above this threshold were minimal. Therefore, to preserve interpretability and to facilitate bedside application, we dichotomized SBP at 100 mmHg, defining it as a binary variable: SBP < 100 mmHg = 1, SBP ≥ 100 mmHg = 0. The four signs of congestion represent overlapping clinical manifestations of the same pathophysiological process. Including each sign as a binary variable in the score would have introduced multicollinearity and inflated standard errors. To reflect the overall burden of congestion and to preserve statistical robustness, these parameters were integrated into a composite congestion index (range 0-4), representing the number of positive findings. The index was analysed as an ordinal variable, with mortality increasing progressively with higher index values. The LOWESS analysis confirmed an approximately linear relationship between the index and the log-odds of death. NT-proBNP was log10-transformed to account for its right-skewed distribution and wide dynamic range. Thus, treating the congestion index as a single ordinal predictor maintained its additive prognostic value while minimizing collinearity and model overfitting.

All five variables were entered into a multivariable logistic regression model to determine their independent association with in-hospital mortality. The model was fitted using penalized logistic regression with L2 regularization (ridge regression) to reduce coefficient variance and prevent overfitting because the number of events was limited relative to the number of predictors. Discriminative performance was evaluated using receiver operating characteristic (ROC) curve analysis and calculation of the area under the curve (AUC) with 95% confidence intervals obtained by 1000 bootstrap resamples. Model calibration was evaluated using a calibration plot, comparing observed event rates with model-predicted probabilities and by computing the Brier score. Decision curve analysis was applied to assess the clinical net benefit of the multivariable model.

The multivariable model was subsequently transformed into a point-based score (“bedside model”) to enhance bedside applicability, following the Framingham methodology[28]. To preserve the relative contribution of each predictor to the overall risk, each regression coefficient (b) was rescaled to an integer number of points using a constant of 0.5 log-odds per point. For continuous variables, clinically meaningful cut-offs were selected based on ROC-derived thresholds, ensuring a monotonic increase in risk across strata. Binary variables contributed points directly according to their coefficients. The FAST-HF score was calculated by summing all individual points. Its discriminative performance and calibration were compared to those of the multivariable model to verify that the transformation into a simplified point-based score did not compromise its predictive accuracy. Discriminative performance of the bedside model was evaluated using the AUC with 95% bootstrap confidence intervals. Its AUC was compared to that of the multivariable model using the DeLong test. Calibration of the bedside FAST-HF score was evaluated using a calibration plot and the Brier score. Clinical utility was evaluated using decision curve analysis. The distribution of the bedside score was evaluated using histograms.

EHMRG and GWTG-HF, established risk scores for AHF, were calculated for comparative purposes. Due to cohort-specific characteristics, metolazone use was coded as absent for the calculation of the EHMRG score, as this medication is not available in Romania. Similarly, all patients were classified as Caucasians for the race component of the GWTG-HF score. Risk scores requiring variables not routinely available in retrospective records, such as the Barthel index for MEESSI-AHF or the 3-minute walk test for the Ottawa Heart Failure Risk Scale, could not be reliably calculated and were therefore not included in comparative analyses. The discriminative performance of FAST-HF was compared with EHMRG and GWTG-HF using ROC curve analysis and AUC values. Differences in AUCs were evaluated using bootstrap-based comparisons with 2000 resamples.

RESULTS

Baseline characteristics

A total of 123 patients with AHF were included and divided into two groups: 20 patients (16.3%) died during hospitalization and 103 patients (83.7%) survived to discharge. Baseline characteristics of the two groups are summarized in Table 1. The mean age of the study population was 71.34 ± 13.98 years. Patients had three clinical forms of presentation: Acute decompensated HF (ADHF) (73.98%), acute pulmonary oedema (APE) (17.07%) and cardiogenic shock (8.94%). Non-survivors were older than survivors (76.4 ± 10.85 years vs 70.3 ± 14.34 years, P = 0.039). On clinical examination, pleural effusion (55% vs 15.53%, P < 0.001), jugular vein distension (75% vs 39.8%, P = 0.0038) and the presence of pulmonary rales (85% vs 54.36%, P = 0.011) differed significantly between the two groups. LVEF was similar between survivors and non-survivors (41.15% ± 16.58% vs 41.50% ± 18.23%, P = 0.934), indicating no significant difference in baseline systolic function between the two groups. NT-proBNP levels were non-normally distributed. Data are shown as mean ± SD in original units for consistency and descriptive purposes, but P values were computed using Welch’s t-test, with values significantly higher in patients who died in hospital (17770.8 ± 11093.44 pg/mL vs 7569.66 ± 8565.913 pg/mL, P = 0.001). Arterial blood gas analysis showed a lower PaO₂ (50.95 ± 15.22 mmHg vs 68.59 ± 31.5 mmHg, P < 0.001) and higher lactate levels (3.35 ± 1.83 mmol/L vs 2.11 ± 1.47 mmol/L, P = 0.009). Overall, in this cohort, markers of congestion, neurohormonal activation (NT-proBNP) and markers of hypoperfusion (lower pH, elevated lactate) were the strongest predictors of in-hospital mortality.

Table 1 Baseline characteristics of patients with acute heart failure stratified by in-hospital mortality, n (%)/mean ± SD.

Variable	Patients (n = 123)	Survivors (n = 103)	Non-survivors (n = 20)	P value
Age (years)	71.34 ± 13.98	70.3 ± 14.34	76.4 ± 10.85	0.039
Male sex	62 (50.4)	52 (50.4)	10 (50)	0.968
Smoking	28 (22.76)	26 (25.24)	2 (10)	0.137
Hypertension	107 (86.99)	91 (88.34)	16 (80)	0.31
Obesity	49 (39.83)	37 (35.92)	12 (60)	0.044
Dyslipidaemia	113 (91.86)	98 (95.14)	15 (75)	0.003
Diabetes mellitus	49 (39.83)	40 (38.83)	9 (45)	0.606
AF/atrial flutter	56 (45.52)	46 (44.66)	10 (50)	0.661
ADHF	100 (81.3)	86 (83.49)	14 (70)	0.157
APE	23 (18.69)	17 (16.51)	6 (30)	0.157
Cardiogenic shock	11 (8.94)	3 (2.91)	8 (40)	< 0.001
Leg oedema	73 (59.34)	58 (56.31)	15 (75)	0.119
Pleural effusion	27 (21.95)	16 (15.53)	11 (55)	< 0.001
Ascites	5 (4.06)	3 (2.91)	2 (10)	0.142
Pulmonary rales	73 (59.34)	56 (54.36)	17 (85)	0.011
Jugular vein distension	56 (45.53)	41 (39.81)	15 (75)	0.003
SaO2	93.2 ± 5.24	93.94 ± 4.16	89.8 ± 8.24	0.04
SBP (mmHg)	142.59 ± 29.54	144.33 ± 27.43	133.6 ± 38.226	0.244
DBP (mmHg)	85.83 ± 16.68	87.84 ± 15.747	75.45 ± 17.91	0.002
Heart rate (bpm)	100.98 ± 28.12	101.5 ± 29.18	98.3 ± 22.32	0.644
LVEF (%)	41.21 ± 16.78	41.15 ± 16.58	41.5 ± 18.23	0.934
Haemoglobin (g/dL)	12.78 ± 2.21	12.87 ± 2.22	12.35 ± 2.19	0.338
eGFR (mL/minute/1.73 m2)	62.98 ± 22.72	64.86 ± 22.23	53.3 ± 23.33	0.037
Blood urea (mg/dL)	56.94 ± 28.51	54.08 ± 27.77	71.38 ± 28.48	0.013
Serum sodium (mmol/L)	137.12 ± 4.6	137.54 ± 4.1	134.95 ± 6.56	0.103
Serum potassium (mmol/L)	4.31 ± 0.73	4.32 ± 0.73	4.25 ± 0.79	0.682
NT-proBNP (pg/mL)	9228.3 ± 9736.4	7569.6 ± 8565.9	17770.8 ± 11093.4	0.001
pH	7.38 ± 0.07	7.39 ± 0.07	7.34 ± 0.06	0.002
PaO2 (mmHg)	65.72 ± 30.14	68.59 ± 31.5	50.95 ± 15.22	< 0.001
PaCO2 (mmHg)	33.47 ± 8.48	32.69 ± 8.54	37.5 ± 7.07	0.011
Lactate (mmol/L)	2.31 ± 1.59	2.11 ± 1.47	3.35 ± 1.83	0.009

AF: Atrial fibrillation; ADHF: Acute decompensated heart failure; APE: Acute pulmonary oedema; SaO₂: Oxygen saturation; BP: Blood pressure; bpm: Beats per minute; LVEF: Left ventricular ejection fraction; eGFR: Estimated glomerular filtration rate; NT-proBNP: N-terminal prohormone of brain natriuretic peptide; PaO₂: Partial pressure of oxygen in arterial blood; PaCO₂: Partial pressure of carbon dioxide in arterial blood; SBP: Systolic blood pressure; DBP: Diastolic blood pressure.

Development of the FAST-HF multivariable model

All five admission parameters (log-transformed NT-proBNP, arterial lactate, serum sodium, low SBP and the congestion index) included in the multivariable penalized logistic regression were independently associated with in-hospital mortality. High NT-proBNP and lactate levels, low serum sodium, hypotension and higher congestion index values were all associated with increased odds of in-hospital mortality (Table 2).

Table 2 Penalized multivariable logistic regression model for in-hospital mortality.

Variable	β-coefficient	OR	95%CI	P value
Log (NT-proBNP)	0.74	2.10	1.30-3.42	0.002
Lactate (mmol/L)	0.42	1.52	1.10-2.09	0.011
Serum sodium (per mmol/L decrease)	-0.15	0.86	0.77-0.96	0.007
SBP < 100 mmHg (yes vs no)	0.91	2.48	1.18-5.20	0.017
Congestion index (0-4)	0.55	1.73	1.22-2.45	0.002

NT-proBNP: N-terminal prohormone of brain natriuretic peptide; SBP: Systolic blood pressure; OR: Odds ratio.

NT-proBNP was log-transformed prior to analysis to account for its highly skewed distribution and wide biological range, ensuring a linear relationship with the log-odds of mortality. Higher log(NT-proBNP) levels were independently associated with in-hospital mortality [odds ratio (OR): 2.10, 95%CI: 1.30-3.42, P = 0.002]. Arterial lactate was retained as a continuous variable as its relationship with the log-odds of mortality was approximately linear, with higher lactate levels independently associated with an increased risk of in-hospital mortality (OR: 1.52, 95%CI: 1.10-2.09, P = 0.011). Serum sodium was analysed as a continuous variable expressed per 1 mmol/L. It demonstrated an inverse association with in-hospital mortality (OR: 0.86, 95%CI: 0.77-0.96, P = 0.007), indicating that lower sodium levels were linked with higher in-hospital mortality. SBP was analysed as a binary variable, with hypotension (SBP < 100 mmHg) being an independent predictor of in-hospital mortality (OR: 2.48, 95%CI: 1.18-5.20, P = 0.017). The congestion index, summing peripheral oedema, jugular vein distension, pulmonary rales and pleural effusion, was treated as an ordinal variable. Each additional sign of congestion was independently associated with higher in-hospital mortality risk (OR: 1.73, 95%CI: 1.22-2.45, P = 0.002).

Model performance and internal validation

The FAST-HF multivariable model demonstrated excellent discriminative performance for predicting in-hospital mortality. The ROC curve analysis yielded an AUC of 0.889 (95%CI: 0.861-0.894), indicating a high ability to distinguish between survivors and non-survivors (Figure 1). Internal validation using 1000 bootstrap resamples yielded a corrected AUC of 0.874, confirming the model’s stability and low degree of overfitting.

Figure 1 Receiver operating characteristic curve of the Fast Assessment Score for triage in heart failure multivariable model for predicting in-hospital mortality. The model showed excellent discriminative performance for predicting in-hospital mortality, with an area under the curve of 0.889 (95%CI: 0.861-0.894). AUC: Area under the curve.

Calibration analysis showed good agreement between predicted and observed mortality probabilities (Figure 2A), with a calibration slope of 0.97 and an intercept close to zero (-0.04). The Brier score was 0.09, reflecting high overall accuracy of probability estimates. Visual inspection of the calibration plot demonstrated near-perfect alignment between observed and predicted risk across deciles of predicted probability, with only minimal deviation in the highest risk stratum.

Figure 2 Calibration plot. A: The Fast Assessment Score for triage in heart failure (FAST-HF) multivariable model for in-hospital mortality prediction. Calibration analysis showed acceptable agreement between predicted and observed mortality probabilities; B: The FAST-HF bedside score for in-hospital mortality prediction. The bedside FAST-HF score showed acceptable calibration, with observed mortality rates generally.

Decision curve analysis demonstrated that the FAST-HF multivariable model provided a greater net clinical benefit than both default strategies (“treat all” and “treat none”) across a clinically relevant range of threshold probabilities (approximately 5%-30%). Within this interval, using FAST-HF to guide early high-risk decision making would result in more correctly identified patients at risk for in-hospital mortality without generating excessive false positives (Figure 3A).

Figure 3 Decision curve analysis. A: The Fast Assessment Score for triage in heart failure (FAST-HF) multivariable model for in-hospital mortality. Decision curve analysis demonstrated that the FAST-HF multivariable model provided a greater net benefit than both default strategies (“treat all” and “treat none”) across a clinically relevant range of threshold probabilities; B: The FAST-HF bedside score for in-hospital mortality prediction. The bedside score showed a positive net benefit across the evaluated threshold probabilities, suggesting potential clinical usefulness compared with treat-all and treat-none strategies. FAST-HF: Fast Assessment Score for triage in heart failure.

The distribution of predicted risk showed a clear separation between survivors and non-survivors (Figure 4). Survivors were predominantly concentrated at low predicted risk values, while non-survivors were shifted towards higher predicted probabilities, supporting the strong discriminative ability of the FAST-HF multivariable model across the full spectrum of risk.

Figure 4 Distribution of Fast Assessment Score for triage in heart failure-predicted in-hospital mortality probabilities according to survival status. Non-survivors showed higher predicted mortality probabilities than survivors, supporting the model’s ability to separate risk groups.

Derivation of the FAST-HF bedside point score

The FAST-HF multivariable model was transformed into a simplified point score using a Framingham-type approach to facilitate bedside application. Each of the five variables from the penalized logistic regression was categorized into clinically meaningful strata guided by their distribution and Youden’s index. For NT-proBNP, lactate and serum sodium, thresholds were chosen to reflect high-risk values, while SBP was retained as a binary variable and the congestion index was grouped into increasing levels of congestion burden (Table 3). The total FAST-HF bedside score is obtained by summing all points, with a maximum of 36 points.

Table 3 Fast Assessment Score for triage in heart failure bedside score: Predictors, categories and assigned points.

Predictor	Category	Points
NT-proBNP (pg/mL)	< 3000	0
	3000-8000	3
	> 8000	5
Lactate (mmol/L)	< 2	0
	2-4	2
	> 4	3
Serum sodium (mmol/L)	≥ 135	0
	130-134	2
	125-129	4
	< 125	6
SBP (mmHg)	≥ 100	0
	< 100	6
Congestion index (0-4)	0	0
	1	4
	2	8
	3	12
	4	16
	Total	36 points

NT-proBNP: N-terminal prohormone of brain natriuretic peptide; SBP: Systolic blood pressure.

The continuous log-transformed NT-proBNP was categorized into three clinically interpretable strata, using Youden’s index, which revealed two clinically relevant inflection points around 3000 pg/mL and 8000 pg/mL. For each category, points were assigned proportionally to the effect size observed in the continuous model. Following a Framingham-type approach, the b-coefficient (0.74) was divided by the smallest non-zero b in the model (approximately = 0.15), yielding approximately 4.9 b-units. This indicates that the highest NT-proBNP category (> 8000 pg/mL) should be assigned 5 points in the bedside score. The intermediate category (3000-8000 pg/mL) corresponds to approximately half of the effect size on the log scale and was assigned 3 points, while the reference category (< 3000 pg/mL) received 0 points.

Serum lactate was transformed into a three-level categorical predictor based on ROC analysis using Youden’s index, which established discriminatory thresholds at approximately 2 mmol/L and a long right-sided tail above 4 mmol/L. To assign points, the b-coefficient (0.42) was divided by 0.15, which yielded approximately 2.8 b-units, indicating that the highest category (> 4 mmol/L) should be assigned 3 points. Lactate levels between 2-4 mmol/L were assigned 2 points, while levels < 2 mmol/L served as the reference category with 0 points.

Serum sodium was converted into a four-level categorical variable based on the inspection of the cohort distribution and ROC analysis using Youden’s index, with thresholds selected at 135 mmol/L, 130 mmol/L and 125 mmol/L. Point values were derived from the continuous model coefficient (b = -0.15 per mmol/L decrease), where each 1 mmol/L reduction corresponds to approximately one b-unit. To preserve interpretability, sodium was grouped into clinically meaningful categories, with points assigned proportionally to the expected b-unit within each range: 0 points for 135 mmol/L, 2 points for 130-134 mmol/L, 4 points for 125-129 mmol/L and 6 points for < 125 mmol/L.

For SBP, both distribution within the cohort and ROC analysis confirmed a sharp inflection in mortality risk around 100 mmHg. In the continuous multivariable model, SBP < 100 mmHg had a b-coefficient of 0.91, corresponding to approximately 6 b-units when divided by 0.15. Accordingly, SBP < 100 mmHg was assigned 6 points, while values ≥ 100 mmHg received 0 points.

ROC analysis of the congestion index confirmed that treating the index as an ordinal variable preserved the strongest discriminatory performance. In the multivariable model, each 1-point increase corresponded to a b-coefficient of 0.55, equivalent to approximately 4 b-units when divided by 0.15. Consequently, each additional congestion sign was assigned 4 points in the bedside score: 0 points for an index of 0, 4 points for an index of 1, 8 points for an index of 2, 12 points for an index of 3 and 16 points for an index of 4.

Comparative performance of the FAST-HF multivariable and bedside point score

When translated into the integer point-based bedside score, the FAST-HF model preserved most of the discriminative performance of the multivariable model, with the bedside score achieving an AUC of 0.867 compared to an AUC of 0.889 obtained by the multivariable model (Figure 5A). No statistically significant difference in AUC was observed between the two models based on the DeLong test (z = 0.728, P = 0.467). The bedside score closely followed the ROC trajectory of the multivariable model, with minimal loss of prognostic information after categorization and point conversion, confirming that the simplified version retains excellent discrimination and is suitable for bedside application.

Figure 5 Comparison of receiver operating characteristic curves. A: For the Fast Assessment Score for triage in heart failure (FAST-HF) multivariable model and the bedside point score. The bedside Fast Assessment Score for triage in heart failure score showed discrimination close to that of the continuous multivariable model, with area under the curve (AUC) values of 0.867 and 0.889, respectively; B: FAST-HF, Emergency Heart Failure Mortality Risk Grade and Get with the Guidelines-heart failure for predicting in-hospital mortality. FAST-HF showed higher discriminatory performance than the established scores in this cohort, with AUC values of 0.867, 0.720 and 0.619, respectively. FAST-HF: Fast Assessment Score for triage in heart failure; EHMRG: Emergency Heart Failure Mortality Risk Grade; GWTG-HF: Get with the Guidelines-heart failure; AUC: Area under the curve.

Calibration of the FAST-HF bedside model was acceptable, with preserved risk gradation across deciles despite expected attenuation compared with the multivariable model (Figure 2B). While calibration slope and intercept were lower for the bedside score (slope 0.266, intercept -5.70) than for the multivariable model (0.97 and -0.04, respectively), overall accuracy remained nearly identical, supported by a similar Brier score (0.0896 vs 0.090).

The FAST-HF bedside score demonstrated a positive and clinically meaningful net benefit across threshold probabilities of 5%-30%, outperforming both default strategies of treating all or treating none (Figure 3B). Although its net benefit was slightly lower than that of the multivariable model, the bedside score closely approximates its performance throughout the clinically relevant range.

Risk stratification and score distribution

Using the cut-off value identified by Youden’s index at 15 points, the FAST-HF bedside score effectively stratified patients into three risk groups with different mortality rates (Table 4). Mortality was 0.0% in the low-risk group (0-8 points), while in the intermediate-risk group (9-14 points) it increased to 7.5%, reaching 38.6% in the high-risk group (≥ 15 points) (Figure 6).

Figure 6 In-hospital mortality across Fast Assessment Score for triage in heart failure bedside risk groups. Patients were classified as low-risk (0-8 points), intermediate-risk (9-14 points) and high-risk group (≥ 15 points). Mortality increased progressively across categories. FAST-HF: Fast Assessment Score for triage in heart failure.

Table 4 Observed in-hospital mortality across bedside score risk categories.

Risk group	FAST-HF Score range	n	Mortality (%)
Low risk	0-8	39	0.0
Intermediate risk	9-14	40	7.5
High risk	≥ 15	44	38.6

FAST-HF: Fast Assessment Score for triage in heart failure.

The distribution of the FAST-HF bedside scores in the study cohort was right-skewed, with most patients clustering between 4 and 18 points (Figure 7). A smaller proportion had scores above 20 points, while only a few patients achieved higher values (> 25 points).

Figure 7 Distribution of the Fast Assessment Score for triage in heart failure bedside score in the study cohort. The histogram showed a broad distribution among patients with acute heart failure. FAST-HF: Fast Assessment Score for Triage in heart failure.

Comparative discriminative performance of FAST-HF and established risk scores

ROC curve analysis was used to compare the discriminative performance of FAST-HF with EHMRG and GWTG-HF, both established prognostic scores for AHF (Table 5). FAST-HF demonstrated excellent discrimination for in-hospital mortality, achieving an AUC of 0.867 (95%CI: 0.784-0.936, P < 0.001), compared with an AUC of 0.720 for EHMRG (95%CI: 0.571-0.856; P = 0.012). The GWTG-HF score demonstrated limited discrimination, with an AUC of 0.619 (95%CI: 0.444-0.776; P = 0.085) compared to FAST-HF (Figure 5B).

Table 5 Comparison of prognostic score performance for in-hospital mortality.

Score	AUC	95%CI	P value
FAST-HF	0.867	0.784-0.936	< 0.001
EHMRG	0.720	0.571-0.856	0.012
GWTG-HF	0.619	0.444-0.776	0.085

FAST-HF: Fast Assessment Score for Triage in Heart Failure; EHMRG: Emergency Heart Failure Mortality Risk Grade; GWTG-HF: Get with the Guidelines-Heart Failure; AUC: Area under the curve.

Bootstrap-based comparisons confirmed that FAST-HF had higher discriminative performance than EHMRG (mean AUC difference +0.15, 95%CI: 0.02-0.28; P = 0.014) and GWTG-HF (mean AUC difference +0.25, 95%CI: 0.08-0.42; P < 0.001).

DISCUSSION

Principal findings

This study reports the development and internal validation of the FAST-HF score, a rapid bedside tool for predicting in-hospital mortality in patients with AHF using readily available admission variables. The score was derived from a penalized multivariable logistic regression model and subsequently simplified into a point-based system without a statistically significant loss of discriminative performance. The FAST-HF bedside score demonstrated excellent discrimination, good calibration and meaningful clinical utility, allowing early identification of patients at risk of in-hospital mortality.

Clinical meaning of individual predictors

The FAST-HF score combines predictors reflecting distinct but complementary pathophysiological dimensions of AHF. NT-proBNP reflects neurohormonal activation and myocardial wall stress and is consistently associated with short-term mortality in AHF. Serum lactate is a sensitive marker of systemic hypoperfusion and metabolic stress, identifying patients with impaired circulatory reserve even in the absence of overt hypotension. Hyponatremia represents a well-established marker of advanced HF and adverse prognosis. Hemodynamic instability was defined as hypotension (SBP < 100 mmHg), a clinically intuitive and guideline-consistent threshold associated with impaired cardiac output and poor outcomes. Finally, the congestion index was designed to capture the cumulative burden of clinical congestion by integrating peripheral oedema, jugular vein distension, pulmonary rales and pleural effusion into a single ordinal measure, rather than relying on individual physical examination findings in isolation. Together, these predictors provide a rapid, bedside-accessible assessment of cardiac stress, perfusion, neurohormonal activation and volume overload.

From multivariable model to bedside score

An important strength of the FAST-HF score is the rigorous methodology used to translate a continuous multivariable model into a simplified bedside score. By applying a Framingham-type scaling approach to the regression coefficients, the relative prognostic weight of each predictor was preserved while ensuring clinical interpretability. Despite categorization and point assignment, the bedside score maintained discrimination comparable to that of the multivariable model, as confirmed by a non-significant DeLong test. These findings indicate that the FAST-HF bedside score retains the core prognostic information of the multivariable model while substantially improving feasibility for real-time clinical use.

Risk stratification based on the FAST-HF bedside score revealed a pronounced stepwise increase in mortality across predefined risk groups. Notably, the Youden-derived cut-off of 15 points accurately identified a high-risk subgroup with elevated in-hospital mortality, while no deaths occurred in the low-risk group. This clear separation supports the clinical utility of the score for early triage and escalation of care, particularly in ED and intensive care settings.

Clinical implications

The FAST-HF bedside score has several important clinical implications for the early management of patients presenting with AHF. Because all included variables are readily available at admission, the score can be calculated within minutes of presentation, enabling rapid risk stratification in the ED.

Conversely, the identification of a low-risk subgroup with a FAST-HF score ≤ 8 points and no observed in-hospital mortality suggests potential utility in supporting early de-escalation strategies, such as standard ward management or early discharge, when clinically appropriate. The intermediate-risk group may warrant tailored management and repeated reassessment, reflecting the dynamic nature of AHF.

The FAST-HF score is not intended to replace clinical judgement but to complement it by providing an objective estimate of short-term risk. Its simplicity and rapid bedside applicability make it particularly suitable for use in the ED, where timely decisions are critical and complex risk models are often impractical.

Comparison with existing risk scores

Several prognostic scores have been developed in recent years to predict short-term mortality in AHF[29]. While these tools have demonstrated good prognostic performance, many rely on variables that are not immediately available at the time of presentation or require complex calculations, limiting their routine use in time-critical settings. In contrast, the FAST-HF score was specifically designed to prioritize early bedside applicability, using parameters that can be obtained within minutes of admission, including arterial blood gas analysis and focused clinical assessment.

Compared to EHMRG and GWTG-HF, which incorporate a larger number of clinical and laboratory variables and were primarily derived from large administrative and registry-based cohorts, FAST-HF emphasizes parsimony and immediacy[15,21]. The inclusion of arterial lactate represents a distinctive feature of FAST-HF, capturing early systemic hypoperfusion that may not be reflected by blood pressure alone and is not routinely included in other risk scores. In this cohort, this design was associated with higher discriminative performance of FAST-HF compared with both EHMRG and GWTG-HF, supporting the clinical relevance of integrating markers of hypoperfusion and congestion into early risk stratification.

While congestion scores reflect the central pathophysiological mechanism of decompensation, they do not fully capture the heterogeneity of early mortality risk[22]. FAST-HF builds on this concept by integrating congestion with markers of hypoperfusion and hemodynamic compromise to provide a broader, bedside-applicable prognostic assessment.

Importantly, despite its simplicity, the FAST-HF bedside score demonstrated discriminative performance comparable to that of the underlying multivariable model and showed clear risk stratification across clinically meaningful categories. These findings suggest that it offers a complementary approach to existing risk scores, particularly in acute care environments where rapid decision-making is essential and comprehensive risk models may be impractical.

Strengths and limitations

The FAST-HF score was developed using a rigorous two-step modelling strategy that combined penalized logistic regression with L2 regularization and transparent point-based simplification, ensuring both statistical robustness and bedside applicability. Given the limited number of in-hospital deaths, Ridge regression was particularly valuable in reducing coefficient instability and the risk of overfitting through coefficient shrinkage and variance reduction. Together with bootstrap-based internal validation, this approach improved model stability despite the relatively low number of outcome events. All variables included are rapidly available at presentation, enabling early risk stratification in acute care settings.

Although the five FAST-HF predictors are pathophysiologically interconnected, they represent complementary dimensions of AHF severity. NT-proBNP reflects myocardial wall stress and neurohormonal activation, arterial lactate reflects tissue hypoperfusion, serum sodium reflects neurohormonal and fluid imbalance, SBP reflects hemodynamic instability and the congestion index reflects the clinical burden of fluid overload. In this context, L2 regularization helps reduce instability related to correlated predictors through coefficient shrinkage, limiting the disproportionate influence of individual variables while allowing them to remain jointly represented in the model. Thus, FAST-HF integrates partially overlapping but clinically distinct prognostic domains into a multidimensional bedside assessment rather than simply duplicating the same clinical sign.

Several limitations should also be acknowledged. This was a single-centre study with a relatively limited sample size, and external validation in independent cohorts is required before broader implementation. Although internal validation was comprehensive, residual overfitting cannot be entirely excluded. In addition, arterial blood gas analysis may not be routinely performed in all patients with AHF, potentially limiting generalizability. Finally, the observational design precludes assessment of whether FAST-HF-guided management strategies improve clinical outcomes.

CONCLUSION

The FAST-HF score is a simple, rapid and clinically grounded bedside tool for predicting in-hospital mortality in AHF. Derived from a robust multivariable model and internally validated, the score preserves strong discriminative performance and provides meaningful risk stratification using readily available admission parameters. Further external validation in independent cohorts is warranted to confirm its generalizability and potential clinical impact.