Refining risk prediction in geriatric heart failure: the emergency heart failure mortality risk grade‑BI model under scrutiny Authors: Kianoush A. h. Saberi L., Durdona Alimova, Sherzod Niyazov, Kholmurod Qayumov, Parvina Gaffarova, Xamida Juraqobilova, Zaynura Kadirova Published: 2026 ABSTRACT Risk stratification is critical for managing acute heart failure in older adults, yet existing tools often underperform in geriatric populations. Özdemir recently introduced the EHMRG‑BI model, integrating brain natriuretic peptide and endotracheal intubation with the Emergency Heart Failure Mortality Risk Grade score to improve 30‑day mortality prediction. While the model demonstrates improved discrimination in a single‑center derivation cohort, several methodological considerations warrant attention. The inclusion of intubation, a treatment decision rather than a baseline characteristic, may introduce bias and limit its utility as an arrival‑time tool. The absence of key geriatric domains such as frailty, functional status, and cognitive impairment may result in residual confounding. Additionally, modelling BNP as a linear predictor disregards its well‑documented non‑linear association with mortality. This Matters Arising offers a perspective on these challenges and proposes refinements, including external validation, integration of geriatric assessments, and flexible BNP modelling, to strengthen the model’s clinical applicability. To the Editor, We read with great interest the article by Özdemir, which introduces the EHMRG‑BI model for predicting 30‑day mortality in older adults with acute heart failure [1]. While the model demonstrates improved discrimination over the original EHMRG score, three methodological considerations warrant attention: the inclusion of intubation, the absence of key geriatric variables, and the linear specification of BNP despite its non‑linear association with mortality. We offer a perspective on these challenges and suggest directions for future refinement. A treatment‑dependent predictor The EHMRG‑BI model incorporates endotracheal intubation as a predictor, yet intubation is a treatment decision made during the emergency department course, not a baseline characteristic. This creates a fundamental paradox: a variable reflecting clinical judgment and institutional practice is used to refine mortality prediction, embedding decision‑making bias rather than capturing intrinsic patient risk. In the present study, intubation was independently associated with 30‑day mortality (OR 5.576, 95% CI 2.496–12.457), contributing to the apparent AUC improvement from 0.681 to 0.777 [1]. However, because intubation decisions are influenced by physician preference, non‑invasive ventilation availability, ICU admission thresholds, and advance directives, the model may overestimate risk in settings with lower intubation thresholds and underestimate it where non‑invasive strategies are prioritised. Unlike the EHMRG components, intubation is an active intervention that may itself alter disease trajectory [2]. In sensitivity analyses excluding intubation, the reduced EHMRG + BNP model retained significantly better discrimination than EHMRG alone (AUC 0.718 vs. 0.681; p = 0.010), suggesting BNP contributes genuine prognostic value independent of intubation [1]. Yet the full model’s incremental improvement (ΔAUC 0.060 over EHMRG + BNP) may be inflated by the very treatment decisions it seeks to predict. External validation studies highlight context‑specific performance. In a Spanish cohort of 1,553 patients, the EHMRG score demonstrated a c‑statistic of 0.741 for 7‑day mortality, lower than the 0.807 and 0.804 reported in Canadian cohorts [3]. In a separate external validation of 237 patients, the EHMRG score showed good discrimination for 30‑day mortality when physicians were blinded to the scores, suggesting performance is robust when not influencing clinical decisions [4]. However, disposition decisions were made based on standard institutional protocols rather than EHMRG scores, reinforcing that treatment decisions and prognostic models operate in parallel. A model incorporating an ED‑course intervention cannot serve as a purely arrival‑time risk stratification tool. Its utility depends on the clinician having already decided whether intubation is needed a decision the model might ideally inform rather than reflect. Future research should evaluate whether EHMRG‑BI performance can be replicated using only predictors available at initial presentation and whether its utility is preserved across institutions with differing intubation practices. Prospective external validation is essential to determine whether the added discrimination attributed to intubation represents genuine prognostic information or an artefact of treatment variability. Frailty, function, and cognitive status A critical limitation of the EHMRG‑BI model is the absence of key geriatric variables, frailty, functional status, cognitive impairment, and nutritional status, despite their established prognostic value in older adults with acute heart failure. The study cohort had a median age of 79 years [1], yet none of these domains were incorporated into the prediction model. Frailty is independently associated with mortality in older patients with heart failure. In the FRAGIC study of 499 outpatients aged ≥ 75 years with chronic HF, frailty identified by the visual mobility scale was independently associated with 1‑year mortality (HR 2.13; 95% CI 1.08–4.20) after adjustment for confounding variables [5]. Similarly, the Clinical Frailty Scale (CFS) has demonstrated robust prognostic value; among 190 hospitalized HF patients aged ≥ 75 years, high CFS scores were independently associated with 1‑year all‑cause mortality, with cumulative incidence rising from 6.3% in low‑CFS patients to 30.5% in high‑CFS patients (P < 0.001), and CFS showed a strong correlation with the Barthel Index and moderate correlation with the MMSE for cognitive function [6]. A meta‑analysis of the FRAGILE‑HF cohort further confirmed that frailty defined by the FRAIL scale (HR 1.31; 95% CI 1.01–1.71) and Fried phenotype (HR 1.35; 95% CI 1.04–1.75) independently predicted 2‑year all‑cause mortality [7]. Functional impairment and disability are equally important. The Barthel Index for activities of daily living contributes additional prognostic value to HF risk models in elderly patients with acute HF [8]. Furthermore, cognitive impairment, present in nearly half of older HF patients, is independently associated with mortality and may influence treatment decisions, including intubation and code status [6]. Nutritional status and polypharmacy also play key roles in detecting frailty in older patients with cardiovascular disease [9]. The absence of these variables in the EHMRG‑BI model may result in residual confounding. Patients with significant frailty or cognitive impairment may have different treatment trajectories and mortality risks that are not captured by physiological and laboratory parameters alone. We suggest that future iterations of the model incorporate validated frailty screening tools, functional assessments, and that prospective studies evaluate whether adding these domains improves risk stratification and clinical utility. The BNP non‑linearity The EHMRG‑BI model employs BNP as a linear predictor, yet the relationship between BNP and mortality in acute heart failure is non‑linear. Sensitivity analyses confirmed that restricted cubic spline modeling improved AUC from 0.777 to 0.7977, with evidence of nonlinearity (p = 0.0038) [1]. Despite this, the primary model retained a linear specification. Studies in stroke‑heart syndrome demonstrate that the BNP‑mortality association is best captured using turning‑point analysis, with risk thresholds beyond which mortality increases substantially [10]. Evidence from cardiac surgery also indicates a log‑linear relationship between natriuretic peptides and outcomes [11]. Given that BNP is one of only two additional predictors in the model, its functional form meaningfully influences performance. Future iterations should adopt non‑linear specifications, such as restricted cubic splines, to more accurately capture the BNP‑mortality relationship and improve risk stratification. Conclusion The EHMRG‑BI model offers a valuable step toward improving mortality prediction in older adults with acute heart failure, yet several methodological considerations warrant attention. The inclusion of endotracheal intubation, a treatment decision rather than a baseline characteristic, may introduce bias and limit its use as an arrival‑time prognostic tool. The absence of key geriatric domains such as frailty, functional status, and cognitive impairment may result in residual confounding and missed opportunities for risk refinement. Additionally, modelling BNP as a linear predictor disregards its well‑documented non‑linear association with mortality, potentially underutilizing the prognostic information available from this biomarker. Future research should priorities external validation across diverse settings, explore the incremental value of frailty and functional assessments, and adopt flexible modelling approaches for BNP to better capture its relationship with mortality. Until such refinements are completed, the EHMRG‑BI model should be regarded as a promising but preliminary tool.