Association between the HACOR score and mortality in patients with acute hypoxemic respiratory failure: a retrospective cohort study

Abstract

Background

Acute hypoxemic respiratory failure (AHRF) is a common critical condition in the intensive care unit (ICU). Accurate early prediction of the risk of death is of great significance for clinical decision-making. Although HACOR score was initially used to predict the failure of non-invasive ventilation in acute hypoxemic respiratory failure (AHRF) patients, the relationship between the HACOR score and the mortality remains unclear. Therefore, a retrospective cohort study was conducted to evaluate the ability of the HACOR score to predict mortality risk in patients with AHRF.

Method

Patients diagnosed with AHRF were extracted from the MIMIC-IV database and divided into two groups (low‑risk < 6 vs. high‑risk ≥ 6) based on the median HACOR score. The relationship between the HACOR score and all‑cause mortality at 28, 90, and 365 days was evaluated. To investigate the association between HACOR score and mortality in AHRF patients, restricted cubic spline (RCS) analysis and Cox proportional hazards regression were performed.

Result

This study included 2774 patients (median age 68 years, IQR 57–79; 57.9% male). The mortality rates in the high-risk group (HACOR ≥ 6) were significantly higher than those in the low-risk group (< 6) at 28 days, 90 days, and 365 days (28 days: 38.3% vs. 28.2%, P < 0.001; 90 days: 46.1% vs. 35.7%, P < 0.001; 365 days: 52.5% vs. 44.3%, P < 0.001). Kaplan-Meier analysis demonstrated that the survival probability of the high-risk group was lower (log-rank P < 0.001). Cox regression analysis indicated that for every 1-point increase in the HACOR score, the risk of death significantly increased (P < 0.001), and restricted cubic spline analysis revealed a linear association between the score and mortality, with a significant threshold effect at HACOR = 6. Subgroup analysis further showed that a HACOR score ≥ 6 was consistently associated with a 28-day mortality rate across various demographic and clinical characteristics (all P < 0.05), reinforcing the robustness of the score.

Conclusion

This study confirmed that an increase in the HACOR score was significantly associated with an increased mortality rate at 28 days, 90 days, and 365 days in patients with AHRF. For every 1-point increase in the score, the risk of death increased, and the threshold effect was significant when HACOR = 6. These findings highlight the utility of the HACOR score for risk stratification in clinical practice.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12890-026-04538-y.

Introduction

Acute hypoxemic respiratory failure (AHRF) is a prevalent and life-threatening condition necessitating admission to the intensive care unit (ICU). It is associated with high morbidity and mortality rates, placing substantial strain on critical care resources globally [1, 2]. For adult patients diagnosed with AHRF upon ICU entry, the early and accurate prediction of mortality risk is a pressing clinical challenge. A simple and reliable prognostic tool would be invaluable, enabling clinicians to identify high-risk individuals at admission, facilitating timely and aggressive interventions, optimizing resource allocation, and improving counseling for families.

Several general severity-of-illness scores, such as the Acute Physiology and Chronic Health Evaluation (APACHE) II score and the Sequential Organ Failure Assessment (SOFA) score, are routinely used for risk stratification in critically ill patients [3–5]. However, their application in specifically predicting mortality for AHRF is often limited. These scores are relatively complex, incorporate parameters that may not be immediately available, and are designed to assess global multiorgan dysfunction rather than to quantify the specific severity of the respiratory insult itself [6, 7]. Therefore, a more pragmatic and respiratory-focused bedside tool is desirable for this distinct population.

The HACOR score (heart rate, acidosis, consciousness, oxygenation, and respiratory rate) has emerged as a potential solution. This score is a simple, bedside tool composed of five readily obtainable physiological parameters initially developed to predict noninvasive ventilation failure in hypoxemic patients [8]. However, it is important to note that previous studies validating the HACOR score [9–11] have primarily focused on patients who were already receiving noninvasive respiratory support (NIV or HFNC), with the score calculated after the initiation of support rather than at ICU admission baseline. Moreover, those studies generally excluded patients who received invasive mechanical ventilation on admission. In contrast, our study calculated the HACOR score upon ICU admission and included all patients with AHRF, regardless of whether they ultimately received noninvasive or invasive mechanical ventilation. Its simplicity and use of real-time clinical data make it an attractive candidate for the rapid risk stratification of adults with AHRF at ICU admission. In addition to their primary aim of predicting NIV failure among patients on noninvasive support, these three studies also reported that the HACOR score was significantly associated with mortality [9–11].

However, the evidence regarding the utility of the HACOR score as a direct predictor of mortality in a broad cohort of adult patients with AHRF at ICU admission remains insufficient. Many existing studies have focused on its role in predicting noninvasive support failure or have been conducted in specific subpopulations [8–10, 12]. A comprehensive evaluation of its association with all-cause mortality in an unselected, general ICU population with AHRF is needed to confirm its broad prognostic value.

Therefore, we conducted this retrospective cohort study to determine the associations between the HACOR score at ICU admission and all-cause mortality at 28, 90, and 365 days in adults with AHRF. We hypothesized that a higher HACOR score is independently associated with increased mortality rates. The primary objective was to evaluate the predictive performance of the HACOR score for 28-day mortality, with 90-day and 365-day mortality assessed as key secondary outcomes.

Method

Database

Data for this study were derived from the Medical Information Mart for Intensive Care (MIMIC)-IV database (version 3.1), a large, freely accessible single-center database encompassing comprehensive and high-quality data of patients admitted to intensive care units (ICUs) at Beth Israel Deaconess Medical Center (BIDMC) from 2008 to 2022. The data has been de-identified and made publicly accessible, eliminating the requirement for specific informed permission from patients. One author (Lei LI) has completed the CITI examination and received permission to use the database for research purposes (certification number: 15145501). Data extraction was performed via Structured Query Language (SQL) via the PostgreSQL toolkit. All patient information within the MIMIC-IV database has been deidentified in compliance with approval from the Institutional Review Boards (IRBs) of the Massachusetts Institute of Technology (MIT) and BIDMC. All methods and procedures were conducted in accordance with the Declaration of Helsinki.

Study design and patients

We initially extracted data from patients who were identified with an International Classification of Diseases (ICD-9 or ICD-10) code for acute respiratory failure (ARF). The inclusion criteria were as follows: (1) age greater than 18 years and (2) meeting the physiological criteria for acute hypoxemic respiratory failure (AHRF), defined as a PaO₂/FiO₂ ratio < 300 mmHg with a PCO₂ ≤ 45 mmHg. To ensure data quality, we excluded the following patients: those with multiple ICU admissions (retaining only data from their initial admission), those lacking any record of critical physiological parameters (including heart rate, Glasgow Coma Scale score, arterial pH, and respiratory rate) within the first 24 h of ICU admission, and those with an ICU length of stay shorter than 24 h. The final study cohort comprised 2774 patients. Patients were stratified into two groups on the basis of the median HACOR score at ICU admission. Figure 1 shows the patient selection flowchart.

Fig. 1.

Flowchart of patient selection

Data extraction

Patient data were retrospectively extracted from the electronic medical records system via structured query language (SQL). The demographic variables included age, sex, and race. Comorbidities were assessed via the Charlson Comorbidity Index and included myocardial infarction, cerebrovascular disease, congestive heart failure, peripheral vascular disease, chronic pulmonary disease, diabetes mellitus, severe liver disease, chronic kidney disease, and malignancy. Disease severity was evaluated via the Sequential Organ Failure Assessment (SOFA) score, which are calculated within the first 24 h of intensive care unit (ICU) admission. Vital signs (heart rate, respiratory rate, mean arterial pressure, oxygen saturation, and the PaO₂/FiO₂ ratio) and laboratory parameters (white blood cell count, platelet count, serum albumin, electrolytes, glucose, liver and renal function tests, and coagulation profiles) were also recorded during the initial 24-hour period. The HACOR score was calculated as previously described [1]. Patient outcomes included ICU and hospital length of stay, as well as 28-day, 90-day, and 365-day mortality, with follow-up initiated from the date of ICU admission. To address potential bias from missing data, we implemented multiple imputation via the “mice” package in R for variables with < 20% missing values, while covariates exceeding this threshold were excluded from analyses to maintain model robustness.

Outcomes

The primary outcome was 28-day all-cause mortality following intensive care unit (ICU) admission, which was assessed as a measure of short-term prognosis. The secondary outcomes included 90-day and 365-day all-cause mortality, reflecting mid- and long-term prognoses, respectively.

Statistical analysis

Continuous variables are presented as medians with interquartile ranges (IQRs) and were compared via the Mann‒Whitney U test. Categorical variables are expressed as frequencies (percentages) and were analyzed via the chi-square test or Fisher’s exact test, as appropriate.

For survival analysis, we constructed Kaplan‒Meier curves to estimate cumulative mortality rates, with between-group comparisons performed via the log-rank test. Cox proportional hazards regression models were used to evaluate the associations between HACOR scores (analyzed as both continuous and categorical variables) and mortality outcomes at 28, 90, and 365 days, with the results expressed as hazard ratios (HRs) and 95% confidence intervals (CIs).

Potential covariates for multivariable Cox models were selected through a three-step approach: first, we considered variables with established clinical relevance; second, we assessed univariate associations (P < 0.10); and third, we evaluated multicollinearity via variance inflation factors (VIFs), excluding variables with a VIF ≥ 5. The final adjusted models included demographic characteristics (age, race), comorbidities (Charlson Comorbidity Index, malignant cancer, myocardial infarction, cerebrovascular disease, severe liver disease), laboratory parameters (white blood cell count, albumin, renal and hepatic function tests, electrolytes, coagulation profiles), and critical interventions (the use of vasopressors and invasive mechanical ventilation).

To explore potential nonlinear relationships, we employed restricted cubic splines with four knots to model the association between HACOR scores and mortality risk. Additionally, we conducted prespecified subgroup analyses stratified by age (≤ 65 vs. > 65 years), sex, and the use of vasopressors and ventilation support. Two-sided p values less than 0.05 were considered statistically significant. R software (version 4.3.2) and SPSS software (version 22.0) were used for statistical analysis.

Results

Population and baseline characteristics

A total of 19,674 patients with acute respiratory failure were screened from the MIMIC-IV database. On the basis of the inclusion and exclusion criteria stated above, 2774 patients were ultimately included in this study (Fig. 1). The mean age of the included patients was 68 (57, 79), and 57.9% of the patients were male. The median HACOR was 6 (IQR 3–9) in the cohort. Using the median HACOR score as the cutoff, patients were categorized into LOW (< 6) and HIGH (≥ 6) groups.

Table 1 shows the baseline characteristics of the two groups. Patients in the high-risk group (HACOR ≥ 6) were more likely to receive HFNC (4.8% vs. 2.9%, P = 0.014) and conventional oxygen therapy (31.6% vs. 24.5%, P < 0.001) but less likely to receive invasive ventilation (63.0% vs. 71.6%, P < 0.001) at ICU admission, reflecting differences in initial management strategies between the two groups. Both groups were similar in terms of age and sex distribution. Compared with patients with lower HACOR scores, patients with higher HACOR scores had higher admission SOFA scores, higher heart and respiratory rates, and lower MBP, SpO2, PH and PaO2/FiO2. Laboratory parameters, including white blood cell count (WBC), potassium, creatinine, aspartate aminotransferase (AST), alanine aminotransferase (ALT), total bilirubin, international normalized ratio (INR), were significantly elevated in the HIGH group compared with the LOW group. In contrast, the albumin, sodium, and calcium levels were significantly lower in the HIGH group. (all P < 0.05).

Table 1.

Demographic and baseline characteristics of patients with acute hypoxemic respiratory failure according to HACOR score

		Overall(N = 2774)	Low (N = 1324)	High (N = 1450)	P
Age		68(57, 79)	67(56, 78)	68(57, 79)	0.388
Gender
Female		1169(42,1%)	556(42.0%)	613(42.3%)	0.908
Male		1605(57.9%)	768(58.0%)	837(57.7%)
Race
White		1695(61.1%)	782(59.1%)	913(63.0%)	0.097
Black		276(9.9%)	128(9.7%)	148(10.2%)
Asian		93(3.4%)	47(3.5%)	46(3.2%)
Other		187(6.7%)	103(7.8%)	84(5.8%)
Unknown		523(18.9%)	264(20.0%)	259(17.9%)
Severity of the disease
SOFA		8(5, 11)	6(4, 9)	9(6, 12)	< 0.001
CCI		5(3, 8)	5(3, 8)	6(4, 8)	0.170
Comorbidities (%)
Myocardial infarct		587(21.2%)	267(20.2%)	320(22.1%)	0.220
Congestive heart failure		1001(36.1%)	488(36.9%)	513(35.4%)	0.418
Cerebrovascular disease		540(19.5%)	298(22.5%)	242(16.7%)	< 0.001
Chronic pulmonary disease		684(24.7%)	324(24.5%)	360(24.8%)	0.828
Renal disease		773(27.9%)	348(26.3%)	385(26.6%)	0.873
Severe liver disease		301(10.9%)	132(10.0%)	169(11.7%)	0.154
Diabetes		901(32.5%)	439(33.2%)	462(31.9%)	0.467
Malignant cancer		397(14.3%)	176(13.3%)	221(15.2%)	0.143
Vital signs
HR (beats/min)		88.3(76.3, 101.4)	85.9(75.2, 98.0)	90.5(77.5, 104.2)	< 0.001
MBP (mmHg)		75.4(70.0, 82.2)	76.1(71.0, 83.0)	74.6(69.4, 81.4)	< 0.001
RR (breaths/min)		20.6(18.0, 23.9)	19.7(17.4, 22.6)	21.6(18.8, 25.1)	< 0.001
SpO2 (%)		97.3(95.7, 98.6)	97.8(96.4, 98.9)	96.6(95.0, 98.1)	< 0.001
Laboratory tests
WBC (K/uL)		12.7(9.1, 17.5)	12.4(9.01, 16.7)	13.2(9.09, 18.3)	0.009
Albumin (g/dL)		3.0(2.6, 3.5)	3.1(2.6, 3.5)	3.0(2.5, 3.5)	0.002
Potassium (mEq/L)		4.2(3.8, 4.6)	4.1(3.8, 4.6)	4.2(3.9, 4.7)	0.005
Sodium (mEq/L)		139(136, 142)	139(136, 142)	139(135, 141)	< 0.001
Calcium (mg/dL)		8.2(7.7, 8.7)	8.2(7.8, 8.7)	8.1(7.6, 8.7)	0.029
Glucose (mg/dL)		139(114, 178)	138(115, 174)	140(114, 183)	0.262
Creatinine (mg/dL)		1.3(0.9, 2.2)	1.2(0.8, 1.9)	1.4(0.9, 2.3)	< 0.001
AST (IU/L)		52(28, 141)	47(27, 102)	57(30, 188)	< 0.001
ALT (IU/L)		34(18, 85)	32(18, 71)	36(19, 107)	< 0.001
Total bilirubin (mg/dL)		0.7(0.4,1.6)	0.7(0.4, 1.3)	0.8(0.5, 1.8)	< 0.001
INR		1.4(1.2, 1.7)	1.3(1.2, 1.6)	1.4(1.2, 1.9)	< 0.001
Arterial blood gas
PH		7.36(7.28, 7.41)	7.38(7.33, 7.42)	7.32(7.24, 7.40)	< 0.001
PaCO2, mmHg		39(35, 43)	39(36, 42)	39(35, 43)	0.21
PaO2/FiO2		170(115, 229)	216(174, 254)	122(82, 171)	< 0.001
Ventilation status
HFNC		108(3.9%)	39(2.9%)	69(4.8%)	0.014
Invasive Vent		1861(67.1%)	948(71.6%)	913(63.0%)	< 0.001
Non-Invasive Vent		23(0.8%)	13(1.0%)	10(0.7%)	0.41
Supplemental Oxygen		782(28.2%)	324(24.5%)	458(31.6%)	< 0.001
Outcomes
28-day mortality		928(33.5%)	373(28.2%)	555(38.3%)	< 0.001
90-day mortality		1142(41.2%)	473(35.7%)	669(46.1%)	< 0.001
1-year mortality		1347(48.6%)	586(44.3%)	761(52.5%)	< 0.001
Length of hospital stay (d)		12.1(6.8, 21.2)	12.4(7.1, 21.4)	12.0(6.6, 20.6)	0.061
Length of ICU stay (d)		5.6(2.9, 10.5)	5.2(2.9, 10.6)	5.8(3.0, 10.5)	0.499

P‑values are for comparisons between the low HACOR (< 6) and high HACOR (≥ 6) groups

During the follow-up period, 928 patients (33.5%) died within 28 days after admission, 1142 patients (41.2%) died within 90 days, and 1347 patients (48.6%) died within 365 days. A comparison of the outcomes of the two groups revealed statistically significant differences in various mortality rates, including the 28-day mortality rate (28.2% vs. 38.3%, P < 0.001), 90-day mortality rate (35.7% vs. 46.1%, P < 0.001), and 365-day mortality rate (44.3% vs. 52.5%, P < 0.001).

The relationship between HACOR and all‑cause mortality

Kaplan‑Meier analysis revealed significantly lower survival probabilities in the high HACOR score (HIGH) group than in the low HACOR score (LOW) group at the 28-day, 90-day, and 365-day follow-ups (all log-rank P < 0.001; Fig. 2). Cox regression analysis revealed that each 1-unit increase in the HACOR score was associated with significantly greater mortality risk at all time points (all P < 0.001), with consistent results after multivariable adjustment (Table 2). When analyzed categorically, the HIGH group presented significantly greater mortality risks than did the LOW group across all time periods (all P < 0.001).

Fig. 2.

A 28-day KM survival curve; B 90-day KM survival curve; C 365-day KM survival curve. HACOR score: Low < 6, High ≥ 6

Table 2.

Cox proportional hazard ratios (HR) for all-cause mortality

Categories	Unadjusted model P		Adjusted model P
	HR (95% CI)		HR (95% CI)
28-day mortality
HACOR as continuous	1.058(1.045–1.072)	<0.001	1.045(1.031–1.059)	<0.001
Low HACOR	reference		reference
High HACOR	1.510(1.324–1.721)	<0.001	1.427(1.247–1.634)	<0.001
90-day mortality
HACOR as continuous	1.053(1.041–1.066)	<0.001	1.042(1.029–1.055)	<0.001
Low HACOR	reference		reference
High HACOR	1.452(1.290–1.633)	<0.001	1.381(1.224–1.559)	<0.001
365-day mortality
HACOR as continuous	1.045(1.033–1.056)	<0.001	1.035(1.023–1.047)	<0.001
Low HACOR	reference		reference
High HACOR	1.345(1.207–1.498)	<0.001	1.274(1.140–1.4)	<0.001

Adjusted model: adjusted for age, race, CCI, malignant cancer, cerebrovascular disease, severe liver disease, myocardial infarct, white blood cell, albumin, creatinine, ALT, AST, bilirubin, potassium, sodium, calcium, INR, vasopressor and ventilation support

Restricted cubic spline analysis confirmed a significant linear association between the HACOR score and mortality (P values for overall < 0.001; P values for nonlinearity > 0.05; Fig. 3).

Fig. 3.

Restricted cubic spline analysis of the association between HACOR score and the risk of all-cause mortality in patients with acute hypoxemic respiratory failure. A RCS curve for 28-day mortality; B RCS curve for 90-day mortality; C RCS curve for 365-day mortality

We further evaluated the discriminative ability and calibration of the HACOR score for 28‑day mortality. The area under the ROC curve (AUC) was 0.584 (95% CI: 0.562–0.607), indicating limited discriminative ability (Supplementary Fig. S1). The Hosmer‑Lemeshow test showed good calibration (χ² = 5.703, df = 8, P = 0.680), suggesting that the predicted probabilities from the HACOR score are not systematically biased (Supplementary Fig. S2), even though the score alone cannot reliably discriminate survivors from non‑survivors.

Subgroup analysis

Subgroup analyses revealed that the association between HACOR score and mortality remained significant across all predefined subgroups, including stratification by sex, age, vasopressor use, and mechanical ventilation status (P < 0.05 in all subgroups). While the HACOR score showed robust predictive value in most subgroups, formal interaction tests revealed significant effect modification by age for both 90-day (P for interaction = 0.025) and 365-day mortality (P for interaction = 0.015), indicating a stronger association in younger patients (≤ 65 years) than in older patients (> 65 years). No significant interaction effects were observed for sex, vasopressor use, or mechanical ventilation status across 28-day, 90-day and 365-day mortality (all P > 0.05). These findings support the consistent predictive utility of the HACOR score across these patient characteristics. (Fig. 4).

Fig. 4.

Subgroup analysis of the association between HACOR score (High vs. Low) and mortality at 28, 90, and 365 days. Mechanical ventilation includes invasive mechanical ventilation and non‑invasive ventilation. Non‑mechanical ventilation includes high‑flow nasal cannula (HFNC) and conventional oxygen therapy. Hazard ratios (HRs) with 95% confidence intervals (CIs) are shown. P values for interaction are presented for each subgroup

Discussion

This large-scale retrospective cohort study suggested that the HACOR score at ICU admission is a strong and independent predictor of both short- and long-term all-cause mortality in a general population of adult patients with AHRF. Our principal findings can be summarized as follows: First, a higher HACOR score was significantly associated with increased risks of 28-day, 90-day, and 365-day mortality, with a clear dose‒response relationship. Second, the association followed a linear pattern, indicating that the mortality risk increased steadily with each unit increase in the HACOR score. Third, the predictive value of the HACOR score remained robust across most clinically relevant subgroups, including sex, vasopressor use, and mechanical ventilation status, although its association was notably stronger in younger patients (≤ 65 years).

The primary strength of our study lies in its validation of the HACOR score within a large, heterogeneous, and unselected cohort of AHRF patients. While the HACOR score was originally developed to predict NIV failure [8], our findings significantly expand its clinical utility by establishing a clear link with all-cause mortality across multiple time horizons. This aligns with and extends the results of several smaller studies [9, 10]. For instance, Duan et al. initially reported that a HACOR score > 5 was associated with higher NIV failure rates and ICU mortality [8]. Our study, with a substantially larger sample size and longer follow-up, confirms that this relationship holds not only for short-term outcomes but also for mid- and long-term survival in a broader AHRF population, irrespective of the initial respiratory support modality. The majority of our cohort received invasive mechanical ventilation. Therefore, our findings support the use of the HACOR score in invasively ventilated AHRF patients. This extends prior studies that focused primarily on non‑invasively supported patients [9–11]. However, external validation in non‑intubated AHRF populations remains needed. The linear relationship we observed, as depicted by the restricted cubic splines, provides a more nuanced understanding than simple cutoff values do, suggesting that the score can serve as a continuous risk assessment tool.

The robust predictive performance of the HACOR score can be attributed to its physiological rationale. The five components of the score collectively capture the integrated severity of respiratory (respiratory rate, oxygenation), cardiovascular (heart rate), metabolic (acidosis), and neurological (conscious) compromise. This multifaceted nature likely makes it a more sensitive indicator of the overall physiological derangement in AHRF than scores that assess organ systems in isolation [13]. Our baseline characteristics strongly support this, as patients in the high-HACOR group presented significantly worse profiles across nearly all measured physiological and laboratory parameters, reflecting greater illness severity and multiorgan dysfunction.

The modest AUC (0.584) combined with adequate calibration (P = 0.680) deserves comment. An AUC near 0.5 indicates that the HACOR score, when used in isolation, has little ability to separate patients who will die from those who will survive. This is not unexpected, as mortality is a complex outcome influenced by many factors beyond the five physiological parameters captured by the HACOR score. The good calibration, on the other hand, suggests that for the population as a whole, the predicted probability of death derived from the HACOR score approximates the observed mortality rate. In other words, the HACOR score may provide a reasonably accurate average risk estimate for a group of patients with a given score, but it cannot reliably predict which individual patient will die. This further supports our conclusion that the HACOR score demonstrates an association with mortality rather than serving as a robust clinical prediction tool.

An interesting finding from our subgroup analysis was the significant interaction between HACOR score and age. The stronger association between the HACOR score and mortality in younger patients (≤ 65 years) warrants further investigation. We hypothesize that in older patients, mortality is influenced by a greater burden of comorbidities, frailty, and nonrespiratory factors not fully captured by the HACOR score [14–16]. In contrast, for younger patients with fewer competing risks, the acute physiological insult quantified by the HACOR score may be the predominant determinant of survival. This underscores that while the HACOR score is a powerful tool, it should be interpreted in the context of the patient’s overall clinical picture, particularly their age and baseline health status.

From a clinical perspective, the HACOR score offers several practical advantages. Its calculation is based on five readily available bedside parameters, making it simple, immediate, and cost-effective to implement in busy ICU settings. Its ability to stratify patients into distinct risk groups for both short- and long-term mortality upon ICU admission can aid clinicians in several ways. High-risk patients (e.g., HACOR ≥ 6) may prompt more intensive monitoring, earlier consideration of advanced life support strategies, and informed discussions with families regarding a guarded prognosis. Conversely, for low-risk patients, it may help avoid overly aggressive interventions. Furthermore, its linear relationship with mortality risk allows for a more granular assessment than binary classifications.

A practical limitation of the HACOR score is its reliance on PaO₂/FiO₂, which requires arterial blood gas sampling. For patients receiving only nasal oxygen or managed without frequent ABGs, this may hinder real‑time risk stratification. Whether the non‑invasive SpO₂/FiO₂ ratio can be reliably substituted for PaO₂/FiO₂ in the HACOR score, and whether it retains predictive accuracy for mortality, warrants future prospective validation.

Our study has several limitations that must be acknowledged. First, the MIMIC-IV database is a single center database originating from a US tertiary teaching hospital, which inherently limits the generalizability of our findings to other healthcare systems, geographic regions, or lower resource settings. Therefore, external validation in multicenter prospective cohorts across diverse geographic and healthcare settings is urgently needed to confirm the robustness and generalizability of our results. Second, as a retrospective study, despite extensive adjustments for measured confounders, residual confounding due to unmeasured factors (e.g., specific etiologies of AHRF, detailed treatment decisions after ICU admission) cannot be entirely ruled out. Future prospective studies with standardized collection of etiologic data are warranted to address this limitation. Third, although subgroup analysis revealed no significant interaction with mechanical ventilation status, the small sample of patients receiving non-invasive ventilation alone (n = 23, 0.8%) limited our assessment of the HACOR score in this subgroup. Furthermore, the score components (respiratory rate, PaO₂/FiO₂, pH, PaCO₂) may be affected by the type and intensity of baseline respiratory support. Future prospective studies with standardized timing of score assessment relative to support initiation are warranted. Fourth, the HACOR score was calculated only at ICU admission; its trajectory over time and the impact of its dynamic changes on outcomes were not evaluated and represent important areas for future research.

Conclusion

In this large retrospective cohort study, a higher HACOR score at ICU admission was strongly and independently associated with increased short- and long-term all-cause mortality in a general population of patients with AHRF. This simple bedside tool, composed of readily available parameters, has potential utility for early risk stratification. However, given the observational design, causality cannot be inferred, and the clinical benefits of HACOR-guided decisions cannot be confirmed. Future multicenter prospective studies are warranted to externally validate these findings.

Authors’ contributions

Lei Li contributed to the conception and design, and critical revision of important intellectual content. Ke Wang performed the analyses, wrote and revised the manuscript. Yan Ou interpreted the data and critically revised important intellectual content. All authors approved the final version of the manuscript and agreed to be accountable for all aspects of the work.

Funding

None.

Data availability

The datasets analysed during the current study are avilable from the MIMIC-IV database. The data utilized in this study are publicly accessible through the MIMIC-IV database version 3.1 ( [https://physionet.org/content/mimiciv/3.1/](https:/physionet.org/content/mimiciv/3.1) ).

Declarations

Ethics approval and consent to participate

The studies involving humans were approved by the Institutional Review Boards (IRBs) of Beth Israel Deaconess Medical Center (BIDMC) and the Massachusetts Institute (MIT) of Technology. The studies were conducted in accordance with local legislation and institutional requirements. The ethics committee/institutional review board waived the requirement of written informed consent for participation from the participants or the participants’ legal guardians/next of kin due to the use of anonymous clinical data.

Consent for publication

The data has been de-identified and made publicly accessible, eliminating the requirement for specific informed permission from patients.

Competing interests

The authors declare no competing interests.