Early Detection of Clinical Deterioration in ICU Patients With Respiratory Failure Receiving Vasoactive Support: A Machine Learning Approach to Identifying High‐Risk Individuals

ABSTRACT

Background and Aims

Acute respiratory failure requiring intensive care is frequently accompanied by hemodynamic instability, necessitating vasoactive pharmacologic support, and elevated mortality. We aimed to develop and validate a machine learning model to stratify all‐cause in‐hospital mortality risk in ICU patients with respiratory failure receiving vasoactive therapy.

Methods

We performed a secondary analysis of the MIMIC‐IV database, including adult ICU patients (2017–2022) with respiratory failure who received vasoactive medications. A random forest survival model was constructed to estimate individualized survival probabilities. Patients were subsequently stratified into two cohorts based on median predicted survival. A multivariable logistic regression model was used to profile clinical and laboratory characteristics associated with low‐ and high‐risk groups.

Results

The final cohort comprised 1951 adult patients. Using the random forest survival model (concordance index = 0.769), patients were categorized into low‐risk and high‐risk groups, with significantly different survival outcomes (log‐rank test p < 0.001). The logistic regression model (p < 0.001, Nagelkerke R ² = 0.714, Brier score = 0.102) achieved high performance (AUC = 0.924) in identifying high‐risk individuals. High‐risk patients were characterized by biomarkers of multiorgan dysfunction, including renal insufficiency, hepatic impairment, coagulopathy, acid–base disturbances, and systemic inflammation. In contrast, higher levels of hemoglobin, albumin, and arterial oxygen tension were associated with lower risk, indicative of preserved physiological reserve.

Conclusions

A machine learning model incorporating early‐phase ICU data effectively stratified mortality risk in patients with respiratory failure receiving vasoactive support. This approach provides a clinically interpretable framework to inform prognostication, optimize resource allocation, and support data‐driven decision‐making.

1. Introduction

Acute respiratory failure—defined by impaired oxygenation or ventilation—constitutes a frequent cause for ICU admission, with mechanical ventilation required in approximately 25% of cases and mortality rates often ranging between 20% and 40% [1, 2]. Frequently, hemodynamic instability accompanies respiratory compromise, leading to the administration of vasoactive agents [3]. Higher vasoactive exposure has been associated with increased ICU mortality [3]. These patients are prone to multi‐organ dysfunction and deterioration, leading to elevated ICU mortality and significant strain on critical care resources [4, 5]. Conventional ICU monitoring tools—such as static scores like SOFA and APACHE—are reactive, and inadequately capture the complexity of physiologic changes [6, 7, 8]. The observed markers of multiorgan dysfunction in high‐risk patients reflect the complex and interconnected pathophysiology of a critical illness [9].

In these patients, clinical decline often evolves over time, marked by gradual physiologic changes [10]. Machine learning models are suggested to capture nonlinear temporal patterns and high‐dimensional interactions, enabling more accurate early warning predictions in ICU populations [4].

However, few machine learning frameworks have been designed for patients experiencing respiratory failure that could benefit from early predictive analytics [6, 11]. Moreover, there appears to be a lack of machine learning models specifically for risk‐stratifying ICU patients with respiratory failure on vasoactive medications [5, 12]. Existing models primarily focus on broad ICU populations and often treat mechanical ventilation or vasopressor initiation as outcomes, rather than incorporating them as factors that modify a patient's risk of mortality [13, 14, 15]. Furthermore, mortality prediction tools typically generate static clinical risk scores, lacking the capacity to incorporate dynamic, time‐dependent changes in patient risk [6, 16, 17]. Models incorporating ventilatory and hemodynamic parameters used for acute respiratory distress syndrome and sepsis still neglect vasoactive support as a core predictor [5, 16, 17]. These limitations hinder anticipatory clinical decision support in patients with rapidly changing clinical states. Therefore, models that integrate physiological data with administered medications, including vasoactive and diuretic therapies, are needed to improve mortality risk stratification in this subgroup of patients. This study aimed to develop and validate a machine learning model for early risk stratification of critically ill ICU patients with respiratory failure receiving vasoactive support, identifying those at a higher risk of in‐hospital mortality. We hypothesized that integrating early high‐resolution physiologic, laboratory, and therapeutic data using a time‐to‐event framework would improve all‐cause in‐hospital mortality risk stratification in ICU patients with respiratory failure requiring vasoactive support. This approach addresses the need for anticipatory decision support concerning a rapidly evolving pathophysiology. By enabling timely, individualized prognostication, the model could improve ICU triage, optimize treatments, facilitate family communication, and advance precision critical care.

2. Patients and Methods

2.1. Design, Study Population, and Data Resource

This study was a secondary analysis of data extracted from the Medical Information Mart for Intensive Care IV (MIMIC‐IV, version 3.1) database—a large, publicly available critical care data set curated by the Massachusetts Institute of Technology's Laboratory for Computational Physiology in collaboration with the Beth Israel Deaconess Medical Center (Boston, MA, USA) (https://mimic.mit.edu) [18]. MIMIC‐IV contains de‐identified health records for ICU patients from 2008 to 2022, including clinical practice in a tertiary care healthcare center. The database encompasses detailed information, including demographics, physiologic and laboratory parameters, therapeutic interventions, comorbidity profiles, and clinical outcomes. For the present investigation, we identified adult patients (aged ≥ 18 years) undergoing their first ICU admission between 2017 and 2022, with an ICU length of stay of at least 48 h. This timeframe was selected to align with contemporary standards in ICU diagnostics, monitoring, and management. The final study cohort included only individuals with respiratory failure requiring vasoactive support during the ICU admission. This cohort formed the basis for developing a predictive model focused on early mortality risk stratification in critically ill patients with combined respiratory and hemodynamic compromise.

2.2. Outcome and Exposures

We included patients with respiratory failure who received vasoactive support during their first ICU admission. The outcome was the risk of all‐cause in‐hospital mortality following ICU admission. Vasoactive agents are administered in ICU settings to stabilize key hemodynamic parameters, including blood pressure, heart rate, cardiac output, and end‐organ perfusion [19]. In this study, vasoactive medications consisted of either vasopressors or inotropes, or both. Vasopressors included agents used to augment vascular tone and maintain arterial pressure—norepinephrine, epinephrine, phenylephrine, vasopressin, and dopamine. Inotropes, used for increasing myocardial contractility, consisted of dobutamine, milrinone, and dopamine. The analytical data set included baseline demographic characteristics, pre‐existing comorbidities, and indices of organ function, captured through a comprehensive set of laboratory parameters and physiologic markers reflective of cardiovascular, respiratory, renal, hepatic, and hematologic function. Additionally, data on diuretic exposure—specifically bumetanide and furosemide—were incorporated as a binary variable to complement the characterization of pharmacologic hemodynamic support.

2.3. Data Availability

The raw, de‐identified data used in this study were selected from the publicly accessible Medical Information Mart for Intensive Care IV (MIMIC‐IV) database, available at https://mimic.mit.edu [18]. MIMIC‐IV is an open‐access, relational database derived from the electronic health records of patients admitted to the Beth Israel Deaconess Medical Center in Boston, Massachusetts.

2.4. Ethical Considerations

The MIMIC‐IV database contains fully de‐identified health records. This study did not involve direct interaction with human participants, the collection of identifiable data, or any procedures requiring informed consent. We adhered to ethical standards for secondary use of publicly available clinical data. Access to the MIMIC‐IV database requires completion of institutional ethics training, ensuring appropriate and responsible handling of sensitive health information. No raw data were redistributed or republished; all reported findings are based exclusively on aggregated analyses. In order to acquire access to data, we successfully completed the CITI “Data or Specimens Only Research” course (record ID: 53106331) and signed the PhysioNet credentialed health data use agreement 1.5.0 on March 17, 2025. Our study followed the ethical guidelines specified in the Declaration of Helsinki, and ethics approval was obtained from the institutional review board of Shahid Beheshti University of Medical Sciences with the reference number of IR.SBMU.RETECH.REC.1403.897.

2.5. Data Analyses

Initial preprocessing steps included exclusion of variables with more than one‐third (> 33%) missingness and high cardinality, followed by removal of binary predictors with high class imbalance (< 10% prevalence). This ensured data robustness while preserving clinical heterogeneity. Collinearity was evaluated using Spearman's rank correlation coefficients, and among highly correlated pairs, one variable was retained to reduce multicollinearity in subsequent modeling. For survival analysis, a Cox proportional hazards model was initially considered. However, it was statistically rejected because of proportionality assumption violations. Instead, a non‐parametric random forest survival model was then used with a 10‐fold cross‐validation to estimate individualized survival probabilities. Random forest survival analysis is used to analyze time‐to‐event outcomes and censored observations, analogous to Cox proportional hazards modeling. Similar to the Cox model, it accounts for both the timing of events and censoring, enabling inclusion of patients with incomplete follow‐up information. Unlike traditional Cox regression, however, random forest survival models do not rely on the proportional hazards assumption and can flexibly capture complex nonlinear relationships and higher‐order interactions among predictors. These characteristics make the method particularly suitable for critically ill ICU populations with heterogeneous and dynamically evolving physiologic states. This strategy therefore enabled robust modeling of temporal mortality risk using high‐dimensional clinical data. Patients were subsequently stratified into low‐risk and high‐risk groups using the median survival time as a cutoff. To further characterize risk strata, a multivariable logistic regression classifier was trained with a 10‐fold cross‐validation to distinguish between high‐ and low‐risk patients. Model performance was evaluated using a confusion matrix and classification metrics, including accuracy, sensitivity, specificity, positive predictive value, negative predictive value, kappa statistic, and the area under the receiver operating characteristic curve (AUC). The calibration of the logistic model was assessed using the Brier score. The positive likelihood ratio was also calculated to quantify the diagnostic value of the classifier. Patient profiles were analyzed to identify clinical features associated with high‐risk classification. Adjusted odds ratios and corresponding 95% confidence intervals were estimated, with statistical significance evaluated using Wald tests. This combination of temporal survival modeling and supervised classification enabled risk stratification in this highly critical cohort. All analyses were conducted using R (version 4.3.1—R Foundation for Statistical Computing, Vienna, Austria. https://www.R-project.org/) and Python (version 3.11.5—https://www.python.org/) using standard scientific computing libraries.

3. Results

3.1. Preprocessing and Feature Selection

The initial sample (N = 1951) included patients with respiratory failure receiving either vasopressors (1939; 99.4%) or inotropes (237; 12.1%), or both. The median follow‐up time was 7.0 days with an interquartile range of 4.0 to 13.5 days. None of the rows were duplicated. However, 12.6% of the initial data were missing (Figure 1A). We excluded the variable race due to its high cardinality (30 levels). Marital status was excluded from the analysis due to its four‐level categorization and a relatively high proportion of missing data (27%), while it did not appear to meaningfully influence the outcomes in the ICU. Moreover, variables with more than one‐third (33%) of their data missing (cardiac output, C‐reactive protein, direct bilirubin, central venous pressure, and arterial O₂ saturation) were excluded. Binary variables were also assessed for class imbalance, with those showing fewer than 10% positive cases being excluded from the analysis: liver disease (3.1% positive), Type 1 diabetes mellitus (1.2%), cerebrovascular disease (1.0%), and acute myocardial infarction (0.4%). Figure 1B illustrates the heatmap of correlations among the study variables using Spearman's correlation coefficients. There were statistically significant correlations of more than 50% between ALT and AST (ρ = 0.84), serum creatinine (Cr) and BUN (ρ = 0.68), serum chloride and sodium (ρ = 0.68), systolic and diastolic blood pressure (ρ = 0.58), and acute kidney injury (AKI) and Cr (ρ = 0.53), all p < 0.001. Significant correlations were also observed between ionized and non‐ionized calcium (ρ = 0.50), and arterial CO₂ pressure and pH (ρ = −0.50), both p < 0.001. To decrease multicollinearity in our models, one variable from each pair of highly correlated variables was excluded, thereby reducing redundancy. We excluded AST, BUN, chloride, diastolic blood pressure, AKI, non‐ionized calcium, and arterial CO₂ pressure from the corresponding pairs. These exclusions were based on reducing redundancy, prioritizing clinically relevant markers, and maintaining precision with more granular, continuous variables. After preprocessing, only 5.3% of the data were missing, which were imputed using the predictive mean matching method (Figure 1C).

Figure 1.

(A) The patterns and percentages of missing data in the initial study data set. (B) The heatmap of pairwise Spearman's correlations among study variables. (C) The patterns and percentages of missing data in the study data set after preprocessing of data.

3.2. Risk‐Stratification and Profiling

Some covariates violated the proportional hazards assumption, with the global test showing χ ²(25) = 33.448, p < 0.001. Since the Cox model's assumption was violated, a random forest survival model was used instead. The resulting model showed satisfactory predictive performance in estimating patients' survival probabilities based on their clinical characteristics, achieving a concordance index (SD) of 0.769 (0.028). To stratify patients by their survival, we categorized them based on the median survival time estimated by the random forest survival model. This resulted in two patient groups—low‐risk (N = 976) and high‐risk (N = 975)—with significantly different survivals, as confirmed by the log‐rank test (Figure 2). Among the 950 patients who died (48.6% of the total sample), 285 (29.2%) were in the low‐risk group, and 665 (68.2%) were in the high‐risk group (p < 0.001). We then developed a logistic regression model to profile low‐risk versus high‐risk patients based on their clinical features. The logistic model [χ ²(29) = 1497.974, p < 0.001; Nagelkerke R ² = 0.714; calibration error (Brier score) = 0.102] showed high performance in distinguishing high‐risk patients from low‐risk patients (Figure 3). Table 1 presents the performance of the logistic model. The high positive likelihood ratio indicates a strong ability to correctly identify individuals at higher risk of adverse outcomes, highlighting the model's effectiveness in distinguishing between low‐risk and high‐risk patients. Table 2 presents the detailed risk profiling information. Statistically significant adjusted odds ratios identify variables with an independent predictive value for high‐risk categorization. Biomarkers such as elevated creatinine, prothrombin time, glucose, phosphorus, and age were positively associated with high‐risk status, while higher hemoglobin, albumin, BMI, and arterial O₂ pressure were inversely associated. Diuretic administration showed a strong association with lower risk, reflecting therapeutic response. Several features were significant in univariable tests but lost significance in the adjusted model, suggesting interrelationships, potential confounding, or collinearity. Overall, the model showed strong discrimination, supporting its stability and relevance in stratifying patients' risk.

Figure 2.

Kaplan–Meier survival curves illustrating the 30‐day survival probability for low‐risk and high‐risk patient groups. A statistically significant difference in survival was observed between the two groups, as confirmed by the log‐rank test.

Figure 3.

Receiver operating characteristic (ROC) curve illustrating the performance of the logistic model in discriminating between the two risk strata. The area under the curve (AUC) indicates high model performance.

Table 1.

Confusion matrix and performance metrics for the study logistic model (positive class = High‐risk).

Confusion matrix
		Reference
		Low‐risk	High‐risk
Prediction	Low‐risk	873	170
	High‐risk	103	805

Performance metrics
Accuracy (95% CI)			0.860 (0.844–0.875)
No information rate			0.500
[Accuracy > No information rate] p			< 0.001*
Kappa			0.720
Sensitivity			0.826
Specificity			0.894
Positive predictive value			0.887
Negative Predictive Value			0.837
Area Under the Curve			0.924
Positive likelihood ratio			7.79

Note: *Significant at p < 0.05.

Table 2.

Profiling of demographic, laboratory, and clinical characteristics of the study cohort, with comparative analyses between low‐risk and high‐risk patient groups.

Feature	Total	Low‐risk	High‐risk	p	Adjusted odds ratio (95% CI)	p
Creatinine (mg/dL)	1.3 (1.1)	1.1 (0.7)	1.6 (1.4)	< 0.001**	1.562 (1.294, 1.884)	< 0.001**
Male (%)	1182 (60.6)	614 (62.9)	568 (58.3)	0.040*	0.898 (0.658, 1.226)	0.499
BMI (kg/m2)	31.0 (8.6)	32.4 (8.6)	29.7 (8.5)	< 0.001**	0.966 (0.949, 0.984)	< 0.001**
Age (year)	60.8 (14.0)	59.9 (14.0)	61.6 (13.9)	0.008*	1.06 (1.046, 1.075)	< 0.001**
ALT (IU/L)	177.7 (712.2)	167.8 (678.3)	187.6 (744.8)	0.539	1.000 (1.000, 1.000)	0.557
Albumin	2.9 (0.6)	2.9 (0.6)	2.8 (0.7)	< 0.001**	0.661 (0.513, 0.852)	0.001*
Alkaline phosphate (IU/L)	115.3 (126.2)	94.6 (98.5)	136.0 (146.0)	< 0.001**	1.001 (1.000, 1.003)	0.058
Systolic BP (mmHg)	116.0 (30.1)	116.1 (28.8)	116.0 (31.4)	0.938	1.001 (0.996, 1.006)	0.636
Arterial O2 pressure (mmHg)	158.6 (104.5)	173.0 (111.5)	144.2 (94.8)	< 0.001**	0.995 (0.993, 0.996)	< 0.001**
Glucose (mg/dL)	158.1 (88.0)	154.2 (70.2)	162.1 (102.6)	0.049*	1.004 (1.002, 1.006)	< 0.001**
Heart rate (bpm)	93.4 (22.3)	92.2 (21.6)	94.6 (23.0)	0.016*	1.004 (0.997, 1.011)	0.272
Hemoglobin (g/dL)	10.6 (2.6)	11.2 (2.5)	9.9 (2.6)	< 0.001**	0.742 (0.693, 0.795)	< 0.001**
Ca2+ (mmol/L)	1.1 (0.4)	1.1 (0.4)	1.1 (0.3)	0.199	1.013 (0.592, 1.734)	0.963
Magnesium (mg/dL)	2.1 (0.6)	2.1 (0.6)	2.1 (0.6)	0.050	0.902 (0.673, 1.208)	0.489
pH (units)	7.339 (0.1)	7.344 (0.1)	7.333 (0.1)	0.030*	5.566 (1.202, 25.77)	0.028*
Phosphorous (mg/dL)	4.1 (1.8)	3.9 (1.5)	4.4 (2.1)	< 0.001**	1.216 (1.093, 1.353)	< 0.001**
Platelet count (K/uL)	193.7 (112.4)	194.7 (97.2)	192.6 (125.8)	0.672	1.000 (0.999, 1.001)	0.980
Potassium (mEq/L)	4.4 (0.9)	4.4 (0.8)	4.4 (0.9)	0.580	0.944 (0.778, 1.146)	0.563
Prothrombin time (s)	17.6 (10.2)	16.0 (7.5)	19.3 (12.1)	< 0.001**	1.040 (1.020, 1.060)	< 0.001**
Respiratory rate (insp/min)	20.6 (7.1)	20.3 (7.1)	20.9 (7.0)	0.043*	1.024 (1.002, 1.046)	0.030*
Sodium (mEq/L)	138.5 (6.3)	139.0 (5.5)	137.9 (7.0)	< 0.001**	0.995 (0.973, 1.018)	0.683
Total bilirubin (mg/dL)	2.3 (5.2)	1.5 (3.4)	3.1 (6.5)	< 0.001**	1.048 (1.012, 1.086)	0.009*
WBC (K/uL)	14.6 (8.8)	14.7 (7.8)	14.6 (9.7)	0.915	0.982 (0.965, 0.999)	0.036*
Chronic kidney disease (%)	471 (24.1)	187 (19.2)	284 (29.1)	< 0.001**	1.061 (0.731, 1.540)	0.754
Essential HTN (%)	747 (38.3)	416 (42.6)	331 (33.9)	< 0.001**	0.507 (0.363, 0.708)	< 0.001**
Heart failure (%)	746 (38.2)	390 (40.0)	356 (36.5)	0.129	1.203 (0.850, 1.704)	0.297
Sepsis (%)	925 (47.4)	382 (39.1)	543 (55.7)	< 0.001**	1.157 (0.846, 1.584)	0.362
Type 2 diabetes (%)	555 (28.4)	284 (29.1)	271 (27.8)	0.557	0.611 (0.425, 0.879)	0.008*
Diuretics (%)	1319 (67.6)	951 (97.4)	368 (37.7)	< 0.001**	0.002 (0.001, 0.003)	< 0.001**

Note: Univariable analyses were performed to evaluate group differences, while a multivariable logistic regression model was developed to estimate adjusted odds ratios with corresponding 95% confidence intervals (CIs). Continuous variables are reported as mean (standard deviation), and categorical variables as counts (%).

^*Significant at p < 0.05.

^**Significant at p < 0.001.

4. Discussion

This study aimed to develop a data‐driven, temporally informed machine learning framework for risk stratification of ICU patients with respiratory failure requiring vasoactive support. Our evidence showed that high‐resolution physiological and laboratory data from the initial phase of ICU admission could be used to predict all‐cause in‐hospital mortality. The random forest survival model successfully identified a high‐risk subgroup characterized by significantly poorer survival. Subsequent multivariable profiling using logistic regression identified clinical features associated with the elevated mortality risk. Specifically, high‐risk patients showed markers of multiorgan dysfunction, including renal impairment, hepatic dysfunction, coagulopathy, acid–base imbalance, and systemic inflammation. Conversely, higher hemoglobin, albumin, and arterial oxygenation were associated with a protective effect, suggesting preserved physiological reserve. These findings reflect pathophysiological trajectories of critically ill patients experiencing hemodynamic compromise and respiratory failure. They reinforced the prognostic relevance of routine ICU parameters when modeled dynamically. This study integrated heterogeneous data without relying on static thresholds, enabling a more representative identification of at‐risk individuals.

Our findings are consistent with recent studies supporting the practical value of machine learning models for predicting all‐cause ICU mortality, particularly in patients with respiratory failure. For instance, the POSTCARDS study demonstrated that models integrating ventilatory, oxygenation, and organ failure parameters can achieve strong predictive performance in patients with respiratory distress [20]. Similarly, Liu et al. used deep learning to analyze data from mechanically ventilated patients and showed that incorporating ventilation duration and systemic parameters markedly improved predictive power [21]. Also, Jeon et al.'s work in pneumonia cohorts implied that machine learning outperformed traditional severity scores by integrating multiple systemic and contextual parameters [22]. Multiorgan failure is typically driven by dysregulated inflammatory responses, ischemia‐reperfusion injury, and altered microcirculatory flow, which impair tissue perfusion and lead to cellular dysfunction across multiple organ systems [23]. Hepatic dysfunction in particular is often a consequence of ischemic injury or sepsis, with the liver failing to clear inflammatory mediators and produce acute‐phase proteins, exacerbating systemic inflammation and contributing to further organ damage [24]. Furthermore, renal dysfunction in critical illness is frequently due to impaired renal perfusion and inflammatory mediator release that worsens overall prognosis [25]. In contrast, preserved physiological markers such as higher hemoglobin, albumin, and arterial oxygenation are indicative of better metabolic control, enhanced oxygen delivery, and improved nutritional status, which collectively promote organ function and resilience. Higher hemoglobin levels indicate an optimal oxygen‐carrying capacity, while elevated albumin reflects better nutritional status and hepatic function, crucial for modulating systemic inflammation and maintaining fluid balance [26].

In our model, diuretic use emerged as a highly important predictor inversely associated with high‐risk classification. This should be interpreted cautiously, as it may partly reflect residual confounding by indication or differences in physiologic reserve. Patients capable of tolerating diuretic therapy may have had greater hemodynamic stability, preserved cardiac responsiveness, or more favorable volume‐management physiology, rather than a direct protective effect of diuretics themselves. The association observed may also reflect effective volume management or heightened clinical vigilance. This may also reflect mitigation of venous congestion, which is a contributor to renal dysfunction, hepatic impairment, impaired oxygenation, and multiorgan injury in critically ill patients. The strong inverse association of diuretic use with mortality replicates findings that treatment‐related features, though seemingly paradoxical, may act as surrogates for clinical responsiveness or proactive care [27, 28].

Diuretics may also improve respiratory mechanics, particularly in volume‐overloaded patients, by decreasing pulmonary edema and enhancing oxygenation [29]. While sepsis was significantly more prevalent in our high‐risk patients in univariate analysis, it did not independently contribute to risk classification in the multivariable model. This attenuation of predictive value suggests that the model captured the pathophysiologic consequences of sepsis—such as elevated inflammatory markers, impaired perfusion, and multiorgan dysfunction—through a constellation of laboratory and clinical features. Thus, the binary sepsis indicator may be redundant considering granular and temporally responsive variables embedded in the model, particularly within this physiologically complex cohort. This may also be partly attributed to the target population we defined. The high prevalence of sepsis in our cohort of patients with respiratory failure and compromised hemodynamics, and its interaction with other features, may dilute the individual effect size of sepsis in the linear model. Overall, this suggests that while sepsis remains clinically important, its risk signal is likely embedded within a broader clinical profile. Kang et al. reported that systemic features and organ failure scores provided superior predictive value over binary sepsis classification in patients undergoing renal replacement therapy [7]. Ye et al. also suggested that sepsis‐related markers are often represented within other clinical features, reducing their standalone significance in mortality prediction for critically ill patients [30]. The observed inverse association of essential hypertension, type 2 diabetes mellitus (T2DM), and higher BMI with high‐risk classification may be attributed to an ICU‐specific “obesity paradox.” This paradox suggests that patients with higher BMI or chronic conditions may benefit from improved clinical management [31, 32]. They have longstanding medical problems and are more likely to receive early therapeutic interventions and rigorous monitoring, especially during critical illness [31]. Additionally, elevated BMI may indicate metabolic and nutritional reserves that enhance tolerance to physiologic stress, while chronic antihypertensive or antidiabetic therapies may indirectly modulate inflammatory and vascular responses [33, 34].

4.1. Clinical Implications

This study presents a clinically applicable machine learning framework capable of identifying ICU patients with respiratory failure and vasoactive support who are at elevated risk for early in‐hospital mortality. An advantage of this study lies in its deliberate focus on “early detection” of clinical deterioration, achieved by restricting model inputs to routinely collected physiologic and laboratory variables that are reliably available within the first 24–48 h of ICU admission. By including such data, the model enables proactive risk stratification during the early phase of critical illness, supporting timely clinical decision‐making. This approach captures complex, nonlinear interactions and evolving patient trajectories, allowing for a more nuanced assessment of deterioration. Rather than depending on composite indices (e.g., SOFA) or high‐granularity trajectories (such as vasopressor dose trends) that are frequently incomplete or inconsistently reported, we prioritized routinely measurable parameters that are consistently available during early ICU admission. This approach was intended to ensure reproducibility, interpretability, and feasibility for real‐time implementation. The ability to profile high‐risk individuals using routinely collected biomarkers enhances interpretability and facilitates integration into existing ICU workflows. This framework can help clinicians and managers use critical care resources more effectively. Early risk stratification allows advanced monitoring, specialized staffing, and costly interventions to be directed toward patients most likely to benefit, while avoiding unnecessary use in lower‐risk groups. From a healthcare management perspective, this risk‐based approach reflects principles of value‐based care by improving cost‐effectiveness through more precise allocation of resources. By identifying patients at elevated risk for early deterioration using routinely available ICU data, the model may support triage prioritization for advanced monitoring, multidisciplinary assessment, and timely allocation of resources. Early risk stratification may facilitate individualized escalation or de‐escalation of therapies by identifying patients who require intensified treatment versus those more likely to tolerate conservative management. The mortality risk estimation may also support goals‐of‐care discussions and improve communication with patients' families by providing data‐informed prognostic context during complex clinical decision‐making. The integration of individualized physiologic and laboratory profiles into mortality prediction supports precision critical care approaches that tailor treatment intensity according to patient‐specific risk trajectories rather than generalized severity scores. The ability to generate clear, data‐driven risk signals at admission suggests potential to improve efficiency, reduce waste, and strengthen the economic sustainability of critical care, without compromising quality of care. Overall, this work contributes a data‐driven foundation for personalized, anticipatory care in a highly critical population with substantial mortality risk.

4.2. Limitations

This study has several limitations that should be considered. The analysis was based on retrospective cohort data. While high‐resolution structured data were used, unmeasured confounding, such as clinician judgment, may have influenced both treatment decisions and outcomes. Although rigorous inclusion criteria were used to define a clinically homogeneous sample, this selection may reduce external validity in broader ICU populations with respiratory failure. Our analysis included data from a specific geographic region. Whether these associations hold in other populations and whether adjustment for treatment timing or disease severity modifies their impact should be investigated in further research. Because the study cohort was derived from a tertiary‐care ICU population including first admissions with ICU stays ≥ 48 h, the findings may not fully generalize to smaller or community‐based ICUs with different patient case mixes, resource availability, admission practices, and monitoring intensity, highlighting the need for external validation. By restricting inclusion to patients surviving beyond the initial 48 h of ICU admission, the study may have underrepresented fulminant early mortality, potentially introducing survivorship bias toward relatively stabilized early ICU survivors.

Additionally, although both hypoxemic and hypercapnic respiratory failure were included to preserve cohort generalizability, these phenotypes may show prognostic trajectories that were not separately evaluated in the present analysis. Future prospective studies could compare the clinical performance of the present framework against established ICU severity scores such as SOFA and APACHE, using longitudinal evaluations. Incorporation of multidimensional endpoints such as ventilator‐free days and organ‐support–free days may also improve assessment of functional recovery and critical care resource use in critically ill patients. Despite these constraints, the study leveraged robust statistical and machine learning models to generate a predictive framework with high discrimination. Future research should focus on prospective validation across diverse ICU populations to assess effects on clinical decision‐making and patient outcomes.

5. Conclusion

In conclusion, we developed a machine learning approach for early all‐cause mortality risk stratification in ICU patients with respiratory failure receiving vasoactive support, leveraging high‐resolution clinical data from the initial phase of ICU admission. This study identified a high‐risk subgroup characterized by distinct pathophysiologic signatures indicative of multiorgan dysfunction. Patient profiling showed that high‐risk individuals frequently had markers of renal and hepatic impairment, impaired oxygenation, coagulopathy, and systemic metabolic derangement. Specifically, increased age, elevated creatinine, prothrombin time, phosphorus, and glucose levels, as well as lower hemoglobin, albumin, and arterial oxygen tension, were independently associated with increased mortality risk. These variables reflect key domains for organ support and provide interpretable indicators for clinical risk appraisal. The model's ability to incorporate routinely available parameters into a dynamic risk prediction tool provides a clinical advantage over static scoring systems. This framework can support proactive intervention, informed prognostication, and optimized ICU resource allocation. Prospective validation and real‐time integration into critical care workflows are important next steps to translate these findings into improved patient outcomes.

Author Contributions

Mohammad Fathi: conceptualization, writing – review and editing, supervision. Hamed Markazi Moghadam: formal analysis, resources, software, visualization, writing – review and editing. Mahdis Fathi: formal analysis, visualization, writing – original draft. Mohammadreza Hajiesmaeili: investigation, validation, writing – original draft. Navid Nooraei: investigation, validation, writing – original draft. Nasser Malekpour Alamdari: investigation, methodology, writing – original draft. Sanaz Zargar Balaye Jame: writing – review and editing, project administration, methodology. Farhad Hashemnezhad Khataee: conceptualization, validation, writing – review and editing, writing – original draft. Nader Markazi Moghaddam: project administration, conceptualization, formal analysis, software, visualization, writing – review and editing.

Funding

The authors have nothing to report.

Ethics Statement

Ethics approval was obtained from the Institutional Review Board of Shahid Beheshti University of Medical Sciences with the reference number of IR.SBMU.RETECH.REC.1403.897.

Conflicts of Interest

The authors declare no conflicts of interest.

Transparency Statement

The Nader Markazi Moghaddam affirms that this manuscript is an honest, accurate, and transparent account of the study being reported; that no important aspects of the study have been omitted; and that any discrepancies from the study as planned (and, if relevant, registered) have been explained.

Data Availability Statement

The raw, de‐identified data used in this study were selected from the publicly accessible Medical Information Mart for Intensive Care IV (MIMIC‐IV) database, available at https://mimic.mit.edu. MIMIC‐IV is an open‐access, relational database derived from the electronic health records of patients admitted to the Beth Israel Deaconess Medical Center in Boston, Massachusetts.