Abstract
Objectives
Patients with malignant tumors are admitted to the ICU for diverse reasons. However, the clinical utility of serum albumin as a prognostic biomarker remains unclear.
Materials and Methods
Patients with malignant tumors were screened from the Medical Information Mart for Intensive Care IV (MIMIC-IV, v3.1). This study employed Kaplan-Meier curves, Cox proportional-hazards models, restricted cubic splines (RCS), receiver operating characteristic (ROC) curves, and subgroup analyses to evaluate serum albumin associated with all-cause mortality. For mortality-risk prediction, we applied machine-learning algorithms and used SHapley Additive exPlanations (SHAP) to identify the most influential predictors among critically ill cancer patients.
Results
A total of 1,739 patients with malignancy were included. The Kaplan–Meier curves showed significantly higher all-cause mortality in the hypoalbuminemia group (serum albumin < 30 g/L) than in the control group at each time point. Multivariable Cox regression models confirmed that hypoalbuminemia was independently associated with 28-day mortality (HR 1.74; 95% CI 1.34–2.27). Serum albumin exhibited a superior predictive capacity for long-term mortality (90-day and 1-year), with AUCs of 0.676 and 0.664, respectively, notably higher than those of the SOFA score (0.617 and 0.579). External validation using data from Tianjin Cancer Hospital yielded consistent results. The Machine learning model identified BUN, serum albumin, respiratory rate, heart rate, and SOFA as the top predictors for 14- and 28- day mortality.
Conclusion
Hypoalbuminemia was independently associated with increased all-cause mortality. Serum albumin measured at ICU admission serves as a prognostic biomarker for identifying high-risk cancer patient groups.
Keywords: serum albumin, cancer, critically ill patients, machine learning, MIMIC database
Introduction
According to the 2022 global cancer statistics report released by the International Agency for Research on Cancer (IARC), nearly 20 million new cancer cases were diagnosed, resulting in 9.7 million deaths.1 Patients with cancer face a substantially elevated risk of life-threatening complications—such as infections, treatment-related toxicities, and infiltration of vital organs by the primary malignancy—necessitating admission to intensive care units (ICUs).2
A meta-analysis of 90 studies in critically ill patients showed that each 10 g/L decline in serum albumin was associated with a 1.89-fold increase in morbidity, and a 1.71-fold prolongation of hospital stay.3 Padkins et al. similarly reported that decreased serum albumin increased mortality, ICU and hospital stay.4 Hypoalbuminemia (serum albumin < 30 g/L)5 is common in critically ill patients, arising from redistribution between the vascular and interstitial compartments, nutritional deficits, increased albumin losses, and impaired hepatic synthesis.6,7
Cancer patients differ from general ICU populations in ways that can alter prognostic associations. First, cancer cachexia and chronic malnutrition are common, worsening functional status, increasing susceptibility to treatment toxicity, and are associated with a 20–30% increase in cancer‑related mortality.8 Second, cancer patients are also predisposed to coagulation abnormalities, tumorlysis syndrome, mass effect compression, and metabolic disturbances. Surgical complications, sequelae of chemo‑/radiotherapy, profound myelosuppression, and organ dysfunction further worsen prognosis.9 Third, paraneoplastic syndromes paraneoplastic syndromes can selectively compromise vital functions, including respiratory, cardiac, and neurologic systems, via immune‑mediated mechanisms such as Lambert–Eaton myasthenic syndrome and paraneoplastic encephalitis.10 Consequently, these cancer patients are admitted to the ICU for various reasons, including sepsis, acute kidney injury (AKI), and acute respiratory distress Syndrome (ARDS).11
These cancer‑specific factors, which contribute to hypoalbuminemia, are associated with increased risk of cancer-related mortality.12–16 For instance, in patients with gastric cardia adenocarcinoma, normal serum albumin levels correlate with significantly higher 5-year survival rates compared to abnormally low levels (38.4% vs 19.1%, P < .0003).15 Similarly, in metastatic melanoma patients receiving PD-1 blockade therapy, hypoalbuminemia has been identified as an independent biomarker of poor prognosis, with significant associations with both overall survival and progression-free survival.17 Furthermore, research indicates that after adjusting for confounders, each 10 g/L decrease in albumin is an independent predictor of increased cancer-related mortality and poorer survival.18 Thus, hypoalbuminemia appears to be an independent risk factor for adverse outcomes.
However, the clinical utility of serum albumin as a prognostic biomarker in critical care cancer patients remains to be fully elucidated. To examine the association between serum albumin and all-cause mortality (ACM), this study leveraged data from the Medical Information Mart for Intensive Care (MIMIC-IV, v3.1) database and further evaluated this relationship using a machine-learning prediction model. Our findings may provide a foundation for early risk stratification and guide optimization of therapeutic strategies, ultimately improving outcomes for critically ill cancer patients.
Methods
Data source
This retrospective observational study used data from the publicly available MIMIC database, which was accessed after completion of the required NIH Protecting Human Research Participants course and examination. Data for this study were obtained from the MIMIC-IV database (version 3.1), which contains the medical records of all patients admitted to the Beth Israel Deaconess Medical Center. As this study utilized de-identified retrospective data from a publicly available database, neither informed consent nor additional ethical approval was required in accordance with institutional guidelines. An independent validation cohort was derived from cancer patients admitted to the ICU of Tianjin Medical University Cancer Institute & Hospital between 2019 and 2023; this database is de-identified and has institutional ethical approval (bc20253150).
Study population
The diagnosis of malignant solid tumors was performed by manually reviewing ICD-9 and ICD-10 codes (details of diagnostic codes are presented in Supplementary Table S1, available as supplementary data at [JAMIA Open] online). ICD codes are diagnostic codes assigned by clinicians at patient discharge. These codes have been standardized for billing, administrative purposes, epidemiological research, and accurately reflect the final diagnosis for each hospitalization.19 Exclusion criteria were as follows: (1) Age < 18 years at first admission; (2) No diagnosis of solid malignancy (patients with lymphoma or leukemia were also excluded); (3) ICU stay < 24 hours or no ICU admission; (4) Missing serum albumin measurement within the first ICU day. For patients with multiple ICU admissions, only the first ICU stay was analysed. A flow diagram of the patient inclusion process is shown in Figure 1.
Figure 1.
Flowchart for participants in MIMIV-IV(v3.1).
Data extraction
Data were extracted using PostgreSQL 13.20 The following covariates were obtained within 24 hours after ICU admission: (1) Demographics: age, sex, race, height and weight; (2) Clinical related indicators: hemoglobin (Hb), platelet count (PLT), white blood cell count (WBC), blood urea nitrogen (BUN), C-reactive protein (CRP) and Sequential Organ Failure Assessment (SOFA)scores; (3)Vital signs: heart rate, systolic blood pressure, diastolic blood pressure, respiratory rate and SpO2; (4) Comorbidities: heart failure, kidney failure, liver failure, respiratory failure, transient ischemic attack (TIA), sepsis, cerebral infarction and myocardial infarction; (5) Treatments: use of dialysis, vasopressin, respiratory support, or albumin. Outcome measures included ICU and hospital length of stay, in-hospital mortality, and all-cause mortality at 14, 28, 90 days and at 1 year after admission. Body mass index (BMI) was calculated as weight in kilograms divided by the square of height in meters (kg/m2). For consistency with ICU practice, respiratory support was defined as the initiation of either invasive ventilation (including intubation and tracheostomy) or High-Flow Nasal Cannula within 24 hours of ICU admission. Patients receiving only low‑flow supplemental oxygen or no support were classified as not receiving respiratory support.
Statistical analysis
Participants were stratified into two groups based on serum albumin: < 30 g/L (hypoalbuminemia group) and ≥ 30 g/L (control group). Normality was assessed by visual inspection of histograms and Q–Q plots and by calculating skewness. Variables with absolute skewness < 1 were expressed as means ± standard deviations (SDs), while non-normally distributed variables were reported as medians (Q1–Q3). Categorical variables were summarized as frequencies (percentages). Group comparisons were performed using chi-square test for categorical variables, and continuous variables were analyzed with independent t-tests (for normally distributed data) or Mann-Whitney U tests (for non-normally distributed data).
We used cox proportional hazards regression models to evaluate the association between serum albumin and clinical outcomes, expressed as hazard ratios(HRs) with 95% confidence intervals (CIs). Three progressively adjusted regression models were constructed to account for potential confounding: Model 1 (Unadjusted), Model 2 (adjusted for age, sex, BMI and race), and Model 3 (adjusted for age, sex, BMI, race, sepsis, kidney failure, liver failure, albumin administration, vasopressin use and respiratory support). Kaplan-Meier survival analysis with log-rank testing was performed to compare ACM between the two groups. To characterize the potential nonlinear association between serum albumin and mortality risk, we performed restricted cubic spline (RCS) regression analyses at four clinically relevant time points: 14-day, 28-day, 90-day, and 1-year post-admission. We performed receiver operating characteristic (ROC) curve analysis to assess the predictive performance of serum albumin and SOFA scores for ACM at specified time points (14-day, 28-day, 90-day, and 1-year post-admission), with discrimination assessed by the area under the curve (AUC). Furthermore, we performed prespecified subgroup analyses to evaluate the robustness and consistency of the prognostic value of serum albumin across various subgroups, including age, sex, race, BMI, kidney failure, liver failure, sepsis, vasopressin use and respiratory support. Statistical significance was defined as a two-tailed P < .05 for all analyses. To assess model stability, potential multicollinearity was evaluated using variance inflation factors (VIF). Results indicated no serious multicollinearity issues in this study (VIFs ranged from 1.00 to 1.81, with all VIFs < 2.0, indicating no substantial multicollinearity concerns in our analytical models).
Unlike traditional regression models, machine learning can both capture complex, non-linear predictor interactions and enhance interpretability by ranking feature importance. For machine learning-based prediction, feature importance for 14-day and 28-day mortality was first ranked using the Boruta algorithm. We then used eXtreme Gradient Boosting (XGBoost) to explore the importance of the selected features in predicting mortality, and employed SHapley Additive exPlanations (SHAP) to identify the most critical predictors. Feature importance was calculated by the sum of the reduction in error resulting from variable splits, thereby indicating the contribution each variable to classifying patients into the control group versus hypoalbuminemia group. The hyperparameters were optimized using a systematic grid search with 5-fold cross-validation. The optimal parameters based on the highest mean AUC-ROC across all folds. To prevent overfitting, the maximum tree depth was limited to 4, L1/L2 regularization was set to 1, and the ensemble comprised 50 trees to ensure stability and precision in the predictions. To address the class imbalance, we employed class weights (scale_pos_weight= neg_count/pos_count= 2.49), confirming that our method achieved good predictive performance. Data analysis was conducted using R software (version 4.3.3; Posit Software, PBC), STATA software (version 18.0; Stata Corp), and IBM SPSS software (version 26.0).
Results
Participant characteristics
A total of 1739 patients with malignancy were included. Using a clinically relevant cut-off of 30 g/L for hypoalbuminemia, patients were divided into two groups (serum albumin < 30 g/L vs ≥30 g/L).3,7,21 Overall, the serum albumin < 30 g/L group had a lower proportion of males (P = .023), higher CRP levels (P = .004), and higher SOFA scores (P < .001). In patients with hypoalbuminemia, significantly lower levels of Hb and both systolic and diastolic blood pressures were observed (all P < .05). In addition, BUN, WBC, and heart rate were significantly higher in the hypoalbuminemia group. Compared with the control group, the hypoalbuminemia group exhibited a higher prevalence of kidney failure, liver failure, and sepsis (all P < .001). These patients required more vasopressors and respiratory support (P < .001 and P = .013, respectively). They also exhibited longer ICU (P = .022) and hospital stays (P < .001), higher in-hospital mortality, and higher 14-day, 28-day, 90-day, and 1-year mortality rates (all P < .001). Detailed results are presented in Table 1.
Table 1.
The baseline characteristics of participants.
| Characteristics | Overall | ALB < 30 | ALB ≥ 30 | P |
|---|---|---|---|---|
| (N = 1739) | (N = 832) | (N = 907) | ||
| Agea, years | 66.8 ± 12.6 | 66.6 ± 12.3 | 67.2 ± 12.9 | 0.361 |
| Age ≥ 65 years, (n,%) | 1010 (58.1) | 478 (57.5) | 532 (58.7) | 0.612 |
| BMIa, kg/m2 | 27.4 ± 7.2 | 27.4 ± 7.0 | 27.5 ± 7.4 | 0.730 |
| Male, (n,%) | 1035 (59.5) | 472 (56.7) | 563 (62.1) | 0.023 |
| Race White, (n,%) | 1107 (63.7) | 527 (63.3) | 580 (63.9) | 0.793 |
| CRPb, mg/L | 23.8 (5.2–83.8) | 36.7 (7.3–101.0) | 15.3 (3.8–72.4) | 0.004 |
| SOFAb | 5.0 (2.0–7.0) | 6.0 (3.0–8.0) | 4.0 (2.0–6.0) | <0.001 |
| WBCb, 109/L | 10.8 (7.3–15.5) | 11.8 (7.7–18.4) | 10.0 (7.1–13.8) | <0.001 |
| PLTb, 109/L | 195 (130–279) | 192 (119–284) | 197 (138–276) | 0.077 |
| HBa, g/L | 10.1 ± 2.0 | 9.5 ± 1.7 | 10.7 ± 2.1 | <0.001 |
| BUNb, mg/dL | 21.5 (14.0–36.0) | 24.0 (15.0–40.0) | 19.5 (13.0–30.0) | <0.001 |
| Vital sign | ||||
| Heart ratea, beats/min | 89.8 ± 17.2 | 92.4 ± 17.1 | 87.4 ± 16.9 | <0.001 |
| Systolic blood pressurea, mmHg | 115.0 ± 16.3 | 109.7 ± 13.4 | 120.0 ± 17.2 | <0.001 |
| Diastolic blood pressurea, mmHg | 63.9 ± 11.1 | 61.2 ± 10.0 | 66.4 ± 11.4 | <0.001 |
| Respiratory ratea, times/min | 19.7 ± 4.0 | 19.8 ± 4.2 | 19.5 ± 3.9 | 0.070 |
| SpO2a, % | 96.7 ± 2.0 | 96.7 ± 2.0 | 96.8 ± 2.0 | 0.235 |
| Comorbidities, (n,%) | ||||
| Heart failure | 19 (1.1) | 9 (1.1) | 10 (1.1) | 0.967 |
| Kidney failure | 263 (15.1) | 156 (18.8) | 107 (11.8) | <0.001 |
| Liver failure | 51 (2.9) | 36 (4.3) | 15 (1.7) | 0.001 |
| Respiratory failure | 230 (13.2) | 123 (14.8) | 107 (11.8) | 0.066 |
| TIA | 55 (3.2) | 21 (2.5) | 34 (3.7) | 0.145 |
| Sepsis | 109 (6.3) | 84 (10.1) | 25 (2.8) | <0.001 |
| Brain infarction | 60 (3.5) | 22 (2.6) | 38 (4.2) | 0.078 |
| Heart infarction | 51 (2.9) | 25 (3.0) | 26 (2.9) | 0.865 |
| Treatments | ||||
| ALB treatments, (n,%) | 175 (10.1) | 107 (12.9) | 68 (7.5) | <0.001 |
| Dialysis treatments, (n,%) | 33 (1.9) | 19 (2.3) | 14 (1.5) | 0.259 |
| Vasopressin, (n,%) | 145 (8.3) | 110 (13.2) | 35 (3.9) | <0.001 |
| Respiratory support, (n,%) | 568 (32.7) | 296 (35.6) | 272 (30.0) | 0.013 |
| Outcomes | ||||
| Hospital daysb | 8.3 (5.1–13.7) | 9.2 (5.7–15.9) | 7.5 (4.9–11.8) | <0.001 |
| ICU daysb | 2.5 (1.7–4.2) | 2.6 (1.7–4.3) | 2.4 (1.7–4.0) | 0.022 |
| Hospital mortality, (n,%) | 187 (10.8) | 127 (15.3) | 60 (6.6) | <0.001 |
| 14-days mortality, (n,%) | 356 (20.5) | 236 (28.4) | 120 (13.2) | <0.001 |
| 28-days mortality, (n,%) | 493 (28.3) | 318 (38.2) | 175 (19.3) | <0.001 |
| 90-days mortality, (n,%) | 717 (41.2) | 446 (53.6) | 271 (29.9) | <0.001 |
| 1-year death, (n,%) | 995 (57.2) | 572 (68.8) | 423 (46.6) | <0.001 |
The hypoalbuminemia group had greater illness severity, evidenced by higher SOFA scores and CRP levels, as well as increased rates of liver and kidney failure and a higher prevalence of sepsis.
Values are means ± SDs(standard deviations).
Values are medians (Q1–Q3), and else are proportions for categorical variables.
Normality was assessed by visual inspection of histograms and Q–Q plots and by calculating skewness. Variables with approximately normal distributions (absolute skewness < 1) are presented as means ± SDs; other variables are presented as medians (Q1–Q3). Categorical variables were analyzed using the chi-square test, and continuous variables were analyzed with independent samples t-tests for normally distributed or Mann-Whitney tests for non-normal distributions.
Bold values indicate statistical significance at the threshold of P < 0.05.
Serum albumins and mortality risk
Among 1739 patients, 356 (20.5%) died within 14 days, 493 (28.3%) within 28 days, 717 (41.2%) within 90 days, and 995 (57.2%) within 1 year. The Kaplan–Meier curves showed significantly higher all-cause mortality in the serum albumin < 30 g/L group at each time point (all P < .001, Figure 2). Multivariable Cox regression models confirmed that lower serum albumin was independently associated with increased in-hospital mortality (HR 1.75; 95% CI 1.28–2.39) and at 14-day mortality (HR 2.34; 95% CI 1.88–2.92), 28-day mortality (HR 2.25; 95% CI 1.87–2.71), 90-day mortality (HR 2.22; 95% CI 1.88–2.54), and 1-year mortality (HR 1.94; 95% CI 1.71–2.20) in the unadjusted model. These associations remained robust after progressive adjustment for demographics, BMI, comorbidities, vasopressor use, respiratory support, and SOFA score. Detailed results are presented in Table 2.
Figure 2.
Kaplan-Meier curves for all-cause mortality stratified by albumin groups. Kaplan-Meier curves showing the cumulative incidence of all-cause mortality in patients stratified by normal albumin(ALB group=0) and hypoalbuminemia group(ALB group=1) at 14 days (A), 28 days (B), 90 days (C), and 1 year (D). The log-rank test was used to compare differences between groups.
Table 2.
Cox proportional hazard ratios (HRs) for mortality.
| Model 1 |
Model 2 |
Model 3 |
||||
|---|---|---|---|---|---|---|
| HR(95%CI) | P | HR(95%CI) | P | HR(95%CI) | P | |
| Hospital mortality | ||||||
| ALB(continuous) | 0.63 (0.51–0.79) | <0.001 | 0.66 (0.48–0.90) | 0.009 | 0.74 (0.54–1.00) | 0.050 |
| ALB < 30 | 1.75 (1.28–2.39) | <0.001 | 1.69 (1.11–2.57) | 0.014 | 1.43 (0.93–2.19) | 0.100 |
| ALB ≥ 30 | Reference | Reference | Reference | |||
| 14-days mortality | ||||||
| ALB(continuous) | 0.49 (0.42–0.57) | <0.001 | 0.55 (0.44–0.69) | <0.001 | 0.61 (0.48–0.77) | <0.001 |
| ALB < 30 | 2.34 (1.88–2.92) | <0.001 | 1.95 (1.44–2.62) | <0.001 | 1.67 (1.23–2.27) | <0.001 |
| ALB ≥ 30 | Reference | Reference | Reference | |||
| 28-days mortality | ||||||
| ALB(continuous) | 0.48 (0.42–0.55) | <0.001 | 0.51 (0.42–0.62) | <0.001 | 0.57 (0.46–0.69) | <0.001 |
| ALB < 30 | 2.25 (1.87–2.71) | <0.001 | 2.03 (1.57–2.63) | <0.001 | 1.74 (1.34–2.27) | <0.001 |
| ALB ≥ 30 | Reference | Reference | Reference | |||
| 90-days mortality | ||||||
| ALB(continuous) | 0.48 (0.43–0.54) | <0.001 | 0.50 (0.43–0.59) | <0.001 | 0.54 (0.45–0.63) | <0.001 |
| ALB < 30 | 2.22 (1.88–2.54) | <0.001 | 2.09 (1.69–2.58) | <0.001 | 1.88 (1.51–2.33) | <0.001 |
| ALB ≥ 30 | Reference | Reference | Reference | |||
| 1-year mortality | ||||||
| ALB(continuous) | 0.53(0.48–0.58) | <0.001 | 0.55(0.48–0.63) | <0.001 | 0.58(0.50–0.67) | <0.001 |
| ALB < 30 | 1.94(1.71–2.20) | <0.001 | 1.81(1.53–2.15) | <0.001 | 1.66(1.39–1.99) | <0.001 |
| ALB ≥ 30 | Reference | Reference | Reference | |||
Hazard Ratios for continuous ALB are presented per 1 g/dL (equivalent to 10 g/L) increase.
Model 1: Unadjusted.
Model 2: Adjusted age, gender, BMI and ethnicity.
Model 3: Adjusted age, gender, BMI, ethnicity, sepsis, kidney failure, liver failure, ALB treatments, vasopressin and respiratory support.
Bold values indicate statistical significance at the threshold of P < 0.05.
To assess the continuous association between the serum albumin and ACM, performed the RCS analysis in Figure 3. RCS analysis revealed a nonlinear inverse relationship between serum albumin and mortality risk after multivariable adjustment, with the steepest mortality reduction observed at serum albumin concentrations < 30 g/L.
Figure 3.
RCS of serum albumin and all-cause mortality. Restricted cubic spline curves modeling the association between serum albumin and all-cause mortality in ICU cancer patients at 14 days (A), 28 days (B), 90 days (C), and 1 year (D). The curves were adjusted for age (< 65 years old vs. ≥ 65 years old), gender (male/female), race (White/no White), BMI (< 28 vs. ≥ 28 kg/m2), renal failure (yes/no), hepatic failure (yes/no), sepsis (yes/no), vasopressin (yes/no) and respiratory support (yes/no). Solid lines represent the hazard ratios (HRs), and shaded areas indicate the 95% CIs. The median albumin level in the study population was used as the reference (HR = 1.0).
Forecasting all-cause mortality in cancer patients using serum albumin
We compared serum albumin and SOFA for predicting ACM in critically ill cancer patients by plotting ROC curves in Figure 4. Serum albumin achieved AUCs of 0.654 at 14 days, 0.664 at 28 days, 0.676 at 90 days, and 0.664 at 1 year. Albumin demonstrates significantly better predictive performance than SOFA score for 90-day and 1-year mortality (DeLong’s test p < .05). However, for 14-day and 28-day mortality prediction, both of them show comparable discriminative ability with no statistically significant differences (p > .05). These findings underscore albumin’s superior predictive value for long-term mortality and its substantial clinical utility.
Figure 4.
ROC curves of albumin and SOFA for predicting mortality of cancer patients. ROC curves comparing the predictive performance of serum albumin and SOFA score for all-cause mortality in ICU cancer patients at 14 days (A), 28 days (B), 90 days (C), and 1 year (D). The area under the curve (AUC) with 95% CIs were calculated for each predictor. Pairwise comparisons of AUCs were performed using the DeLong test. Extremely small DeLong test p‑values are reported as p < 0.001 when below the software’s display/precision limit (90‑day p = 3.0 × 10 − 4; 1‑year p was reported by the software as 0).
Subgroup analysis
We further investigated the association between serum albumin and ACM at 14 days and 28 days across various subgroups of cancer patients. Subgroup analysis showed that each 10 g/L increase in serum albumin was associated with a reduction in 14-day and 28-day mortality. When stratified by age, sex, race, BMI, kidney or liver failure, sepsis, vasopressor use, and respiratory support, We observed no significant interactions between serum albumin and all subgroups (P > .05, Table 3). These results further substantiate the robustness of our conclusions.
Table 3.
Stratified analyses of ALB and 14-day, 28-day mortality.
| ALB with 14-day mortality |
ALB with 28-day mortality |
|||
|---|---|---|---|---|
| Subgroup | Adjusteda OR (95% CI) | P for interactionb | Adjusteda OR (95% CI) | P for interactionb |
| Age(years) | 0.760 | 0.881 | ||
| Age < 65 | 0.62(0.40–0.93) | 0.51(0.35–0.74) | ||
| Age ≥ 65 | 0.53(0.36–0.76) | 0.48(0.34–0.66) | ||
| Gender | 0.484 | 0.821 | ||
| Male | 0.52(0.36–0.74) | 0.47(0.34–0.65) | ||
| Female | 0.62(0.40–0.96) | 0.51(0.34–0.77) | ||
| Race | 0.197 | 0.094 | ||
| Non White | 0.44(0.27–0.70) | 0.38(0.24–0.57) | ||
| White | 0.64(0.46–0.90) | 0.58(0.42–0.79) | ||
| BMI(kg/m2) | 0.540 | 0.465 | ||
| BMI < 28 | 0.60(0.43–0.83) | 0.47(0.34–0.65) | ||
| BMI ≥ 28 | 0.47(0.28–0.77) | 0.50(0.32–0.77) | ||
| Kidney failure | 0.831 | 0.236 | ||
| Yes | 0.41(0.21–0.78) | 0.52(0.28–0.93) | ||
| No | 0.59(0.43–0.80) | 0.48(0.36–0.63) | ||
| Liver failure | 0.887 | 0.921 | ||
| Yes | 0.21(0.02–1.04) | 0.30(0.05–1.09) | ||
| No | 0.57(0.43–0.75) | 0.50(0.39–0.65) | ||
| Sepsis | 0.478 | 0.884 | ||
| Yes | 0.16(0.02–0.65) | 0.07(0.01–0.31) | ||
| No | 0.58(0.43–0.77) | 0.53(0.41–0.69) | ||
| Vasopressin | 0.458 | 0.394 | ||
| Yes | 0.87(0.43–1.76) | 0.67(0.33–1.32) | ||
| No | 0.50(0.37–0.67) | 0.45(0.34–0.59) | ||
| Respiratory support | 0.243 | 0.199 | ||
| Yes | 0.62(0.41–0.92) | 0.64(0.44–0.92) | ||
| No | 0.49(0.33–0.71) | 0.39(0.27–0.55) | ||
Adjusted ORs were calculated per 1 g/dL increase in ALB. Each stratification was adjusted for age(< 65 years old vs ≥ 65 years old), gender(male/female), race(White/no White), BMI(< 28 vs ≥ 28 kg/m2), renal failure(yes/no), hepatic failure(yes/no), sepsis(yes/no), vasopressin(yes/no) and respiratory support (yes/no). OR: odds ratios.
Used the likelihood ratio test comparing models with and without an interaction term.
External validation
The validation cohort comprised 505 patients from Tianjin Medical University Cancer Institute and Hospital, their baseline characteristics are presented in Supplementary Table S2, available as supplementary data at [JAMIA Open] online. As serum albumin increased, the RCS analysis revealed a downward trend in ICU mortality and 28-day mortality (Supplementary Figure S1, available as supplementary data at [JAMIA Open] online).
Machine-learning model prediction
Variables ultimately included in the machine-learning model were identified through Boruta analysis, with variable importance ranked in descending order from right to left. The algorithm selected 16 and 14 features as the optimal predictors of 14-day and 28-day mortality, respectively, from the full patient cohort. The results of the Boruta analysis across the entire population are presented in Figure 5. Figure 6 presents the scatter plots of risk factors associated with mortality and the bar chart of mean importance in the optimal model. BUN, serum albumin, respiratory rate, heart rate and SOFA were identified as the five most relevant variables. These results indicate that serum albumin demonstrates superior predictive value for 14-day and 28-day mortality in critically ill cancer patients.
Figure 5.
Boruta algorithm feature importance ranking for predicting mortality of cancer patients. Feature importance ranking of potential risk factors for 14-day (A) and 28-day (B) mortality in cancer patients using the Boruta algorithm. Box plots show the Z-scores for each feature across multiple random forest iterations. The purple box plots represent shadow features (min, mean, and max), which are used as a reference to determine importance. Features with importance scores higher than the maxi shadow feature are considered comfirmed important (green boxes).
Figure 6.
SHAP interpretation of XGBoost models for predicting 14-day and 28-day mortality. SHAP summary plots for the 14-day (A and C) and 28-day (B and D) mortality prediction models in critically ill cancer patients. For A and B, the length of each bar represents the mean absolute SHAP value of the corresponding feature, indicating its average impact on the model output magnitude. Features with larger mean |SHAP| values are more influential in the prediction process. For A and B, each point represents the SHAP value of a feature for an individual patient.
Discussion
This retrospective cohort study examined the association between hypoalbuminemia and ACM in critically ill cancer patients. Serum albumin < 30 g/L was independently associated with increased ACM at 14 days, 28 days, 90 days and 1 year, even after adjustment for potential confounders. The Kaplan–Meier analysis, Cox proportional hazards models and RCS analysis revealed significantly higher mortality in patients with serum albumin < 30 g/L at all time points. Furthermore, ROC curve analysis revealed that serum albumin had superior discriminative ability compared to the SOFA score for long-term mortality prediction. Subgroup analyses confirmed the robustness of these findings. XGBoost machine learning models and our external validation within a Chinese cohort corroborated these findings, underscoring the importance of serum albumin as a prognostic marker in critically ill cancer patients.
In 2011, Gomez et al. observed that among 200 critically ill cancer patients, those with serum albumin < 20 g/L exhibited the highest mortality (∼73%) and elevated SOFA scores. These patients more frequently developed sepsis and required vasopressor use and respiratory support, suggesting that hypoalbuminemia exacerbates organ dysfunction and increases infection rates.22 Albumin loss results from haemodilution during resuscitation and capillary leakage proportional to the inflammatory response; patients with the greatest vascular permeability have the worst prognosis.23
The strong association between hypoalbuminemia and mortality in critically ill cancer patients can be explained by several cancer-specific mechanisms. Cancer cachexia syndrome induces negative protein balance through increased catabolism and systemic inflammation, directly depleting visceral proteins including albumin.24 As a negative acute-phase reactant, albumin synthesis is suppressed by inflammatory cytokines (TNF-α, IL-6) that are elevated in malignancy,25,26 explaining the inverse relationship with CRP observed in our cohort. Impaired hepatic synthesis due to liver metastases or treatment-related hepatotoxicity represents another cancer-specific pathway. Patients with extensive liver involvement or chemotherapy-induced injury have reduced synthetic capacity.27 Ultimately, hypoalbuminemia may exacerbates clinical outcomes through the loss of its physiological functions: reduced oncotic pressure promotes tissue edema,28,29 while diminished antioxidant, immunomodulatory, and drug-binding properties may increase susceptibility to complications and reduce tolerance to anticancer therapies.30,31 Traditionally, various scores such as Acute Physiology and Chronic Health Evaluation (APACHE II), the Simplified Acute Physiology Score (SAPS II), and SOFA have been used to predict mortality risk in critically ill cancer patients and to guide ICU clinical decisions.32 In our study, serum albumin achieved AUCs of 0.676, and 0.664 at 90 days and 1 year, significantly outperforming SOFA (DeLong test p < .001). Our model demonstrates that serum albumin has modestly better discrimination ability than SOFA for predicting ACM. Several factors may explain this finding. First, SOFA may be confounded by baseline cancer-related organ dysfunction that is not acutely reversible (e.g., respiratory impairment in lung cancer, elevated bilirubin in liver metastases), whereas albumin reflects the composite burden of inflammation, catabolism, and nutritional status that is universally relevant across cancer types. Second, unlike certain SOFA components such as the Glasgow Coma Scale, albumin measurement is not subject to inter-rater variability or subjective interpretation.
Moreover, BUN, serum albumin, respiratory rate, heart rate, and SOFA were identified as the five most relevant variables. Among these, respiratory rate and heart rate are components of APACHE II,33 while BUN is included in SAPS II.34 The SOFA score quantifies dysfunction across six organ systems: respiratory, cardiovascular, renal, hematologic, hepatic, and neurologic.35 Consequently, integrating serum albumin with established scoring systems warrants further investigation for prognostic assessment in critically ill cancer patients.
As the principal determinant of plasma oncotic pressure (∼75%), albumin is commonly used for volume resuscitation.36 It also serves as a carrier for endogenous and exogenous compounds, and functions as an antioxidant, buffer, immunomodulator, and detoxifier.37 However, the benefit of albumin resuscitation in critically ill cancer patients remains controversial. Norton et al. found lower mortality in sepsis patients receiving 4% albumin,38 suggesting that maintaining serum albumin > 30 g/L may confer survival benefit. Conversely, Hyesuk et al. reported no improvement in 7-day survival when albumin was added to lactated Ringer’s solution in cancer patients with sepsis.39 These discordant findings reflect heterogeneity across studies and support individualized albumin therapy; management of hypoalbuminemia should therefore consider underlying aetiology and pathophysiology alongside comprehensive therapy.
This study is the first to investigate the relationship between serum albumin and prognosis in critically ill cancer patients using the large MIMIC-IV database and to externally validate the findings in a Chinese cohort. However, several limitations should be acknowledged. First, residual confounding is possible despite multivariable adjustment, as the MIMIC database lacks key cancer-specific variables including chemotherapy status, cancer stage, and metastatic sites. Second, albumin was measured only at ICU admission; dynamic changes over time may provide additional prognostic information. Prospective studies with longitudinal monitoring are therefore warranted. Third, the use of all-cause mortality precludes distinction between cancer-related and non-cancer causes of death. Finally, although external validation was performed in a Chinese cohort, generalizability to other non-Western populations remains limited due to differences in healthcare systems and patient demographics.
Conclusion
This study demonstrates a significant association between hypoalbuminemia and ACM in critically ill cancer patients. Serum albumin < 30g/L was independently associated with increased ACM at 14 days, 28 days, 90 days and 1 year, even after adjustment for potential confounders. Serum albumin measured at ICU admission can serve as a clinical biomarker to identify high-risk patient groups.
Supplementary Material
Acknowledgements
We are grateful to the MIMIV-IV participants and staff. This work was funded by Scientific and Technological Project of Tianjin (24ZXGZSY00020, 24ZXZSSS00050).
Contributor Information
Guiyue Wang, Department of Anesthesiology, Tianjin Medical University Cancer Institute and Hospital, National Clinical Research Center for Cancer, Tianjin, 300060, PR China; Tianjin’s Clinical Research Center for Cancer, Tianjin, 300060, PR China.
Limei Yuan, Department of Anesthesiology, Tianjin Medical University Cancer Institute and Hospital, National Clinical Research Center for Cancer, Tianjin, 300060, PR China; Tianjin’s Clinical Research Center for Cancer, Tianjin, 300060, PR China; Tianjin Medical University Cancer Institute and Hospital, State Key Laboratory of Druggability Evaluation and Systematic Translational Medicine, Tianjin, 300060, PR China.
Zhenguo Song, Department of Anesthesiology, Tianjin Medical University Cancer Institute and Hospital, National Clinical Research Center for Cancer, Tianjin, 300060, PR China; Tianjin’s Clinical Research Center for Cancer, Tianjin, 300060, PR China.
Ying Shen, Department of Anesthesiology, Shanghai Eighth People’s Hospital, Shanghai, 200072, China.
Jiaxu Li, Department of Anesthesiology, Tianjin Medical University Cancer Institute and Hospital, National Clinical Research Center for Cancer, Tianjin, 300060, PR China; Tianjin’s Clinical Research Center for Cancer, Tianjin, 300060, PR China; Tianjin Medical University Cancer Institute and Hospital, State Key Laboratory of Druggability Evaluation and Systematic Translational Medicine, Tianjin, 300060, PR China.
Xiaobei Zhang, Department of Anesthesiology, Tianjin Medical University Cancer Institute and Hospital, National Clinical Research Center for Cancer, Tianjin, 300060, PR China; Tianjin’s Clinical Research Center for Cancer, Tianjin, 300060, PR China.
Yuan Li, Department of Anesthesiology, Tianjin Medical University Cancer Institute and Hospital, National Clinical Research Center for Cancer, Tianjin, 300060, PR China; Tianjin’s Clinical Research Center for Cancer, Tianjin, 300060, PR China.
Kaili Yu, Department of Anesthesiology, Tianjin Medical University Cancer Institute and Hospital, National Clinical Research Center for Cancer, Tianjin, 300060, PR China; Tianjin’s Clinical Research Center for Cancer, Tianjin, 300060, PR China.
Chengqi Deng, Department of Anesthesiology, Tianjin Medical University Cancer Institute and Hospital, National Clinical Research Center for Cancer, Tianjin, 300060, PR China; Tianjin’s Clinical Research Center for Cancer, Tianjin, 300060, PR China.
Minhui Yi, Department of Obstetrics and Gynecology, The Seventh People’s Hospital of Shanghai University of Traditional Chinese Medicine, Shanghai, 200137, China.
Kaiyuan Wang, Department of Anesthesiology, Tianjin Medical University Cancer Institute and Hospital, National Clinical Research Center for Cancer, Tianjin, 300060, PR China; Tianjin’s Clinical Research Center for Cancer, Tianjin, 300060, PR China; Tianjin Medical University Cancer Institute and Hospital, State Key Laboratory of Druggability Evaluation and Systematic Translational Medicine, Tianjin, 300060, PR China.
Huiqin Mo, Department of Obstetrics and Gynecology, The Seventh People’s Hospital of Shanghai University of Traditional Chinese Medicine, Shanghai, 200137, China.
Author contributions
Gui-yue Wang (Conceptualization, Data curation, Formal analysis, Resources, Validation, Writing—original draft), Limei Yuan (Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing—original draft, Writing—review & editing), Zhenguo Song(Data curation, Investigation, Resources), Ying Shen(Data curation, Investigation, Resources), Jiaxu Li(Methodology, Resources, Software), Xiaobei Zhang(Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Writing—original draft), Yuan Li(Conceptualization, Data curation, Formal analysis, Resources), Kaili Yu(Conceptualization, Data curation, Writing—original draft), Chengqi Deng(Conceptualization, Data curation, Formal analysis, Writing—original draft), Minhui Yi(Conceptualization, Data curation, Resources, Validation), and Kaiyuan Wang(Conceptualization, Data curation, Funding acquisition, Investigation, Methodology, Resources, Software, Writing—original draft, Writing—review & editing), Huiqin Mo (Conceptualization, Data curation, Formal analysis, Methodology, Resources, Software, Writing—original draft, Writing—review & editing)
Ethical approval
The study was carried out in accordance with the Declaration of Helsinki. Due to the de-identification of the MIMIC repository, sensitive data is not involved, so we informed the Ethics Committee of this situation without a written report. validation cohort was derived from cancer patients admitted to the ICU of Tianjin Medical University Cancer Institute & Hospital between 2019 and 2023. The database is de-identified and has institutional ethical approval (bc20253150).
Supplementary material
Supplementary material is available at [JAMIA Open] online.
Conflicts of interest
None declared.
Funding
This work was funded by Scientific and Technological Project of Tianjin [24ZXGZSY00020, 24ZXZSSS00050].
Data availability
The datasets used and analysed during the current study are available from the corresponding author on reasonable request.
Reference
- 1. Bray F, Laversanne M, Sung HYA, et al. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. 2024;74:229-263. [DOI] [PubMed] [Google Scholar]
- 2. Darmon M, Bourmaud A, Georges Q, et al. Changes in critically ill cancer patients’ short-term outcome over the last decades: results of systematic review with meta-analysis on individual data. Intensive Care Med. 2019;45:977-987. [DOI] [PubMed] [Google Scholar]
- 3. Vincent JL, Dubois MJ, Navickis RJ, et al. Hypoalbuminemia in acute illness: is there a rationale for intervention? A meta-analysis of cohort studies and controlled trials. Ann Surg. 2003;237:319-334. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4. Padkins M, Breen T, Anavekar N, et al. Association between albumin level and mortality among cardiac intensive care unit patients. J Intensive Care Med. 2021;36:1475-1482. [DOI] [PubMed] [Google Scholar]
- 5. Dubois MJ, Orellana-Jimenez C, Melot C, et al. Albumin administration improves organ function in critically ill hypoalbuminemic patients: a prospective, randomized, controlled, pilot study. Crit Care Med. 2006;34:2536-2540. [DOI] [PubMed] [Google Scholar]
- 6. Liumbruno G, Bennardello F, Lattanzio A, et al. Recommendations for the use of albumin and immunoglobulins. Blood Transfusion. 2009;7:216-234. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7. Saner FH, Stueben B-O, Hoyer DP, et al. Use or misuse of albumin in critical ill patients. Diseases (Basel, Switzerland). 2023;11:68. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8. Tan YT, Xue R, Pan YW, et al. Cancer cachexia: molecular basis and therapeutic advances. Signal Transduct Target Ther. 2026;11:16. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9. Reddy DRS, Botz GH. Triage and prognostication of cancer patients admitted to the intensive care unit. Crit Care Clin. 2021;37:1. [DOI] [PubMed] [Google Scholar]
- 10. Graus F, Vogrig A, Muñiz-Castrillo S, et al. Updated diagnostic criteria for paraneoplastic neurologic syndromes. Neurol-Neuroimmunol Neuroinflammation. 2021;8:e1014. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11. Schellongowski P, Sperr WR, Wohlfarth P, et al. Critically ill patients with cancer: chances and limitations of intensive care medicine-a narrative review. ESMO Open. 2016;1:e000018. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12. Phillips A, Shaper AG, Whincup PH. Association between serum albumin and mortality from cardiovascular disease, cancer, and other causes. Lancet (London, England). 1989;2:1434-1436. [DOI] [PubMed] [Google Scholar]
- 13. Ouyang XJ, Dang Y, Zhang F, et al. Low serum albumin correlates with poor survival in gastric cancer patients. Clin Lab. 2018;64:239-245. [DOI] [PubMed] [Google Scholar]
- 14. Lim WS, Roh JL, Kim SB, et al. Pretreatment albumin level predicts survival in head and neck squamous cell carcinoma. Laryngoscope. 2017;127:E437-E442. [DOI] [PubMed] [Google Scholar]
- 15. Lien YC, Hsieh CC, Wu YC, et al. Preoperative serum albumin level is a prognostic indicator for adenocarcinoma of the gastric cardia. J Gastrointest Surg. 2004;8:1041-1048. [DOI] [PubMed] [Google Scholar]
- 16. Yang ZY, Zheng YD, Wu Z, et al. Association between pre-diagnostic serum albumin and cancer risk: results from a prospective population-based study. Cancer Med. 2021;10:4054-4065. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17. Leek LVM, Notohardjo JCL, de Joode K, et al. Multi-omic analysis identifies hypoalbuminemia as independent biomarker of poor outcome upon PD-1 blockade in metastatic melanoma. Sci Rep. 2024;14:11244. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18. Tang Q, Li X, Sun CR. Predictive value of serum albumin levels on cancer survival: a prospective cohort study. Front Oncol. 2024;14:1323192. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19. Johnson AEW, Bulgarelli L, Shen L, et al. MIMIC-IV, a freely accessible electronic health record dataset. Sci Data. 2023;10:219. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20. Zoellner H, Hou JY, Lovery M, et al. Inhibition of microvascular endothelial apoptosis in tissue explants by serum albumin. Microvasc Res. 1999;57:162-173. [DOI] [PubMed] [Google Scholar]
- 21. China L, Freemantle N, Forrest E, et al. A randomized trial of albumin infusions in hospitalized patients with cirrhosis. N Engl J Med. 2021;384:808-817. [DOI] [PubMed] [Google Scholar]
- 22. Ñamendys-Silva SA, González-Herrera MO, Texcocano-Becerra J, et al. Hypoalbuminemia in critically III patients with cancer: incidence and mortality. Am J Hosp Palliat Care. 2011;28:253-257. [DOI] [PubMed] [Google Scholar]
- 23. Brenner DA, Buck M, Feitelberg SP, et al. Tumor necrosis factor-alpha inhibits albumin gene expression in a murine model of cachexia. J Clin Invest. 1990;85:248-255. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24. Wyart E, Carrà G, Angelino E, et al. Systemic metabolic crosstalk as driver of cancer cachexia. Trends Endocrinol Metabol. 2025;36:815-826. [DOI] [PubMed] [Google Scholar]
- 25. McFarland DC, Breitbart W, Miller AH, et al. Depression and inflammation in patients with lung cancer: a comparative analysis of acute phase reactant inflammatory markers. Psychosomatics. 2020;61:527-537. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26. Gremese E, Bruno D, Varriano V, et al. Serum albumin levels: a biomarker to be repurposed in different disease settings in clinical practice. J Clin Med. 2023;12:6017. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27. Núñez KG, Sandow T, Patel J, et al. Hypoalbuminemia is a hepatocellular carcinoma independent risk factor for tumor progression in low-risk bridge to transplant candidates. Cancers (Basel). 2022;14:17. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28. Yu YT, Liu J, Hu B, et al. Expert consensus on the use of human serum albumin in critically ill patients. Chin Med J (Engl). 2021;134:1639-1654. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29. Allison SP, Lobo DN. The clinical significance of hypoalbuminaemia. Clin Nutr. 2024;43:909-914. [DOI] [PubMed] [Google Scholar]
- 30. Soeters PB, Wolfe RR, Shenkin A. Hypoalbuminemia: pathogenesis and clinical significance. JPEN J Parenter Enteral Nutr. 2019;43:181-193. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31. Wiedermann CJ. Hypoalbuminemia as surrogate and culprit of infections. Int J Mol Sci. 2021;22:25. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32. Pirracchio R, Petersen ML, Carone M, et al. Mortality prediction in intensive care units with the super ICU learner algorithm (SICULA): a population-based study. Lancet Respir Med. 2015;3:42-52. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33. Giangiuliani G, Mancini A, Gui D. Validation of a severity of illness score (APACHE II) in a surgical intensive care unit. Intensive Care Med. 1989;15:519-522. [DOI] [PubMed] [Google Scholar]
- 34. Le Gall JR, Lemeshow S, Saulnier F. A new simplified acute physiology score (SAPS II) based on a european/North American multicenter study. JAMA. 1993;270:2957-2963. [DOI] [PubMed] [Google Scholar]
- 35. Vincent JL, Moreno R, Takala J, et al. The SOFA (sepsis-related organ failure assessment) score to describe organ dysfunction/failure. On behalf of the working group on sepsis-related problems of the European Society of Intensive Care Medicine. Intensive Care Med. 1996;22:707-710. [DOI] [PubMed] [Google Scholar]
- 36. Weil MH, Henning RJ, Puri VK. Colloid oncotic pressure: clinical significance. Crit Care Med. 1979;7:113-116. [DOI] [PubMed] [Google Scholar]
- 37. Quinlan GJ, Margarson MP, Mumby S, et al. Administration of albumin to patients with sepsis syndrome: a possible beneficial role in plasma thiol repletion. Clin Sci (Lond). 1998;95:459-465. [PubMed] [Google Scholar]
- 38. Finfer S, Bellomo R, Boyce N, et al. A comparison of albumin and saline for fluid resuscitation in the intensive care unit. N Engl J Med. 2004;350:2247-2256. [DOI] [PubMed] [Google Scholar]
- 39. Park CHL, de Almeida JP, de Oliveira GQ, et al. Lactated ringer’s versus 4% albumin on lactated ringer’s in early sepsis therapy in cancer patients: a pilot Single-Center randomized trial. Crit Care Med. 2019;47:E798-E805. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
The datasets used and analysed during the current study are available from the corresponding author on reasonable request.