Lactate/albumin ratio predicts mortality in critically ill COVID-19 patients: a retrospective machine learning study

Abstract

Severe COVID-19 often progresses to critical illness, requiring accurate prognostic biomarkers. Lactate-to-albumin ratio (LAR) has been proposed as a novel indicator to estimate the likelihood of death. Using data from the MIMIC database, this retrospective study assessed lactate-to-albumin ratio (LAR) effectiveness in forecasting outcomes among severely ill COVID-19 patients. Patients were grouped into four quartiles based on their lactate-to-albumin ratio (LAR) values. Analysis using the Kaplan–Meier method revealed a clear difference in survival outcomes among the groups, with individuals with higher levels of LAR indicating a higher observed mortality rate. The RCS analysis identified a distinct nonlinear link between LAR values and overall 28-day death rates (P < 0.001), which retained statistical significance after covariate adjustment (P < 0.001). Multivariate Cox regression verified that LAR independently correlates with 28-day death risk (HR = 1.309, 95% CI: 1.113–1.540). Subgroup analyses consistently indicated increased mortality risks in Q4 across most strata. Both the Boruta and LASSO algorithms identified LAR as a key determinant of 28-day mortality. Among the machine learning models evaluated, the Random Survival Forest (RSF) model demonstrated strong overall predictive performance for 14-day (AUC: 0.948) and 28-day (AUC: 0.887) mortality prediction in both the training and test datasets. The prediction model incorporating LAR achieved outstanding results across multiple algorithmic methods, serving as an effective and simple clinical tool that enables risk stratification and guides therapeutic decision-making through the integration of multidimensional parameters.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-20468-x.

Introduction

Since it first appeared in late 2019, the SARS-CoV-2 virus has resulted in roughly 777 million recorded infections around the world, contributing to a global death toll of over 7 million and an overall case fatality rate near 0.91%^1,2. Critical care epidemiology shows a distinct stratification of disease severity: 14% of confirmed infections are classified as severe and 5% as critical, with critical cases exhibiting a mortality rate of 63.5%^3,4. Notably, 5% of infected individuals with comorbidities require admission to the ICU^3,5. ICU management of severe COVID-19 necessitates continuous monitoring due to the clinical progression characterized by multi-organ involvement. Rapid progression of the disease can give rise to complications such as acute respiratory distress syndrome (ARDS), systemic inflammatory response syndrome (SIRS), and in severe instances, advance to multiple organ dysfunction syndrome (MODS)^3,6,7. Intensive care clinical decision-making faces numerous challenges, including rapid disease evolution necessitating time-sensitive interventions, multifactorial pathogenesis that complicates prognostic assessments, and the lack of standardized multidimensional outcome prediction metrics. These barriers hinder the creation and refinement of effective treatment protocols tailored for patients in critical condition.

For patients experiencing critical illness, the lactate-to-albumin ratio (LAR) has gained recognition as a novel integrated marker reflecting both systemic inflammation and metabolic burden.The underlying mechanisms involve the interplay of elevated lactate and hypoalbuminemia, each with distinct yet complementary pathophysiological implications⁸. Lactate accumulation is primarily driven by cellular hypoxia and a shift to anaerobic glycolysis under conditions of oxygen supply-demand imbalance⁹. Moreover, impaired hepatic and renal perfusion reduces lactate clearance, and inflammatory cytokine storms disrupt mitochondrial function, further exacerbating lactate levels^10,11. Concurrently, hypoalbuminemia reflects suppressed hepatic synthesis mediated by interleukins such as IL-6 and TNF-α, increased capillary permeability due to endothelial injury, and oxidative stress-induced albumin consumption. Hypoalbuminemia reflects the inhibition of liver synthesis mediated by interleukins (such as IL-6, TNF -α), increased capillary permeability caused by endothelial damage, and albumin depletion induced by oxidative stress. Lactic acid amplifies the inflammatory cascade by activating hypoxia inducible factor-1 alpha (HIF-1α), while hypoalbuminemia weakens antioxidant defense and exacerbates tissue damage¹². By integrating these two indicators, LAR can simultaneously capture acute metabolic disorders and chronic nutritional reserve status. Therefore, in complex clinical scenarios such as COVID-19 critical illness, LAR may have better prognostic predictive ability than a single indicator.

This research utilized data from the MIMIC-IV database to construct and examine a predictive framework for assessing 28-day mortality in patients suffering from severe COVID-19, with particular attention given to the lactate-to-albumin ratio (LAR) as a standalone prognostic indicator.

Data and methods

This research utilized the MIMIC-IV version 3.1 dataset, a publicly available resource curated by the MIT Laboratory for Computational Physiology. The current release, updated in October 2024, comprises comprehensive records of intensive care unit admissions spanning from 2008 to 2022, incorporating data related to COVID- 19 cases. To obtain access, the investigators successfully completed the required Collaborative Institutional Training Initiative (CITI) certification and passed the corresponding assessment (ID: 61651777). All patient data within the MIMIC-IV database have been anonymized, ensuring privacy protection. As the study relies solely on this de-identified dataset, it received exemption from institutional review board (IRB) approval and did not require individual patient consent. The dataset can be freely accessed through the PhysioNet platform at https://physionet.org/content/mimiciv/3.1/.

Inclusion and exclusion criteria

COVID-19 cases were selected from the MIMIC-IV version 3.1 dataset based on ICU admission records and identified exclusively using ICD-10 diagnostic classifications. Participants were included if they met the following conditions: (1) aged 18 years or older; (2) had a confirmed COVID-19 diagnosis according to ICD-10 codes; and (3) were first-time admissions to both the hospital and intensive care unit. Exclusion criteria comprised the following: (1) individuals who passed away on the same day as their ICU admission; (2) entries with over 20% missing value across key variables; and (3) cases lacking documented measurements of blood lactate and serum albumin within the first 24 h after ICU entry.

Data extraction

Data utilized in this study were extracted from the MIMIC-IV version 3.1 dataset using PostgreSQL (v14.5-1) and Navicat Premium 16. The extraction process covered an extensive array of clinical variables and was executed using Navicat Premium (v16.3.5), SQL scripts, and PostgreSQL utilities (v13.7.1) to obtain raw patient-level information from the initial 24-hour period following ICU admission.Extracted parameters included demographic characteristics (e.g., age, sex), physiological indicators (such as heart rate, systolic/diastolic blood pressure, respiration), and biochemical markers (e.g., PaO2, PaCO2, bilirubin, electrolytes like sodium and potassium, as well as lactate and serum albumin). Coagulation-related tests (APTT, ALT, AST, PT), blood components (platelet counts, WBC, RBC), and administered treatments (e.g., continuous renal replacement therapy, mechanical ventilation, corticosteroids, antivirals, antibiotics, and vasopressors) were also gathered.Information on existing medical conditions (including prior cardiovascular events, diabetes, chronic obstructive pulmonary disease, heart failure, kidney impairment, etc.) and clinical scoring metrics (e.g., SOFA and Charlson Comorbidity Index) was recorded. For any variables recorded repeatedly during the ICU stay, only the initial values within the first day post-admission were analyzed.

Statistical analysis

This retrospective study excluded variables with more than 20% missing data. For variables with fewer than 20% missing values, multiple imputation techniques were applied. To address multicollinearity, the variance inflation factor (VIF) was calculated, and features with VIF values greater than 5 were removed.Participants were categorized into quartiles based on their lactate-to-albumin ratio (LAR). Continuous variables following a normal distribution were expressed as mean ± standard deviation and compared using one-way ANOVA. For variables not normally distributed, non-parametric tests such as the Kruskal–Wallis test were employed. Categorical variables were presented as proportions and analyzed using the chi-squared (χ²) test or Fisher’s exact test, depending on sample characteristics. Use restricted cubic spline (RCS) to explore the dose-response relationship with primary outcome risk. Kaplan–Meier survival analysis were used to compare 28-day survival across the LAR quartiles.Feature selection was performed using the Boruta algorithm, followed by 10k-fold cross-validation with the LASSO method. Cox proportional hazards regression was used to estimate hazard ratios (HRs) and corresponding 95% confidence intervals (CI) for each outcome. The dataset was randomly split into training and validation sets at a 7:3 ratio. Predictive modeling was conducted using eight machine learning algorithms, including random survival forest (RSF), gradient boosting (GBM), LASSO-Cox, CoxBoost, survival support vector machines (survivalSVM), XGBoost, SuperPC, and plsRcox. Model discrimination was evaluated using receiver operating characteristic (ROC) curves, while clinical utility was assessed via decision curve analysis (DCA). All statistical procedures were performed using R software version 4.5.0, P value below 0.05 was interpreted as statistically significant.

Result

This study extracted clinical data from 1,127 COVID-19 patients from the MIMIC-IV database (Fig. 1) and excluded cases that did not meet the inclusion and exclusion criteria. The results of the variance inflation factor (VIF) analysis indicated that there was no significant multicollinearity among the variables (Table.S1).

Fig. 1.

Selection of the study population from the MIMIC-IV database.

Baseline variables

The baseline demographic and clinical characteristics of patients, stratified by LAR quartiles, are presented in Table 1. The median age was 67 years (IQR: 58–78), with patients in the highest quartile being significantly older than those in lower quartiles (p = 0.008). A higher proportion of males was observed in the upper quartiles (p = 0.008). The Charlson Comorbidity Index (CCI) increased progressively with LAR (p < 0.001), and Sequential Organ Fail ure Assessment (SOFA) scores were significantly higher in the top quartile (p = 0.004). Laboratory parameters also differed across quartiles, with elevated white blood cell counts (p < 0.001), reduced red blood cell counts (p = 0.041), lower platelet counts, and altered metabolic markers including decreased albumin and increased lactate levels in the highest LAR group. All baseline variables showed significant differences across LAR quartiles (p < 0.05).

Table 1.

Patient demographics and baseline characteristics.

Variable	Overall N = 593	Q1 N = 148	Q2 N = 148	Q3 N = 148	Q4 N = 149	P value
Age (years)	65 (55,75)	62 (52,73)	64 (52,75)	67 (56,77)	67 (58,78)	0.008
gender, n (p%)						0.008
Female	231 (38.95%)	75 (50.68%)	50 (33.78%)	55 (37.16%)	51 (34.23%)
Male	362 (61.05%)	73 (49.32%)	98 (66.22%)	93 (62.84%)	98 (65.77%)
Weight (kg)	84.3 (71.0,101.2)	87.4 (73.90,102.60)	84.55 (71.80,103.53)	86.15 (70.45,102.25)	79.80 (67.85,97.10)	0.086
Charlson	4.0 (2.0,6.0)	3.0 (1.0,6.0)	4.0 (2.0,5.0)	4.0 (3.0,6.0)	5.0 (3.0,7.0)	< 0.001
Sofa	6.0 (3.0,9.0)	5.5 (3.0,8.0)	6.0 (3.0,8.0)	6.0 (3.0,9.0)	7.0 (4.0,10.0)	0.004
SIRS	3.00 (2.00–3.00)	2.00 (2.00–3.00)	3.00 (2.00–3.00)	3.00 (2.00–3.00)	3.00 (2.00–4.00)	< 0.001
Vital sings
Respiratory rate	23.0 (19.0,28.0)	22.5 (18.0,28.0)	23.0 (19.0,28.0)	24.0 (20.0,28.0)	23.0 (19.0,27.0)	0.645
PO2 (mmHg)	83.72 ± 54.86	86.37 ± 54.47	77.60 ± 43.35	81.59 ± 53.67	89.26 ± 65.44	0.666
PCO2 (mmHg)	43.00 (37.00,51.00)	42.00 (37.00,52.50)	42.00 (37.00,51.00)	45.00 (39.00,52.50)	42.00 (36.00,51.00)	0.289
Laboratory results
WBC (K/uL)	12.55 ± 14.03	11.16 ± 21.84	10.61 ± 8.47	12.58 ± 9.98	15.84 ± 11.25	< 0.001
RBC (m/uL)	3.89 ± 0.88	3.76 ± 0.84	3.95 ± 0.89	4.02 ± 0.78	3.83 ± 0.99	0.041
Neu (K/uL)	8.57 (5.49,12.83)	7.13 (4.64,9.71)	8.19 (4.98,12.47)	8.58 (5.66,13.05)	10.55 (6.88,15.81)	< 0.001
Lym (K/uL)	2.30 ± 18.46	3.02 ± 21.48	0.98 ± 0.76	3.99 ± 29.99	1.22 ± 2.35	0.370
BAS (K/uL)	0.03 ± 0.04	0.02 ± 0.03	0.03 ± 0.06	0.03 ± 0.04	0.04 ± 0.05	0.047
EOS (K/uL)	0.80 ± 1.48	0.89 ± 1.65	0.79 ± 1.55	0.73 ± 1.22	0.78 ± 1.47	0.875
PLT (K/uL)	217 (151,289)	219 (170,286)	215 (142,293)	242 (174,301)	183 (130,281)	0.005
RDW (%)	14.87 ± 2.37	14.99 ± 2.32	14.71 ± 2.11	14.44 ± 2.15	15.34 ± 2.78	0.010
Hemoglobin (g/dL)	11.40 (9.40,13.10)	10.90 (9.05,12.60)	11.65 (9.70,13.15)	11.55 (10.40,13.50)	11.50 (8.70,13.30)	0.016
Hematocrit (%)	35.05 ± 7.38	33.72 ± 6.67	35.40 ± 7.29	36.22 ± 6.51	34.87 ± 8.70	0.031
ALB (g/dL)	2.89 ± 0.58	3.17 ± 0.50	3.03 ± 0.53	2.77 ± 0.48	2.57 ± 0.61	< 0.001
Lactate (mmol/L)	1.94 ± 1.54	0.96 ± 0.21	1.39 ± 0.27	1.82 ± 0.39	3.59 ± 2.29	< 0.001
LAR	0.54 (0.39,0.83)	0.31 (0.26,0.36)	0.46 (0.42,0.50)	0.65 (0.59,0.73)	1.15 (0.93,1.50)	< 0.001
Blood glucose (mg/dL)	176.21 ± 123.87	144.50 ± 57.45	166.39 ± 65.16	176.77 ± 85.82	216.91 ± 209.27	< 0.001
Creatinine (mg/dL)	1.71 ± 1.75	1.83 ± 2.33	1.65 ± 1.61	1.52 ± 1.49	1.85 ± 1.43	0.004
INR	1.38 ± 0.51	1.26 ± 0.29	1.31 ± 0.45	1.40 ± 0.57	1.55 ± 0.62	< 0.001
APTT (s)	39.60 ± 25.26	36.48 ± 20.06	36.93 ± 21.74	38.32 ± 22.13	46.64 ± 33.60	0.034
PT (s)	15.19 ± 7.79	13.63 ± 3.02	14.28 ± 4.92	15.96 ± 12.54	16.87 ± 6.81	< 0.001
sodium (mEq/L)	138 (135,141)	138 (136,141)	138 (135,141)	138 (134,142)	137 (134,142)	0.538
Potassium (mEq/L)	4.2 (3.8,4.7)	4.2 (3.7,4.7)	4.2 (3.9,4.6)	4.2 (3.9,4.7)	4.3 (3.8,4.8)	0.631
Chloride (mEq/L)	102.0 (98.0,106.0)	102.0 (99.0,105.5)	101.5 (98.0,105.5)	101.0 (98.0,106.0)	102.0 (98.0,106.0)	0.946
Anion gap (mEq/L)	13.0 (11.0,16.0)	13.0 (11.0,16.0)	13.0 (11.0,15.5)	13.0 (11.0,15.0)	14.0 (11.0,17.0)	0.011
AST (IU/L)	127.37 ± 634.32	53.49 ± 82.83	73.52 ± 159.22	133.74 ± 876.24	247.91 ± 889.84	< 0.001
ALT (IU/L)	83.36 ± 395.84	42.01 ± 77.41	53.66 ± 96.93	90.97 ± 540.61	146.36 ± 560.84	0.007
Comorbidities, n(%)
Hypertension	189.00 (31.87%)	41.00 (27.70%)	51.00 (34.46%)	61.00 (41.22%)	36.00 (24.16%)	0.009
CKD	114.00 (19.22%)	30.00 (20.27%)	22.00 (14.86%)	26.00 (17.57%)	36.00 (24.16%)	0.211
Malignant tumor	45.00 (7.59%)	8.00 (5.41%)	8.00 (5.41%)	17.00 (11.49%)	12.00 (8.05%)	0.153
TDM	209.00 (35.24%)	42.00 (28.38%)	57.00 (38.51%)	60.00 (40.54%)	50.00 (33.56%)	0.124
Heart failure	107.00 (18.04%)	28.00 (18.92%)	22.00 (14.86%)	31.00 (20.95%)	26.00 (17.45%)	0.579
MI	62.00 (10.46%)	9.00 (6.08%)	13.00 (8.78%)	18.00 (12.16%)	22.00 (14.77%)	0.076
IHD	135.00 (22.77%)	26.00 (17.57%)	30.00 (20.27%)	42.00 (28.38%)	37.00 (24.83%)	0.121
COPD	68.00 (11.47%)	17.00 (11.49%)	18.00 (12.16%)	21.00 (14.19%)	12.00 (8.05%)	0.414
Therapies, n (%)
VP	461.00 (77.74%)	104.00 (70.27%)	119.00 (80.41%)	117.00 (79.05%)	121.00 (81.21%)	0.087
ATB	551.00 (92.92%)	133.00 (89.86%)	139.00 (93.92%)	139.00 (93.92%)	140.00 (93.96%)	0.425
GC	448.00 (75.55%)	108.00 (72.97%)	111.00 (75.00%)	119.00 (80.41%)	110.00 (73.83%)	0.443
Viral drugs	133.00 (22.43%)	26.00 (17.57%)	36.00 (24.32%)	32.00 (21.62%)	39.00 (26.17%)	0.311
Interventions, n (p%)
Ventilation	529.00 (89.21%)	131.00 (88.51%)	139.00 (93.92%)	135.00 (91.22%)	124.00 (83.22%)	0.022
CRRT	121.00 (20.40%)	27.00 (18.24%)	30.00 (20.27%)	32.00 (21.62%)	32.00 (21.48%)	0.881

LAR: Quartile 1 (0.16–0.39), Quartile 2 (0.39–0.54), Quartile 3 (0.54–0.82), and Quartile 4 (0.82–7.59).

Abbreviation: GC: Glucocorticoid; RR (respiratory rate); Po2 (arterial oxygen pressure); Ventilation (Mechanical Ventilation); GC (Glucocorticoids); ABX (antibiotics use); SIRS (systemic inflammatory response syndrome); APTT (activated partial thromboplastin time); Viral_Drugs (antiviral drugs use); Gender (biological sex); BAS (basophil count); Weight (body weight); Glu (blood glucose); EOS (eosinophil count); CRRT (continuous renal replacement therapy); K (serum potassium); Plt (platelet count); LAR (lactate/albumin ratio); RDW (red cell distribution width); Pco2 (carbon dioxide pressure); NEU (neutrophil count); Anion_gap (serum anion gap); LYM (lymphocyte count); SOFA (sequential organ failure assessment); RBC (red blood cell count); Age (patient age); VP (vasopressor use); CCI (Charlson comorbidity index); WBC (white blood cell count); SA (serum albumin); Scr (serum creatinine); Chloride (serum chloride); Na (serum sodium); PT (prothrombin time); INR (international normalized ratio for prothrombin time); HB (hemoglobin); HCT (hematocrit); ALT (alanine aminotransferase); AST (aspartate aminotransferase).

Kaplan Meier survival analysis

The Kaplan-Meier survival curve shows that the stratification of LAR quartiles reveals a significant gradient in survival probability among groups (Fig. 2). During the 28-day observation period, patients in the highest LAR quartile (0.82–7.59) demonstrated the lowest survival probability, with stepwise improvements observed in the lower quartiles. A significant inverse association was noted between LAR and survival outcomes, wherein elevated LAR levels corresponded to poorer prognosis. The log-rank test confirmed the statistical significance of these differences (p < 0.001).

Fig. 2.

28 day KM survival curve. The KM curve shows the survival rate of each quartile at 28 days. LAR: Quartile 1 (0.16–0.39); Quartile 2 (0.39–0.54); Quartile 3 (0.54–0.82); And quartiles 4 (0.82–7.59).

Filter variables

In this study, we applied three feature selection method, the Boruta algorithm(Fig. 3), LASSO (Fig. S1), and univariate analysis (Table. S2), ultimately narrowing the feature selection scope by identifying the intersection through a Venn diagram (Fig. S2). The selected features exhibited no significant differences in the prediction model compared to those identified by other methods (Fig. S3), thus fully demonstrating their predictive value and consistency.

Fig. 3.

Feature selection based on the Boruta algorithm. The plot illustrates the Z-score distribution of variables, with the horizontal axis representing each feature name and the vertical axis showing their respective Z-scores. Variables are color-coded according to their final decision status: Confirmed (Green), Tentative (Yellow), Rejected (Red) and Shadow (Blue).

Association between LAR levels and clinical outcomes

As shown in Table S3, this table presents the Cox regression model results for 28-day mortality across five progressively adjusted models. In the unadjusted model (Model 1), elevated LAR was significantly associated with increased mortality risk (HR = 1.61; 95% CI: 1.41–1.83; p < 0.001). After sequential adjustments for demographic factors, comorbidities, clinical severity scores, and therapeutic interventions (Models 2–5), LAR remained an independent predictor of 28-day mortality with HRs ranging from 1.45 to 1.66 (all p < 0.001). When LAR was analyzed by quartiles, patients in the Q4 group consistently showed the highest risk of 28-day mortality compared to Q1. In the fully adjusted model (Model 5), the HR for Q4 patients was 2.85 (95% CI: 1.74–4.68; p < 0.001), and Q3 patients also had a significantly elevated risk (HR = 2.29; 95% CI: 1.40–3.76; p = 0.001). Conversely, Q2 patients did not reach statistical significance. The trend test across quartiles confirmed a significant dose-response relationship between increasing LAR levels and mortality risk (all p for trend < 0.001). These findings indicate that higher LAR levels are independently and robustly associated with 28-day all-cause mortality in critically ill patients with hyperlipidemia, even after adjusting for multiple potential confounders.

Restricted cubic spline analysis

A potential non-linear association between lactate-to-albumin ratio (LAR) and 28-day survival was evaluated using a restricted cubic spline (RCS) model, and adjust covariates based on Age, Gender, SOFA score, and Charlson comorbidity index (Fig. 4). In the unadjusted analysis (Fig. 4A), LAR shows an inverted “L” - shaped nonlinear relationship with 28 day mortality rate (P < 0.001). Notably, a threshold effect was observed: the LARthresholds for 28-day mortality were 0.874. The risk of death significantly increases before the threshold, and the increase slows down after exceeding the threshold. After adjusting for covariate factors, RCS analysis still observed a significant non-linear relationship between LAR and 28 day mortality rate (P < 0.001), and LAR threshold = 0.883.

Fig. 4.

RCS analysis of 28-day mortality. LAR (Lactate-to-Albumin Ratio); Y-axis: Mortality Risk. (A): Unadjusted model, (B): Adjusted for covariates: Age, Gender, SOFA score, Charlson comorbidity index, Hypertension (HTN), Chronic Kidney Disease (CKD), Cancer (CA), Diabetes Mellitus (TDM), Heart Failure (HF), Myocardial Infarction (MI), Ischemic Heart Disease (IHD), Chronic Obstructive Pulmonary Disease (COPD), Glucocorticoids (GC), Antibiotics (ABX).

Regression and subgroup analysis

Subgroup analyses of 28-day mortality were conducted to explore the prognostic relevance of the lactate-to-albumin ratio (LAR) across various clinical contexts (Fig. 5). To evaluate the independent association between LAR and mortality within specific clinical subgroups (e.g., age, gender, SOFA score, Charlson Comorbidity Index), the population was stratified accordingly. Across most subgroups, increasing quartiles of LAR were consistently associated with higher hazard ratios (HRs) for mortality, reinforcing its robust prognostic value.The chart illustrates the impact of stratified groups (Q1–Q4) on outcome risk across multiple clinical subgroups (e.g., age, gender, comorbidities). In several subgroups, the HRs for higher quartiles are statistically significant, as their confidence intervals do not cross 1, supporting the robustness of the findings. Interaction analysis revealed no statistically significant effect modification across most subgroups, indicating the association was largely consistent; however, a significant interaction was observed in the mechanical ventilation subgroup (P = 0.043). We will conduct further research on mechanically ventilated patients in the future.

Fig. 5.

Subgroup analysis of patient hazard ratios (HR) stratified by different clinical characteristics. Hierarchical variables include: Age, Gender, underlying diseases (Chronic Kidney Disease (CKD), Cancer (CA), Diabetes Mellitus (TDM), Heart Failure (HF), Chronic Obstructive Pulmonary Disease (COPD)), SOFA score, Charlson Comorbidity Index (CCI), Diabetes, COPD, implementation of mechanical ventilation, and use of glucocorticoid therapy.

Comparative predictive performance of LAR and other index

The predictive accuracy of the lactate-to-albumin ratio (LAR) for 28-day mortality was evaluated and compared against several established clinical index and scoring systems using ROC curve analysis (Fig. 6). Among all indicators, the SOFA score demonstrated the highest discriminatory ability with an area under the curve (AUC) of 0.677. LAR followed closely with an AUC of 0.662, outperforming other inflammatory and composite indices.Specifically, LAR exhibited superior predictive capability compared to the neutrophil-to-lymphocyte ratio (NLR, AUC = 0.583), red cell distribution width-to-platelet ratio (RPR, AUC = 0.629), and red cell distribution width-to-albumin ratio (RAR, AUC = 0.654). Other indices, including the platelet-to-albumin ratio (PAR, AUC = 0.439), platelet-to-lymphocyte ratio (PLR, AUC = 0.495), and systemic immune-inflammation index (SII, AUC = 0.520), demonstrated notably weaker predictive performance.

Fig. 6.

ROC curves comparing the predictive performance of SOFA score, lactate-to-albumin ratio (LAR), neutrophil-to-lymphocyte ratio (NLR), platelet-to-albumin ratio (PAR), platelet-to-lymphocyte ratio (PLR), systemic immune-inflammation index (SII), red cell distribution width-to-platelet ratio (RPR), and red cell distribution width-to-albumin ratio (RAR) for 28-day mortality.

Construction and validation of machine learning prediction models

The predictive performance of clinical and laboratory variables is illustrated in Fig. 6. Among them, the lactate-to-albumin ratio (LAR, AUC = 0.662) exhibited a higher area under the curve compared to other predictors. To construct the prognostic model, data were randomly split into training and validation sets at a 7:3 ratio. Eight machine learning algorithms were implemented, including Cox proportional hazards (CoxPH), DeepSurv, random survival forest (RSF), and extreme gradient boosting survival (XGBoost) models. These models were developed to estimate mortality risk in patients with severe COVID-19. Model performance was evaluated using receiver operating characteristic (ROC) curves and their corresponding AUC. Decision curve analysis (DCA) assessed clinical benefit, while calibration curves measured prediction accuracy for absolute risk. The RSF model demonstrated superior predictive power, with consistently higher AUC values at both 14-day and 28-day time points (Fig.S4, Fig. 7). As shown in Fig. 8, the predicted overall survival (OS) probabilities closely matched observed OS across both training and test cohorts, particularly in the higher-risk range, indicating robust calibration of the RSF model. To evaluate the clinical utility of the machine learning models, we performed decision curve analysis (DCA) for both 14-day and 28-day mortality prediction (Fig.S5 and Fig.S6). Time-dependent AUC curves were used to further assess the discriminative ability of the models across multiple follow-up time points (Fig.S7). A variable importance analysis was conducted on the Random Forest model, which demonstrated the best performance in the evaluation (Fig.S8).

Fig. 7.

28-day mortality ROC curves comparing eight survival models for predicting mortality in critically ill COVID-19 patients. (A) Training cohort; (B) independent test cohort. Models include GBM, random survival forest (RSF), LASSO-Cox, CoxBoost, survival-SVM, XGBoost-Cox, SuperPC, and plsRcox. The diagonal dashed line indicates no discrimination. AUC with 95% CI are reported in the legend; GBM and RSF show the highest discrimination in both cohorts.

Fig. 8.

Calibration plots comparing the random survival forest (RSF)–predicted overall survival (OS) probabilities with the observed OS at 14 days (blue) and 28 days (red).

Discussion

COVID-19 viral pneumonia has caused millions of deaths worldwide since 2020 and remains a significant public health issue. While vaccines and antiviral treatments have reduced transmission, the rise of new variants and unequal access to medical resources have kept the pandemic ongoing. In this retrospective analysis of 594 critically ill COVID-19 patients from the MIMIC-IV database, the lactate-to-albumin ratio (LAR) was identified as an independent predictor of 28-day mortality (HR = 1.309; 95% CI: 1.113–1.54). This study highlights the combined prognostic value of lactate and albumin and improves the clinical use of LAR through survival analysis methods, restricted cubic spline, and machine learning approaches. These methods collectively enhance the clinical utility of LAR as a risk stratification and outcome prediction tool for critically ill patients.

Lactate is a metabolic byproduct generated under conditions of cellular hypoxia and anaerobic glycolysis, and is widely utilized as a biomarker of tissue hypoxia, impaired perfusion, and metabolic acidosis¹³. In COVID-19 patients, the immune and inflammatory responses become aberrantly activated, resulting in elevated levels of pro-inflammatory cytokines such as IL-6 and TNF-α. These alterations lead to mitochondrial dysfunction and microcirculatory disturbances, contributing to organ dysfunction and metabolic dysregulation, which subsequently increases lactate levels in the bloodstream^14,15. Blood lactate levels are frequently utilized as a hemodynamic target in fluid and vasopressor management, and demonstrated a significant association with poor prognosis in critically ill patients^16,17. However, as a standalone marker, lactate is influenced by various physiological and pathological factors, including exercise, metabolic disturbances, liver dysfunction, and pharmacological interventions, limiting its utility as a singular prognostic tool¹⁸. In critically ill patients, albumin levels typically decrease, which may result from increased protein catabolism, reduced synthesis, or enhanced loss. Hypoalbuminemia is strongly correlated with disease severity and poor prognosis^19–21. Nevertheless, albumin levels can also be influenced by liver function, inflammation, and nutritional status, adding complexity to its application. To overcome the limitations of using individual parameters in clinical evaluations, recent studies have recommended the lactate-to-albumin ratio (LAR) as a composite biomarker. By integrating multidimensional information such as tissue oxygen metabolism, inflammatory response, and nutritional status, this index provides a comprehensive assessment of the patient’s current pathophysiological condition.

Numerous studies have identified the lactate-to-albumin ratio (LAR) as a reliable biomarker associated with adverse outcomes across a range of critical illnesses. LAR was divided into quartiles (Q1–Q4) to allow comparison of survival outcomes across clinically interpretable risk groups, and to facilitate consistency with previous critical care and COVID-19 prognostic studies that used quartile-based biomarker stratification to identify mortality patterns^22–24. In septic populations, Ryoo et al.²³ found that LAR outperformed lactate clearance in forecasting 28-day mortality, exhibiting superior predictive accuracy (AUC: 0.70 vs. 0.65; P < 0.01). Complementing these findings, Cakir demonstrated that integrating lactate and albumin levels into a composite index yielded greater prognostic precision than using either parameter in isolation²⁵. In the context of acute kidney injury (AKI), particularly among patients receiving continuous renal replacement therapy (CRRT), Liu et al.²⁶ emphasized LAR’s pronounced predictive utility. Similarly, Lichtenauer et al.¹⁹ applied LAR to assess disease severity in septic individuals, revealing that both lactate and albumin independently contributed to mortality prediction with substantial accuracy. Beyond sepsis, LAR has garnered attention in the setting of COVID-19, where systemic inflammation plays a pivotal role in disease progression and lethality²⁷. Inflammatory indices such as the neutrophil-to-lymphocyte ratio (NLR), platelet-to-lymphocyte ratio (PLR), and C-reactive protein (CRP) have already demonstrated significant prognostic capabilities²⁸. These markers, reflective of the host’s immune-metabolic status, have been linked to escalation of disease severity, respiratory failure, and increased fatality risk. Positioned at the intersection of metabolic dysregulation and inflammation, LAR has emerged as a compelling biomarker for risk assessment in critically ill COVID-19 patients. Continued exploration of LAR may enhance prognostic modeling and support clinical decision-making in intensive care environments.

We analyzed the four LAR groups using Kaplan-Meier survival curves, and the results demonstrated a progressive decrease in survival rates from LAR Q1–Q4 represent progressively increasing LAR risk categories, which may aid in clinical decision-making and highlights the clinical significance of LAR in prognostic stratification for COVID-19 patients. Wang et al.²⁹ demonstrated that in patients with acute myocardial infarction, LAR was strongly associated with all-cause mortality, exhibiting a linear relationship with a threshold of 1.0. Some studies have also reported non-linear associations. For instance, while Liu et al.²⁴ identified a threshold of 1.641 for AKI patients undergoing CRRT, our analysis revealed a slightly lower threshold, suggesting that COVID-19 pathophysiology may affect biomarker sensitivity. Other studies³⁰ have indicated that in ICU patients, the threshold for 30-day and 180-day mortality is approximately 0.65, indicating potential variation in thresholds across populations, while maintaining consistently high predictive value. This study observed a significant non-linear relationship between LAR and mortality risk using restricted cubic spline regression (P < 0.001), with a threshold of 0.874 without adjusting for covariates and 0.883 after adjusting for covariates. Although the risk of death in critically ill patients is multifactorial, our stratified analysis considers the main mixed factors. As indicated in Table S3, LAR was consistently associated with mortality across all models, with hazard ratios ranging from 1.45 to 1.66 (all p < 0.001), and an HR of 2.85 (95% CI: 1.74–4.68) in the fully adjusted model. The observed trend reinforces a dose-response relationship between elevated LAR and increased mortality risk. In subgroup analysis, we examined the role of LAR (Fig. 5), where higher quartiles consistently demonstrated significantly elevated hazard ratio (HR), further reinforcing the strong predictive value of LAR as a prognostic biomarker.

We demonstrated that the lactate-to-albumin ratio (LAR) provides meaningful prognostic value for predicting 28-day mortality in critically ill patients, particularly in the context of COVID-19. Through ROC curve analysis, surpassing several widely studied inflammatory indices such as NLR, PLR, SII, and PAR. Notably, its predictive performance was second only to the SOFA score, a comprehensive clinical severity index, highlighting LAR’s potential as a practical and efficient alternative in resource-limited settings. In multi-index prediction (Fig. 6), the LAR demonstrated superior performance compared to other conventional laboratory indicators (AUC = 0.662, 95% CI: 0.616–0.708). RAR and RPR also showed relatively high AUC values, highlighting their potential as biomarkers for predicting mortality in critically ill patients with COVID-19. In addition, RAR and RPR also demonstrate strong predictive value in other inflammatory diseases, reflecting their ability to assess disease severity and prognosis under different inflammatory conditions^31,32. The weaker performance of PAR, PLR, and SII in our cohort may reflect limitations in their sensitivity to dynamic physiological changes in critically ill COVID-19 patients. The ease of LAR computation and its foundation on well-understood metabolic and nutritional parameters make it especially attractive for clinical integration.

The feature selection employs a hybrid approach, combining Boruta (Fig. 3), 10k-fold LASSO regression (Fig. S1), and single-factor analysis. Following the intersection of Venn diagrams, 8 core predictive factors were identified and screened. The Cox model demonstrated (Fig.S3) that the model retained its discriminative ability, with dimensionality reduced by 61%. The survival machine learning model results demonstrated that the area under the curve (AUC) for the XGBoost and Random Survival Forest (RSF) models was 0.834 (95% CI: 0.794–0.874) and 0.886 (95% CI: 0.853–0.919), respectively. The RSF model exhibited strong predictive performance on both the training and testing datasets for 14-day (Fig.S4) and 28-day (Fig. 7) prognosis. Furthermore, the time-dependent AUC analysis (Fig.S7) demonstrates the durability of model performance across different follow-up periods. This temporal robustness is essential in critical illness, where patient trajectories evolve rapidly and risk assessment must adapt accordingly. The DCA findings (Fig.S5, Fig.S6) highlight that RSF not only achieve strong discrimination but also offer meaningful net clinical benefit across varying decision thresholds. Finally, variable importance analysis (Fig.S8) revealed that while traditional predictors such as age and SOFA score remain dominant, LAR ranks among the top determinants of 28-day mortality. This reinforces the central hypothesis of our study that LAR is not only independently prognostic but also synergizes with multidimensional clinical indices within machine learning frameworks. Unlike composite scores, LAR is simple to calculate and reflects both metabolic stress and systemic inflammation, making it a practical biomarker for bedside use.

However, this study has several limitations. Firstly, the data used in this study comes from a single center public database, which may introduce selection bias or data quality issues. In the future, more multi center data will be added for external validation. Although all patients were admitted to the ICU, data bias caused by viral mutations during the COVID-19 pandemic was not taken into account. But we conducted subgroup analysis and discussion on disease severity and comorbidities to minimize their impact on outcomes. We adjusted some confounding variables in the statistical analysis, but some unknown factors may still affect the results, which highlights the necessity of further investigation. Additionally, although this study demonstrated the short-term predictive ability of LAR, the lack of long-term follow-up data may limit a comprehensive assessment of patient recovery.

Conclusion

This study highlights the prognostic utility of the lactate-to-albumin ratio (LAR) in predicting short-term mortality among critically ill patients with COVID-19. Compared to conventional biomarkers, LAR demonstrated more efficient and interpretable predictive performance, while also enhancing clinical stratification of disease outcomes. Leveraging a Random Survival Forest (RSF) machine learning model, we observed superior sensitivity and specificity for LAR-based prediction. Subsequent analysis further identified a robust positive association between LAR levels and mortality risk. Collectively, these findings suggest that LAR serves as a clinically valuable biomarker for outcome prediction and may support personalized decision-making in high-risk patient cohorts.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

C.F.L. conceived the study, conducted the main analysis, wrote the manuscript, and prepared the figures. Y.K. supervised the study and provided critical revisions. Y.K. and X.J.Z. obtained funding. X.Y.Z. curated the data and assisted with early analysis. Q.Q.Z. contributed to methodology development and validation. J.N.G. supported software implementation and model testing. Y.J.W. reviewed the statistical analyses and provided feedback on data interpretation. Z.Q.W. revised the manuscript language and performed final proofreading. All authors reviewed and approved the final manuscript.

Funding

This work was supported by the Zhengzhou Collaborative Innovation Special Project (Grant Nos. 518-23240003/518-23240004) from Zhengzhou University, the Key Research Project of the Henan Provincial Department of Science and Technology (Grant No. 221111311800) entitled ‘Evaluation of COVID-19 Susceptibility Factors and Development of Post-rehabilitation Quality of Life Prediction Models,’ and the Henan Provincial-Ministry Jointly Built Major Project (SBGJ202301003).

Data availability

The datasets were accessible from the MIMIC-IV (version 3.1) database. Corresponding author will provide the datasets upon reasonable request.These data can be found at https://physionet.org/content/mimiciv/3.1/.

Declarations

Competing interests

The authors declare no competing interests.

Ethical approval and consent to participate

The use of the MIMIC-IV database was approved by the review committee of Massachusetts Institute of Technology and Beth Israel Deaconess Medical Center. Based on our use of public datasets, this study was exempted from approval by the Ethics Committee of the Henan Provincial People’s Hospital.