Improved 28-day mortality prediction by combining Lactate Dehydrogenase-to-Albumin Ratio(LAR) to lactate-enhanced Sequential Organ Failure Assessment (SOFA) score in patients with sepsis: a derivation and validation study

Abstract

Background

Fast risk stratification is essential for patients with sepsis, a life-threatening condition associated with high mortality, as it guides clinical decision-making and management. In this study, we aimed to evaluate the predictive performance of a novel model—termed SOLAR—which integrates the lactate dehydrogenase-to-albumin ratio (LAR) with the Sequential Organ Failure Assessment (SOFA) score and lactate for 28-day mortality in sepsis.

Methods

Data for the derivation cohort were collected from the Medical Information Mart for Intensive Care IV database. The validation cohort was from the Department of Critical Care Medicine at the First Affiliated Hospital of Soochow University. Patients admitted to ICU for sepsis were enrolled. Model performance was analyzed by using multivariable logistic regression and Harrell’s c-index. Reclassification ability of the model was evaluated by using net reclassification index (NRI) and Integrated discrimination improvement (IDI). Benefits in clinical decision making was estimated by using decision curve analysis. Super learner analysis was used to identify the optimal machine learning algorithm for model construction. Nomograms and an online calculator were used to deploy the model for point-of-care utilization.

Results

The derivation cohort comprised 4936 patients with sepsis and sufficient SOFA, LAR, and lactate data were obtained from MIMIC database. The validation cohort contained 371 patients with sepsis admitted to the First Affiliated Hospital of Soochow University. In both cohorts, an increasing trend of 28-day mortality was observed with elevating SOFA scores, lactate levels, and LAR quartiles. Compared to the SOFA-lactate score, the SOLAR score demonstrated improved 28-day mortality discrimination, reflected by an increase in c-index (Derivation cohort: 0.695 [0.680–0.711] vs. 0.660 [0.644–0.676], P < 0.001), Validation cohort: 0.706 [0.651–0.761] vs. 0.638 [0.581–0.695], P < 0.001). The SOLAR score also displayed a superior reclassification ability over the SOFA-lactate score (NRI: Derivation cohort: 0.393 [0.335–0.451], P < 0.001), Validation cohort: 0.506 [0.302–0.709], P < 0.001, IDI: Derivation cohort: 0.034 [0.029–0.039], P < 0.001), Validation cohort: 0.044 [0.026–0.062], P < 0.001). Super learner analysis showed that the logistic regression model achieved the highest c-index (0.749 [0.694–0.803]) based on the SOLAR score. A simplified online calculator of the SOLAR score was deployed for point-of-care risk stratification.

Conclusion

The SOLAR score combining LAR, SOFA score and lactate provides improved accuracy and sensitivity over the SOFA-lactate score for 28-day mortality prediction in sepsis.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12879-026-13971-7.

Introduction

Sepsis is a life-threatening condition caused by dysregulated response to infection that may lead to a potentially fatal multi-organ dysfunction syndrome [1, 2]. In 2017, there were 48.9 million estimated cases of sepsis worldwide, including 11 million sepsis related deaths, accounting for 19.7% of total global deaths [3]. Given the high morbidity and mortality of sepsis, a prognostic tool is required for timely recognition and risk stratification. Therefore, the SOFA score has been developed for this purpose and widely utilized in clinical practice for nearly 3 decades [4–6]. Nevertheless, the SOFA score only achieved mild performance to predict mortality events in sepsis according to recent large-scale studies [7].

Despite its important role in evaluating critically ill patients, the SOFA score is facing new challenges due to limited sensitivity and precision in the era with evolving diagnostic and therapeutic methods for sepsis [8, 9]. Recently, several studies showed that novel biomarkers may improve the performance of SOFA in the risk estimation of sepsis. Liu et al. reported that adding vasoactive drug doses and lactate values significantly improved mortality discriminant power of the original SOFA score in suspected infection, sepsis, and septic shock [10]. Another study demonstrated that adding procalcitonin (PCT) enhanced the discriminative capacity of SOFA for risk prediction in sepsis with an increased area under the curve (AUC) from 0.86 to 0.91 [11]. Combining N-terminal pro-brain natriuretic (NT-proBNP) the original SOFA score also provides better hospital mortality discrimination [12].

However, most of these updated scores rely on sophisticated equipment and additional laboratory tests that are either time consuming or expensive. APACHE II and SAPS II require a larger number of physiologic and laboratory variables, and they are typically calculated over a defined early ICU time window.qSOFA is a rapid bedside tool, but it was originally designed mainly to identify infected patients at risk of poor outcomes, particularly outside the ICU, whereas our study focused on mortality prediction among patients with sepsis using organ dysfunction and laboratory markers.Therefore, a point-of-care and economical approach, such as lactate test, has become a method of choice to improve the SOFA score, as shown by Wright SW et al. in a recent study. Being a cost-effective approach, the SOFA-lactate score outperformed the SOFA score in a vast spectrum of conditions. Likewise, utilizing available information from conventional laboratory tests may provide an alternative way to achieve the improvements.

Lactate dehydrogenase (LDH) to albumin ratio (LAR), which is ready for calculation from routine biochemistry tests, has shown prognostic value for a number of inflammatory diseases. Notably, LDH is a key enzyme in glycolysis that is enhanced by anaerobic metabolism, a common pathophysiological change in tissue hypoxia associated with sepsis [13, 14]. Hypoalbuminemia, a hallmark of excessive inflammation and malnutrition, has been associated with poor survival in critically ill patients [15–17]. Reflecting multiple aspects underlying the development and progression of sepsis, LAR has been proved as an independent prognostic marker for critically ill patients. Liang M et al. showed that LAR was associated with increased mortality risk in patients with sepsis and acute kidney injury [18]. Also, Jeon et al. reported that LAR could be used to determine the necessity for triaging to ICU admission in patients with severe infection [19].

To date, whether LAR could enhance the SOFA-lactate score in mortality discrimination for patients with sepsis remains unknown. Since LAR is easily obtainable and may reflect the combined burden of tissue damage and systemic inflammation, we hypothesized that incorporation of LAR into the lactate-enhanced SOFA score could improve discrimination of 28-day mortality in sepsis. Therefore, this study aimed to develop and validate an LAR-enhanced lactate-SOFA model for risk stratification in patients with sepsis.

Methods

Study design and ethics

The study included patients from two retrospective cohorts. Data from the Medical Information Mart for Intensive Care IV (MIMIC IV) was obtained as a derivation cohort. MIMIC IV is a database released on November 14, 2022, containing de-identified information of patients who received critical care at the Beth Israel Deaconess Medical Center between 2008 and 2019. The establishment of this database was approved by the Massachusetts Institute of Technology (Cambridge, MA) and Beth Israel Deaconess Medical Center (Boston, MA), and informed consent was obtained for the original data collection. Therefore, the ethical approval statement and the need for informed consent were waived. Patients admitted to the ICU of the First Affiliated Hospital for the diagnosis of sepsis from 8/2017 to 9/2022 were retrospectively enrolled as the validation cohort. Patient identity remained anonymous and the requirement for informed consent was therefore waived according to the observational nature of the study. For the external validation cohort, electronic health records were retrieved under IRB approval (IRB No. (2023) LRPA No. 017). Authors had no access to directly identifying information during or after data collection; all datasets were de-identified prior to analysis.

Patient-selection criteria

From the MIMIC-IV database, patients with a diagnosis of sepsis were initially screened, among whom lacked lactate measurements. In addition, patients with sepsis were diagnosed from the Department of Critical Care Medicine of the First Affiliated Hospital of Soochow University. Missing values were emulated using multiple imputation (Supplemental Table 1). Inclusion criteria were: (1) diagnosis of sepsis based on the Sepsis-3.0 criteria [20], (2) age ≥ 18 years. Exclusion criteria were: (1) medical history of previous ICU admission, (2) diagnosis of any malignant tumor, (3) absence of LDH, albumin or lactate values. For patients who had multiple ICU stays, only their initial admission was analyzed. Ultimately, the MIMIC database included 4,936 patients with sepsis, as well as 371 patients with sepsis who were hospitalized in the Intensive Care Unit of the First Affiliated Hospital of Soochow University from August 2017 to September 2022. These patients were all included in the study.

Data extraction

For both derivation and validation cohorts, patient sex, age, weight, heart rate (HR), systolic blood pressure (SBP), diastolic blood pressure (DBP), histories of hypertension, diabetes, liver disease, chronic kidney disease (CKD), chronic obstructive pulmonary disease (COPD), SOFA score, white blood cell (WBC) count, lymphocyte count, monocyte count platelet count, hemoglobin, lactate acid, pH, partial pressure of carbon dioxide (PCO₂), partial pressure of oxygen (PO₂), bicarbonate, blood urea nitrogen (BUN), creatinine, glucose, potassium, alanine aminotransferase (ALT), aspartate aminotransferase (AST), bilirubin, LDH, albumin, calcium, chloride, sodium, potassium, prothrombin time (PT), activated partial thromboplastin time (APTT), and 28-day mortality were obtained. Admission SOFA score was recorded for all patients in the derivation and external validation cohorts. The LAR was calculated as the ratio of LDH (IU/ml) to albumin (g/L).

Definition and outcomes

In both derivation and validation cohorts, the primary outcome of this study was 28-day mortality. Survival data were extracted from the MIMIC-IV database. In the validation cohort, mortality data were collected via electronic health records for hospitalized patients or telephone follow up for those discharge alive.

Statistical analysis

Numeric variables with skewed distribution were presented as medians with 95% confidence intervals and compared using Mann-Whitney’s U test. Categorical variables were presented as counts with percentages and compared using κ² test or Fisher’s exact test when appropriate. Missing covariate data were handled using multiple imputation by chained equations. Five imputed datasets were generated. Predictive mean matching was used for continuous variables, and appropriate regression models were used for categorical variables. Sensitivity analyses were used to determine the robustness of the model. In both derivation and validation cohorts, the lactate concentration was categorized as low (< 2.0 mmol/L), medium (≥ 2.0 to < 4.0 mmol/L), and high (≥ 4.0 mmol/L) for the modelling process. The SOFA score was defined as three levels: low (≥ 2.0 to < 4.0), medium (≥ 4.0 to < 6.0), and high (≥ 6.0). The LAR was categorized by its quartile values based on the derivation cohort. The association between each variable with mortality outcome was analyzed using univariate logistic regression and visualized in matrix plots.

The concordance of the models for predicting 28-day mortality was visualized using the receiver operator characteristic (ROC) and precision-recall (PR) curves, and quantified using Harrell’s c-index. Improvement in predictive performance from combing LAR to the SOFA-lactate score was estimated using the Delong’s test and reclassification analysis. A 10-fold internal cross-validation with bootstrapping was performed for model comparison in the derivation cohort. Model calibration plots of expected and actual events with indexes including F1 and integrated calibration index (ICI) were compared. Potential improvement in clinical decision making was assessed using decision curve analysis. The relative contributions of LAR, lactate concentration, and SOFA classes to the prediction model were visualized using a nomogram. A simplified score was derived to facilitate clinical interpretation. Super learner analysis was used to compare the performance of different machine learning algorithms based on the prediction model. The study protocol was conformed to the TRIPOD guidelines and a checklist was provide as data supplement. Statistical analyses were performed using the R software (Version 4.1.2, Vienna, Austria).

Results

Enrollment of patients

A total of 4936 patients from the MIMIC database diagnosed with sepsis were enrolled as the derivation cohort after applying the inclusion and exclusion criteria. Three hundred and seventy-one patients hospitalized for sepsis in the department of Critical Care in the First Affiliated Hospital of Soochow University from August 2017 to September 2022 was enrolled as the validation cohort. Diagnosis of sepsis was determined according to the Sepsis-3 definitions [20]. The process of patient enrollment was summarized in a flowchart (Fig. 1). Distribution of missing values were presented in Supplemental Fig. 1. Sensitivity analysis of modelling variable and model performance after multiple imputation were showed in Supplemental Table 1.

Fig. 1.

Patient enrollment

Baseline characteristics

Demographic and clinical features of the derivation and validation cohorts were summarized in Table 1. In the derivation cohort, patients who died by 28 days displayed higher SOFA score (4.00 [3.00–6.00] vs. 3.00 [2.00–5.00], P < 0.001), lactate level (3.50 [1.90–7.45] vs. 2.10 [1.40–3.70], P < 0.001), and LAR level (15.41 [9.22–30.57] vs. 9.67 [6.76–16.06], P < 0.001) compared to the survived subjects. The validation cohort showed a similar trend of increase in SOFA score (11.00 [8.00–15.00] vs. 8.00 [6.00–11.00], P < 0.001), lactate level(2.90 [1.60–5.90] vs. 1.70 [1.00-2.90], P < 0.001), and LAR level (19.43 [12.59–34.96] vs. 10.56 [7.51–15.83], P < 0.001) in patients who died compared with those who were alive by 28 days from enrollment (Supplemental Table 2).

Table 1.

Baseline characteristics

n	Derivation cohort	Validation cohort	p
	4936	371
Age (year)	65.30 [53.77, 76.93]	64.00 [51.00, 76.00]	0.030
Gender			< 0.001
Male(%)	2790 (56.5)	147 (39.6)
Female(%)	2146(43.5)	224(60.4)
Hypertension(%)	2352 (47.6)	198 (53.4)	0.038
Diabetes(%)	1045 (21.2)	106 (28.6)	0.001
COPD(%)	507 (10.3)	17 (4.6)	0.001
Liver Disease(%)	1096 (22.2)	18 (4.9)	< 0.001
CKD(%)	1236 (25.0)	35 (9.4)	< 0.001
HR (bpm)	92.00 [78.00, 108.00]	101.00 [88.50, 120.00]	< 0.001
SBP (mmHg)	113.00 [99.00, 132.00]	114.00 [99.50, 130.50]	0.766
DBP (mmHg)	64.00 [54.00, 75.00]	67.00 [56.00, 78.00]	0.008
PH(units)	7.41 [7.36, 7.45]	7.40 [7.31, 7.46]	0.016
PO2(mmHg)	137.00 [80.00, 248.00]	95.20 [71.15, 121.75]	< 0.001
PCO2 (mmHg)	46.00 [38.00, 54.00]	34.10 [28.40, 40.90]	< 0.001
Lac(mmol/L)	2.40 [1.50, 4.80]	2.00 [1.20, 3.70]	< 0.001
WBC (109/L)	14.60 [9.90, 20.80]	11.45 [7.01, 19.01]	< 0.001
Lymphocyte count (109/L)	1.07 [0.62, 1.72]	0.64 [0.37, 1.13]	< 0.001
Monocyte count (109/L)	0.64 [0.36, 1.05]	0.38 [0.16, 0.67]	< 0.001
Platelet count (109/L)	195.00 [127.50, 279.00]	118.00 [56.00, 202.00]	< 0.001
PT (sec)	16.20 [13.60, 22.20]	16.50 [14.70, 18.95]	0.187
APTT (sec)	36.50 [29.60, 56.50]	44.50 [37.45, 52.20]	< 0.001
Bilirubin (umol/L)	15.40 [8.60, 41.00]	16.60 [10.20, 35.85]	0.117
ALT (IU/L)	38.00 [20.00, 110.00]	32.00 [16.95, 78.90]	0.002
AST (IU/L)	63.00 [31.00, 192.00]	47.60 [25.45, 125.85]	< 0.001
Albumin (g/L)	31.00 [26.00, 36.00]	30.10 [26.45, 33.00]	< 0.001
BUN (mmol/L)	11.00 [7.00, 18.00]	13.10 [8.50, 19.70]	< 0.001
Creatinine (umol/L)	123.80 [79.60, 221.00]	107.60 [61.85, 211.60]	< 0.001
LDH (IU/L)	330.00 [233.00, 570.00]	360.20 [243.80, 593.10]	0.172
Glucose (mmol/L)	9.00 [7.00, 13.00]	8.52 [6.04, 11.62]	< 0.001
Potassium (mEq/L)	4.60 [4.10, 5.20]	3.86 [3.46, 4.25]	< 0.001
Sodium (mEq/L)	140.00 [137.00, 143.00]	140.50 [137.05, 144.60]	0.143
Chloride (mEq/L)	106.00 [101.00, 110.00]	104.60 [100.60, 108.65]	0.001
Phosphorus (mEq/L)	2.10 [2.00, 2.30]	2.01 [1.91, 2.09]	< 0.001
SOFA (points)	4.00 [2.00, 5.00]	8.00 [6.00, 12.00]	< 0.001
28-day morality, n (%)	1671 (33.9)	138 (37.2)	0.210
LAR	11.26 [7.38, 20.23]	12.30 [8.28, 20.91]	0.030

HR: heart rate, SBP: systolic blood pressure, DBP: diastolic blood pressure, CKD: chronic kidney disease, COPD: chronic obstructive pulmonary disease, WBC: white blood cell count, Lac: lactate acid, PCO₂:partial pressure of carbon dioxide, PO₂: partial pressure of oxygen, BUN: blood urea nitrogen, ALT: alanine aminotransferase, AST: aspartate aminotransferase, LDH: Lactate dehydrogenase, PT: prothrombin time, APTT: activated partial thromboplastin time, LAR: Lactate dehydrogenase (LDH) to albumin ratio

Association between SOFA score, lactate, and LAR with 28-day mortality

Clinical management is frequently guided by distinct risk categories rather than continuous prognostic estimates. To enable swift risk stratification—which is critical for timely intervention in sepsis—and to explicitly examine the role of lactate and SOFA scores across patient subgroups with varying illness severity, we elected to categorize these continuous predictors based on established, clinically meaningful thresholds. Distribution of patients and mortality events according to SOFA score class (low: 2–4, medium: 4–6, high > 6), lactate level (low: 0–2; medium: 2–4, high > 4,) and LAR quartiles (Quartile 1: <7.378, Quartile 2: 7.378–11.260, Quartile 3: 11.261–20.233, Quartile 4: >20.233) in the derivation cohort were summarized in Fig. 2A progressively increment of mortality rates was correlated with higher SOFA scores (Fig. 2A), lactate levels (Fig. 2B), and LAR quartiles (Fig. 2C). In the validation cohort, the mortality rates were higher in those with elevating SOFA (Fig. 2D), lactate levels (Fig. 2E), and LAR (Fig. 2F) quartiles. The number and proportion of patients who died by 28 days in combinations of SOFA score, lactate level, or LAR quartiles was summarized in matrix plots. In both derivation and validation cohorts, mortality rate, SOFA score class, lactate levels, and LAR quartiles were positively associated (Supplemental Fig. 2). Significant correlation were observed between ordinal SOFA score class (Spearman’s r = 0.2, P < 0.001), lactate levels (Spearman’s r = 0.3, P < 0.001) and LAR quartiles.

Fig. 2.

Proportion of patients and 28-day mortality by SOFA score classes, lactate levels, and LAR quartiles in the derivation and validation cohorts. The white table represents the porportion of patients and the pink table represents the 28-day mortality rate

We employed univariate logistic regression to determine the association between different modelling variables with the mortality outcome. Using the lowest class as reference, increased SOFA classes were associated with a higher risk of 28-day mortality in both derivation (OR: medium 1.52 [1.32–1.76], P < 0.001; high 2.43 [2.10–2.81], P < 0.001). In the validation cohort, an increasing trend of death risk was observed although no statistical significance was reached (OR: medium 0.94 [0.31–2.92], P = 0.910; high 2.27 [0.99–5.88], P = 0.067) cohorts. Similarly, higher lactate levels were associated with an increased risk of 28-day death in the derivation cohort (OR: medium 1.60 [1.37–1.87], P < 0.001; high 3.55 [2.10–2.81], P < 0.001). An increasing trend of mortality risk was observed in the validation cohort (OR: medium 1.47 [0.88–2.46], P = 0.138; high 3.61 [2.11–6.27], P < 0.001). A consistent increase of mortality risk was observed with elevating quartiles of LAR (Derivation: Quartile 2, 1.51 [1.45–2.12], P < 0.001; Quartile 3, 2.40 [2.00-2.89], P < 0.001; Quartile 4, 5.02 [4.19–6.03], P < 0.001. Validation: Quartile 2, 1.51 [0.72–3.28], P = 0.279; Quartile 3, 2.76 [1.37–5.48], P = 0.006; Quartile 4, 6.41 [3.17–13.69], P < 0.001) (Supplemental Table 3).

Improved prediction of 28-day mortality by combining LAR to the lactate-enhanced SOFA score

The combination of SOFA score, serum lactate, and LAR was designated as the SOLAR score. Compared with the SOFA-lactate score, the SOLAR score demonstrated better discrimination for 28-day mortality in the derivation cohort, with a higher AUC (0.695 [0.680–0.711] vs. 0.660 [0.644–0.676], P < 0.001). A similar improvement was observed after 100-times bootstrap cross-validation, with a delta AUC of 0.033 (95% CI 0.024–0.045) (Fig. 3A; Table 2). In the validation cohort, the SOLAR score also outperformed the SOFA-lactate score (AUROC 0.706 [0.651–0.761] vs. 0.638 [0.581–0.695], P < 0.001), and the increase in AUC after bootstrap cross-validation was 0.068 (95% CI 0.030–0.105) (Fig. 3B; Table 2). Compared with the SOFA-lactate score, the SOLAR score showed an increased P-R curve(Precision Recall Curve) for predicting 28-day mortality in the derivation (F1 score 0.430 vs. 0.432, Supplemental Fig. 3A) and the validation (F1 score 0.500 vs. 0.436, Supplemental Fig. 3B) cohorts. In addition, a GiViTI calibration curve for the validation cohort (Supplemental Fig. 3C) were plotted, showing good concordance between observed and predicted probabilities by the SOLAR score. Further analysis based on a risk threshold of death by 30% and 50% supported better reclassification performance of the SOLAR score compared with the SOFA-lactate score (Derivation cohort: NRI 0.056 [0.020–0.091], P = 0.002; IDI 0.034 [0.029–0.039], P < 0.001; Validation cohort: NRI 0.506 [0.302–0.709], P < 0.001; IDI 0.044 [0.026–0.062], P < 0.001) (Supplemental Table 4). From the original risk categories defined by the SOFA score, the SOLAR score reclassified 21%, 51%, and 41% patients in the derivation cohort (Supplemental Fig. 3D). In the validation cohort, 19%, 52%, and 23% patients were reclassified to their corresponding risk levels (Supplemental Fig. 3E). The discriminating performance of the SOLAR model in both cohorts was visualized in confusion matrix plots (Supplemental Fig. 3F-G).

Fig. 3.

Receiver operating characteristics curves for mortality discrimination in the derivation cohort (A) and validation cohort (B)

Table 2.

Mortality discrimination by models

		Cohort AUROC	P value	Cross-validation AUROC
Derivation cohort
1	SOFA-lactate score	0.660(0.644–0.676)	< 0.001	0.660(0.637–0.680)
2	SOLAR	0.695(0.680–0.711)	< 0.001	0.694(0.672–0.715)
Validation cohort
1	SOFA-lactate score	0.638(0.581–0.695)	< 0.001	0.628(0.553–0.712)
2	SOLAR	0.706(0.651–0.761)	< 0.001	0.686(0.618–0.751)

The estimated clinical benefits from SOLAR score in mortality discrimination was then evaluate in a DCA curves. As shown in Supplemental Fig. 3H, the SOLAR score showed a higher net benefit compared to the SOFA-lactate score when the risk threshold was between 18% and 66% in the derivation cohort. In the validation cohort, the improvement in net benefit was observed for a risk threshold of death from 10% to 70% (Supplemental Fig. 3I). The models showed similar clinical benefits beyond the aforementioned risk thresholds.

Optimal prediction model selection

Super learner has been developed as a nonparametric technique for constructing the optimal prediction algorithm from a given set of candidate algorithms. To explore the optimal algorithm supporting the SOLAR model, we conducted a super learner analysis to compare model performance among a variety of machine learning methods [21]. Cross-validated AUC analysis showed that the logistic regression model (general linear model, GLM) achieved the highest AUC for predicting 28-day mortality (AUC 0.749 [0.694–0.803]) (Supplemental Fig. 4). An public version of the prediction model was posted online for convenient bedside utilization (https://szumlmd.shinyapps.io/SOLAR/).

Model simplification and deployment

According to the super learner results, the binary logistic model was chosen to construct the final SOLAR score. To facilitate point-of-care use, a simplified SOLAR score (sSOLAR) was constructed using coefficients from the logistic model (Fig. 4A-B). The original points were divided by 10 and rounded to get the close integers as the simplified points (Fig. 4B). Death risk discrimination by the sSOLAR score was close to the original SOLAR score (AUC: Derivation 0.689 [0.674–0.705] vs. 0.695 [0.680–0.711], P = 0.027, Validation 0.687 [0.632–0.743] vs. 0.706 [0.651–0.761], P = 0.071) (Fig. 4C-D). To facilitate risk stratification, probability of 28-day death was categorized as low, mediate, and high risk thresholding at 30% and 50%. Horizontal bar plots showed an increased mortality rate according to increasing sSOLAR risk categories (P < 0.05) (Fig. 4E).

Fig. 4.

Mortality risk stratification by the nomogram. A: The nomogram based on the levels of SOFA, Lactie acid and LAR. B: A simplified score based on SOLAR nomogram (sSOLAR). C: ROC of the simplified SOLAR(sSOLAR) model and the SOLAR model in the Derivation cohort. D: ROC of the simplified SOLAR(sSOLAR) model and the SOLAR model in the Validation cohort. E: The simplified scoring strategy was summarized with low, mediate, and high risk

Discussion

In this study, we investigated whether combining LAR could improve the performance of the lactate-enhanced SOFA score for 28-day death prediction in sepsis. We found that higher SOFA score, lactate level, and LAR were each associated with increased short-term mortality in both derivation and validation cohorts. More importantly, the SOLAR score consistently outperformed the conventional SOFA-lactate score in discrimination, reclassification, and clinical utility analyses. These findings suggest that LAR provides additional prognostic information and may serve as a useful complement to established sepsis risk stratification tools.

LAR has been reported as a prognostic marker for mortality outcomes in severely ill patients. A study showed that increased LAR could be used as a prognostic biomarker in patients with severe ARDS due to COVID-19 [22]. A Korean study reported that LAR predicts in-hospital mortality in patients with severe infection requiring intensive care [19].In pulmonary infection, higher LAR was associated with elevated risk in-hospital mortality [23]. LAR also showed prognostic power in other inflammatory diseases. For instance, several studies demonstrated that higher LAR is associated with higher mortality in kidney injury [18]. Meanwhile, increased LAR has also been observed in tumor and cerebrovascular disease [24–30]. Mechanistically, LAR may reflect the overall inflammatory burden of the patient and the severity of the disease, thereby providing prediction for clinical outcomes.

SOFA score has originally been developed to evaluate critically ill patients with sepsis by indicating organ function status. Recent advances in clinical research are challenging the efficacy of SOFA in various diseases. SOFA scoring involves multiple organ systems and requires a variety of physiological and biochemical indicators to be collected [8]. Interestingly, another study suggested that omitting the hepatic content may even improve the performance of the SOFA score [31]. However, whether the simplification may improve SOFA in a more common conditions remains uncertain. Moreover, the SOFA score may not be sensitive enough to detect the early stages of sepsis and not always able to accurately reflect the rapid changes in the condition of sepsis patients, which may lead to delays in diagnosis and treatment. Some patients may have a low SOFA score despite having evolving sepsis, which could lead to delayed treatment and worse outcomes [4, 32]. Also, enhancement by additional biomarkers has been shown to extend the prognostic role of SOFA in sepsis and ARDS. Of note, the lactate-enhanced SOFA score has been proved as a cost-effective replacement of SOFA score in low-income areas [33]. Another modified cardiovascular version of SOFA (CV SOFA) has been developed and validated based on Korean cohorts, showing better mortality discrimination over the original CV SOFA after incorporating lactate and NE equivalent dose [10]. Respiratory parameters may also be valuable improvements to the SOFA score [34].However, some patients with sepsis may not receive vasopressors, thereby CV SOFA could not be calculated. We used LAR, a simple marker easy to calculate from routine biochemistry test with nearly no additional cost, to further optimize the lactate-enhance SOFA and found modest improvement in mortality prediction. In a UK cohort including critically ill patients, age, lowest recorded temperature, highest recorded heart rate and hematocrit was added to the SOFA score and shown to enhance the prognostic efficacy for 28-day and 6-month mortality. This study was confined to postoperative patients with fecal peritonitis (FP) and conducted in a relatively small sample size for the derivation cohort [35]. In pediatric patients, age-adjusted heart rate, respiratory rate, capillary refill, and consciousness level was shown to improve the performance of quick SOFA (qSOFA) for triaging febrile children at risk for CC admission and identifying sepsis-related mortality [36]. Another Italian study reported altered mentation and respiratory index as effect addons to qSOFA in death prediction [37]. However, some of these extended variables may not be determined on admission. Therefore, the score may not be obtained at baseline. Also, the proposed value may not be obtained in a universal time window and some markers are determined objectively, therefore may reduce the robustness of the results. A study in west China cohort reported TIPS as an effective enhancement to the qSOFA score in predicting 28-day mortality. The TIPS was calculated based on D-dimer and PCT [38]. Obtaining these markers may be limited by instrumental conditions and occasionally beyond test limitations. Moreover, the PCT test are relatively expensive and may not be available in some primary care settings. Other addon markers, such as heparin-binding protein [39] and heart rate variability [40], are not readily available on admission or beyond conventional laboratory tests. They may less likely provide a timely resolution for critically ill patients. In our study, LAR is a timely and cost-effective marker that is easy to obtain at baseline, thereby providing a clinically ready approach for SOFA optimization. Other studies based on large cohorts incorporated more sophisticate and expensive markers but did not show higher AUC values compared to our model [10, 39]. In addition, we developed the SOLAR score derived from a public large cohort comprising different ethnic groups, performed external validation showed good concordance and calibration.And the Super Learner model achieved a higher AUC than the final parsimonious model. This difference is expected because Super Learner is a flexible ensemble approach that optimizes predictive performance across candidate algorithms, whereas the final model was intentionally restricted to a small number of clinically interpretable variables for bedside use. Therefore, the Super Learner result should be interpreted as an exploratory benchmark of achievable discrimination rather than as the final clinical model. Based on this we also determined the best algorithm to interpret the SOLAR model and subsequently generated a clinical friendly online version for point-of-care application. The robustness of the SOLAR score was further verified in sensitivity analysis with multiple imputation of the missing data.

There are some limitations to our study. First of all, the SOLAR model established in this study was based on the MIMIC database from the United States, which may limit its generalization to other global regions, such as underdeveloped areas. As in some landmark articles, we excluded patients with malignant tumors because of the potential impact on LAR [29, 39].And because SOFA, lactate, and LAR were categorized for clinical interpretability, some prognostic information may have been lost. Although the categorized model showed comparable performance to the continuous-variable model (Supplemental Table 5), future studies may further evaluate continuous or spline-based modeling strategies, particularly for electronic risk calculators where interpretability does not require categorization.Therefore, whether the SOLAR score could be useful in this specific population remains to be elucidated. In addition, we did not analyze the dynamic trace of SOFA, lactate, and LAR in prediction mortality. As the time-dependent bundle of SOLAR score components may provide additional information, especially the effect of early treatment, our future study will explore the role of dynamic SOLAR score as a prognostic tool for sepsis.

Conclusions

SOFA combined with LAR (SOLAR) is more accurate than SOFA score alone in predicting the short-term prognosis of patients with sepsis and can provide a good reference value for the diagnosis and treatment of sepsis. The simplified score based on SOLAR nomogram (sSOLAR) can easily predict the prognosis of patients with sepsis.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

All authors contributed to the study concept and design. WJ provided administrative support. WZY, LSF performed data acquisition. WJ, WZY, YT, XX and LYJ drafted the manuscript. All authors critically revised the manuscript. YT, WJ, and WZY are responsible for the statistical analysis. All authors read and approved the final manuscript.

Funding

The study was supported by Science Foundation of Jiangsu Commission of Health (M2022086)、Suzhou Basic Research Project (SSD2024053) and Natural Science Foundation of Jiangsu Province (BK20241798).

Data availability

Data for the derivation cohort were conducted using a free and open clinical database, namely, the Medical Information Mart for Intensive Care IV (MIMIC-IV).(https://physionet.org/content/mimiciv/0.4/). The data in the validation cohort can be obtained from the corresponding author if needed.

Declarations

Ethical approval

Derivation cohort: the establishment of this database was approved by the Massachusetts Institute of Technology (Cambridge, MA) and Beth Israel Deaconess Medical Center (Boston, MA), and consent was obtained for the original data collection. Therefore, the ethical approval statement and the need for informed consent were waived for the derivation cohort of the manuscript. Validation cohort: the validation cohort protocol was approved by the Ethic Committee of the first affiliated hospital of Soochow University (Approval Number (2023) LRPA No. 017). As this was a retrospective, non-interventional study involving only the review of existing medical records, the requirement for obtaining individual patient informed consent was waived by the Ethics Committee. The waiver was granted on the grounds that the study posed no additional risk to the patients and could not practicably be carried out without the waiver. All patient data were anonymized to ensure confidentiality. The study was conducted in accordance with the principles of the Declaration of Helsinki.

Competing interests

The authors declare no competing interests.