Early arterial lactate trajectories and mortality risk in critically ill heart failure: a two-cohort trajectory analysis

Abstract

Lactate is widely used as a biomarker of tissue hypoperfusion and illness severity in critically ill patients with heart failure (HF). While static lactate levels have prognostic value, dynamic changes in lactate over time may offer deeper insights into metabolic stress and clinical outcomes. However, the prognostic utility of lactate trajectories remains poorly defined in HF populations. We conducted a retrospective cohort study using the MIMIC-IV (n = 5,261) and MIMIC-III (n = 906) databases to identify distinct early arterial lactate trajectories in ICU-admitted HF patients. Latent class mixed model were used to categorize 72-hour lactate patterns, and association with in-hospital, 28-day, and 1-year mortality were assessed using multivariable logistic and Cox regression models. External validation was performed in the MIMIC-III cohort. Three distinct lactate trajectory classes were identified: low-stable (Class 1, 86.4%), early rise with delayed decline (Class 2, 4.1%), and early decline followed by re-elevation (Class 3, 9.6%). Compared with Class 1, Class 2 had higher in-hospital mortality (OR 6.88, 95% CI 4.86–9.74), 28-day mortality (HR 3.88, 95% CI 3.17–4.75), and 1-year mortality (HR 3.16, 95% CI 2.65–3.78; all P < 0.001). Class 3 also showed higher risks versus Class 1 (OR 3.03, 95% CI 2.32–3.98; 28-day HR 2.20, 95% CI 1.83–2.66; 1-year HR 1.83, 95% CI 1.56–2.15; all P < 0.001). Risks showed a consistent gradient (Class 2 > Class 3 > Class 1) across cohorts. Findings were consistent in the validation cohort. Sensitivity and subgroup analyses confirmed the robustness of these associations. Early arterial lactate trajectories were independently associated with both mortality in critically ill patients with HF. Trajectory-based profiling provides more nuanced prognostic insight than initial lactate values alone and may inform early risk stratification and ICU decision-making.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-27105-7.

Introduction

Heart failure (HF) remains a leading cause of morbidity and mortality worldwide and is frequently complicated by acute decompensation requiring intensive care unit (ICU) admission¹. In acute decompensated HF, metabolic disturbances, particularly impaired tissue perfusion and anaerobic glycolysis, are common and often manifest as elevated arterial lactate levels, a well-established marker of systemic hypoxia and illness severity². While single lactate measurements are routinely used for early risk stratification, recent literature has shifted attention toward the dynamic behavior of lactate over time³. Specifically, changes in lactate concentration during the early phase of critical illness, may better reflect therapeutic response and prognosis compared to isolated values. Lactate clearance, defined as the percentage reduction in lactate over time, has been associated with improved survival across a range of conditions including sepsis, cardiogenic shock, and postoperative states^4–7. However, these analyses often rely on simple two-point comparisons and may miss non-linear or heterogeneous trajectories.

Recent advances in ICU research have leveraged trajectory modeling approaches such as Latent Class Mixed Model (LCMM) to uncover patient subgroups with distinct biomarker dynamics. Prior studies using MIMIC cohorts have demonstrated that temporal trajectories of glucose, sodium, and platelet counts are associated with differential outcomes in sepsis and acute kidney injury^8–11. Similarly, recent research has applied trajectory modeling to electrolyte profiles such as serum sodium and potassium in critically ill patients with heart failure, uncovering distinct patterns that are associated with short-term mortality^12,13. However, although lactate is a cornerstone of hemodynamic and metabolic assessment, trajectory-based analyses of its dynamic behavior remain largely unexplored in HF populations.

Beyond its role as a static biomarker, lactate is increasingly recognized as a dynamic metabolic signal involved in myocardial energetics and stress responses. In HF, the failing myocardium can utilize lactate as an alternative fuel via monocarboxylate transporters (MCTs), particularly under hypoxic conditions¹⁴. Beyond hypoperfusion, lactate in HF may reflect multiple, non-mutually exclusive processes—including altered myocardial fuel use, lactate-driven gene regulation (lysine lactylation), vascular homeostasis, mitochondrial integrity, buffering/acid–base effects, and systemic stress responses. Accordingly, we interpret trajectories as risk phenotypes rather than surrogates of a single mechanism. Emerging evidence links lysine lactylation to myocardial remodeling, underscoring the biological plausibility of heterogeneous lactate trajectories in HF^15,16. However, chronic cardiac stress may induce MCTs-mediated lactate efflux, leading to intracellular depletion and potentially impairing adaptive responses¹⁷. These mechanistic insights underscore the complexity of lactate metabolism in HF and highlight the potential value of evaluating its temporal trajectories rather than isolated measurements to obtain a more comprehensive view of risk.

Because a single lactate value may incompletely reflect early metabolic dynamics, we examined whether 72-hour arterial lactate trajectories identify reproducible risk phenotypes in critically ill HF. In this study, we aimed to characterize early arterial lactate trajectories in critically ill HF patients using LCMM based on the MIMIC-IV database, and to assess the relationship between these trajectories and mortality. External validation was performed using the MIMIC-III cohort. We hypothesized that distinct lactate kinetic patterns would be independently associated with short-term and long-term mortality, thereby providing novel insights for prognostic stratification in HF.

Materials and methods

Data source

This research utilized data extracted from the non-overlapping sections of the publicly available the Medical Information Mart for Intensive Care-III Carevue and MIMIC-IV (version 3.1) databases, which are maintained by the Beth Israel Deaconess Medical Center^18,19. Access to these databases was granted to the corresponding author upon successful completion of the Collaborative Institutional Training Initiative (CITI) program and agreement to the PhysioNet Data Use Agreement. As all patient information in these databases is fully de-identified, the study was exempt from institutional review board oversight and the requirement for informed consent was waived.

Participants

Patients who met the following criteria were included in the study: (1) The patient was admitted with a diagnosis of HF; (2) The number of blood gas lactate measurements was greater than two in the first three days; (3) The patient was older than 18 years; (4) The patient was in ICU for more than one day; and (5) The records of the first ICU admission were obtained.

Data extraction and outcomes

We extracted detailed clinical information on the first day of ICU admission, including baseline information, comorbidities, laboratory parameters, interventions, etc., and only the first measurement results were taken for multiple measurements. The primary outcomes included in-hospital mortality rate, 28-day mortality rate and 1-year mortality rate. The secondary outcomes included the length of stay in the ICU and the hospital.

Statistical analysis

Missing values were addressed using the Multiple Imputation by Chained Equations (MICE) package²⁰ (v3.18.0) with m = 20 imputations and 20 iterations per chain (random seed fixed for reproducibility). For continuous variables we used predictive mean matching (PMM, k = 5); for binary and multicategory variables we used logistic and multinomial models, respectively. Estimates were combined using Rubin’s rules. Only covariates were imputed. Arterial lactate time-series (trajectory exposure) were not imputed, as the LCMM accommodates irregular timing under a MAR assumption. Variable-level missingness and the corresponding imputation model are summarized in Supplementary Fig. 1. Extreme outliers were identified and handled to minimize their impact on the results. For continuous covariates, extreme values were handled via winsorization at the 0.5th and 99.5th percentiles to limit undue leverage while preserving rank order; ordinary outliers were not deleted, and no additional trimming was applied. The distribution of continuous variables was assessed using the Kolmogorov-Smirnov test. Continuous variables following a normal distribution were presented as mean ± standard deviation (SD), while those not conforming to normality were reported as median (interquartile range). Categorical variables were summarized as counts (percentages). For comparisons between groups, the Kruskal-Wallis test was applied to continuous variables, and the Pearson chi-square test was used for categorical variables. The data were fitted using a Latent Class Mixed Model (LCMM) implemented in R (lcmm package, v2.2.1)²¹, which assumes that the patient population is heterogeneous, composed of G potential categories and has different trajectories (Supplementary Methods). We used all lactate data measured during the first 72 h of patient admission. The classification evaluation follows the following criteria: (1) Akaike information criterion (AIC), Bayesian information criterion (BIC), and sample-size adjusted BIC (SABIC), with lower values indicating better overall fit. SABIC applies a sample-size-adjusted complexity penalty and is commonly reported alongside AIC/BIC in latent class trajectory modeling, serving as a complementary index when added classes yield only marginal gains in fit; (2) Entropy (higher is better) to assess class separation; (3) The minimum proportion of each classification is not less than 1% to ensure that the classification is meaningful; (4) The average posterior probability is greater than 0.9. The association between different classifications and in-hospital mortality was assessed using logistic regression analysis. To evaluate the impact on 28-day and 1-year mortality, Cox proportional hazards models were employed, with group comparisons conducted via the log-rank test. To minimize confounding, several models were constructed: Model 1 adjusted for baseline characteristics; Model 2 further adjusted for comorbidities and laboratory parameters; and Model 3 included adjustments for interventions in addition to the variables in Model 2. Prespecified subgroup analyses were also carried out. All statistical analyses were performed using R software (version 4.3.3), and a two-sided P value of less than 0.05 was considered statistically significant.

In order to facilitate rapid clinical identification and timely intervention of lactate trajectory abnormalities, a machine learning-based classification model was developed²². In this study, the MIMIC-IV cohort was used as the training set, and the near-zero variance variables were removed to eliminate redundant information, and then the data were normalized to reduce the feature scale differences. Based on this, JMIM (Joint Mutual Information Maximization) algorithm was applied to rank the importance of features, and the key features with scores over 0.8 were selected for model construction. The TabPFN algorithm was used to construct the model, and its discriminant performance was evaluated by receiver operating characteristic (ROC) curve. Finally, the MIMIC-III cohort was used to externally validate the model and comprehensively evaluate its generalization ability and clinical applicability. The contribution of each feature to the prediction results of the model was quantified by the method of Shapley Additive ex Planations (SHAP) to enhance the interpretability of the model.

Results

Baseline characteristics of the study population

A total of 5,261 critically ill patients with HF were included from the MIMIC-IV database for the primary analysis, with an in-hospital mortality rate of 20.74%. An additional 906 patients from the MIMIC-III database were used as an external validation cohort, with a mortality rate of 33.44% (Table 1 and Supplementary Table 1). In the MIMIC-IV cohort, the median age was 73.2 years (IQR: 63.4–81.7), and 59.9% were male. The most prevalent comorbidities were hypertension (77.9%), atrial fibrillation (55.2%), diabetes (42.6%), and chronic kidney disease (38.4%). The median SOFA score was 7.0 (IQR: 5.0–9.0), and the Charlson comorbidity index was 7.0 (IQR: 5.0–8.0). The median initial arterial lactate level was 1.7 mmol/L (IQR: 1.2–2.7).

Table 1.

Baseline characteristics of critically ill patients with heart failure by in-hospital mortality status (MIMIC-IV).

Variables	Total (n = 5,261)	Survival (n = 4,170)	Death (n = 1,091)	P
Basic characteristics
Age (years), M (Q₁, Q₃)	73.16 (63.42, 81.70)	72.02 (62.16, 80.51)	77.48 (68.18, 84.80)	< 0.001
Weight (kg), M (Q₁, Q₃)	81.00 (67.80, 97.60)	81.60 (68.00, 98.27)	79.00 (66.25, 94.95)	< 0.001
Gender, n (%)				0.01
Female	2,112 (40.14)	1,637 (39.26)	475 (43.54)
Male	3,149 (59.86)	2,533 (60.74)	616 (56.46)
Race, n (%)				< 0.001
White	451 (8.57)	374 (8.97)	77 (7.06)
Black	1,392 (26.46)	1,052 (25.23)	340 (31.16)
Other	3,418 (64.97)	2,744 (65.80)	674 (61.78)
Comorbidities
Atrial fibrillation, n (%)				0.046
No	2,359 (44.84)	1,899 (45.54)	460 (42.16)
Yes	2,902 (55.16)	2,271 (54.46)	631 (57.84)
Hypertension, n (%)				0.655
No	1,160 (22.05)	914 (21.92)	246 (22.55)
Yes	4,101 (77.95)	3,256 (78.08)	845 (77.45)
Myocardial infarction, n (%)				0.075
No	3,281 (62.36)	2,626 (62.97)	655 (60.04)
Yes	1,980 (37.64)	1,544 (37.03)	436 (39.96)
Pulmonary disease, n (%)				0.148
No	3,506 (66.64)	2,799 (67.12)	707 (64.80)
Yes	1,755 (33.36)	1,371 (32.88)	384 (35.20)
Renal disease, n (%)				< 0.001
No	3,239 (61.57)	2,652 (63.60)	587 (53.80)
Yes	2,022 (38.43)	1,518 (36.40)	504 (46.20)
Liver disease, n (%)				< 0.001
No	4,631 (88.03)	3,757 (90.10)	874 (80.11)
Yes	630 (11.97)	413 (9.90)	217 (19.89)
Diabetes, n (%)				0.992
No	3,018 (57.37)	2,392 (57.36)	626 (57.38)
Yes	2,243 (42.63)	1,778 (42.64)	465 (42.62)
Malignant cancer, n (%)				< 0.001
No	4,823 (91.67)	3,877 (92.97)	946 (86.71)
Yes	438 (8.33)	293 (7.03)	145 (13.29)
Illness severity
Charlson comorbidity score, M (Q₁, Q₃)	7.00 (5.00, 8.00)	6.00 (5.00, 8.00)	8.00 (6.00, 9.00)	< 0.001
SOFA, M (Q₁, Q₃)	7.00 (5.00, 9.00)	6.00 (4.00, 9.00)	8.00 (6.00, 11.00)	< 0.001
Laboratory
WBC (×109/L), M (Q₁, Q₃)	12.10 (8.50, 17.00)	12.00 (8.40, 16.60)	13.10 (9.30, 18.50)	< 0.001
RBC (×109/L), M (Q₁, Q₃)	3.47 (2.92, 4.11)	3.47 (2.92, 4.12)	3.49 (2.94, 4.09)	0.971
Platelet (×109/L), M (Q₁, Q₃)	180.00 (132.00, 246.00)	178.00 (133.00, 242.00)	188.00 (129.50, 261.50)	0.041
RDW (%), M (Q₁, Q₃)	14.90 (13.80, 16.50)	14.70 (13.70, 16.20)	15.80 (14.40, 17.70)	< 0.001
Anion gap (mmol/L), M (Q₁, Q₃)	15.00 (12.00, 18.00)	14.00 (12.00, 18.00)	17.00 (13.00, 20.00)	< 0.001
Bicarbonate (mmol/L), M (Q₁, Q₃)	22.00 (19.00, 25.00)	22.00 (20.00, 25.00)	21.00 (18.00, 25.00)	< 0.001
BUN (mg/dL), M (Q₁, Q₃)	27.00 (18.00, 45.00)	25.00 (17.00, 41.00)	37.00 (24.00, 57.00)	< 0.001
Calcium (mg/dL), M (Q₁, Q₃)	8.40 (7.90, 8.90)	8.40 (7.90, 8.90)	8.30 (7.80, 8.80)	0.067
Chloride (mmol/L), M (Q₁, Q₃)	103.00 (98.00, 108.00)	104.00 (99.00, 108.00)	102.00 (97.00, 106.00)	< 0.001
Creatinine (mg/dL), M (Q₁, Q₃)	1.30 (0.90, 2.00)	1.20 (0.90, 1.90)	1.60 (1.10, 2.70)	< 0.001
Glucose (mg/dL), M (Q₁, Q₃)	136.00 (109.00, 182.00)	133.00 (109.00, 176.00)	150.00 (113.00, 202.00)	< 0.001
Sodium (mmol/L), M (Q₁, Q₃)	138.00 (135.00, 141.00)	138.00 (135.00, 141.00)	138.00 (134.00, 141.00)	0.416
Potassium (mmol/L), M (Q₁, Q₃)	4.30 (3.90, 4.80)	4.30 (3.90, 4.80)	4.40 (3.90, 5.00)	< 0.001
Magnesium (mg/dL), M (Q₁, Q₃)	2.10 (1.80, 2.40)	2.10 (1.80, 2.50)	2.10 (1.80, 2.30)	< 0.001
Lactate (mmol/L), M (Q₁, Q₃)	1.70 (1.20, 2.70)	1.70 (1.20, 2.50)	2.20 (1.50, 3.80)	< 0.001
Vital signs
PO2 (mmHg), M (Q₁, Q₃)	111.00 (55.00, 303.00)	128.00 (59.00, 336.00)	81.00 (47.00, 153.50)	< 0.001
PCO2, (mmHg) M (Q₁, Q₃)	42.00 (36.00, 49.00)	42.00 (37.00, 48.00)	42.00 (35.00, 52.00)	< 0.001
PH, M (Q₁, Q₃)	7.37 (7.30, 7.41)	7.37 (7.31, 7.42)	7.33 (7.25, 7.39)	0.118
HR (/min), M (Q₁, Q₃)	86.00 (76.00, 101.00)	86.00 (76.00, 99.75)	90.00 (77.00, 107.00)	< 0.001
SBP (mmHg), M (Q₁, Q₃)	114.00 (100.00, 130.00)	114.00 (100.00, 130.00)	112.00 (98.00, 132.00)	< 0.001
DBP (mmHg), M (Q₁, Q₃)	62.00 (52.00, 75.00)	62.00 (52.00, 74.00)	63.00 (52.00, 76.00)	0.101
RR (/min), M (Q₁, Q₃)	19.00 (16.00, 23.00)	18.00 (16.00, 23.00)	21.00 (17.00, 25.00)	0.368
Temperature (℃), M (Q₁, Q₃)	36.56 (36.22, 36.94)	36.56 (36.22, 36.94)	36.61 (36.28, 36.95)	< 0.001
SpO2, M (Q₁, Q₃)	98.00 (95.00, 100.00)	98.00 (95.00, 100.00)	97.00 (94.00, 100.00)	0.238
Urineoutput (ml), M (Q₁, Q₃)	1,325.00 (750.00, 2,145.00)	1,422.50 (880.00, 2,270.00)	841.00 (353.50, 1,612.50)	< 0.001
Intervention during ICU
Mechanical ventilation, n (%)				0.041
No	1,821 (34.61)	1,472 (35.30)	349 (31.99)
Yes	3,440 (65.39)	2,698 (64.70)	742 (68.01)
RRT, n (%)				< 0.001
No	4,991 (94.87)	4,016 (96.31)	975 (89.37)
Yes	270 (5.13)	154 (3.69)	116 (10.63)
Vasoactive drugs, n (%)				0.117
No	2,188 (41.59)	1,757 (42.13)	431 (39.51)
Yes	3,073 (58.41)	2,413 (57.87)	660 (60.49)
Antibiotic, n (%)				0.305
No	1,155 (21.95)	903 (21.65)	252 (23.10)
Yes	4,106 (78.05)	3,267 (78.35)	839 (76.90)
ACEI, n (%)				< 0.001
No	5,020 (95.42)	3,954 (94.82)	1,066 (97.71)
Yes	241 (4.58)	216 (5.18)	25 (2.29)
Beta blocker, n (%)				< 0.001
No	3542 (67.33)	2739 (65.68)	803 (73.60)
Yes	1719 (32.67)	1431 (34.32)	288 (26.40)
Glucocorticoid, n (%)				< 0.001
No	4534 (86.18)	3687 (88.42)	847 (77.64)
Yes	727 (13.82)	483 (11.58)	244 (22.36)
Anticoagulant, n (%)				< 0.001
No	2025 (38.49)	1751 (41.99)	274 (25.11)
Yes	3236 (61.51)	2419 (58.01)	817 (74.89)
Diuretic, n (%)				< 0.001
No	2531 (48.11)	1955 (46.88)	576 (52.80)
Yes	2730 (51.89)	2215 (53.12)	515 (47.20)
Clinical outcomes
Hospital length of stay (days), M (Q₁, Q₃)	10.99 (7.00, 17.72)	11.79 (7.78, 18.47)	7.86 (3.72, 14.84)	< 0.001
ICU length of stay (days), M (Q₁, Q₃)	4.10 (2.30, 7.25)	3.98 (2.27, 6.84)	4.96 (2.55, 9.29)	< 0.001

Comparison of demographic, clinical, and laboratory variables between survivors and non-survivors during hospital stay in the MIMIC-IV cohort.

M: Median, Q1: 1st Quartile, Q3: 3st Quartile.

ACEI: angiotensin-converting enzyme inhibitor. WBC: white blood cell; RBC: red blood cell; RDW: red cell distribution width; BUN: blood urea nitrogen; HR: heart rate; RR: respiratory rate; SBP: systolic blood pressure; DBP: diastolic blood pressure; SpO₂: percutaneous arterial oxygen saturation; SOFA: Sequential Organ Failure Assessment score; RRT: renal replacement therapy.

Compared with survivors, non-survivors had significantly older age (77.5 vs. 72.0 years, P < 0.001), higher initial lactate levels (2.2 vs. 1.7 mmol/L, P < 0.001), and higher levels of BUN, RDW, creatinine, anion gap, and glucose (all P < 0.001). They also exhibited lower urine output and bicarbonate, and were more likely to receive mechanical ventilation and renal replacement therapy. Similar trends in age, lactate, comorbidity burden, and organ function markers were observed in the MIMIC-III cohort, reinforcing the consistency and generalizability of the baseline characteristics across populations.

Measurement frequency, timing, and representativeness

Within 0–72 h of ICU admission, lactate sampling clustered into three categories—Day 1 only, Days 1–2 (≥ 1 draw on both days), and Days 1–3 (≥ 1 draw on each day)—with Days 1–3 showing the highest measurement counts (Supplementary Fig. 4). As expected, patients sampled on Days 1–3 had greater illness severity and more organ support than those with < 3 draws, reflecting closer monitoring (Supplementary Table 7). These patterns contextualize generalizability: the trajectory analysis pertains to patients with sufficient repeated assessments and documents real-world sampling heterogeneity.

Identification of lactate trajectory classes

Based on the LCMM, three distinct arterial lactate trajectory classes within the first 72 h of ICU admission were identified among critically ill patients with HF in the MIMIC-IV cohort (Fig. 1A–C). Class 1 (n = 4,544; 86.37%) exhibited consistently low and stable lactate levels throughout the initial 72-hour period (Low-Stable type). Class 2 (n = 214; 4.07%) showed a moderate baseline lactate concentration with an initial upward trend followed by a gradual decline, indicating a rising-then-falling pattern. Class 3 (n = 503; 9.56%) started with relatively high lactate levels, exhibited an early decline, but subsequently rebounded in the later phase, forming a falling-then-rising pattern. Model fit indices (Table 2) favored a 3-class solution with lower AIC/BIC/SABIC, high entropy, and adequate class sizes. Posterior probabilities for class assignment were all above 0.9 (Supplementary Table 4), indicating good model discrimination. External validation in the MIMIC-III cohort confirmed similar patterns and temporal trends in lactate dynamics (Fig. 1D–F), supporting the reproducibility of trajectory phenotypes. Model fit statistics and posterior classification probabilities for the validation cohort are presented in Supplementary Tables 5 and 6, respectively. The overall study design and workflow are illustrated in Fig. 2. To assess robustness, we repeated the LCMM derivation without excluding any lactate values. The same three trajectory shapes and the same mortality risk gradient were observed, with only modest shifts in class proportions (Supplementary Fig. 5). These findings indicate that the primary conclusions are unchanged and support the prespecified extreme-value handling for clinical plausibility and model stability. These findings support the existence of clinically relevant and biologically distinct lactate trajectory classes in critically ill HF patients.

Fig. 1.

Identification and validation of early arterial lactate trajectory classes in critically ill patients with heart failure using a LCMM. Three distinct classes of early arterial lactate trajectories were identified using a LCMM based on serial lactate measurements during the first 72 h of ICU admission in the MIMIC-IV cohort (Panels A–C), and externally validated in the MIMIC-III cohort (Panels D–F): Class 1 (green): Low-Stable — persistently low and stable lactate levels. Class 2 (blue): Moderate-Increasing — moderately elevated lactate levels with a rising trend. Class 3 (yellow): High-Decreasing — initially high lactate levels followed by a declining pattern. The x-axis represents time in days from ICU admission (0 to 3 days), and the y-axis denotes arterial lactate concentration (mmol/L). Each line represents the model-estimated mean trajectory with 95% confidence intervals (shaded). All included patients had ≥ 3 lactate measurements within 72 h of ICU admission. The apparent upper bound reflects the pre-specified exclusion of extreme lactate values; sensitivity analyses without exclusion yielded similar trajectory shapes and risk gradients.

Table 2.

Model fit statistics for lactate trajectory classification in critically ill heart failure patients from MIMIC-IV.

Class	Log likelihood	AIC	SABIC	BIC	Entropy	Proportion of people (%)
						Class1	Class2	Class3	Class4	Class5	Class6
1	-69416.1016	138846.2032	138869.936	138892.1797	1	100
2	-65573.29246	131170.5849	131211.2698	131249.4018	0.831643549	87.11271621	12.88728379
3	-63856.32402	127746.648	127804.2849	127858.3053	0.885980368	86.37141228	9.560919977	4.067667744
4	-63205.81519	126455.6304	126530.2192	126600.1281	0.76238103	71.29823228	20.45238548	3.022239118	5.227143129
5	-62867.65365	125789.3073	125880.8482	125966.6454	0.793051346	17.98137236	72.70480897	3.68751188	1.444592283	4.181714503
6	-62593.74109	125251.4822	125359.9751	125461.6606	0.759372065	63.73313058	24.78616233	2.946207945	1.444592283	5.208135335	1.881771526

Statistics for choosing the best number of classes. AIC: Akaike Information Criterion; BIC: Bayesian Information Criterion; Entropy: measure of classification certainty (range: 0–1); SABIC: Sample-size Adjusted BIC. Lower AIC/BIC/SABIC indicate better fit after penalizing complexity; higher entropy indicates better class separation.

Fig. 2.

Flowchart of study participants recruitment and heart failure patients subclasses development/validation. MIMIC: Medical Information Mart for Intensive Care; ICU: Intensive Care Unit; HF: heart failure.

Clinical characteristics by lactate trajectory

Key clinical characteristics varied significantly across the three lactate trajectory classes (Table 3; Supplementary Table 2). Patients in class 1 (low-stable trajectory) exhibited the least severe illness, with the lowest SOFA scores, minimal organ support requirements, and lower inflammatory and metabolic derangements. In contrast, class 2 (moderately elevated with rising trend) had the highest SOFA scores and required more frequent mechanical ventilation (61.7%), vasopressors (57.0%), and renal replacement therapy (24.8%). These patients also had the lowest urine output and highest initial lactate levels (median 5.1 mmol/L), as well as elevated white blood cell count and RDW. Class 3 (initial decrease followed by re-elevation) showed intermediate clinical severity across most parameters. These trends were consistently reproduced in the MIMIC-III cohort, supporting the external validity of trajectory-based stratification. Collectively, these findings suggest that dynamic lactate patterns reflect underlying organ dysfunction and are closely associated with acute illness severity in critically ill patients with HF. These baseline and clinical differences suggest that lactate trajectory classes reflect distinct pathophysiological profiles that may underlie the observed variation in short- and long-term mortality.

Table 3.

Baseline characteristics and covariates of the study population stratified by lactate trajectory class (MIMIC-IV).

Variables	Total (n = 5,261)	Class 1 (n = 4,544)	Class 2 (n = 214)	Class 3 (n = 503)	P
Basic characteristics
Age, M (Q₁, Q₃)	73.16 (63.42, 81.70)	73.12 (63.41, 81.74)	73.60 (62.51, 81.17)	73.08 (64.04, 81.47)	0.987
Weight, M (Q₁, Q₃)	81.00 (67.80, 97.60)	81.00 (67.80, 98.00)	80.00 (67.80, 93.92)	80.00 (68.00, 96.15)	0.54
Gender, n (%)					0.746
Female	2,112 (40.14)	1,817 (39.99)	91 (42.52)	204 (40.56)
Male	3,149 (59.86)	2,727 (60.01)	123 (57.48)	299 (59.44)
Race, n (%)					0.025
White	451 (8.57)	376 (8.27)	20 (9.35)	55 (10.93)
Black	1,392 (26.46)	1,178 (25.92)	65 (30.37)	149 (29.62)
Other	3,418 (64.97)	2,990 (65.80)	129 (60.28)	299 (59.44)
Comorbidities
Atrial fibrillation, n (%)					0.302
No	2,359 (44.84)	2,056 (45.25)	88 (41.12)	215 (42.74)
Yes	2,902 (55.16)	2,488 (54.75)	126 (58.88)	288 (57.26)
Hypertension, n (%)					0.3
No	1,160 (22.05)	990 (21.79)	56 (26.17)	114 (22.66)
Yes	4,101 (77.95)	3,554 (78.21)	158 (73.83)	389 (77.34)
Myocardial infarction, n (%)					0.537
No	3,281 (62.36)	2,846 (62.63)	127 (59.35)	308 (61.23)
Yes	1,980 (37.64)	1,698 (37.37)	87 (40.65)	195 (38.77)
Pulmonary disease, n (%)					0.014
No	3,506 (66.64)	2,994 (65.89)	152 (71.03)	360 (71.57)
Yes	1,755 (33.36)	1,550 (34.11)	62 (28.97)	143 (28.43)
Renal disease, n (%)					0.504
No	3,239 (61.57)	2,792 (61.44)	127 (59.35)	320 (63.62)
Yes	2,022 (38.43)	1,752 (38.56)	87 (40.65)	183 (36.38)
Liver disease, n (%)					< 0.001
No	4631 (88.03)	4070 (89.57)	160 (74.77)	401 (79.72)
Yes	630 (11.97)	474 (10.43)	54 (25.23)	102 (20.28)
Diabetes, n (%)					0.609
No	3,018 (57.37)	2,609 (57.42)	128 (59.81)	281 (55.86)
Yes	2,243 (42.63)	1,935 (42.58)	86 (40.19)	222 (44.14)
Malignant cancer, n (%)					0.015
No	4,823 (91.67)	4,185 (92.10)	193 (90.19)	445 (88.47)
Yes	438 (8.33)	359 (7.90)	21 (9.81)	58 (11.53)
Illness severity
Charlson comorbidity score, M (Q₁, Q₃)	7.00 (5.00, 8.00)	7.00 (5.00, 8.00)	7.00 (5.00, 9.00)	7.00 (5.00, 9.00)	0.004
SOFA, M (Q₁, Q₃)	7.00 (5.00, 9.00)	6.00 (4.00, 9.00)	10.00 (7.00, 13.00)	9.00 (7.00, 12.00)	< 0.001
Laboratory
WBC, M (Q₁, Q₃)	12.10 (8.50, 17.00)	12.00 (8.50, 16.60)	12.45 (7.90, 18.88)	14.30 (9.60, 20.00)	< 0.001
RBC, M (Q₁, Q₃)	3.47 (2.92, 4.11)	3.47 (2.93, 4.10)	3.52 (3.00, 4.15)	3.48 (2.84, 4.17)	0.561
Platelet, M (Q₁, Q₃)	180.00 (132.00, 246.00)	182.00 (135.00, 247.00)	183.00 (120.25, 254.50)	170.00 (112.50, 232.00)	< 0.001
RDW, M (Q₁, Q₃)	14.90 (13.80, 16.50)	14.80 (13.70, 16.40)	15.35 (14.00, 17.17)	15.20 (13.90, 17.10)	< 0.001
Anion gap, M (Q₁, Q₃)	15.00 (12.00, 18.00)	14.00 (12.00, 17.00)	17.00 (14.00, 20.75)	19.00 (15.00, 23.00)	< 0.001
Bicarbonate, M (Q₁, Q₃)	22.00 (19.00, 25.00)	22.00 (20.00, 25.00)	20.00 (16.00, 24.00)	19.00 (15.00, 22.00)	< 0.001
BUN, M (Q₁, Q₃)	27.00 (18.00, 45.00)	26.00 (18.00, 44.00)	29.50 (19.25, 48.00)	29.00 (20.00, 46.00)	0.004
Calcium, M (Q₁, Q₃)	8.40 (7.90, 8.90)	8.40 (7.90, 8.90)	8.40 (7.70, 9.00)	8.30 (7.80, 8.90)	0.817
Chloride, M (Q₁, Q₃)	103.00 (98.00, 108.00)	103.50 (99.00, 108.00)	103.00 (97.00, 108.00)	103.00 (98.00, 107.00)	0.029
Creatinine, M (Q₁, Q₃)	1.30 (0.90, 2.00)	1.30 (0.90, 2.00)	1.50 (1.10, 2.40)	1.50 (1.10, 2.20)	< 0.001
Glucose, M (Q₁, Q₃)	136.00 (109.00, 182.00)	134.00 (109.00, 177.00)	140.50 (114.25, 189.50)	160.00 (116.00, 234.50)	< 0.001
Sodium, M (Q₁, Q₃)	138.00 (135.00, 141.00)	138.00 (135.00, 141.00)	139.00 (135.25, 142.00)	138.00 (135.00, 142.00)	0.253
Potassium, M (Q₁, Q₃)	4.30 (3.90, 4.80)	4.30 (3.90, 4.80)	4.40 (4.00, 4.90)	4.40 (3.85, 5.00)	0.076
Magnesium, M (Q₁, Q₃)	2.10 (1.80, 2.40)	2.10 (1.80, 2.40)	2.00 (1.80, 2.40)	2.10 (1.80, 2.50)	0.158
Lactate, M (Q₁, Q₃)	1.70 (1.20, 2.70)	1.60 (1.20, 2.40)	2.50 (1.70, 4.88)	5.20 (2.00, 7.55)	< 0.001
Vital signs
PO2, M (Q₁, Q₃)	111.00 (55.00, 303.00)	113.00 (57.00, 311.00)	87.50 (50.25, 240.25)	101.00 (49.00, 246.00)	< 0.001
PCO2, M (Q₁, Q₃)	42.00 (36.00, 49.00)	42.00 (37.00, 49.00)	41.00 (35.00, 49.00)	40.00 (33.00, 48.00)	< 0.001
PH, M (Q₁, Q₃)	7.37 (7.30, 7.41)	7.37 (7.31, 7.42)	7.34 (7.25, 7.40)	7.32 (7.21, 7.39)	< 0.001
HR, M (Q₁, Q₃)	86.00 (76.00, 101.00)	86.00 (76.00, 100.00)	93.00 (82.00, 109.75)	90.00 (79.50, 107.00)	< 0.001
SBP, M (Q₁, Q₃)	114.00 (100.00, 130.00)	115.00 (101.00, 131.00)	108.00 (94.25, 122.00)	109.00 (95.25, 127.00)	< 0.001
DBP, M (Q₁, Q₃)	62.00 (52.00, 75.00)	62.00 (52.00, 75.00)	63.00 (52.00, 71.88)	59.00 (50.00, 75.00)	0.068
RR, M (Q₁, Q₃)	19.00 (16.00, 23.00)	18.00 (16.00, 23.00)	20.00 (16.00, 25.00)	20.00 (16.00, 25.00)	< 0.001
Temperature, M (Q₁, Q₃)	36.56 (36.22, 36.94)	36.60 (36.28, 36.94)	36.56 (36.18, 36.93)	36.50 (36.00, 36.83)	< 0.001
SpO2, M (Q₁, Q₃)	98.00 (95.00, 100.00)	98.00 (95.00, 100.00)	98.00 (94.00, 100.00)	98.00 (94.00, 100.00)	0.001
Urineoutput, M (Q₁, Q₃)	1,325.00 (750.00, 2,145.00)	1,385.00 (825.00, 2,215.00)	595.00 (212.00, 1,264.75)	952.00 (412.50, 1,662.50)	< 0.001
Intervention during ICU
Mechanical ventilation, n (%)					< 0.001
No	1,821 (34.61)	1,621 (35.67)	53 (24.77)	147 (29.22)
Yes	3,440 (65.39)	2,923 (64.33)	161 (75.23)	356 (70.78)
RRT, n (%)					< 0.001
No	4,991 (94.87)	4,374 (96.26)	169 (78.97)	448 (89.07)
Yes	270 (5.13)	170 (3.74)	45 (21.03)	55 (10.93)
Vasoactive drugs, n (%)					< 0.001
No	2,188 (41.59)	1,957 (43.07)	65 (30.37)	166 (33.00)
Yes	3,073 (58.41)	2,587 (56.93)	149 (69.63)	337 (67.00)
Antibiotic, n (%)					0.016
No	1,155 (21.95)	1,027 (22.60)	40 (18.69)	88 (17.50)
Yes	4,106 (78.05)	3,517 (77.40)	174 (81.31)	415 (82.50)
ACEI, n (%)					< 0.001
No	5,020 (95.42)	4,315 (94.96)	212 (99.07)	493 (98.01)
Yes	241 (4.58)	229 (5.04)	2 (0.93)	10 (1.99)
Beta blocker, n (%)					< 0.001
No	3542 (67.33)	2992 (65.85)	163 (76.17)	387 (76.94)
Yes	1719 (32.67)	1552 (34.15)	51 (23.83)	116 (23.06)
Glucocorticoid, n (%)					< 0.001
No	4534 (86.18)	3976 (87.50)	155 (72.43)	403 (80.12)
Yes	727 (13.82)	568 (12.50)	59 (27.57)	100 (19.88)
Anticoagulant, n (%)					< 0.001
No	2025 (38.49)	1794 (39.48)	55 (25.70)	176 (34.99)
Yes	3236 (61.51)	2750 (60.52)	159 (74.30)	327 (65.01)
Diuretic, n (%)					< 0.001
No	2531 (48.11)	2129 (46.85)	136 (63.55)	266 (52.88)
Yes	2730 (51.89)	2415 (53.15)	78 (36.45)	237 (47.12)
Clinical outcomes
Los hospital	10.99 (7.00, 17.72)	11.05 (7.21, 17.50)	8.12 (2.35, 20.70)	10.71 (5.30, 19.23)	< 0.001
Los icu	4.10 (2.30, 7.25)	4.06 (2.29, 7.06)	3.88 (1.72, 9.48)	4.77 (2.73, 8.70)	< 0.001
Hospital death, n (%)					< 0.001
No	4170 (79.26)	3799 (83.60)	79 (36.92)	292 (58.05)
Yes	1091 (20.74)	745 (16.40)	135 (63.08)	211 (41.95)

Clinical and biochemical profiles across the three latent lactate trajectory classes identified using latent class growth modeling in the MIMIC-IV cohort. M: Median, Q1: 1st Quartile, Q3: 3st Quartile.

ACEI: angiotensin-converting enzyme inhibitor. WBC: white blood cell; RBC: red blood cell; RDW: red cell distribution width; BUN: blood urea nitrogen; HR: heart rate; RR: respiratory rate; SBP: systolic blood pressure; DBP: diastolic blood pressure; SpO₂: percutaneous arterial oxygen saturation; SOFA: Sequential Organ Failure Assessment score; RRT: renal replacement therapy; LOS: length of stay.

Mortality trends across lactate trajectory classes

Kaplan-Meier survival analysis demonstrated significant differences in short- and long-term mortality among the three lactate trajectory classes in the MIMIC-IV cohort (Fig. 3). Patients in class 2, characterized by an early rising and late-declining lactate curve, exhibited the worst prognosis, with substantially reduced 28-day and 1-year survival probabilities (P < 0.001). In contrast, class 1 (low-stable lactate levels) had the most favorable outcomes, while class 3 (initial decline followed by secondary elevation) showed intermediate risk.

Fig. 3.

Kaplan–Meier survival curves stratified by early lactate trajectory classes in the MIMIC-IV cohort. Kaplan-Meier survival analysis of critically ill heart failure patients from the MIMIC-IV cohort, stratified by early arterial lactate trajectory classes identified through LCMM. (A) displays 28-day survival curves for the three classes. (B) shows 1-year survival curves for the same classes.The log-rank test was used to compare survival distributions. Class 2 exhibited the worst short- and long-term survival, while Class 1 had the most favorable prognosis.

These patterns were robustly replicated in the MIMIC-III validation cohort (Supplementary Fig. 2), where class 2 again showed the poorest survival and class 1 the most favorable. The consistent gradient of mortality risk across trajectory groups underscores the prognostic relevance of dynamic lactate profiling in critically ill patients with HF.

Multivariable regression analysis

In-hospital mortality

Multivariable logistic regression revealed significant associations between lactate trajectory classes and in-hospital mortality (Table 4). In the MIMIC-IV cohort, class 2 showed the highest adjusted odds of death compared to class 1 (OR: 6.88; 95% CI: 4.86–9.74; P < 0.001), followed by class 3 (OR: 3.03; 95% CI: 2.32–3.98; P < 0.001). Similar trends were observed in the MIMIC-III cohort, where class 2 remained a strong predictor of mortality (OR: 4.35; 95% CI: 2.27–8.33; P < 0.001), while the association for class 3 was not statistically significant.

Table 4.

Association between lactate trajectories and in-hospital mortality: multivariable logistic regression.

Lactate		Model1		Model2		Model3
		OR (95% CI)	P	OR (95% CI)	P	OR (95% CI)	P
MIMIC-IV	Class 1	1.00 (Ref.)		1.00 (Ref.)		1.00 (Ref.)
	Class 2	9.76 (7.23 ~ 13.16)	< 0.001	7.27 (5.16 ~ 10.24)	< 0.001	6.88 (4.86 ~ 9.74)	< 0.001
	Class 3	3.91 (3.20 ~ 4.78)	< 0.001	3.06 (2.34 ~ 4.00)	< 0.001	3.03 (2.32 ~ 3.98)	< 0.001
MIMIC-III	Class 1	1.00 (Ref.)		1.00 (Ref.)		1.00 (Ref.)
	Class 2	5.49 (3.16 ~ 9.56)	< 0.001	4.40 (2.32 ~ 8.36)	< 0.001	4.35 (2.27 ~ 8.33)	< 0.001
	Class 3	1.62 (1.15 ~ 2.28)	0.006	1.31 (0.84 ~ 2.06)	0.236	1.31 (0.83 ~ 2.06)	0.251

Adjusted odds ratios (ORs) and 95% confidence intervals (CIs) for in-hospital mortality according to lactate trajectory classes in MIMIC-IV and MIMIC-III cohorts, across three models of increasing covariate adjustment. Model 1 adjusted for baseline characteristics; Model 2 further adjusted for comorbidities and laboratory parameters; and Model 3 included adjustments for interventions in addition to the variables in Model 2.

28-day and 1-year mortality

In fully adjusted Cox models (Table 5), class 2 was associated with markedly increased risk of 28-day (HR: 3.88; 95% CI: 3.17–4.75) and 1-year mortality (HR: 3.16; 95% CI: 2.65–3.78) compared to class 1 (both P < 0.001). Class 3 also conferred elevated risk (28-day HR: 2.20; 1-year HR: 1.83; both P < 0.001), though to a lesser extent than class 2. In the MIMIC-III validation cohort, class 2 remained significantly associated with increased 28-day and 1-year mortality, whereas class 3 showed weaker and non-significant associations (Supplementary Table 3).

Table 5.

Association between lactate trajectories and 28-day and 1-year mortality: multivariable Cox regression in MIMIC-IV cohort.

Outcomes	Lactate	Model1		Model2		Model3
		HR (95%CI)	P	HR (95%CI)	P	HR (95%CI)	P
28 day death	Class 1	1.00 (Ref.)		1.00 (Ref.)		1.00 (Ref.)
	Class 2	5.40 (4.50 ~ 6.48)	< 0.001	3.97 (3.25 ~ 4.84)	< 0.001	3.88 (3.17 ~ 4.75)	< 0.001
	Class 3	2.77 (2.39 ~ 3.21)	< 0.001	2.26 (1.88 ~ 2.73)	< 0.001	2.20 (1.83 ~ 2.66)	< 0.001
1 year death	Class 1	1.00 (Ref.)		1.00 (Ref.)		1.00 (Ref.)
	Class 2	4.08 (3.45 ~ 4.81)	< 0.001	3.25 (2.72 ~ 3.88)	< 0.001	3.16 (2.65 ~ 3.78)	< 0.001
	Class 3	2.15 (1.89 ~ 2.44)	< 0.001	1.85 (1.58 ~ 2.18)	< 0.001	1.83 (1.56 ~ 2.15)	< 0.001

Adjusted hazard ratios (HRs) and 95% confidence intervals (CIs) for short- and long-term mortality by lactate trajectory classes in MIMIC-IV cohort, using three models with sequential covariate adjustment. Model 1 adjusted for baseline characteristics; Model 2 further adjusted for comorbidities and laboratory parameters; and Model 3 included adjustments for interventions in addition to the variables in Model 2.

Exploratory machine-learning classification

To facilitate rapid clinical identification of trajectory abnormalities, we trained a supervised classifier in MIMIC-IV using JMIM-selected features (score > 0.8) and TabPFN (Fig. 4A)²³. Discrimination in one-vs-rest cross-validation was high (AUCs 0.87/0.89/0.82 for Classes 1/2/3; Fig. 4B) and remained acceptable in MIMIC-III external validation (0.84/0.72/0.76; Fig. 4C). Model interpretation with SHAP confirmed clinically coherent drivers of prediction, with lactate, pH, and bicarbonate among the leading contributors (Fig. 4D). The classifier outputs per-patient class-membership probabilities to support early risk flagging; findings are exploratory and require prospective evaluation before clinical use.

Fig. 4.

Development and validation of a machine-learning classifier for early identification of lactate trajectory classes. (A) Feature ranking by JMIM in the training cohort (MIMIC-IV). The dashed line marks the prespecified selection threshold (score > 0.8) used for model construction. (B) One-vs-rest ROC curves and AUCs for predicting membership of Class 1 (blue), Class 2 (red), and Class 3 (green) in MIMIC-IV. (C) External validation ROC/AUCs in MIMIC-III using the same preprocessing and selected features. (D) SHAP summary plot (validation cohort) showing the direction and magnitude of feature contributions to the predicted probability of class membership; higher SHAP values indicate greater positive impact on the model output. Top contributors included lactate, pH, bicarbonate, SOFA, pCO₂. The classifier outputs per-patient class probability vectors to support rapid risk triage; this analysis is exploratory and does not alter the primary LCMM results. ACEI: angiotensin-converting enzyme inhibitor. WBC: white blood cell; RBC: red blood cell; RDW: red cell distribution width; BUN: blood urea nitrogen; HR: heart rate; RR: respiratory rate; SBP: systolic blood pressure; DBP: diastolic blood pressure; SpO₂: percutaneous arterial oxygen saturation; SOFA: Sequential Organ Failure Assessment score; RRT: renal replacement therapy; JMIM: Joint Mutual Information Maximization; SHAP: Shapley Additive ex Planations.

Subgroup analyses

Subgroup analyses in the MIMIC-IV cohort (Fig. 5) revealed consistent associations between trajectory class and mortality across strata of age, sex, SOFA score, and vasopressor use, without significant interaction. Findings were replicated in the MIMIC-III cohort (Supplementary Fig. 3).

Fig. 5.

Subgroup analysis and clinical characteristics comparison across early arterial lactate trajectory classes in the MIMIC-IV cohort. Subgroup analysis comparing key clinical features among the three latent trajectory classes of early arterial lactate levels identified via LCMM in critically ill patients with heart failure from the MIMIC-IV. AF: Atrial fibrillation; F: female; M:male; MI: myocardial infarction; RRT: renal replacement therapy.

Model performance and classification validity

Model fit indices are shown in Table 2, supporting the 3-class model as optimal based on BIC, AIC, SABIC, and entropy values. The mean posterior probability for class assignment exceeded 0.9 in all classes, indicating strong classification reliability (Supplementary Table 4). External validation model fit and classification accuracy in MIMIC-III (class enumeration with the same composite criteria likewise supported the 3-class model) are shown in Supplementary Tables 5–6.

Discussion

In this large retrospective cohort study leveraging two independent critical care databases (MIMIC-IV and MIMIC-III), we identified three distinct early arterial lactate trajectory classes in critically ill patients with HF using LCMM. These trajectories, characterized as low-stable, early rising with delayed decline, and early decline followed by re-elevation, were robustly and independently associated with in-hospital, 28-day, and 1-year mortality, even after adjusting for baseline severity, comorbidities, and interventions. To our knowledge, this is the first study to comprehensively characterize early lactate kinetics in HF patients during ICU admission and demonstrate the prognostic utility of dynamic, trajectory-based phenotyping. Our findings suggest that traditional single-timepoint lactate assessments may underestimate the complexity of metabolic stress responses in HF, whereas trajectory-based modeling offers a more nuanced and clinically meaningful stratification of patient risk.

Traditionally, lactate has been interpreted as a static biomarker reflecting tissue hypoperfusion, anaerobic metabolism, or shock severity^24–26. However, this snapshot view neglects the dynamic course of metabolic stress in critical illness. Emerging literature from sepsis⁸, acute kidney injury²⁷, and surgical cohorts²⁸ suggests that temporal patterns of biomarkers outperform isolated values in predicting outcomes. In HF specifically, studies have explored temporal changes in sodium¹³, potassium²⁹, and urine chloride¹², but no prior research has delineated lactate trajectory subphenotypes or examined their prognostic significance in a data-driven manner. A recent prospective study showed that dynamic lactate changes within the first 24 h are predictive of mortality in acute HF³⁰. Building on this, we applied unsupervised LCMM analysis to multiple early lactate readings and identified latent classes with distinct risk profiles, demonstrating that both the shape and timing of lactate changes carry prognostic information beyond initial values.

To further contextualize these findings, we examined the clinical and pathophysiological implications of each trajectory class. Patients in Class 1, with low-stable lactate levels, had the most favorable outcomes and likely represent individuals with preserved perfusion, intact metabolic flexibility, and mild illness. Class 3, characterized by initial elevation followed by decline, showed intermediate outcomes. Although lactate decreased over time, these patients still had significantly higher mortality than Class 1, suggesting residual metabolic stress or delayed recovery. Intriguingly, Class 2, defined by early rise and delayed normalization, had the worst prognosis across all time points. Despite having a similar initial lactate level to Class 3, their sustained elevation and delayed decline likely reflect ineffective resuscitation, progressive tissue hypoxia, or persistent mitochondrial dysfunction.

The divergence between Class 2 and Class 3, despite comparable baseline lactate levels, underscores the limitations of relying solely on admission values and highlights the utility of trajectory-informed risk stratification. These findings are consistent with prior studies in sepsis demonstrating that rising or non-clearance lactate trends are independently associated with increased mortality^31–33.

Lactate in HF is not merely a byproduct of anaerobic metabolism. It also serves as an alternative fuel for cardiomyocytes under stress, especially through MCT1-mediated transport into cells under ischemic conditions^14,34. However, MCT4 is upregulated under chronic catecholamine stimulation, which is commonly observed in decompensated HF, thereby facilitating lactate efflux from cells into the circulation^14,35,36. Such a metabolic shift results in intracellular lactate depletion and extracellular accumulation, which not only compromises its physiological buffering role but also contributes to oxidative stress and metabolic dysregulation^35,37.

Furthermore, recent studies suggest that lactate modulates gene expression through lysine lactylation, a post-translational modification implicated in cardiac remodeling, mitochondrial homeostasis, and sarcomere stabilization^16,38. Persistent hyperlactatemia may disrupt this regulatory axis, leading to impaired mitochondrial respiration and maladaptive hypertrophy. This may partly explain why Class 2 and 3 phenotypes—despite partial lactate decline—still exhibit increased mortality risk^14,39.

Clinical implications and integration into practice

Our results provide compelling support for using lactate trajectories—not static levels—as a dynamic biomarker for prognostic assessment in critical care. While absolute thresholds (e.g., > 2 mmol/L) are widely used, they lack sensitivity for early detection of worsening perfusion or subclinical metabolic failure.

For example, both Class 2 and Class 3 had elevated baseline lactate, but only Class 2 showed continued elevation and markedly higher mortality. This suggests that serial lactate trajectories may aid risk stratification beyond admission values. In practice, recognizing such a pattern should primarily trigger earlier hemodynamic reassessment and heightened surveillance.

Moreover, trajectory-based modeling offers potential utility in clinical trial design. For instance, patients with rising lactate trends may benefit from inclusion in studies targeting metabolic modulators or perfusion optimization strategies, whereas stable patients may be safely observed with conventional management. In this retrospective setting, trajectory classes are best interpreted as etiology-agnostic risk signals that should prompt structured clinical reassessment rather than prescriptive therapy.

Strengths and methodological rigor

This study has several notable strengths. First, it utilized two large, well-characterized critical care datasets (MIMIC-IV and MIMIC-III), enabling external validation and robust inference. Second, we employed LCMM, which accommodates irregular sampling and individual variability, providing a more realistic and clinically interpretable stratification than dichotomized or linear change metrics. Third, using arterial rather than venous lactate measurements enhances the precision of our assessment in the ICU setting.

Limitations

However, several limitations should be acknowledged. First, this was a retrospective cohort study using secondary data from electronic health records, and residual confounding or selection bias may persist despite multivariable adjustment⁴⁰. Second, since lactate measurements were obtained as part of routine care without a standardized protocol, the variability in timing and frequency may have influenced trajectory classification, although LCMM is generally robust to irregular sampling intervals. In addition, requiring ≥ 3 measurements within 0–72 h likely enriched for more closely monitored, higher-acuity patients, so generalizability to those with sparse sampling should be cautious. Third, our analysis focused exclusively on lactate dynamics within the first 72 h of ICU admission; changes occurring later in the clinical course may also carry prognostic value but were not evaluated in this study. We also did not measure mechanistic biomarkers (e.g., bioenergetic, microcirculatory, or lactylation markers) nor evaluate treatment effects; any discussion of alternative etiologies or clinical workflows is interpretive and hypothesis-generating and will require prospective validation. Finally, although we externally validated our findings using the MIMIC-III database, both cohorts were derived from a single healthcare system in the United States. While this minimizes heterogeneity due to varying data structures or institutional protocols, it may limit generalizability to other populations or international healthcare settings. Future prospective multicenter studies are needed to confirm these findings and evaluate the real-time applicability of trajectory-based risk stratification in diverse clinical environments.

Conclusions

In conclusion, this study identified three distinct early lactate trajectory classes among critically ill HF patients, each associated with differential short- and long-term mortality. Notably, patients with early rising and late-declining lactate patterns had the poorest prognosis, independent of baseline characteristics and interventions. These findings support a potential role for trajectory-based lactate assessment in risk stratification of critically ill HF. Prospective, multicenter validation is needed to confirm predictive performance and determine whether integrating lactate trajectories into workflows improves decision-making or patient outcomes.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

P-F W: Conceptualization, Methodology, Software, Data curation, Formal analysis, Writing—original draft & editing.; C-J G: Data curation, Formal analysis, Methodology, Software, Visualization, Writing—original draft.; Q C: Data curation, Software, Visualization.; H M: Software, Validation, Writing—original draft.; B X: Data curation, Methodology, Project administration, Supervision, Writing—Review & editing.; Y-L C: Conceptualization, Supervision, Funding acquisition, Project administration, Writing—Review & editing.

Data availability

The data underlying this study were sourced from the MIMIC database (https://mimic.physionet.org) under restricted access protocols. While public sharing is prohibited, qualified researchers may request access through institutional channels, and the corresponding author will facilitate data provision upon formal MIMIC approval.

Declarations

Competing interests

The authors declare no competing interests.

Ethics statement

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.

Declaration of generative AI and AI-assisted technologies in the writing process

During the preparation of this work, the authors used ChatGPT in order to assist with language refinement and editing. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article.