Lipidomic signatures align with inflammatory patterns and outcomes in critical illness

Lipid profiling of plasma from patients with severe trauma

To determine the dynamics changes in circulating lipids after severe injury in humans, we carried out a quantitative analysis of plasma lipid levels in samples obtained during the PAMPer trial22. This prospective, multi-institutional, pragmatic trial enrolled seriously injured humans suffering polytrauma at risk for hemorrhagic shock. Only patients that were transported by helicopter to a Level 1 trauma center were included and randomization took place in the pre-hospital setting. Patients in the treatment arm received two units of TP initiated during helicopter transport, while the control group was assigned randomly to standard-of-care, which did not include TP in the pre-hospital setting. The use of pre-hospital TP was associated with a 9.8% reduction in 30-day mortality (p = 0.03)22. A total of 193 of the original 523 patients were selected for lipidome analysis (Supplementary Fig. 1). This cohort included both non-survivors (n = 83) and survivors (n = 110) selected to represent the overall cohort. Samples were obtained at admission to the trauma center (0 h) and at 24 and 72 h after admission. Only the time 0 h sample was obtained in the early (died within the 72 h) non-survivors (n = 51). A group of 17 non-fasting healthy subjects was used as controls for baseline values. The detailed demographic information of healthy subjects and patients is shown in Table 1. Since underlying medical conditions and medication history can influence circulating lipid profiles, we also provide this information (Supplementary Data 4). Chronic health conditions and medications were rare in the trauma patient population and evenly distributed across the outcome groups (Supplementary Data 1).

Table 1 Demographic characteristics of the patients by outcome

The overall data analysis workflow is shown in Fig. 1a. Liquid chromatography mass spectrometry (LC-MS) was used to carry out targeted lipidomic analysis on the plasma samples. In total, 996 lipids were quantified using internal standards. In the quality control analysis, the median relative standard deviation (RSD) for the lipid panel was 4%. Lipids are named according to sub-class and acyl chains detected. For example, PE (16:0_18:2) has a phosphatidylethanolamine (PE) backbone and two acyl chains comprised of palmitic acid (C16:0) and linoleic acid (C18:2). The representation of lipids from 14 sub-classes is shown in Fig. 1b. Triglyceride (TAG) (glycerol backbone + three acyl chains) was the most abundant lipid class identified in the plasma (n = 518). Phosphatidylethanolamine (PE), phosphatidylcholine (PC), and diacylglycerols (DAG) all containing 2 acyl chains were the next most abundant classes (n = 128, 121, 58 respectively).

Fig. 1: Temporal patterns in the circulating lipidome after severe trauma.
figure 1

a Scheme of overall analysis strategy. b Representation of 996 lipid species detected in the lipidomic platform grouped by lipid classes. c Uniform Manifold Approximation and Projection (UMAP) plot shows the distribution of healthy subjects (n = 17) and patients with trauma (n = 193), grouped by sampling timepoints (0 h, 24 h, 72 h after admission). d Heatmap shows relative levels of 996 lipid species for healthy subjects and trauma patients, grouped by sampling timepoints using z-score normalized concentrations. Lipid species are clustered by Hierarchical clustering. e Quantitative comparison of circulating total lipid concentration among healthy controls (HC, n = 17) and trauma patients (n = 193), grouped by sampling timepoints. Asterisks indicate statistical significance based on the Kruskal–wallis test with post-hoc analysis using the Dunn test. The p value was adjusted by the Benjamini–Hochberg method: *<0.05; **<0.01; ***<0.001. Box and whisker plots represent mean value, standard deviation, maximum and minimum values, and outliers. TAG triacylglycerol, DAG diacylglycerols, MAG monoacylglycerols, PE phosphatidylethanolamine, PC phosphatidylcholine, PI phosphatidylinositol, LPE Lysophosphatidylethanolamine, LPC Lysophosphatidylcholine, CER Ceramides, HCER hexosylceramides, LCER lactosylceramide, DCER dihydroceramides, CE cholesterol ester. Source data are provided as a Source Data file.

We first explored the dynamic changes in the global pattern of the circulating lipidome in trauma patients. Uniform Manifold Approximation and Projection (UMAP) is a non-linear method for dimension reduction that can identify the global structure of multi-dimensional data. In Fig. 1c, each dot represents a single subject and the distance between dots in the UMAP plot reflects the global similarity/ differences in overall lipid profiles between samples23. We observed that trauma patients at 0 h were quite dispersed and partially overlapping with healthy subjects, suggesting an early and rapidly evolving response pattern immediately post-injury. There was excellent separation across the three time points on UMAP, underscoring the role of time in the major changes in lipid patterns after trauma.

To depict the differences between the healthy controls and patients across time, we projected relative levels of all lipids assayed on a heatmap (Fig. 1d). Compared to healthy controls, most lipid species were persistently lower after trauma. This dramatic shift between healthy controls and injured humans was also observed when total lipid concentrations were compared (Fig. 1e).

Association between lipidome pattern and outcome of trauma patients

We next investigated the association between the circulating lipidome and patient outcomes. The three outcomes used for this analysis included (1) early non-survivors (death within 3 days of admission), (2) non-resolving patients (survivors with duration of intensive care unit [ICU] stay ≥7 days or patients that died after day 3 following admission), and (3) resolving patients (survivors with duration of ICU stay <7 days). UMAP plots of the global lipidomic patterns indicated enrichment of early non-survivors in the region encircled in red at 0 h and an enrichment of the non-resolving patients in the region encircled by the blue line at 72 h (Fig. 2a, b). Furthermore, we observed a dramatic drop in the levels of nearly all major lipid species at 0 h for early non-survivors compared to the other patient groups or healthy controls (Fig. 2c). Patients in both the resolving and non-resolving groups at 0 h also exhibited a drop in most lipid species compared to healthy controls, but not to the degree seen in the non-survivors. Patients in the resolving group exhibited a persistent suppression in most lipids at 24 and 72 h (Fig. 2d, e). Remarkably, patients in the non-resolving group at 72 h demonstrated an increase in a subset of lipids. Further characterization of lipid class and fatty acid types indicated that all 14 classes, including both saturated and unsaturated fatty acids, were suppressed in injured patients at 0 h. However, there was selective elevation of TAG, DAG, PE, and ceramides (CER) at 72 h in the non-resolving cohort. A quantitative time-series analysis showed that total lipid levels were higher at 72 h in the non-resolving patients and that unsaturated fatty acids predominated in TAG and DAG, while PE and CER contained a mixture of saturated and unsaturated fatty acids (Fig. 2f). Interestingly, there was excellent correlation across the elevated lipids from the TAG, DAG, and PE classes (Supplementary Fig. 3a, b). The interconnections between biochemical pathways involved in the synthesis of these lipid classes are shown in Supplementary Fig. 3c. Our findings point to a rapidly evolving pattern in the circulating lipidome after severe injury that includes a loss of all classes of lipids in the circulation that is evident early after injury. This process is exaggerated in patients that die early, suggesting that an abrupt loss of circulating lipids contributes to adverse outcomes. Furthermore, there is a selective increase in four lipid classes by 72 h in patients that remain critically ill or die later in their clinical course.

Fig. 2: Association between temporal patterns of the circulating lipidome and outcome.
figure 2

Uniform Manifold Approximation and Projection (UMAP) plot shows the distribution of healthy control subjects (n = 17) and trauma patients (n = 193), grouped together (a) and separated (b) by outcome and sampling timepoints. Heatmaps show relative levels of 996 lipid species (c); 14 lipid classes (d) and 28 fatty acids labeled by carbon number: double bonds (e) for healthy subjects and trauma patients, grouped by outcome and sampling timepoints. z-score represents normalized concentrations. Rows are clustered by method of hierarchical clustering. f Quantitative comparison of circulating total lipid concentrations among healthy controls (HC) and trauma patients. Lipids are grouped by classes and fatty acids (saturated or unsaturated) identified as the acyl chains in the lipid classes. Patients are grouped by outcome and sampling timepoints. Center dots and error bars represent median value and median absolute deviation, respectively. SFA saturated fatty acid, USFA unsaturated fatty acid. Asterisks indicate statistical significance based on Kruskal–wallis test among 3 groups at 0 h with post-hoc analysis of Dunn test. The P value was adjusted by Benjamini–Hochberg method: *<0.05; **<0.01. Number sign indicates statistical significance based on 2-way AVOVA test of time-series analysis of resolving and non-resolving groups. Pairwise Comparisons were conducted by Estimated Marginal Means test. The P value was adjusted by Benjamini–Hochberg method: #<0.05; ##<0.01; ###<0.001, ####<0.0001. Source data are provided as a Source Data file.

We next examined the impact of injury severity reflected by injury severity scores (ISS) on lipid levels and profiles. Patients were separated into minimal (ISS < 10), moderate (ISS 10–25), or severe (ISS ≥ 25) injury (Supplementary Fig. 2a). Exploration of the lipid profiles by either UMAP or heatmap demonstrated no major impact of ISS on the post-injury lipid patterns (Supplementary Fig. 2b). We also observed poor correlation between ISS and total lipids concentrations of either saturated or unsaturated fatty acids (Supplementary Fig. 2c, d, 0 h timepoint shown). Thus, while injury induces major changes in the circulating lipidome, in this cohort of patients with shock on presentation, ISS alone does not associate with lipid patterns.

Pre-hospital TP enhances lipid levels early after severe injury

The key observation of the PAMPer trial was the demonstration that initiating TP administration in the pre-hospital setting reduced early mortality when compared to standard care22. To assess for an impact of TP, we compared lipid profiles in patients in the treatment arm to those in the standard-of-care arm. UMAP plots demonstrated a skewing in the lipid profiles towards the healthy controls in the TP treatment group at 0 h (Fig. 3a, b). However, this preservation of lipid levels associated with pre-hospital TP was seen to dissipate at 24 and 72 h, with no difference in lipid levels or patterns between the TP and standard-of-care groups at these time points. Both the qualitative and quantitative analysis revealed that patients receiving TP had less of a drop in the levels of most classes of circulating lipids at time 0 h, with higher levels of TAG, DAG, and MAG compared to standard-of-care patients (Fig. 3c, Supplementary Fig. 4A). We then assessed the relationship between the predicted mortality, calculated from the Trauma and Injury Severity Score (TRISS), and lipid levels in the two cohorts (Fig. 3d). Average lipid levels were higher in the TP group across all TRISS values. All unexpected deaths (low TRISS Score: predicted mortality rate less than 50%) were in the standard-of-care patients and 11/14 had lipid levels below the mean for the overall cohort. Deaths seen in the TP group were limited to those with a high expectation for death for all except one patient (high TRISS Score: predicted mortality rate of greater than 75%). A Forest plot of log-odds ratios from a multi-variable logistic regression (generalized estimating equation) is shown in Fig. 3e. This analysis revealed that lower lipid levels at 0 h significantly favored mortality within the first 72 h while TP administration favored survival (OR:2.50, Cl: 1.24–5.01). Only TRISS had a higher association with early mortality than TP or lipid levels even when traumatic brain injury (TBI) and sex were added to the model.

Fig. 3: Potential causal effect for thawed plasma (TP), Lipid concentration and early mortality.
figure 3

Uniform Manifold Approximation and Projection (UMAP) plot shows the distribution of healthy subjects (HC, n = 17) and trauma patients (n = 193) (a), separated by treatment arms with sampling timepoints (b). c Heatmap shows relative levels of 996 lipid species for healthy subjects and trauma patients, grouped by treatment arms and sampling timepoints. Exp, z-score normalized concentration. Rows are clustered by hierarchical clustering. d Relationship of predicted mortality and total lipid concentration at 0 h upon admission. Trauma patients are grouped by treatment arms; tendency lines are modeled by loess methods for 2 groups separately, dash line in the x-axis means 0.5 and y-axis means the median concentration. d indicates patients who died less than 72 h after admission. e Forest plot showing odds ratios from logistic regression (generalized estimating equatio) of clinical factors; Lipid concentration; TP effect for early-nonsurvivors (n = 51) versus others (n = 142). Error bars: 95% confidence interval. f Correlation heatmap showing correlation among cytokines, biomarkers, clinical variables, total lipid concentration and outcome. r: Spearman correlation coefficient. g Causal network among factors in e constructed by FCI (see also methods) in patients with complete lipid and biomarker data (n = 170). The presence of “edges” or connections between nodes in the graph correspond to conditional dependencies relationships. Detailed interpretation of the edges can be found in Methods. Abbreviations: TRISS Trauma and injury severity score, TP thawed plasma, TBI traumatic brain injury, ISS injury severity score, GCS Glasgow coma score, PH Prehospital, INR international normalized ratio. Asterisks in e indicate statistical significance in multi-variable logistic regression model: *<0.05; **<0.01. Asterisks in f indicate statistical significance for correlation coefficient. P-values are approximated by using the t distributions: *<0.05; **<0.01; ***<0.001.

We next carried out correlation analysis to identify the factors that associate with circulating lipid levels in the early response to severe injury. Included in the analysis were 21 inflammatory and immune mediators, 6 markers of endotheliopathy/ tissue injury, and 2 measures of coagulation abnormalities, all measured at time 0 h. Also included in the analysis were typical measures of injury severity and interventions associated with adverse outcomes. Interestingly, the mediators segregated into three subsets, each with strong internal correlation (Fig. 3f). These included a subset represented by pro-inflammatory cytokines and chemokines that mostly positively correlated with early death, injury severity, endotheliopathy, and abnormal coagulation (Subset 1: IL-6, IL-8, IL-10, MCP-1/CCL2, IP-10/CXCL10, and MIG/CXCL9) and two subsets that correlated inversely with the pro-inflammatory mediators and adverse outcomes including, mediators associated with type 2 and 3 immune responses (Subset 2: IL-2, IL-4, IL-5, IL-7, IL-17A, and GM-CSF) and mediators associated with either tissue protection/ repair or lymphocyte regulation (Subset 3: IL-9, IL-22, IL-25, IL-27, IL-33 and IL-21, IL-23). The relationships between these three mediator subsets remained mostly consistent at 24 and 72 h (Supplementary Fig. 7a, b). However, low lipid levels at time 0 h positively correlated only with standard-of-care, early death, coagulation abnormalities and the endotheliopathy marker, sVEGFR, and not with any of the mediator subsets (Fig. 3f).

We next used probabilistic graphical models for mixed data types24,25 to infer potential direct (cause-effect) relationships within the multi-modal observational data included in Fig. 3f. These features were loaded into the algorithm and nodes and edges projected onto a graph with early mortality as the endpoint of interest (Fig. 3g). The α-value of 0.2 for the conditional independence tests of the algorithm was selected using nested leave-one-out cross-validation to select the model with the best predictive performance of patient outcome (see Methods). Circulating lipid concentrations, coagulopathy (including INR), volume of crystalloid used in first 24 h and the pro-inflammatory mediators (via MIG) were identified as direct causal factors contributing to early death (demonstrated by red arrows). The sequential edges connected TP administration to circulating lipid concentrations, coagulopathy, INR, and volume of crystalloid used in first 24 h. These connections indicated a potential mixed causal relationship linking TP with all these factors and fewer early deaths. Other features known to be important to early mortality, including patient and injury characteristics, endothelial and tissue injury, and subset 2 and 3 mediators were indirectly linked to outcomes. Thus, correlation analysis and causal modeling related an interaction between INR and lipid concentration to early death and identified a direct impact of TP on both of these causative factors. These findings further support the notion that a rapid loss of circulating lipids contributes to the early pathogenic state cause by severe injury with shock.

Confirmation of outcome-based changes in the plasma lipidome in trauma and patients with critical illness due to COVID-19

To determine if our findings could be recapitulated in an independent trauma patient cohort, we conducted an in-depth comparison between the PAMPer dataset and a separate trauma dataset26 (Trauma dataset-2:TD-2, n = 86). Because there were differences in the methods used to quantify lipid species across the datasets, we only carried out indirect comparisons of the relative changes (Z-scores) of lipids species within each dataset across the datasets. A total 75 lipids from 9 sub-classes were found to be in common between the PAMPer and TD-2 datasets (Supplementary Fig. 5a, b). There was remarkable consistency in the relative changes of the early drop and late increase in most lipids over time and based on outcome group. The elevated lipids in the non-resolving patients at 72 h were almost entirely in the PE, MAG and DAG classes in both the PAMPer (23/26) and TD-2 (18/19) datasets. TAG, LPE, LPC, and DCER were not measured in TD-2 and therefore, are not included in this comparison.

To further generalize our findings of outcome-associated changes in circulating lipids to another cause of acute critical illness, we analyzed two public datasets derived from COVID-19 patients16,17. Unlike trauma, the onset of critical illness in Covid-19 patients can be highly variable relative to the onset of infection and the time the infection started is often unclear. To assist with the comparison between the trauma and COVID-19 datasets, we set the 0 timepoint in the COVID-19 datasets as the day of symptom onset for non-severe patients and day of progression for severe patients. A total of 29 lipids were identified in common among the 4 datasets (Fig. 4a–d, Supplementary Data 2). Of these, only a subset of PE species were found to be significantly elevated from baseline during critical illness (Supplementary Data 5). We identified eight PE species and one PC specie significantly higher in the non-resolving group (72 h) in PAMPer dataset, while three, six, and five of these PE were elevated in the TD-2, Covid-19 (Guo et al.)16, and Covid-19 (Shui et al.)17 datasets, respectively. Eight of these PE species could be identified when combining PAMPer with any single other database, five PE species were in common when combining PAMPer with any two of the other databases, and a single PE was found to be significantly elevated during critical illness in common across all four databases (Supplementary Data 5). Thus, increases in PE consistently associate with critical illness in trauma and COVID-19.

Fig. 4: Comparison of temporal patterns of common lipids for patients with trauma or COVID-19.
figure 4

a, d Heatmaps show the relative levels of 29 common lipid species from four major classes across patients. Data comes from trauma patients from the PAMPer lipidomics dataset (a) and TD-2 untargeted metabolomics dataset (b); COVID-19 patients from untargeted metabolomics dataset (Guo et al. Cell, 2020) (c) and lipidomics dataset (Shui et al., Cell metabolism, 2020) (d). Patients are grouped by outcome and sampling timepoint (except for d). Asterisks indicate lipids with statistical significance (p value <0.05) and log2 fold change >0.4 by two-sided Wilcoxon Rank Sum test between non-resolving and resolving trauma patients at 72 h (a); non-resolving and resolving trauma patients at D2-D5 (b); severe and non-severe Covid-19 patients (c); severe and mild Covid-19 patients (d). Abbreviations: PE phosphatidylethanolamines, PC phosphatidylcholines, PI phosphatidylinositols, SM sphingomyelins.

Generation and evaluation of a lipid reprogramming score

We next sought to determine if a combination of PE species common to the four trauma and COVID-19 patient datasets could be optimized to generate a Lipid Reprogramming Score (LRS) (Fig. 5a, see also methods for a detailed description). Briefly, the eight PE species detected across four datasets were selected as the starting pool. All the eight PE species were highly correlated with the 37 other lipids (mostly TAG species, Supplementary Data 3) identified as significantly higher in non-resolving PAMPer patients (72 h) by logistic regression taking into account cofounders, including ISS, age, and treatment (Supplementary Fig. 6a, b). A sensitivity analysis identified a model comprised of five PE showing the best performance (Supplementary Data 6). Thus, we defined the LRS as the mean z-score of five PE species (PE (16:0_18:2), PE (16:0_20:4), PE (16:0_22:6), PE (18:0_18:1), PE (18:0_22:6)) representative of PE from all four trauma and COVID-19 datasets. To further technically validate the results, we utilized another platform (LC-HRMS, Platform2) from matched trauma patients (n = 29) and healthy controls (n = 8) to quantify the concentrations of 5 PE species (Supplementary Data 8). All of them were highly correlated between the two platforms (LC-MS/MS, Platform 1, PF1; LC-HRMS, Platform 2, PF2) (Supplementary Fig. 9a) and selectively up-regulated in non-resolving trauma patients at 72 h (Supplementary Fig. 9b).

Fig. 5: Lipid Reprogramming Score (LRS) is an independent risk factor for outcome after trauma or COVID-19.
figure 5

a Graphical scheme of generation and evaluation of LRS. b Comparison of LRS from patients with trauma (n = 142). Patients are grouped by outcome and sampling timepoint. Center dots and error bars represent median value and median absolute deviation, respectively. c Forest plot showing hazard ratio of clinical factors and LRS score for recovery using a Cox regression mixed effect model in patients surviving at 72 h (n = 142). Error bars: 95% confidence interval. d ROC curve for three prognostic models in training cohort from Standard-of-care arm in the PAMPer dataset (trauma patients, n = 73). e Comparison of LRS for patients with COVID-19. Healthy Subjects (n = 25), Non-COVID (n = 25) and COVID-19 patients (n = 45) are grouped with diseases outcome and sampling timepoint. Center dots and error bars represent median value and median absolute deviation, respectively. f Forest plot showing odds ratio of clinical factors from logistic regression and LRS score for Non-severe (n = 25) versus Severe COVID-19 patients (n = 20). Error bars: 95% confidence interval. g Comparison of prognostic value of LRS, PE (16:0_22:6), lymphocyte count, and CRP for Non-severe (n = 25) versus Severe (n = 20) outcome for the COVID-19 cohort (C1) from Guo. et al by ROC curve. ISS injury severity score, Lym lymphocyte count, CRP C-reaction protein. Asterisks in b indicate statistical significance in based on 2-way AVOVA test of time-series analysis of resolving and non-resolving groups. Pairwise Comparisons was conducted by Estimated Marginal Means test. The P value was adjusted by Benjamini–Hochberg method: ****<0.0001. Asterisks in e indicate statistical significance based on Kruskal–wallis test among 6 groups of COVID-19 patients with post-hoc analysis of Dunn test. The P value was adjusted by Benjamini–Hochberg method: *<0.05. Asterisks in d and g indicate statistical significance in multi-variable regression model: *<0.05; **<0.01. Source data are provided as a Source Data file.

We next calculated the LRS for each patient across the three timepoints and plotted these in a UMAP plot (Supplementary Fig. 6c) in order to further reveal their relationships with global lipidome patterns. We found that the gradient in the LRS increased from left-to-right along the x-axis in the UMAP plot, which was consistent with the outcome-based pattern at 72 h. We then transformed the score into a categorical variable with three thresholds based on tertiles (Low, Medium, High) for all PAMPer patients surviving at 72 h (Supplementary Fig. 6c). When displayed on a UMAP plot, the separation of patients into low, medium, and high LRS tertiles distributed the patients similarly to that seen using the continuous LRS. Thus, both the continuous and categorical LRS values represent the magnitude of global changes in the circulating lipidome and may be useful for correlating the lipidomic changes with other patient features. We also explored the relationship between the LRS and either the BMI or early lipid levels in 89 PAMPer patients (Supplementary Fig. 6d, e). There was weak relationship between BMI and total lipid levels (r = 0.18, p = 0.094). The LRS was independent of BMI (r = −0.03, p = 0.73). Thus, the LRS could be a representative marker of the changes in circulating lipids in critically trauma patients.

Risk assessment using LRS for patients with trauma or COVID-19

We next investigated whether the LRS was associated with outcomes in trauma or COVID-19 patients. A time-series analysis demonstrated that non-resolving trauma patients exhibited dramatic increases in the LRS at 24 to 72 h post-trauma compared to resolving patients (Fig. 5b). Recovery analysis revealed that LRS-high and LRS-medium groups experienced a longer period to recovery than patients in the LRS-low group (Supplementary Fig. 6f). In addition, trauma patients with medium or high LRS were associated with higher injury severity, lower admission blood pressure, mass transfusion, higher INR, and higher incidence of NI and MOF (Supplementary Data 4). High LRS was also associated with lower probability of recovery (HR:0.73, Cl:0.56–0.95) even when adjusted for age, ISS, TBI, and treatment effect in a Cox regression mixed effect model (Fig. 5c). To confirm our finding using a second trauma population, we adopted the same strategy to construct the LRS using the TD-2 dataset, which was dominated by resolving trauma patients. The recovery curve, and Cox regression model all showed similar correlations of LRS with outcomes in TD-2 as seen in PAMPer trial patients (Supplementary Fig. 6h, i). Therefore, the LRS showed an independent relationship with persistent critical illness after trauma.

We next explored the prognostic value of the LRS and the five individual PE species that comprise the LRS for predicting whether trauma patients would progress to a non-resolving pattern (Supplementary Data 7). Here, we set the standard-of-care arm in PAMPer dataset as the training set (n = 73). The TP arm from PAMPer dataset was set as an internal test set (n = 69) and the TD-2 dataset was set as an external test set (n = 86). Compared to the reference model27 (ISS + IL6, AUC = 0.798), adding the LRS moderately improved the performance of discrimination (AUC = 0.816, added AUC = 0.018) in the training set (Fig. 5d, Supplementary Fig. 7a). Interestingly, of the five PE that comprise the LRS, PE (18:0_18:1) (RSD:9.82%) also greatly improved the performance of discrimination (AUC = 0.873, added AUC = 0.075) in the training set (Fig. 5d, Supplementary Fig. 7a). We further utilized an established two-step machine learning approach28,29,30 to identify a minimal set of predictive lipid biomarkers and clinical features for predicting the outcome in the PAMPer dataset (standard-of-care arm). This approach was based on feature selection (L1 Regularization – LASSO to avoid overfitting) followed by classification (Support Vector Machine) using the down-selected features. The results suggested that PE (18:0_18:1) was the top selected feature among lipids and only IL6 and ISS ranked higher overall (Supplementary Fig. 7b). The performance of calibration (Supplementary Fig. 7c) was also improved by adding either the LRS or the single PE (18:0_18:1) to ISS + IL6 (Brier Score: ISS + IL6, 0.177; ISS + IL6 + LRS, 0.166; ISS + IL6 + PE (18:0_18:1), 0.139). The results were consistent in the internal (Supplementary Fig. 7D, AUC: ISS + IL6, 0.876; ISS + IL6 + LRS, 0.916; ISS + IL6 + PE (18:0_18:1), 0.900) and external test sets (Supplementary Fig. 7E, AUC: ISS + IL6, 0.797; ISS + IL6 + LRS, 0.814; ISS + IL6 + PE (18:0_18:1), 0.841).

We then tested whether we could generalize the LRS for the two COVID-19 patient datasets using a similar approach. The Shui, et al.17 COVID-19 dataset lacked detailed clinical data, therefore, we only compared differences in LRS among the four outcome groups defined by the authors of the study. We found that mild, moderate and severe COVID-19 patients had a higher LRS compared to healthy subjects (Supplementary Fig. 6g). Consistent with these findings, the LRS was also significantly higher in the severe group when compared to the non-severe COVID-19 patients in the dataset of Guo, et al.16 (Fig. 5e). We also observed an upward trend in the LRS during the time window preceding progression (<48 h and D6-D14 after progression, Fig. 5e). Finally, multi-variable logistic regression suggested that LRS is an independent risk factor for COVID-19 patients (OR: 9.88, Cl: 2.09–78.5, Fig. 5f). C-reactive protein (CRP) and lymphocyte count (Lym) are known to correlate with worse outcomes in COVID-19 patients31. We compared the LRS and its five individual PE species with these two variables to classify severe versus non-severe patients in both a training set (C1, n = 45) and test set (C2, n = 10) (Supplementary Data 7). We found that the LRS alone moderately improved the performance of discrimination (AUC = 0.814, added AUC = 0.028), however a single PE specie from the LRS (PE (16:0_22:6), RSD:5.51%) alone greatly improved the performance (AUC = 0.862, added AUC = 0.076) (Fig. 5g, Supplementary Fig. 7a). The two-step machine learning approach also revealed that the PE (16:0_22:6) were the top selected features (Supplementary Fig. 7C). The performance of calibration was also improved by using PE (16_22:6) (Brier Score: Lym+CRP:0.177; LRS: 0.166; PE (16:0_22:6): 0.139; Supplementary Fig. 7G). The performance of PE (16:0_22:6) also had the highest AUC compared to other two models in the test set (Supplementary Fig. 7f, AUC: CRP + Lym, 0.917; LRS, 0.833, PE (16:0_22:6), 0.958). Interestingly, we also noticed that PE (16:0_22:6) showed similar performance for prognostication as the random forest model based on 17 proteins and 9 metabolites reported in the original manuscript. It is notable that only one patient (XG43) was mislabeled in the test cohort (C1) using PE (16:0_22:6) for prognostication (Supplementary Fig. 7g). Thus, similar to our observations in trauma, a single PE specie derived from the LRS performed well of prognostication for severe disease in COVID-19.

Association between LRS and systemic markers of inflammation and endothelial dysfunction in trauma patients

We next sought to determine if the LRS correlated with circulating markers of inflammation or endothelial and tissue damage. A correlation matrix was constructed using data from the 121 PAMPer patients alive at 72 h that had complete data for lipids, 21 cytokines and chemokines, endotheliopathy markers, and tissue injury markers across time after injury (Time 0 h: Fig. 6a, Times 24 and 72 h: Supplementary Fig. 8a, b). Across the three time points, LRS correlated positively with pro-inflammatory cytokines/chemokines (defined above as subset 1), as well as endotheliopathy and tissue injury biomarkers. Conversely, LRS correlated negatively with subset 2 (lymphocyte-related) and subset 3 (protective/ reparative) cytokines and an adipokine (Adiponectin). These findings suggest that the changes in the circulating lipidome at 72 h, represented by an elevated LRS, associates with biological process that drive worse outcomes (e.g., inflammation, endotheliopathy, and tissue injury), and therefore, may contribute to or be part of the pro-inflammatory/ tissue injury processes that are known to contribute to adverse outcomes in trauma.

Fig. 6: Association between LRS and circulating biomarkers.
figure 6

a Heatmap showing correlation of LRS and circulating biomarkers at 0 h in trauma patients (n = 121), measured by Spearman correlation coefficients. Asterisks in a and b indicate statistical significance for correlation coefficient. Unadjusted p-values are approximated by using the two-sided t distributions: *<0.05; **<0.01; ***<0.001. b Schematic of proposed paradigm showing the relationship between circulating lipid levels and outcomes after severe injury. Early loss of circulating lipids correlates with adverse outcomes while failure to resolve critical illness is associated with the selective increase in glycerolipids and PE.