Multi-Biomarker Prediction Models for Multiple Infection Episodes Following Blunt Trauma.

Severe trauma predisposes patients to multiple independent infection episodes (MIIEs), leading to augmented morbidity and mortality. We developed a method to identify increased MIIE risk before clinical signs appear, which is fundamentally different from existing approaches entailing infections' detection after their establishment. Applying machine learning algorithms to genome-wide transcriptome data from 128 adult blunt trauma patients' (42 MIIE cases and 85 non-cases) leukocytes collected ≤48 hr of injury and ≥3 days before any infection, we constructed a 15-transcript and a 26-transcript multi-biomarker panel model with the least absolute shrinkage and selection operator (LASSO) and Elastic Net, respectively, which accurately predicted MIIE (Area Under Receiver Operating Characteristics Curve [AUROC] [95% confidence intervals, CI]: 0.90 [0.84-0.96] and 0.92 [0.86-0.96]) and significantly outperformed clinical models. Gene Ontology and network analyses found various pathways to be relevant. External validation found our model to be generalizable. Our unique precision medicine approach can be applied to a wide range of patient populations and outcomes.

Introduction

The high value of tailored approaches in the care of patients is increasingly appreciated (Chaussabel, 2015; Parikh et al., 2016; Sweeney and Wong, 2016). A method to expedite the timeline of threat detection to before infection happens could yield valuable time for early prophylactic and therapeutic interventions. Moreover, the ability to identify patients at high risk for repeated infections, or infection-related morbidity and mortality, might be considered an important measure to fairly allocate resources such as medication, personal protective equipment, or another high-value scarce intervention (The Commonwealth of Massachusetts, 2020; Institute of Medicine, 2013; University of Pittsburgh, 2020).

Trauma is one of the leading causes of morbidity and mortality worldwide (Heron, 2018; Krug et al., 2000). Severe trauma induces various immune-related responses acutely—it can trigger a state of immunosuppression (Islam et al., 2016; Ward, 2005), prolonged inappropriate immune response (Heffernan et al., 2012; Huber-Lang et al., 2018), leukocytosis (Paladino et al., 2010), and the elevation of specific subpopulations of myeloid cells (Cuenca et al., 2011). Among trauma patients, infections and infections-related complications contribute to substantial mortality and morbidity and prolonged hospital stay, significantly adding to healthcare costs (Cole et al., 2014; Dutton et al., 2010; Glance et al., 2011; Hashmi et al., 2014). Infections and infections-related outcomes vary across individuals, suggesting the importance of considering individual patients' underlying susceptibility and the degree of immunosuppression, or inappropriate immune response experienced.

Methods to identify trauma patients with particularly increased risk of infection could be advantageous for ensuring timely and appropriate delivery of preventative measures (such as early immune-modulating nutrition, microbiome modulation, early mobilization, early removal of lines/tubes, taking all transmission-based precautions), improving surveillance, and promoting antibiotic stewardship to limit the emergence of multi-drug resistance bacteria, reduce toxicity to patients, and decrease healthcare costs. Previous studies have evaluated the use of injury severity scores, such as the Acute Physiology and Chronic Health Evaluation (APACHE) II (Knaus et al., 1985), Injury Severity Score (ISS) (Baker et al., 1974), and New Injury Severity Score (NISS) (Osler et al., 2010) as predictors of infection, in addition to their intended use to predict mortality (Cheadle et al., 1989; Jamulitrat et al., 2002). Using genome-wide transcriptomic information from leukocytes provided at triage to assess the underlying susceptibility, well before the onset of infections, is expected to significantly improve the accuracy of identifying patients who are most at-risk of multiple independent infection episodes. A recent study described that employing a combination of predictors could be more effective than using a single predictor with strong statistical significance, further suggesting that multi-biomarker panels could be highly effective (Lo et al., 2015). Previous studies have utilized transcriptome data in the trauma setting to find transcripts that correlate with poor outcome (Desai et al., 2011) or sepsis (Sweeney et al., 2015). The objective of this study is distinct, as we aim to develop a method to predict multiple infections prior to classic clinical signs of infection. And thus, our approach focuses on the prevention of infections, aiming to predict the outcome before its onset, using early blood samples.

In a previous study among burn trauma patients, we developed a blood transcriptomic multi-biomarker panel for predicting multiple independent infection episode (MIIE) outcomes during the course of recovery (Yan et al., 2015). We employed the least absolute shrinkage and selection operator (LASSO) machine learning algorithm to select probe sets that together (i.e., multi-transcriptome panel) resulted in highly accurate prediction. This model performed significantly better than those based on injury severity assessments at triage and demographic information (Yan et al., 2015). Here, we employed a new approach of combining the use of two algorithms, LASSO and Elastic Net regression, to investigate MIIE outcome. LASSO and Elastic Net were used to reduce the complexity of regression models, in conjunction with cross-validation to select the optimal parameter for reducing the number of predictors. These techniques are highly beneficial in cases such as transcriptome data where the number of potential predictors is extremely large, to overcome the problem of overfitting. LASSO regression reduces highly correlated predictors and selects a minimal panel of predictors, compared to Elastic Net that includes some correlated predictors. We investigated blunt trauma patients in the multi-center Inflammation and Host Response to Injury (“Glue Grant”) cohort. This cohort enrolled a high number of patients, generated genome-wide transcriptome data, and collected data longitudinally, allowing us to effectively build a predictive model. Our approach employing unbiased analyses of genome-wide information in the identification of patients at increased risk for MIIE before clinical signs of infection may also be advantageous for providing new insights into the molecular pathways that characterize the pathophysiology underlying hypersusceptibility to infections.

Results

Patient Demographics and Baseline Characteristics

Baseline characteristics of the 128 blunt trauma patients included in the study (Figures 1A and 1B) are presented in Table 1. Motor vehicle collisions were the most frequent injury mechanisms (Table S1). The overall study population consisted of approximately 61.7% males and 38.3% females, with the median age [interquartile range, IQR] of 34 [25-44] years (Table 1). Common demographic factors were not significantly different between non-cases (≤1 infection episode) and MIIE cases (≥2 infection episodes). However, baseline injury severity scores were significantly higher for MIIE cases (APACHE II score: 27 [24-31] vs. 29 [26–33.5], p = 0.008; ISS: 29 [18-41] vs. 36 [27–41.5], p = 0.041; NISS: 34 [27-43] vs. 41 [30-50], p = 0.045, respectively), while automatic identification system (AIS) (highest score in any body region) was comparable.

Figure 1

Description of the Patient Population and Study Design

(A and B) (A) Patients who were included/excluded in the study and (B) the study design.

Table 1

Baseline Characteristics, for the Outcome of Multiple Independent Infection Episodes (MIIEs)

Variable Description	All Patients (N = 128)	Non-case, ≤ 1 Infection Episode (n = 85)	MIIE Case, ≥2 Infection Episodes (n = 43)	p value
Demographic information/clinical descriptors
Age	34 [25–44]	35 [26–44]	32 [23.5–46]	0.718a
Sex
Female	49 (38.3%)	30 (35.3%)	19 (44.1%)	0.328c
Male	79 (61.7%)	58 (64.7%)	23 (55.8%)
BMI	29.3 ± 7.0	28.8 ± 7.0	30.1 ± 6.9	0.325b
BMI categories
Underweight	1 (0.8%)	0 (0%)	1 (2.3%)	0.336d
Healthy	37 (28.9%)	28 (32.9%)	9 (20.9%)	0.216d
Overweight	39 (30.5%)	26 (30.6%)	13 (27.9%)	1.000d
Obese	51 (39.8%)	31 (36.5%)	20 (46.5%)	0.340d
Injury characteristics
Crush injury	19 (14.8%)	16 (18.8%)	3 (7.0%)	0.113d
Severe head injury	13 (10.2%)	8 (9.4%)	5 (11.6%)	0.760d
Severity scores
APACHE II	28 [24–32]	27 [24–31]	29 [27–33.5]	0.008a
ISS	34 [22–41.25]	29 [18–41]	36 [27–41.5]	0.041a
NISS	37 [27–43.5]	34 [27–43]	41 [30–50]	0.045a
AIS (highest of any body part)	4 [3–5]	4 [3–5]	4 [3.5–5]	0.587a
Clinical events
Post-trauma interventions
Craniotomy	2 (1.6%)	1 (1.2%)	1 (2.3%)	0.402d
Thoracotomy	8 (6.3%)	5 (5.9%)	3 (7.0%)	1.000d
Laparotomy	61 (47.7%)	35 (41.2%)	26 (60.5%)	0.042d
Orthopedic	98 (76.6%)	61 (71.8%)	37 (86.0%)	0.081d
Vascular	30 (23.4%)	19 (22.4%)	11 (25.6%)	0.667d
Clinical outcomes
Mortality	5 (3.9%)	3 (3.5%)	2 (4.7%)	1.000d
Days in ICU	11 [6–18] ∗	8 [5–14.75] ∗	20 [15–29] ∗	<0.0001a
	11.5 [6–18.25]	8 [5–15]	20 [14.5–28]	<0.0001a
Discharge day since injury	23 [16.5–33] ∗	19 [14–26.75] ∗	35 [27–47] ∗	<0.0001a
	23 [16–33]	19 [13–26]	34 [25.5–47]	<0.0001a
Days on ventilator in ICU	8 [4–15]	6 [3–11]	15 [10–23]	<0.0001a
Non-infection complications	75 (58.6%)	40 (47.1%)	35 (81.4%)	0.0002c
Maximum denver 2 score	2 [1–3]	2 [0–3]	3 [2–3.5]	<0.0001a
Maximum marshall score	5.8 [4.0–7.2]	4.7 [3.3–6.4]	6.9 [5.8–8.1]	<0.0001a
Maximum central nervous system score	4 [4–4]	4 [4–4]	4 [4–4]	0.064a
Maximum cardio score	2.8 [2.0–3.6]	2.4 [1.7–3.2]	3.3 [2.8–4.0]	<0.0001a
Maximum respiratory score	2.8 [2.0–3.6]	2.4 [1.7–3.2]	2.8 [2.1–3.1]	<0.0001a
Maximum renal score	0.8 [0.7–1.0]	0.8 [0.7–0.9]	0.9 [0.7–1.4]	0.063a
Maximum hepatic score	0.0 [0.0–1.2]	0.0 [0.0–0.8]	0.7 [0.0–1.5]	0.002a
Maximum hematologic score	0.5 [0.0–1.1]	0.5 [0.0–1.0]	0.6 [0.2–1.1]	0.224a

∗ICU days and discharge day since injury when calculated only among survivors.

Median [q1-q3], or mean ± SD for continuous variables and n (%) for categorical variables are reported. p values calculations are indicated as follows.

Mann-Whitney U two-tailed test.

unpaired equal variance two-tailed Student's t-test.

Chi-square.

Fisher's Exact two-tailed test.

As expected, orthopedic procedures were the most frequent surgical interventions that patients received overall (76.6%), followed by laparotomy (47.7%), vascular procedures (23.4%), thoracotomy (6.3%), and craniotomy (1.6%) (Table 1). Apart from the proportion of patients having undergone laparotomy, which was significantly higher among MIIE cases than among non-cases (41.2% vs. 60.5%, p = 0.04), other procedures were similar between non-cases and MIIE cases.

There were five total patients who did not survive, and the cause of death was different for each (Table S2). Mortality was similar between non-cases (3.5%) and MIIE cases (4.7%, p = 1.00). Among survivors, MIIE cases had significantly longer hospital stay than non-cases (discharge at day 19 [14–26.75] vs. 35 [27-47], p < 0.0001). MIIE cases also had a higher proportion of those experiencing non-infection complications than non-cases (47.1% vs. 81.4%, p = 0.0002). Maximum Denver and Marshall scores were significantly higher for MIIE cases than for non-cases (Denver score: 2 [0-3] vs. 3 [2–3.5], p < 0.0001; Marshall score: 4.7 [3.3–6.4] vs. 6.9 [5.8–8.1], p < 0.0001). Among the sub-categories that together determine the total Marshall Score, maximum scores of the following were significantly higher for MIIE cases than for non-cases: cardio (2.4 [1.7–3.2] vs. 2.8 [2.8–4.0], p < 0.0001), respiratory (2.4 [1.7–3.2] vs. 2.8 [2.1–3.1], p < 0.0001), and hepatic score (0.0 [0.0–0.8] vs. 0.7 [0.0–1.5], p < 0.001). The maximum central nervous system, renal, and hematologic scores were not significantly different.

Characteristics of Infection Episodes

Patient Case Outcomes and Timing of Infection Onset

Among the 128 patients in the study, there were 85 non-cases—42 with no infection and 43 with one infection episode—and 42 were MIIE cases (i.e., ≥2 infection episodes). The median [IQR] day for detection of the first infection episode among those who experienced at least one infection episode (i.e., excluding 42 patients those who had no infection episode, for a total of 86 patients) was 8 [5-12] days (Table 2).

Table 2

Numbers Represent Individuals Who Have Experienced the Indicated Outcomes

Variable Description	All Patients (N = 128)	0 and 1 Infection Episode (Non-cases) (n = 85)	1 Infection Episode Only (Non-cases) (n = 43)	≥2 Infection Episodes (MIIE Case) (n = 43)
Characteristics of infection
First infection day since injury	8 [5–12]a	8 [5.5–12]a	8 [5.5–12]	7 [4–11.5]
Surgical site infections	36 (28.1%)	7 (8.2%)	7 (16.3%)	29 (67.4%)
Superficial incisional	7 (5.5%)	2 (2.4%)	2 (4.7%)	5 (11.6%)
Deep incisional	30 (23.4%)	5 (5.9%)	5 (11.6%)	25 (62.8%)
Other infection sites	80 (62.5%)	37 (43.5%)	37 (86.0%)	43 (100%)
Pneumonia	50 (39.1%)	20 (23.5%)	20 (46.5%)	30 (69.8%)
Ventilation-associated pneumonia	46 (35.9%)	20 (23.5%)	20 (46.5%)	26 (60.5%)
Bloodstream infection	22 (17.2%)	4 (4.7%)	4 (9.3%)	18 (41.9%)
Urinary tract infection	24 (18.8%)	7 (8.2%)	7 (16.3%)	17 (39.5%)
Pseudomembranous colitis	5 (3.9%)	3 (3.5%)	3 (7.0%)	2 (4.7%)
Catheter-related bloodstream infection	5 (3.9%)	1 (1.2%)	1 (2.3%)	4 (9.3%)
Empyema	3 (2.3%)	2 (2.4%)	2 (4.7%)	1 (2.3%)
Other	7 (5.5%)	2 (2.4%)	2 (4.7%)	5 (11.6%)
Organism incidence
Gram-positive bacteria	51 (39.8%)	20 (23.5%)	20 (46.5%)	31 (72.1%)
Staphylococcus aureus	25 (19.5%)	8 (9.4%)	8 (18.2%)	17 (39.5%)
Enterococcus species	16 (12.5%)	5 (5.9%)	5 (11.4%)	11 (25.6%)
Coagulase-negative staphylococci	8 (6.3%)	1 (1.2%)	1 (2.3%)	7 (16.3%)
Clostridium species	6 (4.7%)	3 (3.5%)	3 (7.0%)	3 (7.0%)
Streptococcus pneumoniae	3 (2.3%)	1 (1.2%)	1 (2.3%)	2 (4.7%)
Streptococcus viridans	3 (2.3%)	1 (1.2%)	1 (2.3%)	2 (4.7%)
Gram positive NOS	3 (2.3%)	1 (1.2%)	1 (2.3%)	2 (4.7%)
Gram-negative bacteria	57 (44.5%)	23 (27.1%)	23 (53.5%)	34 (79.1%)
Enterobacter species	21 (16.4%)	5 (5.9%)	5 (11.4%)	16 (37.2%)
Acinetobacter species	17 (13.3%)	4 (4.7%)	4 (9.1%)	13 (30.2%)
Pseudomonas aeruginosa	12 (9.4%)	5 (5.9%)	5 (11.4%)	7 (16.3)
Haemophilus influenza	8 (6.3%)	2 (2.4%)	2 (4.5%)	6 (14.0%)
Escherichia coli	7 (5.5%)	5 (5.9%)	5 (11.4%)	2 (4.7%)
Bacteroides species	4 (3.1%)	0 (0%)	0 (0%)	4 (9.3%)
Klebsiella pneumoniae	4 (3.1%)	1 (1.2%)	1 (2.3%)	3 (7.0%)
Neisseria	3 (2.3%)	0 (0%)	0 (0%)	3 (7.0%)
Proteus	3 (2.3%)	1 (1.2%)	1 (2.3%)	2 (4.7%)
Serratia marcescens	3 (2.3%)	1 (1.2%)	1 (2.3%)	2 (4.7%)
Stenotrophomonas species	2 (1.6%)	2 (2.4%)	2 (4.6%)	0 (0%)
Gram negative NOS	6 (4.7%)	1 (1.2%)	1 (2.3%)	5 (11.6%)
Fungi	10 (7.8%)	4 (4.7%)	4 (9.3%)	6 (14.0%)
Candida species	9 (7.0%)	4 (4.7%)	4 (9.1%)	5 (11.6%)
Fungi NOS	1 (0.8%)	0 (0%)	0 (0%)	1 (2.3%)
Unknown	5 (3.9%)	2 (2.4%)	2 (2.3%)	3 (7.0%)

The 42 non-cases who had 0 infection episode, without recordings for first infection day were omitted for these calculations. Recorded species are indicated or described as not otherwise specified (NOS).

(ii) Incidence of Surgical Site Infections versus Other Nosocomial Infections

Among all 128 patients, 36 (28.1%) experienced surgical site infections, compared to 80 (62.5%) who experienced other nosocomial infections (Table 2). Comparing specific subtypes of nosocomial infections, pneumonia (39.1% overall) was highest, followed by urinary tract infection (18.8%), blood infection (17.2%), pseudomembranous colitis (3.9%), catheter-related bloodstream infection (3.9%), empyema (2.3%), and other unspecified infections (5.5%).

(iii) Microorganism Detection

When comparing the incidence of various microorganisms among non-cases with one infection episode versus MIIE cases, relatively higher proportion was found for MIIE cases specifically for Gram positives of Staphylococcus aureus (18.2% vs. 39.5%), Enterococcus species (11.4% vs. 25.6%), coagulase-negative staphylococci (2.3% vs. 16.3%), and Streptococcus pneumoniae and viridans (2.3% vs. 4.7% for both). The incidence of the Gram-positive Clostridium species was the same for non-cases and MIIE cases (both 7.0%) (Table 2). For Gram-negative bacteria, the incidences of the following microorganisms were higher for MIIE cases than for non-cases: Enterobacter species (11.4% vs. 37.2%), Acinetobacter species (9.1% vs. 30.2%), Pseudomonas aeruginosa (11.4% vs. 16.3%), Haemophilus influenzae (4.5% vs. 14.0%), Bacteroides species (0% vs. 9.3%), Klebsiella pneumoniae (2.3% vs. 7.0%), Neisseria (0% vs. 7.0%), Proteus (2.3% vs. 4.7%), Serratia marcescens (2.3% vs. 4.7%), and Gram-negative, not otherwise specified (NOS) (2.3% vs. 11.6%). The incidence was higher among non-cases than among MIIE cases for Escherichia coli (11.4% vs. 4.7%) and Stenotrophomonas (4.6% vs. 0%). Fungi incidences were higher among MIIE cases than among non-cases: Candida species (9.1% vs. 11.6%) and unspecified fungi (0% vs. 2.3%).

(iv) Timing of Microorganism Detection

The median time to the first day of detection of different microorganisms ranged widely (Figure 2). The Gram positives, Streptococcus pneumoniae, and Streptococcus viridans were found first (at median day 4 and 5, respectively), followed by various Gram negatives and the Gram-positive, Clostridium species, between median day 7 and 9.5. Microorganisms that appeared relatively later (day 11–12) include Candida species, Enterobacter species, and Serratia marcescens, Stenotrophomonas, Enterococcus, Coagulase-negative staphylococci, and Staphylococcus aureus.

Figure 2

Timing of Detection of Microorganisms

Box plot of the median day of the first infection since the injury, where white boxes indicate Gram positives, light gray boxes indicate Gram negatives, and dark gray boxes indicate fungi and unknown. Recorded species are indicated or described as not otherwise specified (NOS).

Gene Ontology Reports of 1.5-Fold Changed Probe Sets

We identified 137 probe sets showing at least 1.5-fold upregulated or downregulated difference in expression levels between non-cases and MIIE cases (Figure 3A) and performed Gene Ontology (GO) analyses to find enriched biological processes and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway terms (Table S3). As expected, terms relevant to various immune response pathways were among the top fold enrichment. They included interleukin-4 secretion, regulation of interferon-gamma secretion, cytolysis, regulation of natural killer cell-mediated cytotoxicity, viral genome replication, cellular defense response, adaptive immune response, humoral immune response, T cell co-stimulation, regulation of immune response, T cell receptor signaling pathway, response to tumor necrosis factor, response to virus, immune response, and innate immune response. Other biological process terms with high fold enrichment were signaling pathways with relevance to cell proliferation and differentiation, including regulation of p38 MAPK kinase, calcium-mediated signaling, regulation of fat cell differentiation, organ regeneration, MAP kinase activity, and phosphatidylinositol 3-kinase (PI3K) signaling. Similarly, KEGG pathway terms with high fold enrichment included those relevant to immune response, including natural killer cell-mediated cytotoxicity, malaria, hematopoietic cell lineage, and T cell receptor signaling pathway. Additionally, regulation of cancer, which may have overlapping signaling pathways with infections-related terms, and related to tissue regeneration, also appeared among the enriched terms. Although according to the false discovery rate (FDR)-adjusted p values, statistical significance in enrichment was detected only for a small number of terms, the fold enrichment ranking consistently points to the relevance of immune response terms, as expected.

Figure 3

Identification of Probe Sets with 1.5-Fold Change Expression Level Comparing MIIE Cases to Non-cases, and Pathway Analysis

(A) Volcano plot using the initial 25,567 probe sets included in the analyses, where black dotted lines mark the log2(1.5) threshold, black dots indicate probe sets with ≥1.5-fold change (137 probe sets), orange squares with labels mark data points corresponding to the 15-probe set panel, and blue triangles mark the additional probe sets, that together with the 15 probe sets, result in the 26-probe set panel.

(B) Ingenuity pathway analysis (IPA) of the 26 probe sets that make up the comprehensive biomarker panel. Green marks downregulated and red marks upregulated probe sets. Solid lines indicate known direct interactions, and dotted lines indicate indirect interaction.

Prognostic Biomarker Panel Selection

To identify a multi-biomarker panel that collectively predicts the outcome of MIIEs, we analyzed the 137 differentially regulated probe sets by employing a machine learning pipeline that we previously developed and successfully used in our previous study among burn patients (Yan et al., 2015). In the current study, we further added to our analysis pipeline by utilizing a combination of LASSO and Elastic Net regression methods. The LASSO regression that reduces redundancy in predictor selection would allow for a narrow selection of a minimal biomarker panel, which is expected to be more practical. Elastic Net regression that includes correlated predictors would allow for a more comprehensive discovery of additional probe sets that are potentially biologically relevant. With LASSO, 15 probe sets were selected, mostly relevant to immune functions and signaling cascades for cellular proliferation and differentiation (Tables 3 and S4; Figures 3A, S2A, and S2B). Upregulated probe sets in the minimal 15-probe set panel included those associated with immune function terms, hepatocyte growth factor (HGF) and kelch repeat and BTB domain containing 7 (KBTBD7); signaling cascades, adenosine A3 receptor (ADORA3), ADP-ribosylation factor-like GTPase 4A (ARL4A), epiplakin 1 (EPPK1), zinc finger protein 354A (ZNF354A), SH3 and PX domains 2B (SH3PXD2B); and those with no function term assigned, RNase A family 1 (RNASE1), and BTB domain containing 19 (BTBD19). Downregulated probe sets included those with immune functions, zeta chain of T cell receptor associated protein kinase 70kDa (ZAP70), ER aminopeptidase 2 (ERAP2), CD96 molecule (CD96), membrane metallo-endopeptidase (MME), and killer cell lectin-like receptor subfamily F, member 1 (KLRF1); and a non-coding transcript, nuclear paraspeckle assembly transcript 1 (NEAT1) (Tables 3 and S4; Figures 3A).

Table 3

Fold Change and P values of the 15-Probe Set Panel from LASSO and 26-Probe Set Panel from Elastic Net, in Order of Magnitude of upregulated to Downregulated

Probe Set	Gene Symbol	Gene Name	Fold Change (MIIE/non-cases)	In Both 15- and 26-Probe Set Panel, or Uniquely in 26-Probe Set Panel
210998_s_at	HGF	hepatocyte growth factor	1.85	15 and 26
238488_at	IP O 11	importin 11	1.80	26
1557049_at	BTBD19	BTB (POZ) domain containing 19	1.76	15 and 26
211372_s_at	IL1R2	interleukin 1 receptor, type II	1.72	26
205003_at	DOCK4	dedicator of cytokinesis 4	1.61	26
223660_at	ADORA3	adenosine A3 receptor	1.59	15 and 26
229970_at	KBTBD7	kelch repeat and BTB (POZ) domain containing 7	1.59	15 and 26
203962_s_at	NEBL	Nebulette	1.58	26
205020_s_at	ARL4A	ADP-ribosylation factor-like GTPase 4A	1.57	15 and 26
203543_s_at	KLF9	Kruppel-like factor 9	1.56	26
204438_at	MRC1	mannose receptor, C type 1	1.56	26
201785_at	RNASE1	ribonuclease, RNase A family 1	1.54	15 and 26
232164_s_at	EPPK1	epiplakin 1	1.52	15 and 26
205427_at	ZNF354A	zinc finger protein 354A	1.51	15 and 26
231823_s_at	SH3PXD2B	SH3 and PX domains 2B	1.50	15 and 26
214032_at	ZAP70	zeta chain of T cell receptor-associated protein kinase 70kDa	0.67	15 and 26
202208_s_at	ARL4C	ADP-ribosylation factor-like GTPase 4C	0.66	26
227462_at	ERAP2	ER aminopeptidase 2	0.66	15 and 26
206761_at	CD96	CD96 molecule	0.66	15 and 26
203434_s_at	MME	membrane metallo-endopeptidase	0.65	26
238320_at	NEAT1	nuclear paraspeckle assembly transcript 1 (non-protein coding)	0.65	15 and 26
203435_s_at	MME	membrane metallo-endopeptidase	0.65	15 and 26
204633_s_at	RPS6KA5	ribosomal protein S6 kinase, 90kDa, polypeptide 5	0.63	26
1555691_a_at	KLRK1	killer cell lectin-like receptor subfamily K, member 1	0.62	26
220646_s_at	KLRF1	killer cell lectin-like receptor subfamily F, member 1	0.55	15 and 26
206666_at	GZMK	granzyme K	0.50	26

With Elastic Net, a total of 26 probe sets were selected that included the 15 probe sets from LASSO and 11 additional ones (Tables 3 and S4; Figures 3A, S2C, and S2D). The additional upregulated probe sets consisted of those with immune functions, interleukin 1 receptor, type II (IL1R2) and mannose receptor, C type 1 (MRC1); those involved with signaling including, importin 11 (IP O 11), dedicator of cytokinesis 4 (DOCK4) and Kruppel-like factor 9 (KLF9); as well as nebulette (NEBL) that is important for muscle filament assembly. The additional downregulated probe sets included a different probe set for MME, and other immune function genes, ribosomal protein S6 kinase, 90kDa, polypeptide 5 (RPS6KA5), killer cell lectin-like receptor, subfamily K, member 1 (KLRK1) and granzyme K (GZMK); and the signaling molecule, ADP-ribosylation factor-like GTPase 4C (ARL4C) (Table 3; Figures 3A and Table S4).

Pathway Network Suggests the Relevance of Major Pathways

We assessed the molecular network connection among the 26 probe sets that were selected by Elastic Net (Figure 3B). The major nodes with the most extensive edges that were central to the connections consisted of major signaling pathway components that are key regulators of inflammation, mitogenic response, and tissue regeneration. These notable nodes included tumor necrosis factor (TNF), transforming growth factor beta-1 (TGFβ-1), and various chemokine (C-X-C motif) ligand (CXCL) members and chemokine (C-C motif) ligand (CCL) members. TNF and TGFβ-1 are major cytokines that can act both synergistically or antagonistically to each other, depending on the cell context. Further substantiating the involvement of these factors, we found p38 mitogen-activated protein kinase (MAPK), which can act downstream of both TNF-α and TGFβ-1 pathways. The other major nodes, extracellular signal-regulated kinases 1/2 (ERK1/2), and mothers against decapentaplegic homolog 3 (SMAD3) are also key downstream components of the TGFβ pathway. A key downstream transcriptional regulator for the canonical Wnt pathway, β-catenin (CTNNB1), another major pathway known to cross talk with TGFβ, was also found as a major node in this network.

Performance in Prediction of MIIEs Using the Multi-Biomarker Panel versus the Clinical Severity Models

The Area Under Receiver Operating Characteristics Curve (AUROC) [95% confidence intervals, CI] of the logistic regression model for predicting MIIE outcomes with the 15-probe set biomarker panel developed with LASSO was 0.90 [0.84–0.96] (Figure 4A). For the 26-probe set panel developed with Elastic Net, there was marginal AUROC increase (0.92 [0.86–0.96], p = 0.11) (Figure 4A). Given the objective of the study of developing measures taken early upon arrival at the hospitals for predicting future outcome during recovery, we evaluated common injury severity scores. Compared to the biomarker models, the AUROC of the various injury severity models was notably lower (0.64 [0.54–0.74] for APACHE II, 0.611 [0.51–0.71] for ISS, and 0.61 [0.50–0.71] for NISS, all p < 0.0001 compared to the 15-probe set panel model) (Figure 4A). The odds ratios (ORs) [95% CI] for all the covariates of each of the models were also found (Tables S5–S7).

Figure 4

Comparisons of the Performance of the Multi-Biomarker Panel versus the Clinical Severity Models

ROC curves of the various models constructed and the respective AUROC [95% CI], 10-fold cross-validation (CV) AUROC [95% CI], and p value of total AUROC difference compared to the 15-probe set biomarker panel.

(A) Various models: 15-probe set biomarker panel, 26-probe set biomarker panel, APACHE II, ISS, NISS.

(B) Various combined biomarker and clinical score models: 15-probe set biomarker + APACHE, 15-probe set biomarker + ISS, 15-probe set biomarker + NISS, 26-probe set biomarker + APACHE, 26-probe set biomarker + ISS, 26-probe set biomarker + NISS.

The 15- and 26-probe set panel models had sensitivity [95% CI] of 0.74 [0.59–0.86] vs. 0.79 [0.64–0.90]; specificity [95% CI] of 0.94 [0.87–0.98] vs. 0.94 [0.87–0.98]; positive predictive value (PPV) of 0.86 [0.71–0.95] vs. 0.87 [0.73–0.96]; and negative predictive value (NPV) of 0.88 [0.79–0.94] vs. 0.90 [0.82–0.95], respectively (Table S8). For the various injury severity scores (APACHE II, ISS, NISS), the sensitivity, specificity, PPV, and NPV were generally lower compared to the multi-biomarker panel models (Table S8).

Moreover, we constructed multivariate logistic regression models combining the 15- or the 26-probe set panel with each of the clinical injury severity scores. The AUROC [95% CI] of the 15 probe set panel combined with APACHE II was 0.90 [0.84–0.96], was 0.902 [0.84–0.96] with ISS, and was 0.902 [0.84–0.96] with NISS (Figure 4B). For the 26-probe set panel, the combination with APACHE II was 0.90 [0.86–0.97]; with ISS, it was 0.92 [0.86–0.97]; and with NISS, it was 0.92 [0.86–0.97] (Figure 4B). None of the combined models were significantly different from their respective biomarker only models, suggesting that the addition of clinical score information does not improve the prediction.

External Validation of the Multi-Biomarker Panel in a Different Cohort of Severe Trauma Patients

We performed external validation of our biomarker panel using a severe blunt trauma patient cohort from a previous publication by Cabrera et al. (Cabrera et al., 2017), which had post-admission blood transcriptome log2 expression values available. The microarray platform used in this study was different from the one used in the Glue Grant, and the measure of one of the genes in the 15-probe set panel was missing. Despite the differences in measurement technology, the lack of measurement for one of the panel, and outcome resolution between the Glue Grant and the Cabrera et al. data sets (as described in detail in the methods section), the multi-biomarker panel achieved a relatively high AUROC of 0.76 [0.57–0.96] for the 24-hr post-admission data set and 0.81 [0.62–1.00] for 72-hr post-admission (Figures 5A and 5B). On the other hand, the model with ISS (the only injury severity score that was shared with the Cabrera et al. data set) had a much lower AUROC of 0.64 [0.42–0.86], which was not significantly above 0.5. These results provide evidence for the generalizability of our multi-biomarker panel to different trauma cohort.

Figure 5

The AUROC Curve When Applying the Biomarker Panel Model Constructed using the Glue Grant, to the Cabrera et al. Data Set as the External Validation Test Set

For both the (A) 24 hr and (B) 72 hr time point of the Cabrera et al. data set, our biomarker panel model conferred significant prediction, as evidenced by the AUROC [95% CI] significantly above 0.5. On the other hand, the prediction model with (C) Injury Severity Score (ISS) did not provide significant prediction.

Discussion

Our study shows that employing novel prognostic models based on early blood transcriptome profiling following severe trauma is an effective method for identifying patients who are particularly at high risk for MIIE and thus hypersusceptible to infections. That the transcriptome information provides much better prediction than injury severity information argues for the importance of considering each patient's underlying susceptibility and of elucidating relevant molecular mechanisms. The comprehensive data set we used had genome-wide transcriptome and clinical data collected longitudinally from a large number of patients, providing the opportunity to assess early susceptibility. Notably, our results suggest that by measuring the biomarkers among our panel at admission, patients at increased risk for MIIE could be identified before any clinical signs of infection appear. The biomarker panel models in our study had particularly high specificity and NPV measures, while also exerting good sensitivity and PPV. Moreover, when applied to an external validation cohort, it still performed decently. On the other hand, none of the injury severity scores used often at triage in the trauma setting were effective in predicting the MIIE.

The objective of the study was to provide proof-of-concept results for developing a method to gain additional insights into patients' course of recovery from a simple blood draw at admission. Further prospective studies that entail blood draws at admission and measuring the biomarkers we describe here to compare with the severity scores and physiological measures taken normally would provide additional confirmation of the notion that transcriptome data has the potential to improve outcome predictions in the clinical setting. This study focused on the outcome of MIIEs and the potential prevention of infections; however, it is conceivable that the biomarker discovery method described here can be applied to develop prediction methods for other outcomes. A personalized medicine approach and rapid identification of patients with high risk of specific outcomes based on a simple blood draw at admission is expected to improve surveillance, facilitate decision-making and adequate resource allocation, and improve prevention and management before outcomes occur.

Our findings could potentially facilitate clinical decision-making by effectively discriminating between those who are expected to develop multiple infections and those who are not. A proposed approach could be that patients who are found not to be hypersusceptible to MIIEs could continue to receive the currently established standard of care, whereas those who are identified to be at high risk could benefit from increased surveillance and additional preventative measures to be taken early. Additional interventions for the high-risk group may include increased surveillance for early mobilization and removal of lines/tubes, coating IV lines and urine catheters with antimicrobials and/or antibiotics, immunomodulatory nutrition therapies (Aghaeepour et al., 2017; Lorenz et al., 2015), and microbiome alterations (Harris et al., 2017; Tosh and McDonald, 2012). Such additional measures would incur unnecessary costs if implemented in all trauma patients; however, it could be cost-effective when used in this targeted set of patients. Efforts aimed at increased prevention have the potential to contribute to alleviating the current antibiotic resistance crisis, toxicity of antibiotics, and the imposed burden on healthcare costs. It is also conceivable that accurate outcome prediction and risk stratification methodologies, such as the one we describe here, could be valuable amid crisis situations that result in severe hospital overload with critically ill patients and scarcity of medical resources. The ability to identify patients at low risk for specific morbidity and mortality could aid in informed prioritization of resource allocation to patients with better potential for recovery and survival (The Commonwealth of Massachusetts, 2020; Institute of Medicine, 2013; University of Pittsburgh, 2020).

Having applied both LASSO and Elastic Net regression methods allowed us to construct a highly predictive model from a minimal set of predictors, meanwhile also more comprehensively assess underlying biological mechanisms by allowing additional transcripts to be included. The LASSO approach selects a stringent set of predictors with less redundancy, which is advantageous in the clinical setting, where a device requiring fewer measurements is more practical and easier to implement. The Elastic Net approach that allows for correlated predictors to be selected found additional transcripts for a more comprehensive discovery of biological mechanisms. The probe sets selected consisted of transcripts with GO terms relevant to infections, as expected, and signaling pertinent to oncogenesis and cancer progression.

In our study, HGF was the transcript showing the highest upregulation among MIIE patient blood and in both the 15-probe set and 26-probe set panels. HGF and Met expression levels have been suggested as a putative biomarker for monitoring infections, as it is well established that the HGF-Met signaling pathway deregulation promotes the growth and invasion by various pathogens (Imamura and Matsumoto, 2017). Another upregulated transcript, ADORA3, has also been implicated in the clinical setting, and agonists have been developed and shown to induce anti-inflammatory effects by altering the Wnt and NF-κB pathways. As such, the agonists are considered for purposes of treating cancers and inflammatory diseases such as rheumatoid arthritis and psoriasis (Fishman et al., 2012). NEAT1 is a non-coding RNA that is shown to colocalize with MALAT1, a long non-coding RNA often associated with metastatic cancer, at many genomic sites to transcriptionally regulate target genes (West et al., 2014). CD96 is highly expressed in T and NK cells and well established to be a regulator of immune responses during infection and cancer (Georgiev et al., 2018).

Elastic Net selected two probe sets corresponding to MME, providing further support for its importance in MIIE outcome. Studies on its molecular mechanisms and clinical use of inhibitors to its protein product, Neprilysin has been conducted widely, including in Alzheimer's, heart failures, hypertension, and renal diseases (Riddell and Vader, 2017). Our study suggests that its potential role in immunity among patients warrants further investigation. KLRK1 and KLRF1, both killer cell lectin-like receptor subfamily members, were found by Elastic Net, providing evidence of their relevance in infections in the blunt trauma setting. These receptors are abundant on NK cells, and it is well established that they play crucial roles in innate immunity (Barten et al., 2001). These previous findings provide additional confidence in the relevance of our methodology. The pathway analysis found key signaling pathway components among the central nodes having extensive edges, including the major cytokines, TNF, TGFβ-1, CCLs, and CXCLs, as well as key signaling components, p38 MAPK, ERK1/2, SMAD3, and CTNNB1. These components represent the chief signaling pathways that regulate inflammation, mitogenic response, and tissue regeneration, which are also often dysregulated in cancer. Notably, our results suggest that the TNF, TGFβ-1, and Wnt signaling pathways, which are known to also cross talk with one another through downstream cascades, may be important central pathways that explain the interconnection between the prognostic biomarkers identified. These results may suggest that these signaling pathways may represent new host immunomodulatory targets that warrant future mechanistic studies. Follow-up studies in model organisms and controlled studies would aid in establishing whether the genes identified in this study drive susceptibility and in uncovering further mechanistic insights.

It is noteworthy that when comparing the current biomarker panels in the blunt trauma setting with that from our previous study among burn patients (Yan et al., 2015), we observed that none of the transcripts in the panels were shared. These differences may indicate that increased risk depends on the interaction of the type of trauma with each patient's underlying susceptibility to MIIE. Such observation may suggest the need for developing different multi-biomarker panels catered to different types of trauma.

Our study describes methods toward the development of precision medicine tools and offers the possibility of analyses also for outcomes other than multiple infections. The failure of drug trials targeting sepsis (Marshall, 2014; Mitka, 2011) highlights the importance of further studies elucidating the underlying molecular mechanisms and components of heterogeneity in susceptibility to infection and infection-related morbidity within a population. It is conceivable that measuring our biomarker panel to triage patients according to susceptibility to multiple infections will strategically guide prophylactic patient management and help reduce the incidence of infections to limit sepsis (Boomer et al., 2011; Chaussabel, 2015; Parikh et al., 2016). Moreover, the analysis process we describe in this study can potentially also be applied toward biomarker development for sepsis outcome.

This study provides for the first time prediction models for hypersusceptibility to infections, which is highly relevant for critically injured trauma patients, using a machine learning approach. A concern in general for prediction model building is that models may overfit to a specific data set, making them less generalizable. However, using the multi-biomarker panel derived from the Glue Grant population to make predictions in the Cabrera et al. population yielded a relatively high AUROC, demonstrating the broader applicability of our biomarker panel model. These two populations had comparable injury severity; however, they were considerably different in their geographic locations and healthcare systems. Moreover, the gene expression levels were measured by two different transcriptome technologies, and the Cabrera et al. data set was very small in sample size. Despite these differences, our multi-biomarker panels still conferred prediction, providing additional assurance in the validity of our results and evidence for the generalizability of our model. Additional large prospective studies would more rigorously test the validity and generalizability of the multi-biomarker panel identified in this study. Nevertheless, this study provides the first step toward the idea of developing novel approaches for predicting outcomes from blood transcriptome information at admissions.

The value of early MIIE identification, prior to any clinical sign of infection, could be an indispensable tool in other types of trauma and to a wide range of clinical settings. Uncovering biomarkers of increased susceptibility to infections may open new avenues for novel therapeutic targets, as well as contribute to standardizing populations in clinical trials. Although predictive algorithms cannot eliminate medical uncertainty, our analysis method is expected to be widely applicable to other susceptible populations, such as those with diabetes or cardiac disease, the frail elderly population, those treated with immunosuppressive medication, as well as others. The described methodology of multi-biomarker panel development has the potential to be applied to outcomes and clinical contexts other than MIIE and trauma, providing additional value.

Limitations of the Study

This study entailed a secondary analysis, with external validation using a small data set. Additional external data sets with larger sample sizes, and moreover, a large prospective study would provide additional concrete evidence for the validity and utility of our biomarker panel.

Resource Availability

Lead Contact

Further information and requests for resources should be directed to and will be fulfilled by the Lead Contact, Laurence G. Rahme (rahme@molbio.mgh.harvard.edu).

Materials Availability

This study did not generate new unique reagents.

Data and Code Availability

Data set requests should be made to the Glue Grant Consortium, due to human study IRB restrictions. The code for the analyses is available from the lead contact upon request.

Methods

All methods can be found in the accompanying Transparent Methods supplemental file.