A total of 8239 patients were screened in the CanTBI study (Fig. 1); 3465 patients screened positive for TBI (42%). After informed consent, 466 adult and pediatric patients with mild, moderate, and severe TBI were enrolled into the prospective CanTBI biobank and database for TBI study. There were 300 adult patients with TBI and 59 of these patients (19.6%) were diagnosed with severe TBI (sTBI) and included in this study. These 59 patients had a mean age of 50 years ± 20.6 (SD). For detailed patient and injury characteristics see Table 1. Figure 1 shows the patient flow chart with the numbers of patients with follow-up data at 3- and 12-months post-injury. The clinical prognostic model results for each clinical variable are shown in Table 2. Only, age and injury severity score (ISS) were significantly (p < 0.05) associated with unfavorable outcome at 3 months, but not significant at 12 months. There was a significant difference in age and ISS between patients who died (n = 21) and those who survived (n = 23) at 3 months (Table S1), with older age and higher ISS associated with unfavorable outcome. The cut points of ISS and age were determined at ≥ 75 and ≥ 49, respectively, the significant predictors to separate non-survivors from survivors at 3 months. Also, the cut points were calculated for Marshall classification = 4 and GCS = 6 between non-survivors and survivors. These two variables were not statistically significantly different between the two cohorts.
-
Metabolomics for the prognosis of 3- and 12-month outcomes of sTBI
Prediction models show that a serum metabolic biosignature can be used to prognosticate GOSE outcome at 3 and 12 months and the mortality outcome at 3 months.
Unsupervised PCA showed a relatively good grouping between cohorts with unfavorable and favorable outcome using all metabolites detected in serum samples collected on days 1 and 4. PCA revealed a high level of variability (R2X > 0.5) of metabolites suggesting a biosignature between two cohorts (Fig S1-S3). Metabolic biosignatures obtained by DI/LC-MS/MS using samples on day 4 presented clearer groupings between unfavorable and favorable cohorts compared with 1H-NMR and samples on day 1. PLS-DA-based analysis demonstrated a good predictive (Q2 > 0.5), highly significant (p < 0.001) and highly sensitive and specific (> 99%) prediction model to discriminate between patients with unfavorable vs favorable outcomes using a serum metabolic biosignature on day 4 obtained by DI/LC-MS/MS (Table 3 and Fig. 2). Nonetheless, day 1 metabolic biosignatures were also significant predictors for GOSE outcomes (Fig S4). The permutation analysis (200 times permuted, not shown) verified that the models are valid and unlikely to be overfit. Artificial neural network analysis, a machine learning-based analysis (ANN, Tables S4-S5) indicated the higher predictability (AUC > 0.90) for the prognosis of GOSE outcome among patients with unfavorable outcome compared with patients with favorable outcome at 3 months and higher predictability (AUC > 0.90) for the prognosis of GOSE outcome among patients with favorable outcome compared with patients with unfavorable outcome at 12-month. This is based primarily on DI/LC-MS/MS data on day 4. See supplementary data for more details.
Further analyses were performed to investigate the relative correlation of the most differentiating metabolites between unfavorable and favorable outcomes for each prediction model-based list of metabolites including 9 to 26 metabolites for different models (Figs. S5-S16). A predictive metabolic biosignature to predict GOSE outcome at 3-months was characterized by an increased in lysoPCs, propionic acid (C3:1), stearic acid (C18), oleic acid (C18:1), linoleic acid (C18:2), and myristic acid (C14) on the 1st day post-injury yielding an unfavorable outcome (Figs. S5A and B). Also, a decrease in methionine-sulfoxide, glutamate, histidine, citrulline, isoleucine, glutamine, phenylalanine, and asparagine were associated with an unfavorable outcome on the 1st day post-injury (Figs. S5A & B). Interestingly, a predictive metabolic biosignature on day 4 (Figs. S6 and S8) showed increased glutamate (excitotoxicity), propionic acid, linoleic acid, valeric acid (C5), indole acetic acid, ɑ-ketoglutaric acid, ɑ-aminoadipic acid, alanine, lysoPCs (18:2, 18:0 & 17:0), tyrosine, NAA, aspartate, and valine in those with an unfavorable outcome, while these metabolites were decreased on day 1 post-injury. For prognosis of GOSE at 12-months, patients with unfavorable outcome were characterized by increased lysoPCs (14:0, 20:3 & 28:1), short chain ACs (C5OH, C3, C0, C4). ornithine, sphingomyelin (16:1), valine, serine, leucine, lactate, and a decrease in trans-hydroxyproline, serine, serotonin, citrulline spermine methionine-sulfoxide, acetylornithine and medium chain acylcarnitines on day 1 post-injury (Figs. S9 and S11). In addition, 12-month unfavorable outcome was associated with increased lysoPCs (28:1, 14:0), tryptophan, caproic acid (C6), oleic acid, tyrosine, creatinine, alanine, histidine, valine, and leucine on day 4 post-injury (Figs. S10 and S12). To predict 3-month mortality, metabolomic analysis showed increased glucose, PCs (38:0aa, 40:6 ae), acylcarnitines (C3:1, C10:1, C14:1 C14, C10, C16:2 C8), betaine, 3-hdroxyisovalerate, citrate, O-phosphocholine, formate, fumarate, and pyruvate on day 1 in patients predicted to die by 3 months (Figs S13 and S14). Those patients predicted to die showed decreased glutamine, and branched chain amino acids, citrulline and histidine on day 1 (Fig. S13 and S15). Increased ɑ-ketoglutaric acid, hippuric acid, indole acetic acid, ornithine tryptophan, ɑ-aminoadipic acid, PCs (38:0aa, 36:0aa), branched chain amino acids, creatine, creatinine, tyrosine threonine was found in those patients predicted to die based on the day 4 metabolic profile (Figs S14 and S16). Univariate T-test analysis showed remarkable similarities to PLS-based prediction models to identify predictive biomarkers (Figs. S5-S16). See supplementary data for more details.
Metabolite heatmap plot (Fig. 3 & S17-S18) directly visualized the metabolite alterations on the same days and from day 1 to day 4 for both cohorts with unfavorable and favorable outcomes. These results show a higher level of metabolite alterations from day 1 to day 4 particularly for predicting unfavorable outcomes compared to favorable outcome. Overall, metabolite changes are more considerable among the patients with unfavorable outcomes. Lysophosphatidylcholines (16:0, 16:1, 17:0, 18:0 and 18:2) and lysophosphatidylcholines 14:0, 20:3, 28:1, and 18:0) showed an increase from day 1 to day 4 mostly in patients with unfavorable outcomes. Also, BCAAs, NAA, tyrosine, ornithine, and glutamate increased from day 1 to day 4 predominantly among the patients with unfavorable outcome. Histidine, alanine, serine, citrulline, pyruvate and lactate decreased from day 1 to day 4 mainly among the patients with unfavorable outcomes.
A brief overview highlights that unfavorable outcome was associated with increased metabolites related to lipids and anaerobic metabolism and decreased metabolites related to serotonergic, polyamine metabolism and NMDA receptor integrity in day 1 post-injury. Increased metabolites related to neuroinflammation, excitotoxicity and brain injury specific biomarkers were found on day 4 post-injury. Also, notable was an association of increased metabolites related to acylcarnitine metabolism and energy metabolism with mortality.
-
Clinical variables for the prognosis of GOSE outcome at 3 months, 12 months, and mortality
We investigated whether clinical variables could predict the outcome of sTBI at 3- and 12-months post sTBI. SIMPLS analysis revealed the most differentiating clinical variables for predicting outcomes at 3 months (age, ISS, Marshall classification and hypoxemia) and 12 months (age, GCS, hypoxemia, and loss of consciousness). However, these clinical variables had low prediction capacity (Q2 < 0.16) and less sensitivity (66%), and specificity (86%) compared to metabolomics-based prediction (Table S6). SIMPLS analysis of clinical data revealed that age and ISS are useful predictors (Q2 = 0.37, AUC = 0.86) to prognosticate mortality. However, these clinical variables lack significant sensitivity and specificity (66%-83%) compared to metabolomics data (Table S6 vs. Table 3 and Table S7).
SIMPLS analysis demonstrated that clinical variables could moderately improve the performance of metabolomics-based prediction models to prognosticate only GOSE outcome at 3 months and mortality (Table S7). For the prognosis of GOSE outcome at 12-month, clinical variables were found to minimally improve the metabolomics model (data not shown). However, age was an important clinical predictor of outcome among clinical variables, with a high level of contribution to prediction models, particularly for mortality. Consequently, Marshall classification (3 months outcome) and GCS (12 months outcome) remain important clinical variable (Table S8). Although SIMPLS and PLS-DA use different algorithms, the two approaches showed overall similar predictabilities when metabolites were used to prognosticate sTBI outcomes, with only slight differences (as shown in Table S7 vs Tables S4-6). Importantly, permutation tests (not shown) verified the predictabilities of metabolite-based prediction models and was used to help prevent overfitting of the data.