Abdominal myosteatosis measured with computed tomography predicts poor outcomes in patients with glioblastoma

Abstract

Background

Alterations in cellular metabolism affect cancer survival and can manifest in metrics of body composition. We investigated the effects of various body composition metrics on survival in patients with glioblastoma (GBM).

Methods

We retrospectively analyzed patients who had an abdominal and pelvic computed tomography (CT) scan performed within 1 month of diagnosis of GBM (178 participants, 102 males, 76 females, median age: 62.1 years). Volumetric body composition metrics were derived using automated CT segmentation of adipose tissue, skeletal muscle, and aortic calcification from L1 to L5. Univariable and multivariable Cox proportional hazards models were performed separately in males and females using known predictors of GBM overall survival (OS) as covariates. A sex-specific composite score of predisposing and protective factors was constructed using the relative importance of each metric in GBM OS.

Results

Higher skeletal muscle volume and lower skeletal muscle fat fraction were associated with better OS in the entire dataset. A robust and independent effect on GBM OS was seen specifically for fraction of inter/intramuscular adipose tissue to total adipose tissue after correction for known survival predictors and comorbidities. Worse OS was observed with increased abdominal aortic calcification volume in both sexes. There was a significant difference in GBM OS among participants stratified into quartiles based on sex-specific composite predisposing and protective scores.

Conclusion

The relationship between body composition and GBM OS provides an actionable advancement toward precision medicine in GBM management, as lifestyle and dietary regimens can alter body composition and metabolism and from there GBM survival.

Glioblastoma (GBM) is the most common malignant central nervous system (CNS) pathology in adults with an incidence of 3.5 per 100 000 people.¹ Despite advances in surgical techniques and chemoradiation, the prognosis of GBM is extremely poor with a median survival of 8–9 months and 5-year survival of 6.9%.^1–3 One avenue that could provide new insights into improving outcomes in these patients is by understanding how systems physiology can both influence and be influenced by tumorigenesis within the CNS.

Obesity as measured with body mass index (BMI) or visceral adiposity is associated with mortality in several cancer types including GBM.⁴ Higher BMI, particularly in the context of overweight and obesity, has been linked to conflicting findings: some studies report lower progression-free survival in GBM patients,^5,6 while others suggest higher overall survival (OS) in newly diagnosed GBM patients^7,8 with higher BMI. Higher abdominal adiposity measured as waist circumference is also related to worse OS in GBM.^4,7 Sarcopenia, characterized by reduced muscle volume and elevated intramuscular fat, has synergistic effects with visceral adiposity in predicting increased cancer mortality,^9,10 including GBM mortality when measured in the head.¹¹ Given these observations, it is possible that quantification of abdominal body composition could be a more robust source of metabolism estimation in GBM patients compared to BMI alone, as it allows for simultaneous estimation of visceral, subcutaneous, and intramuscular adipose tissue contents as well as various muscle groups. Furthermore, biomarkers derived from abdominal and pelvic organs could provide insights into alternative mechanisms by which cellular metabolism can affect cancer survival.

Interestingly, sex differences in body composition exist, where males have higher amounts of visceral fat and muscle mass and females have higher amounts of subcutaneous fat.¹² This has implications in cancer outcomes, as sex differences in phenotype and cellular metabolism affect cancer survival,^13–15 and those males with GBM typically have worse survival than females.^13,16,17 Furthermore, our group has used abdominal computed tomography (CT) to demonstrate an interaction between visceral fat and sex predicting survival in renal cell carcinoma, pancreatic cancer, and lymphoma.^14,18,19 This raises the question whether abdominal body composition metrics could be used to predict survival differently in males and females with GBM.^20,21

To our knowledge, the association between abdominal body composition measures and GBM outcomes has not been evaluated. In this study, we examined the association between abdominal body composition metrics, including sex differences, in predicting OS in patients diagnosed with GBM.

Materials and Methods

Participants

Through a retrospective cohort design, we used participants who were recruited from the Multi-Center Intraoperative MRI (iMRI) Neurosurgery Database (I-MiND). I-MiND is a prospective multisite Research Electronic Data Capture (REDCap) registry, supported by an educational grant from the IMRIS intraoperative MRI system, with ongoing recruitment of patients receiving preoperative MRI for resection of brain tumors.²² For this study, we first identified all I-MiND participants who were recruited from referrals to the Barnes-Jewish Hospital System—affiliated with Washington University School of Medicine—between January 2006 and January 2017 and had at least one abdominal and pelvic CT scan at any time in their medical record. The date of brain MRI at the time of recruitment in the study was considered as diagnosis or baseline date. The following additional inclusion criteria were then applied to select participants for the current study: (1) having a postoperative tissue diagnosis of GBM, (2) GBM being the only malignancy on record for that patient, and (3) CT performed with ±1 month of diagnosis or baseline date. This led to the inclusion of 178 participants (102 males, 76 females, median age [Q1–Q3]: 62.1 [56.3–73.1] years). If multiple CT scans were available for 1 participant, the scan closest to the baseline date was selected for analyses. All CT scans were contrast-enhanced. The most common indication for CT scans was postoperative abdominal pain (34.6%), suspected bowel obstruction (23.1%), epigastric pain (8%), and suspected abdominal infection (5.9%).

Abdominal and Pelvic CT Scan Body Composition Analysis

Body composition analysis was performed using the Data Analysis Facilitation Suite (DAFS, version 3) platform provided by Voronoi Health Analytics (https://www.voronoihealthanalytics.com). This software has been validated in the assessment of body composition for outcomes assessment using CT datasets and optimized to work effectively for both contrast and non-contrast images.^9,23–25 In brief, individual axial CT images are fed into a nonlinear image processing algorithm that provides a multi-slice, multi-tissue segmentation of the entire field of view and labeling of each axial slice by the vertebral body present on each slice. For this study, body composition metrics were all measured as a volume (cm³) encompassing the entire L1–L5 vertebral segments. The segmentations and slice labels generated by DAFS were reviewed and manually corrected, if necessary, by a subspecialized radiologist with 15 years of experience in abdominal imaging.

Supplementary Table 1 provides a description of body composition variables included from the DAFS segmentation output. A total of 11 variables in the volumes (cm³) and ratios of body composition metrics were extracted or synthesized from the DAFS output that included the inter/intramuscular adipose tissue (IMAT), visceral adipose tissue (VAT), subcutaneous adipose tissue (SAT), total adipose tissue (TAT), iliopsoas muscles, total skeletal muscle, and abdominal aortic calcifications (AAC). Supplementary Figure 1 demonstrates examples of axial CT images and corresponding DAFS outputs for 2 representative participants with similar demographics and tumor size and location.

Demographic and Clinical Parameters

The total OS was recorded as the time difference, in months, between the date of surgery and the last date the participant was recorded alive. The following demographic and clinical variables were obtained from the electronic medical record: (1) age at diagnosis, (2) sex, (3) race, (4) height, (5) weight, (6) BMI at the time of surgery, (7) history of diabetes mellitus (DM), (8) hypertension (HTN), (9) chronic kidney disease (CKD), (10) current or past history smoking, (11) methylation status of O⁶-methylguanine-DNA methyltransferase (MGMT) gene promoter—which, if methylated, confers favorable response to treatment,²⁶ (12) isocitrate dehydrogenase (IDH) mutation status, (13) extent of tumor resection including gross total resection, subtotal resection, or biopsy only, (14) adjuvant chemotherapy, (15) adjuvant radiation therapy, (16) last follow-up date, and (17) last recorded alive date.

As a standard tool for preoperative assessment of physical status, the American Society of Anesthesiologists (ASA) score measured according to the latest available update was also extracted from the I-MiND database, with ASA scores 1–2 representing a healthy to mildly limited individual and scores 3–4 representing severe systemic illness with or without threat to life.²⁷ Presence or absence of comorbidities like diabetes (DM), hypertension (HTN), and CKD was determined either based on the explicit mention of the condition or the use of related medications, such as glycemic control agents or antihypertensives upon the extensive review of electronic medical records.

Statistical Analysis

Statistical analyses were performed using the R software version 4.3.2 (https://www.r-project.org/). Between group comparisons, univariable, and multivariable Cox proportional hazards models were performed using the R packages dplyr, survival, and survminer. All continuous variables including demographic, clinical, and body composition metrics were tested for the assumption of normality of residuals using the Shapiro–Wilks test and described using median and interquartile range if the normality assumption was not met or mean ± standard deviation if the assumption was met. Between groups, comparisons of males and females were conducted using the Mann–Whitney U test or t-test for continuous variables and the Chi-square or Fisher’s exact for categorical variables.

The R package survival was used to perform Cox proportional hazards univariable regression models to identify variables that significantly predicted OS separately in the entire dataset, males, and females. Analyses were conducted using both continuous form of body composition variables and dichotomized variables based on population median values (50th percentile). A hazard ratio (HR) below 1 indicates a protective effect and better OS. Multivariable analysis through stepwise selection was performed using the step function from package StepReg to select variables for multivariable models by iteratively adding and removing variables based on statistical criteria. A P-value threshold of 0.25 was set for a variable to enter the model and P-value of 0.2 was used to keep a variable in the model. A final multivariable regression model was then fit using variables with P-value of <.1 with the addition of demographic and clinical covariates irrespective of their P-values. These covariates included: age at diagnosis, ASA score, MGMT promoter methylation status, receipt of adjuvant chemotherapy, receipt of adjuvant radiation therapy, extent of resection, history of HTN, DM, or CKD, and history of past or current smoking. Importantly, since the majority of I-MiND participants were recruited before IDH mutation status became standard for glial tumor classification, only 54.5% of our participants had this information available, among which only 1 participant had IDH mutation. Therefore, IDH mutation status was not used in any of the univariable or multivariable survival analyses.

All P-values were corrected for multiple comparisons using the Benjamini–Hochberg method.²⁸ A P-value of <.05 was considered statistically significant.

Composite Score Calculation for Outcomes

Functions xgb.train and xgb.importance from the R package xgboost were used to evaluate the relative importance of significant univariable body composition metrics in predicting GBM OS. Variables that were significant in the univariable analyses were used to craft separate predisposing and protective CT composite scores in males and females. We first used a tree-learning algorithm with 100 boosting rounds to partition the data into training (80%) and test (20%) datasets based on the OS status to generate feature importance matrices for these predisposing and protective variables separately for males and females.²⁹ Next, to construct this composite score, we normalized the selected body composition metrics and multiplied them by their respective normalized gains (ie, weights) from the feature importance matrices and subsequently summed them for each participant. Finally, we used Cox proportional hazards and pairwise log-rank test models to evaluate survival distribution between participants in different quartiles of the composite scores and the utility of 25th and 75th percentile as arbitrary cutoffs.

Results

Patient Demographic and Clinical Characteristics

Table 1 demonstrates the demographic and clinical characteristics of participants stratified by sex showing no difference in age, race, MGMT status, IDH mutation status, follow-up time, or treatment between males and females. Table 2 demonstrates differences in body composition metrics between sexes where males had larger absolute and fractional visceral adipose tissue (VAT and VAT/TAT) and larger skeletal muscle volume including iliopsoas muscle volume. Females, on the other hand, had a larger fractional SAT and skeletal muscle fat fraction (SAT/TAT and SKM fat fraction).

Before investigating body composition metrics, we used univariable Cox proportional hazard models to investigate GBM OS in relation to known predictors of GBM survival. As seen in Table 2 younger age, lower ASA scores (1 and 2), gross total resection of the tumor, receipt of adjuvant chemotherapy, and receipt of adjuvant radiation therapy were associated with improved OS in both males and females. Comorbidities such as DM, CKD, smoking, and overweight (BMI ≥ 25 kg/m²) or obesity (BMI ≥ 30 kg/m²) were not associated with significant differences in OS in our population, while there was a 2 times increased risk of death among females with hypertension (Table 2). Among participants with available IDH mutation status, all but one male participant had wild-type IDH. As a result, survival analyses based on IDH status were not performed.

Higher Fractional IMAT is an Independent Predictor of OS in GBM

Univariable analyses using body composition metrics as continuous variables revealed that the most consistent adipose tissue metric that was associated with significantly poorer GBM OS when all patients (ie, males and females) were considered as a higher absolute and fractional inter/intramuscular adipose tissue (IMAT and IMAT/TAT) (Figure 1 and Table 2 middle panel). Interestingly, when males and females were analyzed independently, this significant finding held in males, but not females. This pattern was also observed for IMAT/TAT when it was considered as a dichotomous variable with the median as a threshold. For example, fractional IMAT (IMAT/TAT) above the median was associated with an increased risk of death in both males (HR [95% CI]: 1.8 [1.2–2.8]; P = .004) and the entire dataset (HR [95% CI]: 1.5 [1.1–2.1]; P = .01) (Figure 1 and Table 2 bottom panel). Additionally, in males only, a higher fraction of subcutaneous adipose tissue (SAT/TAT) above population median, was predictive of better outcome (HR [95% CI]: 0.6 [0.4–0.9]; P = .04) (Table 2 bottom panel).

When considered in a multivariable model, higher fractional IMAT emerged as independently associated with poorer outcomes in the entire dataset and males (P = .007 and P = .001, respectively) (Table 3 and Figure 2). Expectedly, IMAT was not selected during the stepwise regression model in females and therefore not evaluated in the multivariable analyses. SAT/TAT was selected for multivariable analyses in the entire cohort and men but did not retain significance in either group (Table 3). Together, these findings indicate a robust biological effect of IMAT that is associated with worse OS in GBM, especially in males.

Skeletal Muscle Mass and Skeletal Muscle Fat Faction Predict Worse OS in GBM

Because of the significant results with IMAT, muscles were then evaluated using 3 different approaches: the volume of the iliopsoas muscles, the total volume of all abdominal muscles (including the iliopsoas), referred to as total SKM volume, and the SKM fat fraction. In univariable analyses using continuous variables, the most robust marker of survival was the SKM fat fraction which was predictive of worse OS in all patients, as well as in males and females separately (Table 2 middle panel) (HR [95% CI]: 2.8 [1.4–5.5]; P = .03). Previous studies have demonstrated a beneficial effect of muscle mass on brain tumor patients based upon masseter muscle measurements.¹¹ In our cohort, total skeletal muscle volume as well as iliopsoas volume were both associated with better OS in all patients, but not in males or females separately (Figure 1 and Table 2 middle panel).

The iliopsoas muscle volume was the only muscular health marker that survived in the stepwise regression model for the entire cohort and males and was therefore included in the multivariable regression model. However, it did not demonstrate independent significance when adjusted for clinical markers or survival-related comorbidities in either group (Table 3).

Abdominal Aortic Atherosclerosis is Associated With Worse Outcomes in Females With GBM

In addition to fat and muscle, we measured abdominal aortic calcification volume from L1 to L5, which is a marker of cardiovascular health. Using the group medians as cutoff, participants with higher volume of AAC had worse outcomes, both in the entire dataset (HR [95% CI]: 2 [1.04–2.7]; P < .001) and across sex groups (HR [95% CI]: 1.8 [1.2–2.7]; P = .009 in males and HR [95% CI]: 2.2 [1.3–3.8]; P = .002 in females) (Figure 1 and Table 2 bottom panel). AAC were selected for multivariable analyses in females but did not retain significance as an independent predictor of OS (Table 3).

A CT-Assigned Composite Predisposing Score Predicts GBM Survival

We used a tree-learning model to evaluate the relative importance of significant body composition metrics and to craft weighted sums of these metrics as composite predisposing and protective scores separately in the entire cohort, male and female. Variables that were significant in the univariable analyses were used in the composite predisposing and protective factor analyses. For the entire dataset, the composite predisposing score included the following variables: IMAT, IMAT/TAT, SKM fat fraction, and AAC (Table 2). The composite protective score incorporated iliopsoas and total skeletal muscle volumes, with SAT/TAT added specifically for males (Table 2).

In the entire dataset having a composite predisposing score above 75th percentile was associated with 1.9 times higher risk of death (HR [95% CI]: 1.9 [1.3–2.7], P < .001), while scoring below 25th percentile in composite protecting score—lower cumulative protective metrics—was associated with significantly improved OS (HR [95% CI]: 0.6 [0.4–0.9], P = .02). Those with lowest burden of the adverse factors (composite predisposing within first quartile) had 50% higher chance of survival compared to those in fourth quartile (HR [95% CI]: 0.5 [0.2–0.8], P < .001) (Figure 3A). Similarly, a cumulatively higher burden of protective factors—fourth quartile in composite protecting score—conferred 60% higher OS rates compared to those in the first quartile (HR [95% CI]: 0.4 [0.3–0.6], P < .001) (Figure 3B).

In males, having a cumulatively higher burden of predisposing factors—aka. composite predisposing score above 25th percentile—was associated with 90% lower OS in GBM (HR [95% CI]: 1.9 [1.3–2.7], P < .001). Additionally, scoring above the 75th percentile in the composite predisposing score in males was associated with significantly worse OS when compared to those with a composite predisposing score below the 25th percentile (Figure 3C). When the composite protective score was used, there was a statistically significant difference in OS of males, scoring above 25th percentile compared to those below this cutoff (HR [95% CI]: 0.5 [0.3–0.9], P < .03) (Figure 3D).

In females, OS was significantly worse in those with the higher composite predisposing score in the third and fourth quartiles who had a higher burden of predisposing scores—compared to those within the first quartile—lower burden of predisposing scores (Figure 3E). Similarly having a higher composite protective score in quartile 4 was associated with higher chances of survival compared with composite protective score in the first and second quartile (Figure 3F). On the other hand, scoring below 25th percentile in the composite protecting score conferred a 50% increased risk of death when compared to scores below 25th percentile (HR [95% CI]: 0.5 [0.3–0.8], P = .007) (Figure 3F).

Discussion

Genetic and acquired variations in cellular biology and metabolism dynamically shape individual differences in body composition throughout the lifetime. The idea behind using body composition metrics to predict cancer survival is based on the premise that the cellular mechanisms that govern differences in body composition could alter susceptibility to different cancer types, thus affecting cancer cell biology and survival outcomes. Of these, visceral adiposity and sarcopenia are associated with increased overall and cause-specific mortality across several cancer types.^10,30 While studies have linked obesity (measured by BMI and waist circumference) and temporalis muscle sarcopenia to worse OS in GBM,^4,7,11 there are a lack of data on how the distribution and composition of adipose tissue, skeletal muscle, and vascular biomarkers might relate to survival in these patients.

Quantitative imaging of the abdomen and pelvis enables simultaneous measurement of a variety of body composition metrics that cannot be achieved through brain imaging alone. Using participants with newly diagnosed GBM we demonstrate proof of concept for a novel paradigm to predict GBM outcomes using abdominal and pelvic body composition metrics obtained through CT imaging. We report for the first time, a relationship between worse OS in GBM and higher IMAT, lower skeletal muscle mass, and higher volume of AAC in the abdomen and pelvis. Notably, sex differences in the effects of these biomarkers on GBM patient survival may also be present. Using a tree-learning algorithm, we curated composite scores based on the relative importance of predisposing and protective body composition metrics that were able to predict OS in GBM. The utility of our proposed method is further supported by the ubiquitous use of CT imaging in clinical cancer care.

Obesity is not only a risk for all-cause mortality but is also related to higher cancer-related deaths, a relationship that is heavily modified by sex and distribution of adiposity.^31–33 Specifically, higher VAT associated with worse outcomes in cancers.³⁴ Conversely, increased SAT predicts better OS in gastrointestinal, lung, and renal cancers.³⁵ While we observed favorable outcomes in males with SAT/TAT above the 50th percentile, no significant difference in OS was found related to VAT or VAT/TAT. This could be due to the generally lower survival rates in GBM patients leaving little variability that diminishes the impact of these body composition metrics as a result (aka. survival bias).³⁶ Another possible explanation for lack of significant effect with dichotomized values involves the optimization of the cutoff point in our specific population as we used the 50th percentile as an arbitrary cutoff. Future studies can focus on optimizing age and sex-specific cutoffs for our population. Improved OS in males with higher relative SAT is in keeping with the male-dominant all-cause mortality risk related to higher VAT and low SAT reported in the literature.^30,34

We report an effect of total skeletal muscle volume and iliopsoas muscle volume as significant predictors of OS in GBM patients across the entire dataset. Both IMAT/TAT and SKM fat fraction are markers of fatty infiltration in skeletal muscle, normalized to total adipose tissue and skeletal muscle volume, respectively. This finding is consistent with reports linking sarcopenia—especially when combined with obesity, known as sarcopenic obesity—to worse disease-specific survival across various malignancies.^37,38 In GBM specifically, temporalis muscle thickness, a surrogate marker of sarcopenia, is independently related to improved OS.^11,39–41 Our findings, add on to the existing literature by demonstrating skeletal muscle fat fraction can be used as a novel biomarker of skeletal muscle health, outperforming skeletal muscle volume in GBM OS prediction. On the other hand, fatty infiltration of skeletal muscle is shown to predict worse survival in critically ill patients,^42,43 and is associated with comorbidities such as CKD and diabetes.^44,45 The finding that IMAT/TAT is an independent predictor of OS in GBM, even after adjusting for clinical predictors and comorbidities like CKD and DM, suggests a robust and independent influence on GBM survival.

An important finding in our study is the effect of AAC in predicting OS in all patients, and specifically in females with GBM as a continuous variable. AAC are nearly as strong as coronary artery calcifications in predicting cardiovascular events.⁴⁶ Our findings are supported by evidence from the literature that AAC is a robust predictor of mortality from cardiovascular disease, especially in females.^46–48 To the extent that cardiovascular events affect OS in GBM, the relationship between AAC and OS in GBM in our cohort is likely reflective of how much aortic calcifications are representative of the overall atherosclerotic disease burden within the body. In keeping with this theory, internal carotid artery calcium scores are shown to be related to OS in GBM.⁴⁹ The relationship between AAC and OS in our cohort could also to some extent, be indicative of the inflammatory underpinnings of atherosclerosis. Atherosclerosis itself is a chronic inflammatory disease and its progression is worsened by other chronic inflammatory conditions.⁵⁰ Inflammation is also associated with GBM tumorigenesis and progression, as demonstrated by the correlation between traumatic brain injury and GBM, along with the inflammatory characteristics of the tumor microenvironment.^51,52 Consequently, the observed correlation between AAC and GBM OS may stem from shared inflammatory underpinnings.

Finally, we propose a road map to build composite scores based on the relative importance of these metrics that provide a significant predictive value for GBM survival. This is a proof of concept that quantitative assessment of abdominal imaging biomarkers could provide insights into the overall health of the cancer patient and from there, predict outcomes. Given the life-long effects of diet, exercise, and metabolism on body composition, our findings have implications for lifestyle and dietary interventions in conjunction with conventional therapy to improve GBM outcomes.

Limitations

Our study is limited by its retrospective nature. Since the majority of I-MiND participants were recruited before IDH mutation status became standard for glial tumor classification, only 54.5% of our participants had this information available and we could not perform survival analyses based on this important genetic factor, nor we could use it in our multivariable analyses. Moreover, only the overall, and not cancer-specific, survival rate were available limiting any evaluation of the relationship between the body composition metrics and cancer-specific causes of death. A prospective, validation study considering cancer-specific causes of death and/or progression-free survival might help confirm these results.

Funding

This work was supported by National Institutes of Health [R21CA242221, R00CA218869, and UL1TR002345 to J.E.I.]; American Cancer Society [RSG-231156822-01-TBE to J.E.I.]; and Barnes-Jewish Hospital Foundation and Siteman Cancer Center (J.E.I.).

Conflict of interest statement

V.T.Y.C., D.M., K.P., and M.F.B. developed the software through Voronoi Health Analytics, Inc. that was used for this study. This had no influence on the data analysis, data interpretation, or manuscript preparation and submission at any point.

Authorship statement

Conceptualization (J.E.I.), Supervision (J.E.I.), Funding (M.R.C. and J.E.I.), Methodology (F.R., J.E., C.A.R., M.R.C., S.H.K., and J.E.I.), Software (V.T.Y.C., D.M., K.P., S.H.K., and M.F.B.), Formal Analysis (F.R., G.C., O.M., and J.E.I.), Investigation (F.R. and J.E.I.), Data curation (F.R., N.A., D.H.B., S.D., J.E., S.J., M.N., and S.H.K), Writing—draft (F.R. and J.E.I.), Writing—revision (F.R., G.C., O.M., N.A., D.H.B., V.T.Y.C., S.D., J.E., S.J., D.M., M.N., K.P., C.A.R., M.F.B., M.R.C., and J.E.I.).

Ethics approval

This study was performed in line with the principles of the Declaration of Helsinki. Approval was granted by the Institutional Review Board of Washington University School of Medicine in Saint Louis under IRB number 201706171.

Consent to participate

This was a retrospective study that was performed under a waiver of consent.

Consent to publish

Not applicable.

Data availability

The data supporting these findings are available on request to and review by the corresponding author and coauthors. The data are not publicly available due to privacy and ethical restrictions.

References