Augmented intelligence decision tool for stroke prediction combines factors from CHA 2 DS 2 ‐VASc and the intermountain risk score for patients with atrial fibrillation

Introduction: CHA2DS2‐VASc and CHADS2 are computationally simple risk prediction tools used to guide anticoagulation decisions for stroke prophylaxis, but they have modest risk discrimination ability and use static dichotomous variables. The Intermountain Mortality Risk Scores (IMRS) are dynamic decision tools using standard clinical laboratory tests. This study derived new stroke prediction scores using variables from both CHA2DS2‐VASc and IMRS. Methods and Results: In outpatients with first atrial fibrillation (AF) diagnosis at the Intermountain Healthcare (females, n = 26 063 males, n = 29 807), sex‐specific "IMRS‐VASc" scores were derived using variables from CHA2DS2‐VASc, warfarin use, the complete blood count, and the comprehensive metabolic profile. Validation was performed in an independent Intermountain outpatient AF cohort (females, n = 11 021; males, n = 12 641). Stroke occurred among 3.1% and 3.1% of females and 2.3% and 2.5% of males in derivation and validation groups, respectively. IMRS‐VASc stratified stroke with similar ability in derivation (c‐statistics, females: c = 0.703, males: c = 0.697) and validation groups (females: c = 0.681, males: c = 0.685). CHA2DS2‐VASc (females: c = 0.581 and c = 0.605; males: c = 0.616 and c = 0.613 in derivation and validation, respectively) and CHADS2 (females: c = 0.581 and c = 0.608; males: c = 0.620 and c = 0.621 in derivation and validation, respectively) were substantially weaker stroke predictors. IMRS was the strongest mortality predictor (females: c = 0.783 and c = 0.782; males: c = 0.796 and c = 0.794 in derivation and validation, respectively) and all scores were poor at predicting bleeding risk. Conclusions: A temporally dynamic risk score, IMRS‐VASc was derived and validated as a predictor of stroke in outpatients with AF. IMRS‐VASc requires further validation and the evaluation of its use in guiding care and treatment decisions for patients with AF.

Keywords: clinical decision tool; IMRS; IMRS‐VASc; learning health system; risk stratification

Atrial fibrillation (AF) is associated with systemic comorbidities, conduction abnormalities, and a five‐fold increase in thromboembolism risk.[1] This arrhythmia affects about 4 million people in the United States and is predicted to double by the year 2030.[2] AF burden affects older adults, especially individuals with co‐existing comorbidities, such as obesity, hypertension, diabetes mellitus, coronary artery disease, and aging.[3] Approximately 50% of patients with paroxysmal AF will progress to either persistent AF or mortality within 10 years of diagnosis.[4] Currently, oral anticoagulation (OAC) is the cornerstone of thromboembolism prevention in patients with AF. Contemporary guidelines rely on CHADS2 or CHA2DS2‐VASc as stroke risk assessment tools in a clinical setting, where patients with a score of 2 or higher are designated as high‐risk and for whom OACs are recommended.[[5]] Although, the CHADS2 and CHA2DS2‐VASc schemas are simple and easy to memorize, their ability to predict stroke is limited and the scores can only increase over time.[[7]]

A more dynamic risk stratification tool that provides information regarding the need for OAC and other care or treatment decisions would be beneficial in clinical practice. Such a tool could consider factors that are not only commonly available across the continuum of care but that are dynamic—increasing or decreasing based on changes in a patient's health or disease management, or changes in lifestyle choices that may alter long‐term AF‐related comorbidities. The Intermountain Mortality Risk Scores (IMRS) are augmented intelligence decision tools composed using parameters from the complete blood count (CBC) and basic metabolic profile (BMP) and have been extensively validated to stratify morbidity, and mortality in a variety of patient populations in various geographic locations.[[11]] Augmented intelligence is the automated delivery of actionable information that is difficult or impossible for clinicians to ascertain on their own in the normal process of medical practice. IMRS tools are dynamic in that the score components may reveal both improvements and declines in health over repeated measurement, and demonstrate adaptive risk prediction as health, and treatment modalities modify risk over time.[13] IMRS was previously shown to provide additional risk prediction ability in subgroups defined by CHA2DS2‐VASc scores.[14]

This study evaluated the potential combination of risk factors from CHA2DS2‐VASc and IMRS by deriving a new clinical decision tool, IMRS‐VASc, and validating this set of new risk scores. It also evaluated whether the new scores are more predictive of stroke in the AF population than CHA2DS2‐VASc, CHADS2, and IMRS.

The objective of the study was to develop and validate sex‐specific aggregate risk metrics that provide greater ability to predict incident stroke among patients with AF than previously developed risk scores. This risk score derivation was proposed to be conducted using the commonly available and familiar components of the CHA2DS2‐VASc score and the IMRS, including the CBC and either the BMP or, when available, the comprehensive metabolic profile (CMP). Among this newly‐diagnosed AF population, a limited number of patients (11% of females and 9% of males; Table) had a history of warfarin use and this variable was also used in the modeling. The new sex‐specific models were named the IMRS‐VASc risk scores, drawing from the names of both source scores.

1 Baseline characteristics by sex in the validation and derivation groups

1 Abbreviations: ACE, Angiotensin converting enxyme ihibitor; ARB, angiotensin II receptor blockers; CAD, coronary artery disease; CCB, calcium channel blocker; IMRS, intermountain mortality risk scores; MI, myocardial infarction; PVD, peripheral vascular disease; TIA, transient ischemic attack.

2 * IMRS categorizations were made based on previously‐derived sex‐specific thresholds.[11]

This study evaluated N = 79 532 (n = 37 084 females, and n = 42 448 males), community‐dwelling patients with non‐valvular AF who were ≥18 years of age and clinically managed at Intermountain Healthcare. The study included patients seen from January 1990 to May 2013. The index date of the study for each patient was the date of their first AF diagnosis. Only patients with an available CBC or an available BMP/CMP, or both, from the time of or within 6 months before the initial AF diagnosis were included in the study, with the laboratory results closest to but not after the initial AF diagnosis being used in the case that multiple CBC or metabolic profile panels in that time window were available. IMRS‐VASc scores were derived in 70% of this population (n = 26 063 females; n = 29 807 males) to identify the significantly predictive variables and determine weightings (see below). These IMRS‐VASc scores were then validated using the other 30% of the population (n = 11 021 females; n = 12 641 males) that had been held aside for testing, whether the derived scores were predictive in an independent population. This study was approved by the Intermountain Healthcare Institutional Review Board as a data‐only study with a waiver of consent because it posed minimal risk to subjects.

The IMRS‐VASc derivation was generated based on the processes used for the derivation of IMRS previously, but the weightings derived here for IMRS‐VASc are completely different from those in the standalone IMRS that was previously calibrated for mortality.[11] The pragmatic parsimonious risk tool methodology was developed to use common, objective, standardized, inexpensive, electronically‐available variables such as the CBC and CMP that are utilized routinely in clinical practice, are familiar to clinicians, and can be used in calculations within the electronic health record. Other clinically meaningful variables may be used, such as genetic factors, if they meet the criteria of the IMRS derivation approach.

IMRS‐VASc also added the components of the CHA2DS2‐VASc score. The CHA2DS2‐VASc schema is based on the use of the following variables to calculate the score, with each variable adding one point to the score if present (unless otherwise noted): congestive heart failure, hypertension, age 75 years or older (add 2 to the score), diabetes mellitus, stroke or transient ischemic attack or thromboembolism (add 2 to the score), vascular disease (peripheral arterial disease, myocardial infarction, or aortic plaque), age 65 to 74 years, and female. CHA2DS2‐VASc was developed for the prediction of stroke risk and each incremental increase in a score revealed an increase in stroke risk, in a linear fashion.[[5], [15]]

The IMRS‐VASc scoring system was designed from sex‐specific multivariable Cox regression models in the derivation population that initially considered each factor from the CBC panel, each parameter of the CMP, each component used in CHA2DS2‐VASc, and warfarin use history. Descriptions of missing data for each variable and the variables that were included in the final IMRS‐VASc models are provided in Table S1. Comparative weightings of each variable included in the final model were derived from regression β‐coefficients of the multivariable Cox models that were assessed for the primary endpoint of stroke. Incorporation of variables in the Cox model was performed using stepwise selection and also forced entry to build the most predictive but parsimonious models. Every variable from IMRS and CHA2DS2‐VASc was included in regression modeling, while the final model included only those that added independent information regarding risk. Those factors whose weightings were zero for all values of the variable were excluded from the final model because they only added variation and not effect. The final model for each sex was constructed using forced variable entry of those factors that were significant (P ≤ .05) or near‐significant (P < .10) predictors of the dependent outcome or were confounders that caused more than a 10% change in the β‐coefficient of at least one other independent variable in the model.

Mortality was determined from electronic records of all 22 Intermountain hospitals, from electronic Utah death certificates, and from the social security death master file. Utah death certificates and hospital records provided causes of death for the majority of deceased patients. Incidence of stroke and bleeding were established using International Classification of Disease coding (ICD‐9 and ICD‐10) from the Intermountain electronic data warehouse that contains data from Intermountain hospitals and more than 180 clinics. Intermountain provides healthcare services to approximately two‐thirds of the population of Utah and southeastern Idaho. All patients had at least 5‐years of follow‐up. For survival analyses examining stroke outcomes, patients were censored at 2 years after entry into the study if they had no fatal or nonfatal stroke and were censored as non‐events before 2 years if the patient passed away due to a non‐stroke cause. Secondary time points were also examined at 3 and 5 years. IMRS‐VASc was categorized for clinical relevancy of analyses into tertiles, with the tertiles 1, 2, and 3 being considered to be low, moderate, and high‐risk groups, respectively.

A secondary bleeding outcome was also examined in the 2‐, 3‐, and 5‐year time periods in the derivation population using censoring at those years or censoring for mortality as a non‐event as in the stroke analyses. Survival analyses were also performed examining all‐cause mortality. For bleeding and mortality, IMRS‐VASc scores were applied to the analyses as derived for the stroke endpoint, and IMRS,[11] CHADS2, and CHA2DS2‐VASc scores were also all applied to the analyses of stroke, bleeding, and mortality based on their previously derived calculations. IMRS utilizes all CBC parameters (ie, hematocrit, white blood cell count, platelet count, mean platelet volume, red cell distribution width, mean corpuscular hemoglobin concentration, and mean platelet volume; excluded were red blood cell count, hemoglobin, and mean corpuscular hemoglobin which three were collinear with other CBC variables) and all BMP factors (ie, sodium, potassium, bicarbonate, calcium, glucose, and creatinine; excluded due to collinearity were chloride and blood urea nitrogen) and was categorized into three risk groups based on previously published thresholds,[11] and CHADS2 and CHA2DS2‐VASc were categorized into patient groups based on scores ≤1, 2, and ≥3.

The derived IMRS‐VASc scores were applied for validation purposes to the independent set of 30% of patients that were held aside at the beginning of the study. Cox regression was used with IMRS‐VASc as the independent variable and stroke as the dependent variable. Both categorical and continuous IMRS‐VASc variables were evaluated and hazard ratios (HR), 95% confidence intervals (CI), and P values were computed. Kaplan‐Meier survival curves were used to graphically represent the survival estimates by risk category and the log‐rank statistic was used to examine significance. The stroke endpoint was examined with up to 2‐, 3‐, and 5‐year events, and bleeding and mortality outcomes were also examined in those time frames.

To compare the predictive capability of IMRS‐VASc risk scores to other scores, the IMRS, CHADS2, and CHA2DS2‐VASc scores were also examined using the same methods as in the derivation population. These other risk scores were compared with the predictive ability of IMRS‐VASc using c‐statistics, predictive values, accuracy, and the net reclassification improvement (NRI) index to establish to what extent IMRS‐VASc offered greater predictive ability. Statistical significance for the NRI was evaluated using an asymptotic Z test of the null hypothesis that NRI = 0. A P value of P ≤ .05 was designated as significant in the validation population given the independent nature of that patient set.

Baseline characteristics of females and males in the derivation and validation populations are shown in Table. IMRS‐VASc had the following ranges of integer risk score values when categorized into tertiles: tertile 1: ≤5, tertile 2: 6 to 7, tertile 3: ≥8. In both the derivation and validation populations, IMRS‐VASc had a greater discrimination ability for stroke than IMRS, CHADS2, or CHA2DS2‐VASc based on evaluation of the c‐statistic (Table). It achieved a much higher accuracy for stroke prediction than even CHA2DS2‐VASc despite similar sensitivity and negative predictive value (Table S2). The NRI among females was 0.16 (P < .001) for IMRS‐VASc vs CHA2DS2‐VASc in validation group (0.22; P < .001 in derivation). Among males, the NRI was 0.08 (P < .001) for IMRS‐VASc vs CHA2DS2‐VASc (0.07; P < .001 in derivation).

2 C‐statistics with 95% confidence intervals for 2‐y outcomes of the primary (2‐y) IMRS‐VASc and other risk scores for discrimination of stroke, mortality, and bleeding risk

3 Abbreviation: IMRS, intermountain mortality risk scores.

IMRS‐VASc had stroke frequencies in the validation population that were very similar to those found in the derivation population and was highly predictive of the incidence of stroke at 2 years across categories of low, moderate, and high‐risk (Table). In contrast, high‐risk CHA2DS2‐VASc scores of 3 or greater had just 3.2% to 3.6% stroke (with values varying depending on whether the groups were female or male, or in the derivation or validation populations) compared to high‐risk IMRS‐VASc of 8 or greater that had stroke of 4.2% to 6.7% (Table), with sample sizes more evenly distributed by IMRS‐VASc between low, moderate, and high‐risk. Kaplan‐Meier survival curves displaying the separation in risk by IMRS‐VASc in the validation population are shown in Figure. Figure provides forest plots with HR and CI for stroke in the validation population for IMRS‐VASc and the other risk scores (derivation results are given in Figure S1).

3 Frequency of stroke by categories defined by (A) 2‐y IMRS‐VASc risk score tertiles, and (B) CHA 2 DS 2 ‐VASc score categories

3 Frequency of stroke by categories defined by (A) 2‐y IMRS‐VASc risk score tertiles, and (B) CHA 2 DS 2 ‐VASc score categories

4 Abbreviation: IMRS, intermountain mortality risk scores.

In secondary analyses of substrata defined by CHA2DS2‐VASc category, IMRS‐VASc was significantly predictive of stroke at 2 years in each category. For a CHA2DS2‐VASc score of ≤1: the c‐statistics were c = 0.707 (CI = 0.606, 0.809) for females and c = 0.684 (CI = 0.632, 0.735) for males; for 2: c = 0.702 (CI = 0.629, 0.774) for females and c = 0.662 (CI = 0.615, 0.709) for males; and for ≥ 3: c = 0.672 (CI = 0.654, 0.690) for females and c = 0.663 (CI = 0.640, 0.686) for males. Kaplan‐Meier survival curves for stroke outcomes using the primary 2‐year IMRS‐VASc scores in each of the three categories of CHA2DS2‐VASc are provided in Figure S2 for females and Figure S3 for males.

In secondary analyses of 2‐year mortality and 2‐year bleeding outcomes, IMRS‐VASc stratified risk of these outcomes in a highly statistically significant manner (Table). It was weaker, though, at discriminating mortality and bleeding risk than IMRS, CHADS2, and CHA2DS2‐VASc (Table). As expected, IMRS excelled at discriminating risk of mortality with c‐statistics of 0.78 to 0.80 (Table), and both CHADS2 and CHA2DS2‐VASc had intermediate abilities to discriminate mortality risk. None of the risk models did well at discriminating bleeding results (Table).

4 Frequency of mortality and bleeding by categories defined from tertiles of the 2‐y IMRS‐VASc risk scores

5 Abbreviation: IMRS, intermountain mortality risk scores.

The percentages of 3‐ and 5‐years stroke by 3‐ and 5‐years IMRS‐VASc scores are shown in Table S3. HRs and CIs for these scores and these time frames of stroke are provided in Figure. As representative results, the Kaplan‐Meier survival curves for 5‐year IMRS‐VASc in 5‐year stroke are shown in Figure S4. C‐statistics for 3‐ and 5‐years scores are provided in Table S4 for these secondary models. Finally, the association of 3‐ and 5‐year scores with mortality and bleeding outcomes are provided in Table S5. Sensitivity analyses are found in Table S6 for CHA2DS2‐VASc in warfarin‐naïve patients.

In adults with newly‐diagnosed non‐valvular AF, an augmented intelligence decision tool "IMRS‐VASc" was derived using component variables from CHA2DS2‐VASc and IMRS to predict 2‐year incident stroke after AF diagnosis. IMRS‐VASc was validated for 2‐year stroke prediction in an independent internal set of patients with AF, with higher c‐statistics than CHADS2, CHA2DS2‐VASc, and IMRS. IMRS‐VASc weightings of CHA2DS2‐VASc and IMRS variables were different than in the two legacy scores, with some components receiving weightings of zero (eg, heart failure, diabetes, and vascular disease). Secondary IMRS‐VASc scores for 3‐ and 5‐years stroke were also derived and validated, with the legacy IMRS having the best c‐statistics for mortality. Interestingly, the sex‐specific IMRS‐VASc revealed sex differences in risk prediction, with sodium, bicarbonate, bilirubin, mean corpuscular hemoglobin concentration, and red cell distribution width predicting stroke among females but not for males, while other factors discriminated risk only for males (ie, potassium, hematocrit, white blood cell count, and diabetes).

AF is a prevalent and complex disease that reflects systemic disease as well as local adverse atrial remodeling.[[5]] AF is a well‐recognized cause of stroke, and strokes related to the arrhythmia tend to be associated with worse morbidity and mortality.[18] Reduction in stroke risk in patients with AF is most commonly through OAC based on clinical guidelines.[[5]] Choices for OAC use are based upon a risk vs benefit analysis and that they as agents are also associated with elevated risks of severe bleeds and intracranial bleeds.

Guideline‐based care utilizes the CHA2DS2‐VASc score, as is recognized in both clinical guidelines as a critical tool in the development of a patient's care plan.[[5]] Recent European Society of Cardiology guidelines for the treatment of patients with AF utilize the threshold of a CHA2DS2‐VASc score of 1 or higher for the prescription of OAC,[5] while the American guidelines set the threshold at 2.[6] These thresholds provide a high sensitivity, and thus guideline‐driven care includes almost all patients with AF in the OAC treatment group and, potentially, many false positives in patients who do not (or do not yet) need OAC. Other challenges to the use of these scores include their limited utility for stroke prediction.[7] Next, the parameters used in the scores to determine risk level are essentially all binary static variables which, once present, do not change or resolve even if the patient's health improves.[[8], [19]] In addition, the score does not have the ability to differentiate the severity of diseases, such as a patient with poorly controlled high blood pressure with end‐organ injury and another with well treated high blood pressure.

Despite these limitations, the CHA2DS2‐VASc and CHADS2 scores remain the most successful and commonly utilized scores in medicine to guide the use of anticoagulation. The scores benefit from one crucial aspect that causes clinicians to overlook, or at least be willing to deal with, their limitations: they are usable. The scores are easy to compute without the need for devices and the data elements used in that computation are usually gathered in a medical history. A sacrifice is made in the ability to predict stroke accurately to improve simplicity and broad use.[20]

In contrast, IMRS‐VASc improves the ability to predict stroke and, as with CHA2DS2‐VASc, does not add time or complexity in the clinical setting. It does this if it is coded into the electronic health record, which calculates and delivers IMRS‐VASc to clinicians without clinician involvement, as with IMRS.[13] It also can be used dynamically in the management of a chronic disease over time through repeated measurement of CBC and CMP panels that can reveal both worsening as well as improving health. Its use of those laboratory panels also provides the measurement of risk information using low‐cost, ubiquitous, standardized, objective blood tests that in previous work have been shown to provide substantial risk prediction (eg, for mortality[11]). Its utilization of the CBC and CMP also encapsulates a similar degree of risk stratification ability as when risk prediction modeling includes many more variables because these laboratory parameters often account for the risk information of other variables. That is, the same degree of risk prediction can be achieved with fewer variables.[21] IMRS‐VASc did that here by capturing the risk information of the CHA2DS2‐VASc variables heart failure and vascular disease without requiring those variables to be measured.

As in previous risk score applications,[[13]] the clinical use of IMRS‐VASc may improve care through medication management and evaluation of the overall care plan. As with IMRS,[13] the electronic health record will calculate and deliver IMRS‐VASc without requiring physicians, advanced practice providers, or others to do anything different in their medical practice. A simple color of green, yellow, or red for a risk score category of low, moderate, or high‐risk will provide the augmented intelligence that is the actionable information the clinician needs to know how to modify care.[22] For example, the high‐risk group could be included as likely needing OAC with limited need for extensive workup, while the moderate risk group may be better managed by undergoing thoughtful and more in‐depth physician or advanced practice clinician examination to identify who is actually in need of OAC. Within this group, optimization of medical therapies and improved lifestyle changes may alter the moderate risk to a lower risk over time. The low‐risk group, although larger than the comparable CHA2DS2‐VASc group, could be considered a wait‐and‐see population where other standard care may be given but the prescription of OAC would be reconsidered later. This group also may be ideal in considering AF burden or pill‐in‐the‐pocket anticoagulation approaches that do not provide constant exposure to OAC. As such, as with IMRS,[[11]] other therapies, diagnostic modalities, and examinations could also be guided by the use of the risk scores.

An advantage of IMRS‐VASc is that it also performs well in females. Some component variables predicted stroke only in females, including sodium, bicarbonate, bilirubin, mean corpuscular hemoglobin concentration, and red cell distribution width, which may reveal important pathways related to stroke that apply most strongly to females and should be investigated further for possible sex differences in stroke risk. Further consideration clinically is needed for older females and other subpopulations that may be at elevated risk of bleeding or have other concerns that suggest a less aggressive approach to anticoagulation. Having improved ability to discern risk among females may aid in this effort.

This observational study utilized data previously collected in a prospective fashion among patients undergoing clinical care in a primary care setting. Due to its observational nature and the retrospectively chosen hypothesis, the population selection and data gathering may have resulted in unobserved or uncontrolled confounding. The study used disease classification coding to define prospective events and, although Intermountain serves more than two‐thirds of patients in the Utah and Southern Idaho geographies, some patients may have sought follow‐up care elsewhere. The study's approach to data collection may also suffer from missed events due to subclinical outcomes. Further, the time frame of patients studied may include secular trends due to changes in modalities of healthcare over time such as the use of new medications or devices, and evolving standards of care. The study's design may limit the generalizability of findings to other populations. Additional studies are needed to validate IMRS‐VASc in association with stroke in other populations with different racial compositions and distinct baseline stroke rates. If validated in other geographic locales, the evaluation of clinical use of the risk scores will require prospective implementation and evaluation.

Strengths of the study include a novel paradigm for risk scoring that uses practical, parsimonious models that may reduce the work of clinicians by employing the electronic health record for aiding physicians rather than adding burden in which they would need to gather data and calculate a score for each patient. Also, the use of several decades of data and the presence of stroke cases and of controls across each of those years should help balance any bias due to changing care processes. The primary components of the scores were common clinical laboratory tests that are ubiquitous in patient records and have been consistently defined and collected for over fifty years in medicine.

A newly‐derived risk score for predicting stroke, IMRS‐VASc was validated in association analyses with stroke outcomes among a second, independent set of outpatients with AF. It also was shown to have greater predictive ability than other risk scores, including standard risk scores used for the care of patients with AF. While it did predict mortality and bleeding outcomes, it was not as powerful at predicting those outcomes as IMRS (the score with the greatest predictive ability for mortality) or the other scores (the 4 scores studied herein had similarly limited predictive ability for bleeding). IMRS‐VASc is temporally dynamic due to the use of standard clinical laboratory factors that, when measured repeatedly over time, may improve or worsen and, thereby, reveal improvements in health due to changes in lifestyle and benefits of medical therapies, as well as declines in health due to disease progression or poor health behaviors. The IMRS‐VASc augmented intelligence tool should be evaluated further for validation in other populations and for potential use in guiding care and treatment decisions for patients with AF.

BDH and HTM are inventors of clinical decision tools that are licensed to CareCentra. BDH is the PI of grants funded by Intermountain Healthcare's Foundry Innovation program, the Intermountain Research and Medical Foundation, CareCentra, GlaxoSmithKline, and AstraZeneca for the development and/or clinical implementation of clinical decision tools. This study was funded by a grant from the Intermountain Healthcare Foundry Innovation Program. The funding source had no role in the design of the study, the data analysis, the interpretation of the findings, or the writing or publication of the study manuscript.

The authors declare that there is no conflict of interests.

Conception and design: BDH, HTM, and TJB; Analysis of data: HTM; Interpretation of data: BDH, VJ, HTM, KGG, and TJB; Drafting of the manuscript or revising it critically for important intellectual content: BDH, VJ, HTM, KGG, and TJB; Final approval of the manuscript submitted: BDH, VJ, HTM, KGG, and TJB.

GRAPH: Supporting information