By Sagar S. Mungekar, MD
Faculty Peer Reviewed
President Obama’s Affordable Care Act was in the news again this week, with proponents and opponents both voicing their opinions on how the Supreme Court should rule later this month. In March the Court considered the notion of severability; that is, whether certain provisions of the law could stand without the individual mandate or whether the entire law would have to go. The core provision of the act—the so-called individual mandate—would require nearly everyone to possess insurance, either from an employer or from a direct purchase. The decision of which and how much insurance coverage to purchase may be driven by patients’ own conceptions or predictions of how sick they are and whether they would be high- or low-utilizers of the medical system.
We saw several risk-stratification tools in the journals this week, many of which sought to assign patients a number or score to denote a better or worse prognosis. In this PrimeCuts, we will briefly review these prediction tools, as well as the study designs used to develop them.
The first one is a cohort study of heart failure (HF) patients.[1] Indeed, decompensated HF patients are hospitalized often, and during the initial ER presentation, the decision to admit or discharge is often difficult. The authors sought to derive a model to predict acute mortality in HF patients. The data were abstracted from medical charts of some 12,591 Canadian patients who presented to the ER with symptoms of heart failure exacerbation. They excluded those who arrived to the ER as transfers from other hospitals, those with DNR orders, and those who were dependent on dialysis. The patients were divided into derivation and validation cohorts, with approximately two-thirds hospitalized and one-third discharged in each. Using multivariate regression analysis, the authors identified ten variables, weighted each one, and developed the Emergency Heart Failure Mortality Risk Grade (EHMRG), a score centered around 0. Both cohorts behaved similarly with each 20-point rise in EHMRG causing the odds of 7-day death to increase by 41% (OR 1.41; 95% CI:1.34-1.48) and 39% (OR 1.39; 95% CI:1.32-1.47) in the derivation and validation groups respectively.
This sort of cohort study is beneficial at looking at all-comers, as older studies excluded those discharged directly from the ER. However, several factors should be noted before using the EHMRG at one’s own institution. The algorithm adds 60 points for those transported by EMS, regardless of whether such emergent transportation was necessary. Patients in urban centers with higher EMS utilization may have falsely elevated scores. Another component of the score is the absolute creatinine concentration, not the change from the patient’s baseline. While the predictive value of this variable was validated, it should be noted that the cohorts had mean creatinine concentrations of ~1.4 +/- 0.8 mg/dL. This relatively narrow distribution is possibly reflective of patient homogeneity, making medical centers with racially diverse populations weigh this variable differently.
With that last thought in mind, the next article in the same issue caught my eye: “Estimating Equations for Glomerular Filtration Rate in the Era of Creatinine Standardization”.[2] These authors performed a systematic review of studies comparing creatinine-based GFR equations to a reference—again a question of whether a single algorithm could be used to determine the GFR across different populations. Twenty studies were included in this analysis. Neither the Chronic Kidney Disease Epidemiology Collaboration (CKD- EPI) nor the Modification of Diet in Renal Disease (MDRD) study equations worked well for all patients. The former performed better at higher GFRs and the latter at lower. One should note that, in general, these equations were fitted for populations in North America, Europe, and Australia but required modifications for Asian and African populations.
From my own experiences, I have noticed that these calculations can vary significantly: while caring for a patient recently, three members of our team used three different equations and calculated three widely ranging estimates of the GFR, varying by almost a factor of two. Determining the GFR is clearly important when dosing medications and risk-stratifying, but in hospitals with diverse populations, the estimate provided by clinical laboratories as part of the basic metabolic panel may be inaccurate. Physicians may have to “personalize” these equations to estimate the GFR and document the method to maintain consistency.
The idea of tailoring calculations by patient demographics is clearly not a new one. If we flip to the British Medical Journal, we see a new adjustment validation for an old test: the concentration of D-dimers to rule out deep vein thrombosis (DVT).[3] The authors performed a retrospective, cross-sectional, diagnostic analysis to determine whether the conventional cutoff of 500 μg/L could be raised for elderly populations. The data came from just two accuracy studies, but included 1374 patients, comparing initial D-dimer values to evidence of DVTs by ultrasonography. The age-dependent cutoff allowed the researchers to exclude 5.7% more patients (absolute increase [95% CI:4.1-7.8%]). The fixed cutoff of 750 μg/L in patients 60 years of age and older fared similarly, excluding 5.4% more patients (absolute increase [95% CI:3.8-7.4%]). Notably, this cutoff threshold did not miss any extra cases.
The role of measuring the D-dimer concentration is often questioned; this study provides validation that in elderly patients with a measurement above the usual dichotomous cutoff of 500 μg/L could still be useful in ruling out DVTs.
Jumping to the Archives, online readers found another prediction tool: favorable neurological outcomes in survivors of in-hospital cardiac arrest.[4] This was also a cohort study including ~43,000 patients who survived an in-hospital arrest outside of the ER, OR, or procedure areas. The primary outcome was favorable neurological survival to discharge. The predictive variables included age, arrest rhythm, time to defibrillation, pre-arrest cognitive state (a retrospective estimation post-arrest), level of monitoring, hospital location, duration of resuscitation, and comorbid factors, which the authors compiled into the Cardiac Arrest Survival Postresuscitation In-hospital (CASPRI) score. The nomogram comparing the score and the percentage of neurological survival appears to be sigmoidal, with survival’s precipitously dropping as scores rise from 0, and higher scores’ (20-40) having less than 20% favorable outcomes.
While this model performed well in the validation group, the utility of the score is narrow. As noted in the invited commentary, the only factor physicians may be able to control is how and whether the patient was monitored.[5] Furthermore, since the score must be calculated after the arrest and survival, the CASPRI score is not designed to predict the outcome of the arrest itself; rather, it first predicts a favorable neurological outcome at discharge, which must then be extrapolated to the resultant subsequent quality of life.
The last study I’ll highlight was in The Lancet, which I found interesting for its content and its type of analysis.[6] The authors used a Mendelian randomization. This method is relatively new (first described in 1991) and aims to approximate a randomized controlled trial from data obtained observationally. The theory is fairly straightforward: if we assume that genetic polymorphisms assort randomly and we seek to test whether a particular gene or downstream product affects a disease, then inheriting a variation that affects this biochemical pathway should also affect the presence or extent of the disease. This method takes advantage of the inherent genetic randomization that occurs in nature without the need to randomize individuals for a study a priori.
In this study, the authors used single nucleotide polymorphisms (SNPs) associated with plasma HDL concentration to test whether elevated levels protect against myocardial infarction (MI). Some 50,000 participants were studied, of which over 4,000 suffered an MI. They validated both the population and the method by testing a known correlation: high LDL increases the risk of MI. They subsequently tested a specific SNP in the endothelial lipase gene that raises HDL cholesterol without changing other lipid levels. The rise of HDL by virtue of this SNP should have been expected to decrease the risk of MI 13% (95% CI:9-16%), but surprisingly, there was no association between this SNP and MI (OR 0.99 [95% CI:0.88-1.11]). Similarly, 14 other SNPs associated with a rise in HDL were not associated with a decrease in MI risk. These data challenge the dogma that HDL is the good cholesterol that “helps”.
Mendelian randomization obviously has its limitations; not all SNP distributions are random if genes are linked, and many genes are pleiotropic. Furthermore, while the authors were fortunate to have found a specific SNP that affects HDL without affecting several other serum and clinical measures, a candidate gene or polymorphism is not always available. Still, as the human genome repository grows, we grow closer to testing genotypes and markers as surrogates for the patients themselves. These five articles have outlined that while reducing a patient to a numerical score can simplify complex cases, the idea of personalization is still relevant as confounding factors abound. Molecular genetic fingerprinting such as SNPs may allow physicians in the coming years to offer patients even more tailored advice as they seek to assign risk and prognosticate.
Dr. Sagar Mungekar is a 1st year resident at NYU Langone Medical Center
Peer reviewed by Ishmeal Bradley, MD, section editor, Clinical Correlations
Image courtesy of Wikimedia Commons
References:
1. Lee DS, et al. Prediction of Heart Failure Mortality in Emergent Care. Ann Intern Med 2012;156:767-775. http://annals.org/article.aspx?articleid=1170879
2. Earley A, Miskulin DM, Lamb EJ, Levey, AS, Uhlig, K. Estimating Equations for Glomerular Filtration Rate in the Era of Creatinine Standardization. Ann Intern Med 2012;156:785-795. http://annals.org/content/156/11/785.full
3. Schouten HJ, et al. Validation of two age dependent D-dimer cut-off values for exclusion of deep vein thrombosis in suspected elderly patients in primary care. BMJ 2012;344:e2985. http://www.bmj.com/content/344/bmj.e2985
4. Char PS, et al. A Validated Prediction Tool for Initial Survivors of In-Hospital Cardiac Arrest. Arch Intern Med 2012; [Published online]: e2-e7. http://archinte.jamanetwork.com/article.aspx?articleID=1162169
5. Huszti E, Nichol G. Prediction of “mostly dead” vs “all dead” after in-hospital cardiac arrest. Comment on “A validated prediction tool for initial survivors of in-hospital cardiac arrest.” Arch Intern Med 2012. [Published online]: e8. http://archinte.jamanetwork.com/article.aspx?articleid=1162172.
6. Voight, BF, et al. Plasma HDL cholesterol and risk of myocardial infarction: a Mendelian randomisation study. The Lancet, 2012; Epub May 17, 2012, DOI:10.1016/S0140-6736(12)60312-2. http://www.thelancet.com/journals/lancet/article/PIIS0140-6736(12)60312-2/fulltext