Ten commandments of interpreting and applying results of biomarker research.
Before we incorporate their conclusions into our clinical practice, we have to be aware of the limitations of these non-specific biomarkers. The specific limitations of CRP have been described in a review published in this Journal two years ago and will not be discussed further (4). What is probably more important is the conceptual framework through which we should interpret biomarker research. In this editorial, I describe my opinions on how we should analyse and interpret the results of biomarker research.
1. Clinical context can change the utility of a biomarker dramatically
Whether a biomarker is useful clinically depends heavily on the clinical context, or in epidemiological terms the pre-test probability (or prevalence) of the outcome in question before the test. The same biomarker can be very useful in a cohort under a certain clinical context, but not useful when applied to another cohort in a completely different clinical context. For a non-specific biomarker to be clinically useful, the pre-test probability (or prevalence of the outcome) should not be too high (Figure 1). For example, when a diagnostic or prognostic biomarker with a negative and positive likelihood ratio of two is applied to a cohort of patients who have a very high pre-test probability of an outcome or adverse event (e.g. >30%), the post-test probability will remain high (e.g. 15 to 60%) regardless of the results of the biomarker, making the biomarker not very useful in such context. Zhang et al showed that CRP may be more useful as a prognostic marker later in the course of critical illness than on intensive care unit (ICU) admission3. This observation is entirely predictable if we apply this principle and look at the included studies carefully. Almost all the negative studies used admission CRP as a predictor of ICU or hospital mortality that was relatively common. In contrast, using CRP concentrations near the time of ICU discharge to predict the relatively uncommon event--mortality after ICU discharge - was more useful (3,5).
2. P value does not dictate the importance of a biomarker
Kang et al demonstrated that lactate concentration on admission was an independent predictor of mortality after adjusting for other confounders (2). Although lactate concentration on admission was higher in the non-survivors (mean 3.7 mmol/l) than the survivors (mean 3.0 mmol/l, odds ratio 1.28, P=0.002), the difference would be difficult to notice in clinical practice and is unlikely to change the way clinicians manage their patients who have septic shock with liver dysfunction, or inform patients about their risk of mortality. From a predictive modelling perspective, it is the variability in the outcome explained by a biomarker that is more important than the decimal point of the P value (6,7). The chi-square contribution of a biomarker relative to other confounders in explaining the variability of the outcome will give us the clearest picture of the relative importance of a biomarker to other clinical information and this is clinically relevant.
3. Sensitive and specific biomarkers are infrequent
Because sensitive and specific biomarkers are
infrequent, clinicians must be aware of all the possible confounders that may affect the results of biomarkers. These may include a patient's age (8), drugs (e.g. beta-agonists increase serum lactate, corticosteroids suppress serum CRP) and organ function (e.g. liver dysfunction can reduce CRP production and lactate clearance) (9,10).
4. Timing of the test is pivotal in interpreting the results of the test
The appropriate timing of measuring a biomarker depends on the pre-test probability of the outcome and half-life of the biomarker. The timing of the test is pivotal in interpreting the results of the test because inappropriate timing may give misleading results. For example, CRP has a reasonably long half-life (17 hours) and it is common for CRP to increase to a peak at 24 to 48 hours after the initiation of an inflammatory event and improve after the patient has improved clinically. As such, the time lag between the clinical events and CRP concentrations should be considered when a CRP concentration is interpreted.
5. Biomarkers are useful only to good clinical practice
Biomarkers can only complement and cannot replace good clinical history and physical examination of patients. In fact, good clinical history and examination of patients are of paramount importance in assessing the pre-test probability of an outcome in question, and may dictate whether a nonspecific biomarker is indicated. Consequently, biomarkers should be used only as an adjunct after careful consideration of the pre-test probability (or prevalence) of an outcome. When the results of a biomarker are not consistent with other clinical information, the test should be repeated and the clinical condition of the patients should be reassessed.
6. An ideal biomarker should be able to change the management of a patient
A good biomarker should offer information beyond the usual clinical information. This implies that the results of a biomarker should change the post-test probability of an event by an amount that is clinically important (e.g. 30% vs 5%). By stratifying patients' risk for an outcome, clinicians may change the management of their patients and improve their outcome. For example, a high CRP concentration from a patient who has stayed in ICU for significant period of time and is close to ICU discharge may warrant exclusion of nosocomial infections.
7. Association does not confirm causation or reversibility
When a biomarker is associated with a worse prognosis of a disease, this does not imply reducing the concentrations of this biomarker will necessarily improve the prognosis of the patients. Many cytokines are increased in sepsis and high concentrations of many cytokines are associated with a higher mortality. Yet no therapy aimed at reducing one or more of these cytokines can reduce mortality.
8. Cross validation is needed for biomarker research
Similar to randomised controlled trials and other observational research, cross validation of biomarker studies are essential before we incorporate these biomarkers in our clinical practice. The differences in clinical context, confounders and timings of the biomarker are pivotal in summarising the results of biomarker studies. In this regard, by stratifying the timings of measuring CRP concentrations to ICU admission or near ICU discharge, the study by Zhang et al is a small step forward in resolving the conflicting results of the existing studies on prognostic value of CRP in critically ill patients.
9. Never analyse a biomarker as a categorical variable if you can analyse it as a continuous variable
Transforming a continuous variable into a categorical variable will lose statistical power and may mask the complex non-linear relationship between the biomarker concentrations and the outcome. Furthermore, choosing inappropriate cut-points to form different categories can also create false positive and negative results.
10. Justify the cost of measuring a biomarker
Although CRP and lactate concentrations are relatively inexpensive, the same cannot be said for many new biomarkers. A balance between cost and effectiveness of a biomarker is particularly important if the incremental benefit of a biomarker is small, or the number needed to test to avoid an adverse outcome is large.
Using these principles to interpret biomarker research, the two biomarker studies published in this issue have some strengths and weaknesses (2,3). Because the study by Kang et al is a small single-centre study, further validation of the prognostic significance of admission lactate concentrations is needed. Although the systematic review by Zhang et al has included a reasonable number of patients, significant heterogeneity between pooled studies exists, making the stratified meta-analysis prone to bias. Nevertheless, these two studies raise an important question of whether we should incorporate common biomarkers in prognostic models of critically ill patients. Currently almost all prognostic models, such as the Acute Physiology and Chronic Health Evaluation models, do not include admission lactate concentrations as a covariate. There is also no prognostic model at ICU discharge to assist clinicians in predicting ICU readmission or post-ICU mortality. However, it is possible that CRP or lactate concentrations may become one of the important covariates in the prognostic models of critically ill patients in the future.
(1.) Thielmann M, Massoudy P, Neuhauser M, Tsagakis K, Marggraf G, Kamler M et al. Prognostic value of preoperative cardiac troponin I in patients undergoing emergency coronary artery bypass surgery with non-ST-elevation or ST-elevation acute coronary syndromes. Circulation 2006; 114(1 Suppl): 1448-1453.
(2.) Kang YR, Um S-W, Koh W-J, Suh GY, Chung MP, Kim H et al. Initial lactate level and mortality in septic shock patients with hepatic dysfunction. Anaesth Intensive Care 2011; 39:5:862-867.
(3.) Zhang Z, Ni H. C-reactive protein as a predictor of mortality in critically ill patients: a meta-analysis and systematic review. Anaesth Intensive Care 2011; 39:5:854-861.
(4.) Ho KM, Lipman J. An update on C-reactive protein for intensivists. Anaesth Intensive Care 2009; 37:234-241.
(5.) Ho KM, Lee KY, Dobb GJ, Webb SA. C-reactive protein concentration as a predictor of in-hospital mortality after ICU discharge: a prospective cohort study. Intensive Care Med 2008; 34:481-487.
(6.) Ho KM, Leonard AD. Concentration-dependent effect of hypocalcaemia on mortality of patients with critical bleeding requiring massive transfusion: a cohort study. Anaesth Intensive Care 2011; 39:46-54.
(7.) Ho KM, Knuiman M, Finn J, Webb SA. Estimating long-term survival of critically ill patients: the PREDICT model. PLoS One 2008; 3:e3226.
(8.) Wibrow BA, Ho KM, Flexman JP, Keil AD, Kohrs DL. Eosinopenia as a diagnostic marker of bloodstream infection in hospitalised paediatric and adult patients: a case-control study. Anaesth Intensive Care 2011; 39:224-230.
(9.) Ho KM, Tan JA. Benefits and risks of corticosteroid prophylaxis in adult cardiac surgery: a dose-response meta-analysis. Circulation 2009; 119:1853-1866.
(10.) Ho KM. Hyperlactataemia induced by CVVHDF with low lactate bicarbonate-buffered solutions in patients with liver dysfunction. Nephrol Dial Transplant 2006; 21:1096-1099.
Figure 1: Effect of the prevalence (or pre-test probability) of an outcome on utility of a non-specific prognostic biomarker such as Creactive protein or lactate concentrations. A non-specific A very specific A prognostic prognostic biomarker prognostic biomarker biomarker is may be used to (or a gold standard unlikely to be predict the risk of prognostic test) is useful when the an outcome, needed. Non- prevalence of potentially changing specific biomarkers outcome is very subsequent clinical are relatively common management useless in this clinical context 20-30% 50-60% 100% Increasing prevalence of an outcome
K. M. Ho Department of Intensive Care Medicine, Royal Perth Hospital; and School of Population Health, University of Western Australia, Perth, Western Australia
|Printer friendly Cite/link Email Feedback|
|Publication:||Anaesthesia and Intensive Care|
|Date:||Sep 1, 2011|
|Previous Article:||High frequency oscillation in the adult intensive care unit. Nearly mainstream?|
|Next Article:||Hypocoagulability, hypercoagulability and the interpretation of 'normal' kaolin-activated thrombelastographs in the perioperative period.|