Limits...
An integrated approach to prognosis using protein microarrays and nonparametric methods.

Knickerbocker T, Chen JR, Thadhani R, MacBeath G - Mol. Syst. Biol. (2007)

Bottom Line: We found that molecular markers are not uniformly prognostic, but instead vary in their value depending on a combination of clinical variables.This may explain why reports in this area aiming to identify prognostic markers, without taking into account clinical variables, are either conflicting or show that markers have marginal prognostic value.Just as treatments are now being tailored to specific subsets of patients, our results show that prognosis can also benefit from a 'personalized' approach.

View Article: PubMed Central - PubMed

Affiliation: Department of Chemistry and Chemical Biology, Harvard University, Cambridge, MA 02138, USA.

ABSTRACT
Over the past several years, multivariate approaches have been developed that address the problem of disease diagnosis. Here, we report an integrated approach to the problem of prognosis that uses protein microarrays to measure a focused set of molecular markers and non-parametric methods to reveal non-linear relationships among these markers, clinical variables, and patient outcome. As proof-of-concept, we applied our approach to the prediction of early mortality in patients initiating kidney dialysis. We found that molecular markers are not uniformly prognostic, but instead vary in their value depending on a combination of clinical variables. This may explain why reports in this area aiming to identify prognostic markers, without taking into account clinical variables, are either conflicting or show that markers have marginal prognostic value. Just as treatments are now being tailored to specific subsets of patients, our results show that prognosis can also benefit from a 'personalized' approach.

Show MeSH
Predictors based on generalized additive models. The clinical and cytokine predictors assign patients probabilities of death with respect to the current study. (A) Scatter plot of 468 incident dialysis patients, colored according to outcome (red: died within 15 weeks; black: survived more than 15 weeks). (B) Contour plot of the scatter plot shown in panel A. The ‘x' indicates the data centroid and the closed curves contain, from inside out, 30, 50, and 70% of the patients, respectively. (C) Continuous predictor built using a combination of clinical and cytokine data. Numerical values are provided as Supplementary data. (D) Probability of death as a function of cytokine predictor, plotted at four different values of the clinical predictor. If the clinical predictor is low (0.3 or 0.5), cytokines do not provide substantial information. If the clinical predictor is high (0.7 or 0.9), however, cytokines provide further risk stratification. (E) Strategy for patient management. New patients are assigned a risk of mortality based on their clinical parameters. Those that fall in the medium-to-high risk category are further stratified based on their serum cytokine levels.
© Copyright Policy - open-access
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC1911205&req=5

f3: Predictors based on generalized additive models. The clinical and cytokine predictors assign patients probabilities of death with respect to the current study. (A) Scatter plot of 468 incident dialysis patients, colored according to outcome (red: died within 15 weeks; black: survived more than 15 weeks). (B) Contour plot of the scatter plot shown in panel A. The ‘x' indicates the data centroid and the closed curves contain, from inside out, 30, 50, and 70% of the patients, respectively. (C) Continuous predictor built using a combination of clinical and cytokine data. Numerical values are provided as Supplementary data. (D) Probability of death as a function of cytokine predictor, plotted at four different values of the clinical predictor. If the clinical predictor is low (0.3 or 0.5), cytokines do not provide substantial information. If the clinical predictor is high (0.7 or 0.9), however, cytokines provide further risk stratification. (E) Strategy for patient management. New patients are assigned a risk of mortality based on their clinical parameters. Those that fall in the medium-to-high risk category are further stratified based on their serum cytokine levels.

Mentions: Using only seven parameters, the combined model is able to separate patient outcomes effectively. While there are outliers in any human population, the centroids of the two patient populations are well separated (Figure 3B). Since the goal of our approach is to provide a continuous predictor of outcome, we estimated probability densities for death () and survival () using kernel methods. Kernel methods amount to convolving discrete data with a Gaussian window to obtain continuous estimates for densities. In other words, the density estimate at each location is a weighted average of all the discrete samples, with the weight of each sample decreasing with increase in distance between the sample and that location.


An integrated approach to prognosis using protein microarrays and nonparametric methods.

Knickerbocker T, Chen JR, Thadhani R, MacBeath G - Mol. Syst. Biol. (2007)

Predictors based on generalized additive models. The clinical and cytokine predictors assign patients probabilities of death with respect to the current study. (A) Scatter plot of 468 incident dialysis patients, colored according to outcome (red: died within 15 weeks; black: survived more than 15 weeks). (B) Contour plot of the scatter plot shown in panel A. The ‘x' indicates the data centroid and the closed curves contain, from inside out, 30, 50, and 70% of the patients, respectively. (C) Continuous predictor built using a combination of clinical and cytokine data. Numerical values are provided as Supplementary data. (D) Probability of death as a function of cytokine predictor, plotted at four different values of the clinical predictor. If the clinical predictor is low (0.3 or 0.5), cytokines do not provide substantial information. If the clinical predictor is high (0.7 or 0.9), however, cytokines provide further risk stratification. (E) Strategy for patient management. New patients are assigned a risk of mortality based on their clinical parameters. Those that fall in the medium-to-high risk category are further stratified based on their serum cytokine levels.
© Copyright Policy - open-access
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC1911205&req=5

f3: Predictors based on generalized additive models. The clinical and cytokine predictors assign patients probabilities of death with respect to the current study. (A) Scatter plot of 468 incident dialysis patients, colored according to outcome (red: died within 15 weeks; black: survived more than 15 weeks). (B) Contour plot of the scatter plot shown in panel A. The ‘x' indicates the data centroid and the closed curves contain, from inside out, 30, 50, and 70% of the patients, respectively. (C) Continuous predictor built using a combination of clinical and cytokine data. Numerical values are provided as Supplementary data. (D) Probability of death as a function of cytokine predictor, plotted at four different values of the clinical predictor. If the clinical predictor is low (0.3 or 0.5), cytokines do not provide substantial information. If the clinical predictor is high (0.7 or 0.9), however, cytokines provide further risk stratification. (E) Strategy for patient management. New patients are assigned a risk of mortality based on their clinical parameters. Those that fall in the medium-to-high risk category are further stratified based on their serum cytokine levels.
Mentions: Using only seven parameters, the combined model is able to separate patient outcomes effectively. While there are outliers in any human population, the centroids of the two patient populations are well separated (Figure 3B). Since the goal of our approach is to provide a continuous predictor of outcome, we estimated probability densities for death () and survival () using kernel methods. Kernel methods amount to convolving discrete data with a Gaussian window to obtain continuous estimates for densities. In other words, the density estimate at each location is a weighted average of all the discrete samples, with the weight of each sample decreasing with increase in distance between the sample and that location.

Bottom Line: We found that molecular markers are not uniformly prognostic, but instead vary in their value depending on a combination of clinical variables.This may explain why reports in this area aiming to identify prognostic markers, without taking into account clinical variables, are either conflicting or show that markers have marginal prognostic value.Just as treatments are now being tailored to specific subsets of patients, our results show that prognosis can also benefit from a 'personalized' approach.

View Article: PubMed Central - PubMed

Affiliation: Department of Chemistry and Chemical Biology, Harvard University, Cambridge, MA 02138, USA.

ABSTRACT
Over the past several years, multivariate approaches have been developed that address the problem of disease diagnosis. Here, we report an integrated approach to the problem of prognosis that uses protein microarrays to measure a focused set of molecular markers and non-parametric methods to reveal non-linear relationships among these markers, clinical variables, and patient outcome. As proof-of-concept, we applied our approach to the prediction of early mortality in patients initiating kidney dialysis. We found that molecular markers are not uniformly prognostic, but instead vary in their value depending on a combination of clinical variables. This may explain why reports in this area aiming to identify prognostic markers, without taking into account clinical variables, are either conflicting or show that markers have marginal prognostic value. Just as treatments are now being tailored to specific subsets of patients, our results show that prognosis can also benefit from a 'personalized' approach.

Show MeSH