Limits...
A multi-analyte serum test for the detection of non-small cell lung cancer.

Farlow EC, Vercillo MS, Coon JS, Basu S, Kim AW, Faber LP, Warren WH, Bonomi P, Liptay MJ, Borgia JA - Br. J. Cancer (2010)

Bottom Line: The resulting biomarkers were fashioned into a classification algorithm and validated against serum from a second patient cohort.When tested against a second patient cohort, the panel successfully classified 75 of 88 patients.A high rate of false positives was observed within the cohort in which patients had non-neoplastic lung nodules, possibly as a consequence of the inflammatory nature of these conditions.

View Article: PubMed Central - PubMed

Affiliation: Department of General Surgery, Rush University Medical Center, 785 Jelke-Southcenter, 1750 W. Harrison Street, Chicago, IL 60612, USA.

ABSTRACT

Background: In this study, we appraised a wide assortment of biomarkers previously shown to have diagnostic or prognostic value for non-small cell lung cancer (NSCLC) with the intent of establishing a multi-analyte serum test capable of identifying patients with lung cancer.

Methods: Circulating levels of 47 biomarkers were evaluated against patient cohorts consisting of 90 NSCLC and 43 non-cancer controls using commercial immunoassays. Multivariate statistical methods were used on all biomarkers achieving statistical relevance to define an optimised panel of diagnostic biomarkers for NSCLC. The resulting biomarkers were fashioned into a classification algorithm and validated against serum from a second patient cohort.

Results: A total of 14 analytes achieved statistical relevance upon evaluation. Multivariate statistical methods then identified a panel of six biomarkers (tumour necrosis factor-α, CYFRA 21-1, interleukin-1ra, matrix metalloproteinase-2, monocyte chemotactic protein-1 and sE-selectin) as being the most efficacious for diagnosing early stage NSCLC. When tested against a second patient cohort, the panel successfully classified 75 of 88 patients.

Conclusions: Here, we report the development of a serum algorithm with high specificity for classifying patients with NSCLC against cohorts of various 'high-risk' individuals. A high rate of false positives was observed within the cohort in which patients had non-neoplastic lung nodules, possibly as a consequence of the inflammatory nature of these conditions.

Show MeSH

Related in: MedlinePlus

Box plots for the 6 biomarkers identified by the Random Forest algorithm. Box plots for the six selected biomarkers selected by the Random Forest analysis on the discovery cohort. Abscissa labels: 0=surgically resected, non-neoplastic nodules, 1=‘normal' controls, 2=Stage IA NSCLC, 3=Stage IB NSCLC and 4=Stages II and III (node positive) NSCLC. Notes: disease staging is based on pathologic stage; extreme values are not shown in the plots. Significance (Mann–Whitney Rank sum test) is shown with bars above boxes with a=P< 0.001; b=P<0.01 and c=P<0.05.
© Copyright Policy
Related In: Results  -  Collection


getmorefigures.php?uid=PMC2967050&req=5

fig1: Box plots for the 6 biomarkers identified by the Random Forest algorithm. Box plots for the six selected biomarkers selected by the Random Forest analysis on the discovery cohort. Abscissa labels: 0=surgically resected, non-neoplastic nodules, 1=‘normal' controls, 2=Stage IA NSCLC, 3=Stage IB NSCLC and 4=Stages II and III (node positive) NSCLC. Notes: disease staging is based on pathologic stage; extreme values are not shown in the plots. Significance (Mann–Whitney Rank sum test) is shown with bars above boxes with a=P< 0.001; b=P<0.01 and c=P<0.05.

Mentions: A panel of 6 biomarkers was selected from the 14 biomarkers meeting our inclusion criteria for statistical relevance using the Random Forests algorithm, as defined in Materials and methods section. The averaged out-of-bag ‘misclassification errors' as well as the AUC from the range of the 1000 trees of the Random Forest grown for each of their respective sub-panels are shown in Table 3. We found that the continued ‘focusing' of the panel from the 14 individual biomarkers to the six-analyte panel improved our ability to correctly classify patients relative to the pathological NSCLC status. However, after the fifth iteration, the AUC and associated sensitivity and specificity values (data not shown) decreased as the number of biomarkers decreased leading us to select the six-member panel from this fifth iteration as the most optimal combination for detecting NSCLC. Individual ‘box and whisker' plots are shown for these six biomarkers in Figure 1. Next, we defined a classification tree based on a sub-panel of six markers (consisting of TNF-α, CYFRA 21.1, IL-1ra, MMP-2, MCP-1, sE-selectin) selected from the Random Forest algorithm within the RPART software package to provide a convenient and useful algorithm for distinguishing NSCLC from benign controls. The classification tree resulting from this process is represented in Figure 2. This tree correctly classified 127 out of the 133 cases (a correct classification rate of 95%). The ROC curve for this classification tree is shown in Figure 3. Test performance characteristics for this panel boast a 97.9% AUC translating to 99% sensitivity and 95% specificity. As reported in our previous studies using this strategy (Borgia et al, 2009), we observed a substantial gain in our ability to screen for NSCLC when using the multi-analyte panel over any individual biomarker.


A multi-analyte serum test for the detection of non-small cell lung cancer.

Farlow EC, Vercillo MS, Coon JS, Basu S, Kim AW, Faber LP, Warren WH, Bonomi P, Liptay MJ, Borgia JA - Br. J. Cancer (2010)

Box plots for the 6 biomarkers identified by the Random Forest algorithm. Box plots for the six selected biomarkers selected by the Random Forest analysis on the discovery cohort. Abscissa labels: 0=surgically resected, non-neoplastic nodules, 1=‘normal' controls, 2=Stage IA NSCLC, 3=Stage IB NSCLC and 4=Stages II and III (node positive) NSCLC. Notes: disease staging is based on pathologic stage; extreme values are not shown in the plots. Significance (Mann–Whitney Rank sum test) is shown with bars above boxes with a=P< 0.001; b=P<0.01 and c=P<0.05.
© Copyright Policy
Related In: Results  -  Collection

Show All Figures
getmorefigures.php?uid=PMC2967050&req=5

fig1: Box plots for the 6 biomarkers identified by the Random Forest algorithm. Box plots for the six selected biomarkers selected by the Random Forest analysis on the discovery cohort. Abscissa labels: 0=surgically resected, non-neoplastic nodules, 1=‘normal' controls, 2=Stage IA NSCLC, 3=Stage IB NSCLC and 4=Stages II and III (node positive) NSCLC. Notes: disease staging is based on pathologic stage; extreme values are not shown in the plots. Significance (Mann–Whitney Rank sum test) is shown with bars above boxes with a=P< 0.001; b=P<0.01 and c=P<0.05.
Mentions: A panel of 6 biomarkers was selected from the 14 biomarkers meeting our inclusion criteria for statistical relevance using the Random Forests algorithm, as defined in Materials and methods section. The averaged out-of-bag ‘misclassification errors' as well as the AUC from the range of the 1000 trees of the Random Forest grown for each of their respective sub-panels are shown in Table 3. We found that the continued ‘focusing' of the panel from the 14 individual biomarkers to the six-analyte panel improved our ability to correctly classify patients relative to the pathological NSCLC status. However, after the fifth iteration, the AUC and associated sensitivity and specificity values (data not shown) decreased as the number of biomarkers decreased leading us to select the six-member panel from this fifth iteration as the most optimal combination for detecting NSCLC. Individual ‘box and whisker' plots are shown for these six biomarkers in Figure 1. Next, we defined a classification tree based on a sub-panel of six markers (consisting of TNF-α, CYFRA 21.1, IL-1ra, MMP-2, MCP-1, sE-selectin) selected from the Random Forest algorithm within the RPART software package to provide a convenient and useful algorithm for distinguishing NSCLC from benign controls. The classification tree resulting from this process is represented in Figure 2. This tree correctly classified 127 out of the 133 cases (a correct classification rate of 95%). The ROC curve for this classification tree is shown in Figure 3. Test performance characteristics for this panel boast a 97.9% AUC translating to 99% sensitivity and 95% specificity. As reported in our previous studies using this strategy (Borgia et al, 2009), we observed a substantial gain in our ability to screen for NSCLC when using the multi-analyte panel over any individual biomarker.

Bottom Line: The resulting biomarkers were fashioned into a classification algorithm and validated against serum from a second patient cohort.When tested against a second patient cohort, the panel successfully classified 75 of 88 patients.A high rate of false positives was observed within the cohort in which patients had non-neoplastic lung nodules, possibly as a consequence of the inflammatory nature of these conditions.

View Article: PubMed Central - PubMed

Affiliation: Department of General Surgery, Rush University Medical Center, 785 Jelke-Southcenter, 1750 W. Harrison Street, Chicago, IL 60612, USA.

ABSTRACT

Background: In this study, we appraised a wide assortment of biomarkers previously shown to have diagnostic or prognostic value for non-small cell lung cancer (NSCLC) with the intent of establishing a multi-analyte serum test capable of identifying patients with lung cancer.

Methods: Circulating levels of 47 biomarkers were evaluated against patient cohorts consisting of 90 NSCLC and 43 non-cancer controls using commercial immunoassays. Multivariate statistical methods were used on all biomarkers achieving statistical relevance to define an optimised panel of diagnostic biomarkers for NSCLC. The resulting biomarkers were fashioned into a classification algorithm and validated against serum from a second patient cohort.

Results: A total of 14 analytes achieved statistical relevance upon evaluation. Multivariate statistical methods then identified a panel of six biomarkers (tumour necrosis factor-α, CYFRA 21-1, interleukin-1ra, matrix metalloproteinase-2, monocyte chemotactic protein-1 and sE-selectin) as being the most efficacious for diagnosing early stage NSCLC. When tested against a second patient cohort, the panel successfully classified 75 of 88 patients.

Conclusions: Here, we report the development of a serum algorithm with high specificity for classifying patients with NSCLC against cohorts of various 'high-risk' individuals. A high rate of false positives was observed within the cohort in which patients had non-neoplastic lung nodules, possibly as a consequence of the inflammatory nature of these conditions.

Show MeSH
Related in: MedlinePlus