Limits...
STatistically Assigned Response Criteria in Solid Tumors (STARCIST).

Bengtsson T, Sanabria-Bohorquez SM, McCarthy TJ, Binns DS, Hicks RJ, de Crespigny AJ - Cancer Imaging (2015)

Bottom Line: This finding is consistent with published results, but our data shows greater variability.Application of our method to NSCLC patients treated with erlotinib produces results distinct from those based on the EORTC criteria.Based on data presented here as well as previous repeatability studies, the proposed method has greater statistical power to detect a significant %-decrease on SUVmax compared to published criteria relying on symmetric thresholds.

View Article: PubMed Central - PubMed

Affiliation: Biostatistics, Genentech Inc, 1 DNA Way, South San Francisco, CA, 94080, USA. thomasgb@gene.com.

ABSTRACT

Background: Several reproducibility studies have established good test-retest reliability of FDG-PET in various oncology settings. However, these studies are based on relatively short inter-scan periods of 1-3 days while, in contrast, response assessments based on FDG-PET in early phase drug trials are typically made over an interval of 2-3 weeks during the first treatment cycle. With focus on longer, on-treatment scan intervals, we develop a data-driven approach to calculate baseline-specific cutoff values to determine patient-level changes in glucose uptake that are unlikely to be explained by random variability. Our method takes into account the statistical nature of natural fluctuations in SUV as well as potential bias effects.

Methods: To assess variability in SUV over clinically relevant scan intervals for clinical trials, we analyzed baseline and follow-up FDG-PET scans with a median scan interval of 21 days from 53 advanced stage cancer patients enrolled in a Phase 1 trial. The 53 patients received a sub-pharmacologic drug dose and the trial data is treated as a 'test-retest' data set. A simulation-based tool is presented which takes as input baseline lesion SUVmax values, the variance of spurious changes in SUVmax between scans, the desired Type I error rate, and outputs lesion and patient based cut-off values. Bias corrections are included to account for variations in tracer uptake time.

Results: In the training data, changes in SUVmax follow an approximately zero-mean Gaussian distribution with constant variance across levels of the baseline measurements. Because of constant variance, the coefficient of variation is a decreasing function of the magnitude of baseline SUVmax. This finding is consistent with published results, but our data shows greater variability. Application of our method to NSCLC patients treated with erlotinib produces results distinct from those based on the EORTC criteria. Based on data presented here as well as previous repeatability studies, the proposed method has greater statistical power to detect a significant %-decrease on SUVmax compared to published criteria relying on symmetric thresholds.

Conclusions: Defining patient-specific, baseline dependent cut-off values based on the () distribution of naturally occurring fluctuations in glucose uptake enable identification of statistically significant changes in SUVmax. For lower baseline values, the produced cutoff values are notably asymmetric with relatively large changes (e.g. >50 %) required for statistical significance. For use with prospectively defined endpoints, the developed method enables the use of one-armed trials to detect pharmacodynamic drug effects based on FDG-PET. The clinical importance of changes in SUVmax is likely to remain dependent on both tumor biology and the type of treatment.

No MeSH data available.


Related in: MedlinePlus

a) The scatterplot of SUVmax values highlights the strong correlation between baseline and follow-up values across lesions in our training dataset. The dashed regression line has a slope that is not significantly different from one. b) A histogram of the differences in lesion SUVmax between baseline and follow-up, which is approximated by a normal distribution of mean 0 and standard deviation 1.9 (solid curve). The dashed curve shows a t-distribution with 5° of freedom and scale parameter 1.49
© Copyright Policy - open-access
Related In: Results  -  Collection

License 1 - License 2
getmorefigures.php?uid=PMC4522098&req=5

Fig1: a) The scatterplot of SUVmax values highlights the strong correlation between baseline and follow-up values across lesions in our training dataset. The dashed regression line has a slope that is not significantly different from one. b) A histogram of the differences in lesion SUVmax between baseline and follow-up, which is approximated by a normal distribution of mean 0 and standard deviation 1.9 (solid curve). The dashed curve shows a t-distribution with 5° of freedom and scale parameter 1.49

Mentions: In the training data set, as shown in the scatter plot in Fig. 1a, there was a strong correlation (r = .86) between SUVmax values at screening and follow-up, with most lesion values falling close to the line of identity (solid line). The estimated regression line in this plot (dotted) had a slope of 1.05 (SE = .038) and was not significantly different from unity (p = .24). A linear mixed effects analysis showed that the mean change in SUVmax did not vary signifcantly varied across trial sites (or scanners), tumor location, or with uptake time (t2-t1). (The details of these analyses are presented in Appendix 1.) Plotting the raw differences in SUVmax between the two time points for all lesions yields the histogram shown in Fig. 1b. As seen, although the data are slightly more peaked than the Gaussian distribution, the differences in SUVmax values are symmetrically distributed and are reasonably well approximated by a normal distribution with parameters set to the sample mean (−0.23, p < .16) and standard deviation (1.91) (solid curve). With parameters obtained by maximum likelihood (ML), this plot also shows a t-density with 5° of freedom and a scale parameter of 1.49 (dashed curve).Fig. 1


STatistically Assigned Response Criteria in Solid Tumors (STARCIST).

Bengtsson T, Sanabria-Bohorquez SM, McCarthy TJ, Binns DS, Hicks RJ, de Crespigny AJ - Cancer Imaging (2015)

a) The scatterplot of SUVmax values highlights the strong correlation between baseline and follow-up values across lesions in our training dataset. The dashed regression line has a slope that is not significantly different from one. b) A histogram of the differences in lesion SUVmax between baseline and follow-up, which is approximated by a normal distribution of mean 0 and standard deviation 1.9 (solid curve). The dashed curve shows a t-distribution with 5° of freedom and scale parameter 1.49
© Copyright Policy - open-access
Related In: Results  -  Collection

License 1 - License 2
Show All Figures
getmorefigures.php?uid=PMC4522098&req=5

Fig1: a) The scatterplot of SUVmax values highlights the strong correlation between baseline and follow-up values across lesions in our training dataset. The dashed regression line has a slope that is not significantly different from one. b) A histogram of the differences in lesion SUVmax between baseline and follow-up, which is approximated by a normal distribution of mean 0 and standard deviation 1.9 (solid curve). The dashed curve shows a t-distribution with 5° of freedom and scale parameter 1.49
Mentions: In the training data set, as shown in the scatter plot in Fig. 1a, there was a strong correlation (r = .86) between SUVmax values at screening and follow-up, with most lesion values falling close to the line of identity (solid line). The estimated regression line in this plot (dotted) had a slope of 1.05 (SE = .038) and was not significantly different from unity (p = .24). A linear mixed effects analysis showed that the mean change in SUVmax did not vary signifcantly varied across trial sites (or scanners), tumor location, or with uptake time (t2-t1). (The details of these analyses are presented in Appendix 1.) Plotting the raw differences in SUVmax between the two time points for all lesions yields the histogram shown in Fig. 1b. As seen, although the data are slightly more peaked than the Gaussian distribution, the differences in SUVmax values are symmetrically distributed and are reasonably well approximated by a normal distribution with parameters set to the sample mean (−0.23, p < .16) and standard deviation (1.91) (solid curve). With parameters obtained by maximum likelihood (ML), this plot also shows a t-density with 5° of freedom and a scale parameter of 1.49 (dashed curve).Fig. 1

Bottom Line: This finding is consistent with published results, but our data shows greater variability.Application of our method to NSCLC patients treated with erlotinib produces results distinct from those based on the EORTC criteria.Based on data presented here as well as previous repeatability studies, the proposed method has greater statistical power to detect a significant %-decrease on SUVmax compared to published criteria relying on symmetric thresholds.

View Article: PubMed Central - PubMed

Affiliation: Biostatistics, Genentech Inc, 1 DNA Way, South San Francisco, CA, 94080, USA. thomasgb@gene.com.

ABSTRACT

Background: Several reproducibility studies have established good test-retest reliability of FDG-PET in various oncology settings. However, these studies are based on relatively short inter-scan periods of 1-3 days while, in contrast, response assessments based on FDG-PET in early phase drug trials are typically made over an interval of 2-3 weeks during the first treatment cycle. With focus on longer, on-treatment scan intervals, we develop a data-driven approach to calculate baseline-specific cutoff values to determine patient-level changes in glucose uptake that are unlikely to be explained by random variability. Our method takes into account the statistical nature of natural fluctuations in SUV as well as potential bias effects.

Methods: To assess variability in SUV over clinically relevant scan intervals for clinical trials, we analyzed baseline and follow-up FDG-PET scans with a median scan interval of 21 days from 53 advanced stage cancer patients enrolled in a Phase 1 trial. The 53 patients received a sub-pharmacologic drug dose and the trial data is treated as a 'test-retest' data set. A simulation-based tool is presented which takes as input baseline lesion SUVmax values, the variance of spurious changes in SUVmax between scans, the desired Type I error rate, and outputs lesion and patient based cut-off values. Bias corrections are included to account for variations in tracer uptake time.

Results: In the training data, changes in SUVmax follow an approximately zero-mean Gaussian distribution with constant variance across levels of the baseline measurements. Because of constant variance, the coefficient of variation is a decreasing function of the magnitude of baseline SUVmax. This finding is consistent with published results, but our data shows greater variability. Application of our method to NSCLC patients treated with erlotinib produces results distinct from those based on the EORTC criteria. Based on data presented here as well as previous repeatability studies, the proposed method has greater statistical power to detect a significant %-decrease on SUVmax compared to published criteria relying on symmetric thresholds.

Conclusions: Defining patient-specific, baseline dependent cut-off values based on the () distribution of naturally occurring fluctuations in glucose uptake enable identification of statistically significant changes in SUVmax. For lower baseline values, the produced cutoff values are notably asymmetric with relatively large changes (e.g. >50 %) required for statistical significance. For use with prospectively defined endpoints, the developed method enables the use of one-armed trials to detect pharmacodynamic drug effects based on FDG-PET. The clinical importance of changes in SUVmax is likely to remain dependent on both tumor biology and the type of treatment.

No MeSH data available.


Related in: MedlinePlus