Limits...
Estimation and external validation of a new prognostic model for predicting recurrence-free survival for early breast cancer patients in the UK.

Campbell HE, Gray AM, Harris AL, Briggs AH, Taylor MA - Br. J. Cancer (2010)

Bottom Line: Online.When compared with the NPI, the model was able to better discriminate between women with excellent and good prognoses, and it did not overestimate 10-year recurrence-free survival to the extent observed for Adjuvant!

View Article: PubMed Central - PubMed

Affiliation: Health Economics Research Centre, Department of Public Health, University of Oxford, Old Road Campus, Headington, Oxford OX3 7LF, UK. helen.campbell@dphpc.ox.ac.uk

ABSTRACT

Background: We aimed to estimate and externally validate a new UK-specific prognostic model for predicting the long-term risk of a first recurrent event (local recurrence, metastatic recurrence, or second primary breast cancer) in women diagnosed with early breast cancer.

Methods: Using data on the prognostic characteristics and outcomes of 1844 women treated for early breast cancer at the Churchill Hospital in Oxford, parametric regression-based survival analysis was used to estimate a prognostic model for recurrence-free survival. The model, which incorporated established prognostic factors, was externally validated using independent data. Its performance was compared with that of the Nottingham Prognostic Index (NPI) and Adjuvant! Online.

Results: The number of positive axillary lymph nodes, tumour grade, tumour size and patient age were strong predictors of recurrence. Oestrogen receptor (ER) positivity was shown to afford a moderate protective effect. The model was able to separate patients into distinct prognostic groups, and predicted well at the patient level, mean Brier Accuracy Score=0.17 (s.e.=0.004) and overall C=0.745 (95% CI, 0.717-0.773). Its performance diminished only slightly when applied to a second independent data set. When compared with the NPI, the model was able to better discriminate between women with excellent and good prognoses, and it did not overestimate 10-year recurrence-free survival to the extent observed for Adjuvant! Online.

Conclusion: The model estimated here predicts well at both the individual patient and group levels, and appears transportable to patients treated at other UK hospitals. Its parametric form permits long-term extrapolation giving it an advantage over other prognostic tools currently in use. A simple point scoring system and reference table allow 5-, 10-, and 15-year predictions from the model to be quickly and easily estimated. The model is also available to download as an interactive computer program.

Show MeSH

Related in: MedlinePlus

Recurrence hazard functions – for ‘average' Churchill Hospital patient (A) and as predicted by the model (B). For comparability with B, which is evaluated at the mean of the covariates, the hazard contributions in A (estimated as the change in the Nelson–Aalen cumulative hazard between time ti and time ti−1, and smoothed by STATA's default kernel density function) were also calculated at the mean values of the model covariates using STATA's adjustfor(varlist) command.
© Copyright Policy
Related In: Results  -  Collection


getmorefigures.php?uid=PMC2966633&req=5

fig1: Recurrence hazard functions – for ‘average' Churchill Hospital patient (A) and as predicted by the model (B). For comparability with B, which is evaluated at the mean of the covariates, the hazard contributions in A (estimated as the change in the Nelson–Aalen cumulative hazard between time ti and time ti−1, and smoothed by STATA's default kernel density function) were also calculated at the mean values of the model covariates using STATA's adjustfor(varlist) command.

Mentions: Model estimation was also conducted in STATA. A parametric regression-based survival model was estimated on time from initial surgery to a first recurrent event or censoring (patients were censored when they died from causes unrelated to breast cancer without recurrence being first recorded (n=111/1844) or were lost to follow-up, again without any previous diagnosis of recurrence). Parametric models assume survival times and consequently the hazard function follow a particular distribution. Based on the hazard of recurrence for the average woman in the Churchill data set (Figure 1A), we considered models using log-normal, log-logistic, and gamma distributions. All three can model a hazard function which changes direction with time and are supported by the accelerated failure time (AFT) class of model (Bradburn et al, 2003). Accelerated failure time models differ from the more established proportional hazards (PH) models in that rather than estimating a baseline hazard function, AFT models instead estimate a baseline survival function. The estimated covariates are then multiplicative with respect to survival time. When exponentiated (transformed using the formula ex, where e=2.71828 and x is the coefficient value), the coefficients in an AFT model are termed as time ratios. A time ratio greater (less) than one indicates that a covariate increases (decreases) time to recurrence, stretching (shrinking) the baseline survivor function along the time axis.


Estimation and external validation of a new prognostic model for predicting recurrence-free survival for early breast cancer patients in the UK.

Campbell HE, Gray AM, Harris AL, Briggs AH, Taylor MA - Br. J. Cancer (2010)

Recurrence hazard functions – for ‘average' Churchill Hospital patient (A) and as predicted by the model (B). For comparability with B, which is evaluated at the mean of the covariates, the hazard contributions in A (estimated as the change in the Nelson–Aalen cumulative hazard between time ti and time ti−1, and smoothed by STATA's default kernel density function) were also calculated at the mean values of the model covariates using STATA's adjustfor(varlist) command.
© Copyright Policy
Related In: Results  -  Collection

Show All Figures
getmorefigures.php?uid=PMC2966633&req=5

fig1: Recurrence hazard functions – for ‘average' Churchill Hospital patient (A) and as predicted by the model (B). For comparability with B, which is evaluated at the mean of the covariates, the hazard contributions in A (estimated as the change in the Nelson–Aalen cumulative hazard between time ti and time ti−1, and smoothed by STATA's default kernel density function) were also calculated at the mean values of the model covariates using STATA's adjustfor(varlist) command.
Mentions: Model estimation was also conducted in STATA. A parametric regression-based survival model was estimated on time from initial surgery to a first recurrent event or censoring (patients were censored when they died from causes unrelated to breast cancer without recurrence being first recorded (n=111/1844) or were lost to follow-up, again without any previous diagnosis of recurrence). Parametric models assume survival times and consequently the hazard function follow a particular distribution. Based on the hazard of recurrence for the average woman in the Churchill data set (Figure 1A), we considered models using log-normal, log-logistic, and gamma distributions. All three can model a hazard function which changes direction with time and are supported by the accelerated failure time (AFT) class of model (Bradburn et al, 2003). Accelerated failure time models differ from the more established proportional hazards (PH) models in that rather than estimating a baseline hazard function, AFT models instead estimate a baseline survival function. The estimated covariates are then multiplicative with respect to survival time. When exponentiated (transformed using the formula ex, where e=2.71828 and x is the coefficient value), the coefficients in an AFT model are termed as time ratios. A time ratio greater (less) than one indicates that a covariate increases (decreases) time to recurrence, stretching (shrinking) the baseline survivor function along the time axis.

Bottom Line: Online.When compared with the NPI, the model was able to better discriminate between women with excellent and good prognoses, and it did not overestimate 10-year recurrence-free survival to the extent observed for Adjuvant!

View Article: PubMed Central - PubMed

Affiliation: Health Economics Research Centre, Department of Public Health, University of Oxford, Old Road Campus, Headington, Oxford OX3 7LF, UK. helen.campbell@dphpc.ox.ac.uk

ABSTRACT

Background: We aimed to estimate and externally validate a new UK-specific prognostic model for predicting the long-term risk of a first recurrent event (local recurrence, metastatic recurrence, or second primary breast cancer) in women diagnosed with early breast cancer.

Methods: Using data on the prognostic characteristics and outcomes of 1844 women treated for early breast cancer at the Churchill Hospital in Oxford, parametric regression-based survival analysis was used to estimate a prognostic model for recurrence-free survival. The model, which incorporated established prognostic factors, was externally validated using independent data. Its performance was compared with that of the Nottingham Prognostic Index (NPI) and Adjuvant! Online.

Results: The number of positive axillary lymph nodes, tumour grade, tumour size and patient age were strong predictors of recurrence. Oestrogen receptor (ER) positivity was shown to afford a moderate protective effect. The model was able to separate patients into distinct prognostic groups, and predicted well at the patient level, mean Brier Accuracy Score=0.17 (s.e.=0.004) and overall C=0.745 (95% CI, 0.717-0.773). Its performance diminished only slightly when applied to a second independent data set. When compared with the NPI, the model was able to better discriminate between women with excellent and good prognoses, and it did not overestimate 10-year recurrence-free survival to the extent observed for Adjuvant! Online.

Conclusion: The model estimated here predicts well at both the individual patient and group levels, and appears transportable to patients treated at other UK hospitals. Its parametric form permits long-term extrapolation giving it an advantage over other prognostic tools currently in use. A simple point scoring system and reference table allow 5-, 10-, and 15-year predictions from the model to be quickly and easily estimated. The model is also available to download as an interactive computer program.

Show MeSH
Related in: MedlinePlus