A New Predictive Model for Breast Cancer Survival in New Zealand: Development, Internal and External Validation, and Comparison With the Nottingham Prognostic Index

Abstract

Background: Women diagnosed with breast cancer, their doctors, and their families, would find a valid estimate of her prognosis helpful in planning treatment and support. Assessing prognosis is complex as many factors influence it. Several predictive models have been produced, but none has been developed or tested on patients in New Zealand (NZ). Aim: We aimed to develop and validate a NZ predictive model (NZPM) for breast cancer, and compare its performance to a widely used UK-developed model, the Nottingham Prognostic Index (NPI). Methods: We developed a model to predict 10-year breast cancer-specific survival, using data collected prospectively in the largest population-based breast cancer registry in NZ (Auckland, 9182 patients), and assessed its performance in this data set (internal validation) and in an independent NZ population-based series of 2625 patients in Waikato (external validation). The data included all women with primary invasive breast cancer diagnosed from 1 June 2000 to 30 June 2014, with follow-up to death or to 31 December 2014. We used multivariate Cox proportional hazards regression to assess predictors and to estimate the probability of breast cancer mortality within 10 years, and therefore 10-year survival, for each patient. We assessed observed survival by the Kaplan-Meier method. We assessed discrimination by the C-statistic, and calibration by comparing predicted and observed survival rates for patients in 10 groups ordered by predicted 10-year survival. We compared this NZPM with the NPI in the validation data set. Results: The final NZPM used continuous variables of age, tumor size, and number of positive lymph nodes, and categorical variables of ethnicity, tumor stage, tumor grade, ER and PR receptors, HER2 status, and histologic type of tumor. Discrimination was good: C-statistics were 0.84 for internal validity and 0.83 for independent external validity. For calibration, for both internal and external validity, the predicted 10-year survival probabilities in 10 groups of patients, ordered by predicted survival, were all within the 95% confidence intervals (CI) of the observed Kaplan-Meier survival probabilities. The NZPM showed good discrimination even within the prognostic groups defined by the NPI. Conclusion: These results for the NZPM show good internal and external validity, transportability, potential clinical value, and its clear superiority over the NPI. Further research will assess other potential predictors, other outcomes, performance in specific subgroups of patients, and compare the NZPM to other models, which have been developed in other countries and have not yet been tested in NZ.

Combining Molecular, Imaging, and Clinical Data Analysis for Predicting Cancer Prognosis

“…Of the manuscripts analysed in this review, the studies of [61,68,[70][71][72][73][74][75]78,81,82,84,86,91,[97][98][99]106,110] employ an early integration strategy (n = 19). Among the manuscripts discussed in this review, the articles by [69,76,79,80,[87][88][89][90]92,93,95,96,100,[102][103][104][105]107,108] use an intermediate integration strategy (n = 19).…”

“…Out of the 43 studies, 6 studies addressed the model using a conventional statistics approach [68][69][70][71][72][73] and 25 developed ML-based models [61,]. In the remaining 14 studies, a combination of both approaches was used [97][98][99][100][101][102][103][104][105][106][107][108][109][110].…”

“…Table 6 describes several characteristics related to how cancer prognosis predictive models are built, such as the sample size of the data set, the application of dimensionality reduction techniques, the statistical methods and validation techniques used, the metrics of performance, the output of the model, and whether the model is externally validated and/or compared to others. Four studies used the CPH regression model [70][71][72][73], a semi-parametric model able to handle right-censored data whose output is the Hazard Ratio (HR) and that is arguably the most used technique for survival analysis in the medical field [12]. A Lasso-Cox approach [69] and supervised Principal Components regression (superPC) [68] were used in one article.…”

“…Regarding validation techniques, 10-fold CV was applied in three studies [68,69,73], all of them with less than 1000 patients. Bootstrapping was used in [72], which counted 9182 patients. Finally, two studies involving 5738 and 2186 patients did not apply any validation method [70,71].…”

“…Moreover, only two studies provided external validation of their models on independent data sets. The Waikato data sets and the NTBCS, BCOS, and POSH data sets were used in [70,72], respectively.…”