Predict

Information for professionals

Welcome to PREDICT, an online prognostication and treatment benefit tool designed to help clinicians and patients make informed decisions about treatment following breast cancer surgery. The survival estimates, presented both with and without adjuvant therapy (hormone therapy, chemotherapy and trastuzumab), are provided for 5 and 10 years following surgery. Development of the model was a collaborative project between the Cambridge Breast Unit, University of Cambridge Department of Oncology and the Eastern Cancer Information and Registration Centre (ECRIC) and was supported by an unrestricted educational grant from Pfizer Limited.

We welcome any feedback you may have about PREDICT. If you have questions about its development or there are features you would like to have added to the model please let us know by emailing us at info@predict.nhs.uk

Using PREDICT

Model development

Model validation

Model extension: HER2 status (version 1.1)

Model extension: KI67 status (version 1.2)

Model re-fitting (version 2.0)

PREDICT and Oncotype DX™

Using PREDICT

The model is easy to use following data entry for an individual patient including patient age, tumour size, tumour grade, number of positive nodes, ER status, HER2 status, KI67 status and mode of detection. Survival estimates, with and without adjuvant therapy, are presented in visual and text formats. Treatment benefits for hormone therapy and chemotherapy are calculated by applying relative risk reductions from the Oxford overview to the breast cancer specific mortality. Predicted mortality reductions are available for both second generation (anthracycline-containing, >4 cycles or equivalent) and third generation (taxane-containing) chemotherapy regimens.

The Cambridge Breast Unit (UK) uses the absolute 10-year survival benefit from chemotherapy to guide decision making for adjuvant chemotherapy as follows: <3% no chemotherapy; 3-5% chemotherapy discussed as a possible option; >5% chemotherapy recommended.

The relative risk reduction for hormone therapy is based on 5 years of tamoxifen.

Model development

The model was derived from cancer registry information on 5,694 women treated in East Anglia from 1999-2003. Breast cancer mortality models for ER positive and ER negative tumours were constructed using Cox proportional hazards, adjusted for known prognostic factors and mode of detection (symptomatic versus screen-detected). The survival estimates for an individual patient are based on the average co morbidity for women with breast cancer of a similar age. Further information about the model is provided in a paper published in Breast Cancer Research in January 2010.

Model validation

The clinical validity of a prediction model can be defined as the accuracy of the model to predict future events. The two key measures of clinical validity are calibration and discrimination.

Calibration is how well the model predicts the total number of events in a given data set. A perfectly calibrated model is one where the observed (or actual) number of events in a given patient cohort is the same as the number of events predicted by the model.

Discrimination is how well the model predicts the occurrence of an event in individual patients. The discrimination statistic is a number between zero and one. It is generally obtained from the area under a receiver-operator characteristic curve. If a random pair of patients is selected from a dataset - one being a survivor and the other a non-survivor - the discrimination is the probability that the non-survivor will have a higher predicted risk than the survivor.

PREDICT was originally validated using a dateset of over 5000 breast cancer patients from the West Midlands Cancer Intelligence Unit.

We also validated PREDICT using a dataset from British Columbia that had been previously used for a validation of Adjuvant! Online. Predict provided overall and breast cancer specific survival estimates that were at least as accurate as estimates from Adjuvant! The results of this validation were published in the European Journal of Surgical Oncology.

Model extension: HER2 status (version 1.1)

The model was updated in October 2011 to include HER2 status. Estimates for the prognostic effect of HER2 status were based on analysis of 10,179 cases collected by the Breast Cancer Association Consortium (BCAC). A validation of the new model in the British Columbia dataset was published in the British Journal of Cancer. This showed that inclusion of HER2 status in the model improved the estimates of breast cancer-specific mortality, especially in HER2 positive patients.

The benefit of trastuzumab is based on the relative risk reduction of 31 percent in mortality up to five years in published trials.

Model extension: KI67 status (version 1.2)

More recently we have added KI67 status to the model. The prognostic effect of KI67 was taken from published data showing that ER positive tumours that express KI67 are associated with a 30 percent poorer relative survival.

KI67 positivity for the PREDICT model was defined as greater than 10 percent of tumour cells staining positive.

We have validated the version of PREDICT that includes KI67 using a data set from Nottingham. The addition of KI67 led to a small imporvement in calibration and discimination in 1,274 patients with ER+ disease - the area under the ROC curve improved from 0.7611 to 0.7676 (p=0.005). These data have been accepted for publication in BMC Cancer.

Model re-fitting (version 2.0)

While the overall fit of the model has been good in multiple independent case series, PREDICT has been shown to underestimate breast cancer specific mortality in women diagnosed under the age of 40, particularly those with ER positive disease (See publication [5]). Another limitation of the model is the use of discrete categories for tumour size and node status which result in “step” changes in risk estimates on moving from one category to the next. For example, a woman with an 18mm or 19mm tumour will be predicted to have the same breast cancer specific mortality if all the other prognostic factors are the same whereas breast cancer specific morality of women with a 19mm or 20mm tumour will differ. We have therefore refitted the PREDICT prognostic model using the original cohort of cases from East Anglia with updated survival time in order to take into account age at diagnosis and to smooth out the survival function for tumour size and node status. The fit of the model has been tested in three independent data sets that had also been used to validate the original version of PREDICT.

Calibration in ER negative disease validation data set: PREDICT v1.2 over-estimated the number of breast cancer deaths by 10 per cent (observed 447 compared to 492 predicted). This over-estimation was most notable in the larger tumours and in the high-grade tumours. In contrast, the calibration of PREDICT v2.0 in ER negative cases was excellent good (predicted 449).

Calibration in ER negative disease validation data set: The calibration of both PREDICT v1.2 and PREDICT v2.0 was good in ER positive cases (observed breast cancer deaths 633 compared to 643 (v1.2) and 634 (v2.0) predicted). However, as previously described, PREDICT v1.2 significantly under-estimated breast cancer specific mortality in women diagnosed with ER positive disease at younger ages, whereas the fit of PREDICT v2.0 was good in all age groups.

PREDICT and Oncotype DX™

Oncotype DX™ is a prognostic model (breast cancer recurrence) based on a test of gene expression profiles in tumours. It has recently been recommended by NICE (DG10) for use in women with oestrogen receptor positive (ER+), lymph node negative (LN−) and human epidermal growth factor receptor 2 negative (HER2−) early breast cancer to guide chemotherapy decisions if the person is assessed as being at intermediate risk, and where the information on the biological features of the cancer provided by oncotype DX™ is likely to help in predicting the course of the disease.

The oncotype DX™ recurrence score has been shown to be a prognostic factor independent of the other variables included in PREDICT. However, the incremental improvement in discrimination that it would be expected to provide has not been established. We are seeking to identify relevant data that will enable this to be done.