Predictive online calculators are used by clinicians as decision aids in early breast cancer (EBC). While use statistics for these calculators have not been published, as of 2017 NHS Predict was being accessed more than 20,000 times a month. These predictive tools have not had accuracy & benefit of use prospectively confirmed in EBC, yet use of calculators has been encouraged in EBC guidelines. It is important to understand the populations informing model development & validation, to understand how data bias may impact predictions in under-represented subpopulations. This work sought to elucidate the risk of bias in model development & validation for 3 online EBC calculators (NHS Predict, Adjuvant! & Cancermath), in an effort to highlight sub-populations where calculated risk & therefore treatment benefit estimates may be less reliable.
A literature search was conducted in PubMed, search terms were “predict*” “adjuvant” “breast” & “algorithm”. Results were screened for relevance to the three predictive tools under scrutiny & additional references were extracted from relevant papers. Using a modified CHARMS checklist, the relevant sections of the development & validation papers were extracted.
6 development & 24 validation papers were reviewed as summarised in the Table 264PPredict Adjuvant Cancermath Development population size & date range 5694 1977-2008 37,968 1977-2007 499,724 1977-2007 Aged <35 in development population 2% (111) 0 >0.5% Aged >65 in development population 32% (1781) 0 >17% Tumour size >5cm in development population 5% (287) 0 0 Number of validation studies 10 13 3 % retrospective 100 100 100 Total number of patients in validation studies 19,864 19,618 11,203 Age >65 in validations 35% (7134) 42% (8313) 40% (4519) Age <35 in validations 16% (3235) 8% (1518) 9% (1007) Tumour size >5cm in validations 5% (287) 5% (1015) 6% (634) Universal exclusions Multi-focal, inflammatory, male Multi-focal, inflammatory, male Multi-focal, inflammatory, male Neoadjuvant chemotherapy not an exclusion 1 study (121 patients) 0 0 Overall conclusions of validation authors Earlier versions under-predicted mortality in women <35 Poor performance in tumours >5cm. Poor performance in general in: <35 and >65 More advanced disease Malay ethnicity Overly optimistic survival predictions across subgroups in UK population. Poor performance in < 35 Systematically under-predicted mortality, especially for ER-negative tumours.
All 3 predictive tools have under-represented groups in their development cohorts, specifically those under 35 & over 65 years old, as well as larger tumours. Validation studies consistently demonstrate worse performance in these groups. However, due to inconsistent methodology in validation studies, quantitating the summary performance within & across tools is difficult. These predictive tools should be used with caution in under-represented populations. More work is required to look at clinical utility of tools as well as their statistical performance.
The authors.
Has not received any funding.
All authors have declared no conflicts of interest.