There was a large amount of variation across the eight included studies (see ‘Characteristics of included studies’ table: Table 4.1). All were observational studies, with four case-controls (19, 266, 372, 373) and four retrospective cohorts (263-265, 374). Of the four case-control studies, three studies used a matched design (266, 372, 373). Two studies were sub-studies from randomised controlled trials (one nested case-control (19) and one cohort study (374)). The studies ranged in size from 349 (266) to 1066 (263) women. Six studies included women from
128 Western populations (the UK (19), Finland (19), USA (266, 372), Canada (372), Sweden (265, 373) and the Netherlands (374)) and two studies were in women from South Korea (263, 264). Follow-up ranged from 5 years (263) to 14 years (265), and only one study (19) was in the preventive setting, with the rest being in the adjuvant treatment setting for women with breast cancer. Half of the studies assessed tamoxifen treatment only (19, 263, 265, 266), two studies assessed tamoxifen and an AI (264, 374), and two studies were not specific to a particular endocrine therapy, whereby only a subset of women were on endocrine therapy during their adjuvant treatment (372, 373). Two studies included a placebo (19) or control group (265), although the latter study did not compare across the interventions and could only be used in the prognostic review. Therefore, only one study was assessed in the predictive biomarker review (19). There was a mixture of premenopausal and postmenopausal women, and in the treatment setting, the two South Korean studies included women with DCIS or invasive breast cancer, whereas the other studies included women with invasive disease only.
The two sub-studies from clinical trials used visually-assessed density, with one assessing density to the nearest 5% (19) and the other assessing density in 20% Boyd categories (374). Two studies used a machine learning-based density assessment trained on Cumulus images (265, 373), three studies used Cumulus percent density (264, 266, 372) and one study used BI- RADS density (263). Density change cut-points varied greatly, with some adopting a 5% (264), 10% (19, 372, 373) or top tertile (determined by the distribution of controls) (266) absolute percent density reduction cut-point, one using a 20% relative dense area reduction cut-point (265), and another using reduction in BI-RADS category (263). One study did not report their definition of density change (374).
Two studies had recurrence (recurrence-free survival) as their endpoint (263, 264), two studies had mortality as their endpoint (265, 266), two studies looked at incidence of contralateral breast cancer (372, 373), one study assessed incidence of contralateral breast cancer and recurrence as its endpoint (374), whilst the final study assessed risk of developing invasive or DCIS breast cancer in a sample of at-risk women (19).
A detailed description of the included studies can be found in the data capture forms (appendix B.VI).
129
130
Table 4.1: Characteristics of included studies
Cuzick 2011
Methods Nested case-control within a multi-centre international randomised controlled trial (IBIS-I).
Recruitment April 1992-March 2001, diagnosis before 1 October 2007.
Prognostic and predictive biomarker. Prevention setting.
Participants 123 cases from the UK and Finland, 942 controls from the UK. Age 35-70 years at recruitment to IBIS-I trial.
Premenopausal and postmenopausal women.
Approximately twice the population risk of developing breast cancer.
Interventions Tamoxifen 20mg daily (n=507), placebo daily (n=558), 5 years of treatment.
Visually-assessed percent density.
Density reduction 10% or more vs. no change at 12-18 month follow-up mammogram in tamoxifen arm (prognostic biomarker). Density reduction 10% or more vs. less than 10% at 12-18 month follow-up mammogram in tamoxifen arm compared with placebo arm (interaction, predictive biomarker).
Outcomes Incidence of invasive breast cancer and DCIS.
Notes
Item Authors'
judgement Support for judgement
Study participation Low risk Source population described, demographic factors (and tumour characteristics for cases) similar in included and not included women from the IBIS-I main trial.
Study attrition Low risk Only 44 women withdrew who were not included in this sub-study (referenced (200)).
Prognostic factor measurement
Moderate risk
Valid and reliable density measure, although density change measure was determined by the data (the cut- point was chosen as “the minimum change that could be reproducibly detected”), the reader was not blinded to case-control status.
131 Outcome measurement Low risk From trial database.
Study confounding Low risk Adequate adjustment. Statistical analysis and
reporting
Low risk Adequate analysis, although no interaction reported in the paper (worked out from raw data provided by the study authors).
Kim 2012
Methods Retrospective cohort, non-randomised.
Initial ER+ breast cancer diagnosis October 2003-December 2006. Follow-up median 69 months.
Prognostic biomarker. Treatment setting.
Participants Seoul National University Hospital, South Korea. 1065 women.
Age 24-77 years.
No information on menopausal status but likely includes premenopausal and postmenopausal women.
12% DCIS, 88% invasive at first diagnosis.
Interventions Tamoxifen 5 years (n=657), tamoxifen 2-3 years + AIs (total 5 years) (n=41), tamoxifen 5 years + AIs (unknown total time) (n=192), AIs 5 years (n=175), at least 2 years of treatment. Cumulus percent density.
Density reduction 5% or more vs. less than 5% at 8-20 month follow-up mammogram in tamoxifen.
Density reduction 5% or more vs. less than 5% at 8-20 month follow-up mammogram in AIs.
Outcomes Recurrence.
Notes
Item Authors'
judgement Support for judgement Study participation Moderate
risk
No information on source population or key characteristics in women included vs. not included from source population.
Study attrition Moderate risk
No information on participant drop-out, loss to follow-up or reasons for censoring.
132 measurement risk treatment or time between baseline and follow-up
mammogram. Outcome measurement Moderate
risk
No information on start of follow-up or reasons for censoring.
Study confounding High risk No information on when confounding factors were measured, unclear adjustments (if any) for analysis separated by treatment.
Statistical analysis and reporting
High risk Unclear when follow-up started or reasons for censoring, unclear adjustments, unclear if subgroup analyses include women on tamoxifen only, AIs only or women who switched treatment.
Knight 2018
Methods Case-control (matched on follow-up time, geographical area, birth year, diagnosis year and ethnicity), non-randomised.
Initial breast cancer diagnosis 1990-2008, recruitment 2009-2012. Follow-up mean 8 years.
Treatment setting.
Participants WECARE study (USA and Canada).
224 cases and 243 controls with mammograms at both time points. Mean age 46 years at mammogram before or at first diagnosis. Premenopausal and postmenopausal women.
All invasive at first diagnosis.
Interventions Mainly tamoxifen, but specific treatments not reported. Cumulus percent density.
Density reduction 10% or more vs. less than 10% at 6 month-4 year follow-up mammogram.
Outcomes Incidence of a secondary primary breast cancer (e.g. in the contralateral breast).
Notes Cannot include as a prognostic or predictive biomarker because the analysis adjusted for tamoxifen use.
Item Authors'
judgement Support for judgement Study participation Moderate
risk
No information on source population but comparisons conducted on key characteristics between women included vs. not included from source population.
133 Study attrition High risk Potential for survival bias whereby women included
were more likely to have survived at the time of interview than the wider cohort. Unsure about density or outcome in women who died before the study or who did not have available mammograms. "All women had to be alive at the time of contact for interview" and "Women in whom we could not obtain a mammogram in an appropriate time window (see below) were more likely to have an earlier year of first breast cancer diagnosis (65% diagnosed in 1990– 1996 vs. 40% in 1990–1996)".
Prognostic factor measurement
Moderate risk
No information on blinding to treatment, or why 435/467 women were used in the analysis (could have been digital mammograms (instead of film like the rest of the study sample) or poor quality
mammograms).
Outcome measurement Low risk From population registry. Study confounding Moderate
risk
435/467 women used in the analysis (could have been missing data on adjusting factors), risk factors were obtained retrospectively by a telephone survey (potential for recall bias).
Statistical analysis and reporting
High risk Analysis adjusted for treatment so unable to extract the effect as a prognostic or predictive biomarker, the study may have included other endocrine therapies besides tamoxifen.
Ko 2013
Methods Retrospective cohort, non-randomised.
Initial ER+ breast cancer diagnosis January 2003-December 2008. Follow-up mean 59 months.
Prognostic biomarker. Treatment setting.
Participants National Cancer Centre, Goyang, South Korea. 1066 women.
Age 25-78 years.
Unclear information on menopausal status but likely includes premenopausal and postmenopausal women.
134 13% DCIS, 87% invasive at first diagnosis.
Interventions Tamoxifen (all women), at least 2 years of treatment. BI-RADS density.
Reduction of at least 1 category vs. no reduction of at least 1 category at 10-34 month follow-up mammogram.
Outcomes Recurrence.
Notes
Item Authors'
judgement Support for judgement Study participation Moderate
risk
No information on source population or key characteristics in women included vs. not included from source population.
Study attrition Moderate risk
No information on participant drop-out, loss to follow-up or reasons for censoring.
Prognostic factor measurement
Moderate risk
No information on whether restricted to contralateral breast or time between baseline and follow-up mammogram, no test of intra-reader reproducibility. Outcome measurement Moderate
risk
No information on start of follow-up or reasons for censoring.
Study confounding Moderate risk
No information on when confounding factors were measured, no adjustment for chemotherapy although it was associated with mammographic density reduction.
Statistical analysis and reporting
High risk Unclear when follow-up started or reasons for censoring, title says 'premenopausal' women but likely includes postmenopausal women too since age range 25-78 years, no adjustment for confounding factors such as chemotherapy.
Li 2013
Methods Retrospective cohort, non-randomised.
Initial breast cancer diagnosis 1993-1995, follow-up until 31 December 2008.
Follow-up median 14 years. Prognostic biomarker. Treatment setting.
135
Participants Sweden.
974 women.
Median age 62-63 years at diagnosis. Postmenopausal women.
All invasive at first diagnosis.
Interventions Tamoxifen 20mg daily (n=231), tamoxifen 40mg daily (n=123), tamoxifen 20+40mg daily (n=108), tamoxifen 'other' dose daily (n=12), median 60 months of treatment.
Fully-automated area-based method measuring absolute dense area.
Relative dense area reduction more than 20% vs. stable dense area (≤9% increase to ≤10% reduction) at 6-36 month follow-up mammogram.
Outcomes Breast cancer mortality (time to death caused by breast cancer).
Notes
Item Authors'
judgement Support for judgement
Study participation Moderate risk No information on source population or key characteristics in women included vs. not included from source population.
Study attrition Low risk Follow-up information from population registry, participant drop-out and loss to follow-up as a result of emigration likely to be small.
Prognostic factor measurement
Moderate risk Cut-points chosen "a priori" but without justification.
Outcome measurement Low risk From population registry, clear definitions of start of follow-up and reasons for censoring.
Study confounding Low risk Adequate adjustment. Statistical analysis and
reporting
Low risk
Adequate analysis.
Nyante 2015
Methods Case-control (matched on age at diagnosis, diagnosis year and disease stage), non-randomised.
Initial ER+ breast cancer diagnosis 1990-2008, recruitment 1 January 1991-31 December 2010 (end of follow-up).
136 Prognostic biomarker.
Treatment setting.
Participants Kaiser Permanente Northwest, USA. 97 cases and 252 controls.
Age 32-87 years at first diagnosis.
No information on menopausal status but likely includes premenopausal and postmenopausal women.
All invasive at first diagnosis.
Interventions Tamoxifen (all women), at least 1 tamoxifen prescription started within 1 year of diagnosis.
Cumulus percent density.
Density reduction more than 8.7% vs. less than 0.5% at 3-26 month follow-up mammogram.
Outcomes Breast cancer mortality (time to death caused by breast cancer).
Notes
Item Authors'
judgement Support for judgement
Study participation Moderate risk No information on source population or key characteristics in women included vs. not included from source population.
Study attrition Low risk Follow-up information from population registry, reasons for censoring discussed.
Prognostic factor measurement
Low risk Valid and reliable density and density change measures (based on tertiles).
Outcome measurement Low risk From population registry, clear definitions of start of follow-up and reasons for censoring.
Study confounding Low risk Adequate adjustment. Statistical analysis and
reporting
Low risk
Adequate analysis.
Sandberg 2013
Methods Case-control (matched on age and calendar period of first breast cancer diagnosis, adjuvant therapy and follow-up time), non- randomised.
Initial breast cancer diagnosis 1976-2005. Follow-up mean 8 years.
137 Prognostic biomarker.
Treatment setting.
Participants Sweden.
211 cases and 211 controls.
Age at first diagnosis: ≤45 years (n cases =37, n controls=37), 45- 55 years (n cases=68, n controls=68), 55-65 years (n cases=56, n controls=56) and ≥65 years (n cases =50, n controls=50).
Premenopausal and postmenopausal women. All invasive at first diagnosis.
Interventions Endocrine therapy (n cases=87, n controls=87), but specific treatments not reported.
Fully-automated area-based method measuring percentage density. Density reduction 10% or more vs. stable density (<10% increase to <10% reduction) at 1-5 year follow-up mammogram.
Outcomes Incidence of a secondary primary breast cancer (e.g. in the contralateral breast).
Notes
Item Authors'
judgement Support for judgement Study participation Moderate
risk
No information on source population but comparisons conducted on key characteristics between women included vs. not included from source population. Study attrition Moderate
risk
Follow-up information from population registry, but no information on reasons for censoring or loss to follow-up.
Prognostic factor measurement
Moderate risk
Large variability in time between baseline and follow- up mammograms (follow-up mammogram 1-5 years after first breast cancer diagnosis), 66 women excluded if baseline percent density <10% or >90% because they could not undergo some of the defined density changes but these numbers were unknown for the subgroup of women on endocrine therapy. Outcome measurement Low risk From population registry.
Study confounding Moderate risk
Adjusted for age through matching, but other
adjustments are unclear in the subgroup of women on endocrine therapy.
138 Statistical analysis and
reporting
Moderate risk
Appropriate analysis for the study's primary objective, but the analysis in the subgroup of women on
endocrine therapy was a secondary objective. Numbers unknown and unclear adjustments for the subgroup of women on endocrine therapy, cannot separate out endocrine therapies.
van Nes 2015
Methods Retrospective sub-cohort within a multi-centre randomised controlled trial (TEAM).
Start of TEAM trial enrolment in 2001, but unknown time period of sub-cohort study.
Follow-up median 6 years. Prognostic biomarker. Treatment setting.
Participants The Netherlands.
378 women.
Age 45-91 years at baseline. Postmenopausal women. All invasive at first diagnosis.
Interventions Exemestane 25mg daily for 5 years (n=197), tamoxifen 20mg daily for 2-3 years followed by 3-2 years of exemestane (totalling 5 years) (n=181).
Visually-assessed percent density in 20% bands (Boyd categories). Unclear comparison: "change in breast density".
Outcomes Recurrence and incidence of a secondary primary breast cancer (e.g. in the contralateral breast) combined (loco-regional recurrence, distance recurrence or contralateral breast cancer). Notes Cannot include as a prognostic biomarker because there were no
results to extract.
Item Authors'
judgement Support for judgement
Study participation Low risk Source population not described but referenced (375), comparisons conducted between women included vs. not included from source population.
139 risk follow-up or reasons for censoring.
Prognostic factor measurement
Low risk Valid and reliable density and density change measures, although no information on which follow- up mammograms were used for the density change measure (therefore no information on time between baseline and follow-up mammogram).
Outcome measurement Low risk From trial database, clear definition of start of follow-up but unclear reasons for censoring (per protocol analysis so women were censored when they stopped treatment but no information on other
reasons). Study confounding Moderate
risk Unclear adjustments (if any). Statistical analysis and
reporting
High risk Insufficient presentation of data, adjustments and results for density change, both treatment arms combined so unable to separate out endocrine therapies.