Figures and show the model-predicted primary events (CDR and recall rate) associated with baseline mammography, while Figures and summarize predicted secondary performance measures. Table shows the regression equations for incidence, prevalence and model predictions with the associated R2 value. The increase in CDR with age generally tracks the prevalence 2nd order polynomial, while the recall rate is linear and stable with age. Both the primary events and the secondary performance measures have a wide range of values depending on radiologist performance at the 10th or 90th percentiles. Specificity dominates the recall rate by an order of magnitude: the false-positive outcomes (140/1000) and recall rate (147/1000) at age 50 are about 20 times the CDR (7.2/1000). At age 50, the recall rate range from 10th to 90th percentile is 16.3%, but this decreases only to 16.1% when using a median sensitivity at the extremes of specificity. Therefore, equal absolute percentage changes in radiologist sensitivity or specificity would have disparate effects.
Figure 2 Predicted baseline screening mammography primary performance measure: cancer detection rate (CDR). The probability of a true-positive outcome is equivalent to the CDR, or breast cancers detected per 1000 screening mammograms for women ages 35–65. (more ...)
Figure 3 Predicted baseline screening mammography primary performance measure: recall rate (RR). The RR percentage is recalls per 100 (not 1000) screening mammograms for women ages 35–65. Women both with and without cancer are recalled for further diagnostic (more ...)
Figure 4 Predicted secondary performance measures: positive predictive values for screening and diagnostic mammograms (PPVS, PPVD). Increasing cancer prevalence and screening sensitivity with age mean radiologists detect more cancers and the potential absolute (more ...)
Figure 5 Predicted secondary performance measures: total intervention rate and positive biopsy fraction (TIR, PBF). The TIR is for 1000 screening mammograms for women ages 35–65 and includes the interventions of tissue biopsy and needle aspiration. The (more ...)
Best-fit equations for model predictions.
Figure shows how the PPVS and PPVD increase with age. The reciprocal of the predicted value, or the number of screening recall mammograms or positive diagnostic mammograms needed to detect a cancer, consequently decreases with age. The average PPVS or CDR/recall rate at age 50 is 4.9% (range 3.3–8.1), and the reciprocal is 20 screening positives/cancer. A radiologist skilled in finding breast cancer but with average specificity would improve PPVS at age 50 to 5.3%, but a radiologist skilled in calling normal as normal but with average sensitivity would have a PPVS of 9.6%. This doubling highlights the primary importance of specificity in improving secondary performance measures.
Figure also shows that the predicted total baseline mammogram intervention rate for women over 50 varies from 39/1000 at age 50 to 51/1000 at age 65. The PBF at age 45 of 10% (range 7–16) is lower than the NBCCEDP (12.8%) and BCSC modified (15.6%) values, while the PBF of 26% at age 55 (range 19–38) is closer to the corresponding values of 19% and 27%. The negative predictive value of a screening mammogram varies from 99.95% at age 40 to 99.93% at age 50 and 99.79% at age 60. The low prevalence and high specificity overwhelm the effect of sensitivity: at age 50 and 10% sensitivity, the negative predictive value drops to only 99.2%.
Table summarizes the actual BCSC cancer rates and selected performance measures as percentages of the model input incidence, prevalence and corresponding model output predictions. The actual BCSC cancer rate at 9–15 months averages 115% of the model input SEER incidence data. However, the actual BCSC first mammography cancer rate varies between 55% and 77% of the model input prevalence, mean 66%. Even when using the low range of sojourn time, which decreases the model input, the mean is 81%. Consequently, the actual BCSC CDR is below the predicted CDR on average about 72% using one-year sensitivity, and 98% using two-year sensitivity. The actual BCSC recall rate averages 96% of the predicted recall rate, which implies the specificity in the model is reasonably accurate. The actual BCSC PPVS averages 71% of the predicted PPVS, again reflecting the lower actual cancer rate.
Screening mammography BCSC performance data.*