Distribution Of Risk Estimates Research Articles

Abstract P3-09-02: Risk stratification using clinical risk factors and genetic variants in a personalized screening trial

Abstract Introduction: Tailoring breast cancer screening according to individual risk may represent an improvement over the current practice of age-based screening. WISDOM (Women Informed to Screen Depending on Measures of Risk) is an ongoing randomized trial comparing the safety, efficacy, cost, and patient acceptability of personalized versus annual screening. Women in the personalized arm receive screening recommendations based on sequencing of 9 genes associated with hereditary breast cancer and a 5-year risk estimate from the Breast Cancer Surveillance Consortium (BCSC) risk model modified by a polygenic risk score (PRS) comprised of 75 single nucleotide polymorphisms. WISDOM represents the first-ever use of a PRS to prospectively modify risk estimates and allows comparison of risk model performance in a population-based setting. Thus, we evaluated the risk estimates generated by: 1) the Breast Cancer Risk Assessment Tool (BCRAT) based on the Gail model, 2) the BCSC model, and 3) the BCSC model modified by the PRS (BCSC-PRS). Methods: We analyzed participants in the personalized screening arm of the WISDOM Study (NCT02620852). The trial opened in October 2016 and is enrolling participants aged 40-74 years. Participants' self-reported demographic and risk factor information were collected through an online portal. Genotyping of participants in the personalized arm was done using a custom panel from Color Genomics. 5-year risk estimates were generated using the BCRAT (2011 version), BCSC, and BCSC-PRS models. In the latter, the PRS was used as a Bayesian likelihood ratio to modify the BCSC 5-year risk estimate. We compared the distributions of BCRAT, BCSC, and BCSC-PRS risk estimates around a low-risk (<1%) and moderately high-risk (≥3%) threshold using a paired statistical test (McNemar). Results: To date, WISDOM has enrolled 2,065 participants, of whom 1,157 are in the personalized arm and 830 have completed risk assessment. The median age was 57 years (interquartile range, IQR 49-64). 83% were Caucasian, 2% African-American, and 7% Asian. 8% self-reported as Hispanic. The median 5-year risk was 1.7% (IQR 1.1-2.3%) using the BCRAT, 1.6% (IQR 1.1-2.3%) using the BCSC model, and 1.5% (IQR 0.9-2.7%) using the BCSC-PRS model. The BCSC-PRS model classified more women into the low (<1%) and moderately high (≥3%) risk categories compared with the BCRAT (p < 0.001) and BCSC model (p < 0.001), Table. 5-year risk classification according to the BCRAT, BCSC and BCSC-PRS models <1%1-3%≥3% n (%)n (%)n (%)Gail161 (19)556 (67)113 (14)BCSC159 (19)568 (68)103 (12)BCSC-PRS275 (33)379 (46)176 (21) Discussion: Adding a PRS to the BCSC model categorized significantly more women below the low-risk threshold and above the moderately high-risk threshold compared with the BCSC model and BCRAT. Furthermore, the BCSC and BCRAT generated similar distributions of risk estimates. Follow-up with incident breast cancer data is needed to determine whether the reclassification provided by the PRS improves risk stratification and clinical outcomes. However, our preliminary findings suggest that incorporating genetic variants into a validated clinical model is feasible and could enhance risk prediction. Citation Format: Shieh Y, Ziv E, Eklund M, Sabacan L, Firouzian R, Madlensky L, Anton-Culver H, Borowsky A, LaCroix A, Naeim A, Parker B, van't Veer L, Esserman L, Tice J, WISDOM Study and Athena Network Investigators WS. Risk stratification using clinical risk factors and genetic variants in a personalized screening trial [abstract]. In: Proceedings of the 2017 San Antonio Breast Cancer Symposium; 2017 Dec 5-9; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2018;78(4 Suppl):Abstract nr P3-09-02.

Graphical presentation of distributions of risk in screening

The screening performance of tests involving multiple markers is usually presented visually as two Gaussian relative frequency distributions of risk, one curve relating to affected and the other to unaffected individuals. If the distribution of the underlying screening markers is approximately Gaussian, risk estimates based on the same markers will usually also be approximately Gaussian. However, this approximation sometimes fails. Here we examine the circumstances when this occurs. A theoretical statistical analysis. Hypothetical log Gaussian relative distributions of affected and unaffected individuals were generated for three antenatal screening markers for Down's syndrome. Log likelihood ratios were calculated for each marker value and plots of the relative frequency distributions were compared with plots of Gaussian distributions based on the means and standard deviations of these log likelihood ratios. When the standard deviations of the distributions of a perfectly Gaussian screening marker are similar in affected and unaffected individuals, the distributions of risk estimates are also approximately Gaussian. If the standard deviations differ materially, incorrectly assuming that the distributions of the risk estimates are Gaussian creates a graphical anomaly in which the distributions of risk in affected and unaffected individuals plotted on a continuous risk scale intersect in two places. This is theoretically impossible. Plotting the risk distributions empirically reveals that all individuals have an estimated risk above a specified value. For individuals with more extreme marker values, the risk estimates reverse and increase instead of continuing to decrease. It is useful to check whether a Gaussian approximation for the distribution of risk estimates based on a screening marker is valid. If the value of the marker level at which risk reversal occurs lies within the set truncation limits, these may need to be reset, and a Gaussian model may be inappropriate to illustrate the risk distributions.

Errors in Exposure Assessment, Statistical Power and the Interpretation of Residential Radon Studies

To date, epidemiological studies of risk from residential radon have not convincingly demonstrated an association with lung cancer. These case-control studies, however, have inherent limitations due to errors in estimates of exposure to indoor radon. These errors take on special significance because the level of residential risk predicted from studies of underground miners is relatively low and possibly at the limit detectable by current epidemiological methods. To illustrate the problem caused by errors in exposure assessment, a series of case-control studies were simulated and resulting dose-response relationships evaluated. For each of four assumed error distributions for exposure to radon progeny, 10 indoor radon studies of 700 cases and 700 controls were generated randomly from a population with a risk of radon-induced lung cancer based on extrapolations from studies of underground miners. When exposures were assumed as known without error, 6 of 10 studies failed to find a significant dose response, in accord with the theoretical power of the study of 0.47. For simulations in which exposures were measured with error, the situation was worse, as the power of the study was reduced further and it was even less likely that a single study would result in a significant finding. For each error scenario, combining data from the 10 simulated studies did result in a significant dose response. However, the pooled results are somewhat misleading, because the effects of mobility, missing radon measurements, residential occupancy and potential confounding variables such as cigarette smoking were not taken into account. Empirical estimates of power were computed using 1,000 simulated case-control studies. When mobility and missing radon measurements in prior homes were incorporated into the design, the power of the study decreased, reducing the chance of detecting a significant effect of exposure. Enlarging study size to 2,000 cases and 2,000 controls increased the power of the study to 0.90 when exposure error was absent and subjects lived in one home only, but power was below 0.40 under realistic conditions for exposure error and mobility. When studies were generated under an assumption that exposure does not increase risk, up to 15% of simulated studies with 700 cases and 700 controls resulted in an estimated dose-response parameter in excess of the dose response from studies of miners. With increasing mobility and exposure error, it became virtually impossible to distinguish between the distributions of risk estimates from simulated studies based on an underlying excess relative risk of 0.015/working level month from estimates based on no risk from exposure.(ABSTRACT TRUNCATED AT 400 WORDS)