Pioneering New Strategies in Preventive Healthcare: Design and Analysis Innovations

Prognostic models in cancer use patient demographic and tumor characteristics to predict survival and dynamic disease prognosis. Past works in breast cancer and lung cancer have shown that cancer detection method, screen-detected or symptom-detected, has prognostic significance. We recently investigated this phenomenon in the lung component of the Prostate, Lung, Colorectal, and Ovarian (PLCO) screening trial. Patients were randomized to intervention, receiving four annual chest x-rays (CXRs), or to control, receiving usual care. In PLCO, lung cancer detection method has independent prognostic value exceeding that of variables commonly used in lung cancer prognostic models. Issues associated with lead-time bias and length-bias, and surrogate outcomes in cancer screening trials will be discussed in the context.  

In a retrospective analysis of the American Cancer Society Cancer Prevention Study-3 (CPS-3) cohort, GRAIL's multicancer early detection (MCED) test was evaluated for detectability in stored plasma samples collected before cancer diagnosis. In the present work, classical state-transition models were adapted to characterize the natural history of ctDNA-shedding cancers. The terminology of sensitivity for detecting preclinical cancers was discussed in the context of retrospective testing, and a novel Bayesian likelihood method was developed to estimate sensitivity and preclinical duration at early stage (localized and regional) and late stage (distant). We isolated certain state transition and sensitivity assumptions that were not directly testable in retrospective studies. Despite limitations of stored plasma samples, Bayesian analysis of the CPS-3 retrospective data showed that among the twelve prespecified cancers that represent two-thirds of cancer deaths in the United States, the test showed 61% estimated overall sensitivity, 43% estimated sensitivity for early-stage cancer, and 1.4-year average duration of preclinical detectability at early stage. With the limitation of a non-interventional study and untestable modeling assumptions, these results are consistent with the potential of GRAIL's MCED test to detect cancers early in the preclinical phase. 

Screening trials have required very large sample sizes and long time-horizons to definitively demonstrate mortality reductions. We and others have recently demonstrated that statistical power can be greatly increased by testing stored material in the control-arm and considering trial outcomes only among those who ever test positive. In the "targeted" analysis, one only tests control-arm cancer-cases, whereas the "Intended Effect" (IE) analysis tests all control-arm specimens. We extend both analyses to account for loss-of-signal in stored specimens by storing and retesting specimens in the screen-arm. Although the targeted analysis tests many fewer control-arm specimens (saving money), we show that it is sensitive to non-compliance in specimen collections that is differential by arm. We extend the IE analysis to only test a representative subsample of the control-arm while maintaining nearly all statistical power. With careful attention to assumptions, there is much scope for novel analyses to substantially increase power in screening trials.
 

Before implementing a biomarker test for early cancer detection into routine clinical care, the test must demonstrate clinical utility, i.e., the test results should lead to clinical actions that positively affect patient-relevant outcomes. Unlike therapeutical trials for patients diagnosed with cancer, designing a randomized controlled trial (RCT) to demonstrate the clinical utility of an early detection biomarker with mortality and related endpoints poses unique challenges. We propose a generic multistate disease history model. The model links key performance metrics of the test, such as sensitivity, to primary endpoints like the incidence of late-stage cancer and mortality. It also incorporates the practical implementation of the biomarker-testing program in real-world scenarios. Multiple pathways from diagnosis to mortality endpoint were considered to accommodate differential and time-varying screening effects. We show how such a model can be used to calculate justified target accuracy levels for launching a utility trial based on the model's projected cost-benefit ratio of a screening program. We use numerical examples from the National Lung Screening Trial (NLST) to demonstrate the method.