Skip to Main Content

An official website of the United States government

Methodologies for modeling clinical performance and utilities of a cancer screening test using NLST trial as an example

Principal Investigator
Name
James Dai
Degrees
Ph.D.
Institution
GRAIL LLC
Position Title
Biostatistics Senior Director
Email
About this CDAS Project
Study
NLST (Learn more about this study)
Project ID
NLST-1090
Initial CDAS Request Approval
Jul 3, 2023
Title
Methodologies for modeling clinical performance and utilities of a cancer screening test using NLST trial as an example
Summary
GRAIL has developed a microsimulation model for designing cancer screening trials and evaluating a targeted methylation-based muti-cancer early detection (MCED) test. This model simulates an asymptomatic population of participants, each undertaken one or multiple screening rounds of MCED screening on individual cancer trajectories with stochastic variation. The primary goal is to project effect sizes of key study endpoints in a screening trial, e.g., positive predictive value (PPV), stage shift and cancer mortality. Key input parameters in this microsimulation model include test sensitivity and sojourn time for each cancer class targeted by MCED at each stage. From clinical diagnosis, survival and mortality were then simulated for each individual by sampling the empirical cancer type and stage specific survival probabilities from SEER.


With three rounds of annual screening by low-dose CT (LDCT), NLST trial provides an ideal data set from a randomized and controlled trial (RCT) to motivate and validate methodologies for modeling clinical utilities of a cancer screening test, e.g. GRAIL’s MCED test. General topics of interest include, but not limit to, calibrating the aforementioned microsimulation model, developing and validating methods for estimating cancer stage-specific sojourn time and sensitivity, and projecting long-term mortality benefit based on observed stage shift during the screening period. The following aims summarize potential use of NLST data for modeling cancer screening trials.
Aims

1. GRAIL’s microsimulation model simulates the preclinical cancer stages in a backward fashion from the time of clinical diagnosis: the transition time of each preclinical cancer stage was sequentially generated by subtracting the simulated dwell times from an exponential distribution at each cancer stage. The mortality benefit was derived from the commonly used stage-shift model. The NLST data provide a dataset for studying operating characteristics of the microsimulation model and validating modeling structures and assumptions.

2. The key parameters for projecting stage shift by a cancer screening test in an RCT include the stage-specific sojourn time and sensitivity estimates. However, the majority of existing works are focused on the sojourn time for the entire preclinical latency. There is a paucity of estimation methods for stage-specific estimates in a RCT. The clinically detected lung cancers and screen-detected lung cancers in the NLST data will enable estimating these parameters in a maximum likelihood framework.

3. Stage-shift models are commonly used to project the mortality benefit in an RCT based on observed stage shift. Strong assumptions have to be made, e.g., survival benefit of screen detection was added to the counterfactual clinical detection time, and screen-detected cancers had similar survival time as clinically presented cancers. The NLST trial data will be used to evaluate to what degree these assumptions are satisfactory for an aggressive cancer like lung cancer, whether existing stage-shift models can be further improved with novel modification.

Collaborators

Zhaoyu Yin, GRAIL
Dongjing Guo, GRAIL
Rita Lopatin, GRAIL
Xinyi Hou, GRAIL
Bong Chul Chu, GRAIL
Wei Liang, GRAIL
Roger Jiang, GRAIL
Nan Zhang, GRAIL