Skip to Main Content
Principal Investigator
Alex Bui
Position Title
About this CDAS Project
NLST (Learn more about this study)
Project ID
Initial CDAS Request Approval
Nov 3, 2014
Probabilistic risk prediction models and diagnostic prediction for lung cancer screening
The imminent rollout of imaging-based lung cancer screening programs aim to provide earlier detection and treatment. However, the eligibility criteria for these programs are largely derived from predictive risk models for lung cancer based on only a few demographic, smoking, comorbidity, and genetic variables; moreover, these models take a “static” perspective (e.g., calculating risk using only the last set of observations) and do not incorporate the evolving history of the individual. Opportunities now arise to expand the set of biomarkers and scope of prediction to address questions of clinical import, including diagnostic prediction in the setting of the indeterminate lung nodule and prognostic modeling of the biological trajectory of individual lung cancers. Models built from large-scale observational datasets (e.g., electronic health records, EHRs) and that integrate the range of a patient’s current and past clinical, imaging, and pathology information will help drive more accurate predictions with greater specificity in determining risk and downstream events such as lung cancer development, treatment response, and survival. These insights are imperative in guiding future screening policies. The intent of this project is to develop new, integrated risk models using probabilistic frameworks to estimate the likelihood of development of lung cancer; and to use such calculations in novel computational methods that explore optimal screening policies given different costs and consideration models. Specifically, this effort develops and validates the ideas behind designing/learning continuous time belief networks; as well as partially observable Markov decision processes. These probabilistic models will be compared to current risk models, as well as conventional statistical approaches (e.g., logistic regression, decision trees, etc.).

Aim 1. To develop probabilistic inference methods for calculating risk over time via a continuous time belief network (CTBN) derived from the full spectrum of observational lung cancer screening data

Aim 2. To determine phenotypic signatures that predict lung cancer and lung cancer biology in screening populations with indeterminate nodules.

Aim 3. To inform lung cancer screening decisions at the population and the individual level through a partially observable Markov decision process (POMDP).


Denise Aberle, MD
William Hsu, PhD

Related Publications