Impact of deep learning lung cancer risk models on clinical decisions with blood-based diagnostics
Principal Investigator
Name
Michael Kammer
Degrees
Ph.D
Institution
Biodesix, inc
Position Title
Head of Radiomics Research
About this CDAS Project
Study
NLST
(Learn more about this study)
Project ID
NLST-1515
Initial CDAS Request Approval
Jun 8, 2026
Title
Impact of deep learning lung cancer risk models on clinical decisions with blood-based diagnostics
Summary
This project aims to evaluate the performance of existing radiomic and deep learning–based lung cancer risk models using the National Lung Screening Trial (NLST) imaging dataset, and to model how these radiomics tools may influence downstream clinical decision-making when paired with blood-based diagnostic testing. The overarching goal is to optimize multimodal risk stratification strategies for patients with indeterminate pulmonary nodules.
Pulmonary nodules are frequently identified on low-dose CT (LDCT) imaging, either as incidental findings or in screening programs, yet a substantial proportion are indeterminate at presentation, creating diagnostic uncertainty and variability in clinical management. Existing pathways often involve serial imaging, invasive procedures, or empiric decision-making, which can lead to unnecessary interventions or delayed diagnosis. Radiomic and deep learning models, such as Sybil and DeepLungIPN, offer the potential to improve risk stratification by extracting quantitative imaging features and generating malignancy risk estimates directly from CT data. However, their relative performance and clinical utility in the specific setting of indeterminate nodules remain incompletely defined, and their interaction with commercially available blood tests is unclear.
Using the NLST imaging dataset, we will implement and validate several published radiomic and deep learning models, including publicly available frameworks such as Sybil and DeepLungIPN. Analyses will focus on nodules meeting criteria for indeterminate malignancy potential, defined by size, morphology, and clinical context. Model outputs will include individualized malignancy risk scores derived from baseline and, where applicable, longitudinal imaging. Performance will be assessed using discrimination, calibration, and clinical utility metrics, including AUC, sensitivity, specificity, and decision-curve analysis.
A key objective of this study is to evaluate how radiomic risk stratification affects downstream utilization of blood-based diagnostic tests for nodule characterization. We will simulate clinical decision pathways in which patients with indeterminate nodules are triaged to blood-based testing (e.g., proteomic or genomic assays) based on predefined imaging-derived risk thresholds. By incorporating established or published performance characteristics of blood-based assays, we will estimate the impact of combined testing strategies on clinically meaningful outcomes, including reduction of unnecessary invasive procedures, improved diagnostic efficiency, and timeliness of care.
We will also compare alternative diagnostic workflows, including blood-first, and integrated sequential strategies, to determine which approaches provide the greatest clinical and operational benefit.
This study will generate evidence to support the integration of advanced imaging analytics with blood-based diagnostics for the evaluation of pulmonary nodules of indeterminate malignancy potential. The results are expected to inform more precise, efficient, and patient-centered diagnostic pathways, with the potential to reduce unnecessary procedures while improving early detection of lung cancer.
Aims
• Evaluate the performance of existing radiomic and deep learning–based lung cancer risk models using the National Lung Screening Trial (NLST) imaging dataset.
• Optimize multimodal risk stratification strategies for patients with indeterminate pulmonary nodules.
• Model and evaluate how radiomic risk stratification affects downstream utilization of blood-based diagnostic tests for nodule characterization.
• Simulate clinical decision pathways in which patients with indeterminate nodules are triaged to blood-based testing (e.g., proteomic or genomic assays) based on predefined imaging-derived risk thresholds.
• Estimate the impact of combined testing strategies on clinically meaningful outcomes, including reduction of unnecessary invasive procedures, improved diagnostic efficiency, and timeliness of care.
Collaborators
MONIQUE GIVENS Biodesix, inc
Christasia Chavis Biodesix, inc
Srinivas Ravula Biodesix, inc
Michael Kammer Biodesix, inc