Machine-learning Modelling Gateways for Early-Disease Prediction of Lung Cancer
Principal Investigator
Name
Volker Hessel
Degrees
Ph.D.
Institution
Adelaide University
Position Title
Professor
About this CDAS Project
Study
NLST
(Learn more about this study)
Project ID
NLST-1533
Initial CDAS Request Approval
Oct 5, 2026
Title
Machine-learning Modelling Gateways for Early-Disease Prediction of Lung Cancer
Summary
Lung cancer remains a leading cause of cancer mortality, largely because many cases are diagnosed at an advanced stage. Although low-dose computed tomography screening supports earlier detection, improved risk-stratification methods are needed to identify individuals who may require closer follow-up. This project will investigate whether machine learning models based on National Lung Screening Trial (NLST) data can improve early lung cancer risk prediction and support future clinical decision-making.
This study will use a retrospective secondary-data analysis design. The comprehensive collection of 15 standard NLST datasets will be analysed, including relevant demographic characteristics, smoking history, medical risk factors, screening history, structured screening findings, diagnostic procedures, lung cancer diagnosis and characteristics, disease stage, follow-up information, mortality and other available outcomes. The project will use structured NLST data only. Raw CT images, coronary artery calcium images and pathology images are not requested.
The analysis will begin with data-quality assessment, management of missing values, descriptive statistics and exploratory analysis. Relevant variables will be selected according to the research objectives and NLST documentation. Appropriate preprocessing will be applied to categorical and continuous variables. Data will be separated at the participant level into training, validation and test sets to minimise information leakage. Class imbalance will be addressed using appropriate training and evaluation procedures.
Candidate models will include logistic regression, regularised regression, random forest, extreme gradient boosting and support vector machines. Ensemble or stacking methods may also be evaluated. Depending on the selected outcome, the models will predict lung cancer diagnosis within a defined follow-up period or analyse time-to-event outcomes.
Classification performance will be evaluated using the area under the receiver operating characteristic curve, area under the precision-recall curve, sensitivity, specificity, predictive values, F1 score and calibration measures. Survival analyses will use appropriate measures such as concordance and survival calibration. Cross-validation, subgroup analyses and sensitivity analyses will assess model robustness and potential variation across participant groups.
Model interpretability will be examined using feature-importance and SHAP-based methods. Decision-curve analysis and clinically relevant risk thresholds will be used to assess potential clinical utility. Data from a collaborating hospital in Nice, France, may be analysed separately in a later phase, subject to independent approvals. NLST participant-level data will remain within the approved Adelaide University environment and will not be accessed by unapproved collaborators. Findings will be shared through peer-reviewed publications and scientific presentations in accordance with CDAS requirements.
Aims
1. The first aim is to develop and validate machine learning models for early lung cancer risk prediction using NLST demographic, clinical, smoking-history, screening, diagnostic, follow-up and outcome data, together with approved imaging data where applicable. Candidate methods will include logistic and regularized regression, random forest, extreme gradient boosting, support vector machines and ensemble or stacking approaches. The data will be separated at the participant level into training, validation and test sets to minimize information leakage and support reproducible model development.
2. The second aim is to evaluate the discrimination, calibration, robustness and interpretability of the developed models. Model performance will be assessed using appropriate measures, including the area under the receiver operating characteristic curve, area under the precision-recall curve, sensitivity, specificity, positive and negative predictive values, F1 score and calibration measures. Cross-validation, subgroup analyses and sensitivity analyses will be conducted to evaluate model stability and possible differences in performance across relevant participant groups. Explainable artificial intelligence methods, including feature-importance and SHAP-based analyses, will be used to identify important predictors and explain model behavior.
3. The third aim is to assess the potential clinical utility of model-generated risk predictions for lung cancer screening, follow-up and diagnostic decision-making. Decision-curve analysis and clinically relevant risk thresholds will be used to determine whether the models provide additional value beyond conventional risk-assessment approaches. The resulting framework may also provide a foundation for future clinical decision-support and digital twin applications. This project is research conducted with the intention of sharing and publishing its findings. Results will be disseminated through peer-reviewed journal publications and scientific presentations in accordance with CDAS data-use, confidentiality, publication and acknowledgement requirements.
Collaborators
Sakshith Reddy Muthannagari Adelaide University
Arad Haghighatnejad Adelaide University
Volker Hessel Adelaide University
Shu Liang Adelaide University