Abstract 7423: Modelling COPD to lung cancer progression using machine learning frameworks

Abstract Background: Chronic obstructive pulmonary disease (COPD) is a lung condition characterised by chronic respiratory symptoms due to abnormalities of the airways that causes persistent, often progressive airflow obstruction and independent risk factor for lung cancer. However, the mechanisms linking COPD to lung cancer remain poorly understood, particularly the role of mutual risk factors such as air pollution and smoking. Emerging evidence challenges that environmental carcinogens can be purely mutagenic, and suggests that chronic inflammation also facilitates cancer progression and can be compounded in genetically susceptible individuals. This study aimed to investigate any gene-environment interactions contributing to lung cancer development in patients with COPD. Methods: We used machine learning frameworks to solve this problem given the limitations of conventional methods on this high-dimensional and multi-modal data. Using germline SNPs, measurements of air pollution through PM 2.5, and clinical information from the UK Biobank, a cohort of 488, 377 participants, we created models to predict whether patients with COPD would get lung cancer. We tested multiple frameworks optimised for tabular data such as eXtreme Gradient Boosting (XGBoost) and TabNet, in addition to Elastic Net, to be able to independently capture different facets of the data and account for potential non-linear interactions not modelled in linear approaches. As a secondary aim, we focussed on COPD patients with at least a ten-year post-diagnosis survival, minimizing bias from mortality effects on cancer progression. Results: In (n=24, 094) individuals with COPD, 1, 312 progressed to develop lung cancer by the date of censorship. Patients progressing to lung cancer were significantly older at enrolment, had greater exposure to air pollution, and smoked more heavily (all p < 0.001). Using an 80:20 train-test split, XGBoost and Elastic Net models outperformed TabNet (p< 0.01 by DeLong’s test) in predicting lung cancer risk, with XGBoost achieving the highest ROC-AUC on the held-out test set (0.73, 95% CI: 0.72-0.75). Restricting the inclusion criteria to ten-year COPD survivors notably improved prediction accuracy (ROC-AUC = 0.87, 95% CI: 0.85-0.90) and identified SNPs related to lung function and inflammation (such as RAGE) as important risk factor of lung cancer in patients in COPD patients. Conclusions: This study demonstrates that integrating machine learning with epidemiological data enables high-accuracy modelling of lung cancer risk in patents with COPD, emphasizing a potential role of inflammation mediating PM 2.5 gene-environment interactions. These findings underscore the need for research on the biological mechanisms driving inflammation-associated cancer risk in COPD for patient stratification, as well as the potential for applying machine learning frameworks methods to other cancers to explore tumorigenic inflammation. Citation Format: Tej Pandya, Marcellus Augustine, Kevin Litchfield, Nuno Rocha Nene, Charles Swanton. Modelling COPD to lung cancer progression using machine learning frameworks [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 7423.