Authors: Valentina Gogulancea1, Aileen Lonie2, Adrian Moore2, Adrián Otamendi Laspiur3, Eduardo Alonso3, Alba Garin-Muga3,4, Julien Guiot5, Astrid Paulus5, Benoit Ernst5, Marjorie Gangolf6, Jon Eneko Idoyaga Uribarrena7, Eunate Arana Arri7, Francisco Núñez-Benjumea8, Miguel Giraldez-Álvarez8, Frank Guijarro Masero9, Sara Alvarez Munoz9, Jonathan Wallace10, Michaela Black1, Debbie Rankin1

Affiliations: 1 School of Computing, Engineering and Intelligent Systems, Ulster University, Derry~Londonderry, UK, 2 School of Life and Health Sciences, Ulster University, Coleraine, UK, 3 Fundación Vicomtech, Basque Research and Technology Alliance (BRTA), Mikeletegi 57, Donostia-San Sebastián, Spain, 4 Biogipuzkoa Health Research Institute, E-Health Group, Donostia-San Sebastián, Spain, 5 Department of Respiratory Medicine, University Hospital of Liège (CHU Liège), Liège, Belgium, 6 Department of Data Analysis, Projection and Exploitation, University Hospital of Liège (CHU Liège), Liège, Belgium, 7 Biobizkaia HRI, Scientific Coordination, Osakidetza, Barakaldo, Spain, 8 Institute of Biomedicine of Seville, Seville, Spain, 9 BILBOMÁTICA, Bilbao, Spain, 10 School of Computing, Ulster University

Background/ Introduction: Lung cancer is the most frequently diagnosed cancer type and the leading cause of cancer deaths, accounting for 18.7% of all cancer mortality worldwide. While environmental exposures such as air pollution and radon are recognized risk factors, they are not currently incorporated into most lung cancer screening risk models.

Material & Methods: We used machine learning algorithms to predict the risk of developing lung cancer in a cohort consisting of 208,349 patients, of which 23,009 had been diagnosed with lung cancer. The dataset contains a mixture of demographic (age, gender, smoking status) and clinical (history of comorbidities and associated conditions) variables. To the original dataset we appended geospatial variables identified as possible lung cancer risk factors (e.g. air and soil pollution, radon gas). We performed classification analysis using lung cancer diagnosis as the target variable. We tested several algorithms (logistic regression, random forests, decision trees, stochastic gradient descent, XGBoost and a Multi-Layer Perceptron) and fine-tuned them to obtain the best performance metrics (F1 and ROC-AUC scores). The classifiers were trained and tested using a clinical and a geospatially augmented dataset, using 5-fold cross-validation repeated 10 times. A subset of the original dataset was kept for external validation.

Results: The models yield ROC-AUC values ranging between 0.81-0.86, comparable to or exceeding previously reported AI models (0.654–0.90) and traditional statistical approaches (0.68–0.83). The addition of environmental and socio-economic risk factors along with demographic, lifestyle and clinical variables leads to small improvements in model discriminatory power. In the validation efforts, models employing a subset of clinical and geospatial variables performed substantially better than ‘pure’ clinical models.

Conclusion: This study demonstrates that environmental exposure information can be successfully integrated into a machine learning based risk prediction model. The approach can be applied to patients’ electronic health records, without the need for additional patient questionnaires.