Comparative analysis of lung cancer risk prediction models
Danil Ye. Dyadyun, Student. Perm State University
Robert V. Garafutdinov, Cand. Sc. (Econ.), Associate Professor of Information Systems
and Mathematical Methods in Economics Department. Perm State University
Abstract
Low-dose computed tomography reduces lung cancer mortality, but its widespread implementation is limited by a high rate of false-positive results. To improve screening efficiency, predictive models that maintain a specified level of specificity (typically ≥ 90 %) are required. This study compared the sensitivity of the established PLCOm2012 model with machine learning methods (XGBoost, neural network) under a fixed specificity constraint using synthetic data simulating population-level lung cancer risk. The results demonstrated that the neural network achieved the highest sensitivity (69.4 %), outperforming PLCOm2012 (61.1 %) and enabling the detection of 8.3 % more lung cancer cases without increasing the number of false-positive predictions. Thus, ML models show promise for enhancing screening efficacy by capturing complex patterns in the data. Their application could improve lung cancer detection rates without increasing the burden on the healthcare system.
Keywords: lung cancer; screening; LDCT; PLCOm2012; machine learning; discrimination threshold; specificity; recall.
For citation: Dyadyun D.Ye., Garafutdinov R.V. Comparative analysis of lung cancer risk prediction models. Digital Models and Solutions. 2026. Vol. 5, no. 2, pp. 5–18. DOI: 10.29141/2949- 477X-2026-5-2-1. EDN: OYAOEC.
