Development of Lung Cancer Risk Screening Tool with Causal Discovery Model Evaluation Approach

Causal graph discovery approaches in healthcare for detecting high-risk diseases have been more widely applied in the last decade. The main challenge in causal graph discovery in healthcare data is the complexity of big data, which requires appropriate algorithms to reveal causal relationships between variables. This study focuses on evaluating the performance of seven causal discovery models—Peter-Clark (PC), Greedy Equivalent Search (GES), Direct LiNGAM, Directed Acyclic Graph-Graph Neural Network (DAG-GNN), Greedy Sparsest Permutation (GraSP), and Recursive Causal Discovery (RCD)—on opensource healthcare datasets. The model performance was evaluated using the Structural Intervention Distance (SID), Structural Hamming Distance (SHD), Matthews Correlation Coefficient (MCC), and Fobernius Norm (FN) metrics. The evaluation results conclusively show that the GES model performs best on low-complexity datasets. Meanwhile, the DAG-GNN model offers consistent performance on high-complexity data with MCC values ranging from 0.77 to 0.88. The application of the GES model for lung cancer risk screening, based on user question responses, demonstrated effectiveness by measuring MCC, SID, and SHD scores between the reference adjacency metrics and the resulting screening metrics.