Latent Class Analysis

Latent Class Analysis was used as an unsupervised feature selection method that did not use cancer diagnosis as an outcome variable. The LCA score had higher sensitivity than the LASSO score but lower specificity. The four-class model was selected as optimal because it had the lowest BIC, significant test results, high posterior probabilities, and interpretability. The LCA analysis required complete data and therefore used a 1,545-patient subset with complete biomarker records. A three-class LCA solution was selected based on AIC, BIC, and chi-square statistics. Class assignment was robust for the four-class solution. The LCA model grouped patients into latent classes based on categorical biomarker values defined by clinical laboratory cut-offs. The study used latent class analysis to identify unobserved subgroups based on shared categorical risk patterns. LCA's broader biomarker inclusion reflected its unsupervised design and captured overall illness severity rather than cancer alone. The researchers estimated one- through ten-class models without presupposing the optimal number of classes.