High-dimensional genomic feature selection with the ordered stereotype logit model

Sparsity regularization enhances gene selection and leukemia subtype classification via logistic regression

This study investigated the application of sparsity regularization methods to improve the classification of leukemia subtypes using high-dimensional gene expression data. Multinomial logistic regression models with the sparsity methods of Ridge, Lasso, and Elastic Net regularizations were employed to address overfitting and dimensionality issues while enhancing model interpretability. The study used a leukemia cancer dataset from the Curated Microarray Database (CuMiDa), which included gene expression data for 16,383 genes across 281 samples representing seven different types of leukemia. The statistical metrics of Accuracy, Kappa statistics, AUC, and F1-score were used to measure the models' implementation. Besides, the effectiveness and ability of each method in gene selection and dimensional reduction of the models were discussed. Elastic Net regularization was a better technique than the Ridge and Lasso based on overall classification performance; it also reached the highest accuracy along with Kappa values. On the other hand, both Lasso and Elastic Net were making more effective feature selections, creating sparse models that could efficiently discriminate leukemia subtypes. In this way, the results highlighted that the inclusion of sparsity regularization could enhance knowledge and accuracy in such a challenging task of subclass leukemia classification, enabling much more tailored treatments.

Penalized Bayesian forward continuation ratio model with application to high-dimensional data with discrete survival outcomes

While time-to-event data are often continuous, there are several instances where discrete survival data, which are inherently ordinal, may be available or are more appropriate or useful. Several discrete survival models exist, but the forward continuation ratio model with a complementary log-log link has a survival interpretation and is closely related to the Cox proportional hazards model, despite being an ordinal model. This model has previously been implemented in the high-dimensional setting using the ordinal generalized monotone incremental forward stagewise algorithm. Here, we propose a Bayesian penalized forward continuation ratio model with a complementary log-log link and explore different priors to perform variable selection and regularization. Through simulations, we show that our Bayesian model outperformed the existing frequentist method in terms of variable selection performance, and that a 10% prior inclusion probability performed better than 1% or 50%. We also illustrate our model on a publicly available acute myeloid leukemia dataset to identify genomic features associated with discrete survival. We identified nine features that map to ten unique genes, five of which have been previously associated with leukemia in the literature. In conclusion, our proposed Bayesian model is flexible, allows simultaneous variable selection and uncertainty quantification, and performed well in simulation studies and application to real data.