New ML model from Ferreira Jr. et al. boosts CNS tumor classification accuracy by 5%

A team of researchers led by Paulo R. Ferreira Jr. (Universities in Brazil) has published a new machine learning framework for classifying central nervous system (CNS) tumors based on DNA methylation profiling. The method employs Sparse Random Projection to reduce the high dimensionality of methylation data, followed by multinomial logistic regression for classification. This combination is designed to be both computationally efficient and robust, addressing common pitfalls in prior work such as data leakage and insufficient cross-validation.

In rigorous evaluations, the model scored 96% mean accuracy under stratified 3-fold cross-validation on a 2,801-sample reference cohort. On a separate 1,104-sample clinical cohort, it achieved 86% accuracy at the 91-class level and 93% at the methylation class family level. These results represent absolute gains of ~4 and ~5 percentage points over the current state-of-the-art. The paper emphasizes that such a 5-point improvement is clinically meaningful, as even a single misclassification can alter cancer subtype assignment and subsequent treatment pathways.