Evaluating Hard and Soft Discriminant Models for High-Dimensional Leukemia Data

Accurately classifying cancer subtypes using high-dimensional gene expression data is a critical challenge in bioinformatics and clinical diagnostics. This study compares the performance of hard and soft Partial Least Squares Discriminant Analysis (PLS-DA) models in distinguishing acute myeloid leukemia (AML) from acute lymphoblastic leukemia (ALL) using a microarray dataset comprising ۷۲ bone marrow samples and ۷,۱۲۹ gene expression variables. Both models were implemented using the PLSDAGUI software. Hard PLS-DA, based on linear discriminant analysis (LDA), enforces exclusive class assignments, whereas soft PLS-DA, incorporating quadratic discriminant analysis (QDA), allows for probabilistic and overlapping class memberships. The results showed that both approaches performed comparably during the training phase. However, during the test phase, soft PLS-DA occasionally exhibited lower sensitivity, particularly for ambiguous samples. This apparent drawback may, in fact, reflect its more realistic and cautious treatment of class uncertainty, which avoids forced misclassifications. These results underscore the complementary strengths of the two methods and highlight the importance of soft classification in scenarios involving high biological complexity and class overlap.

کلیدواژه ها:

Soft and hard discriminant analysis ، PLS-DA ، Leukemia ، Gene expression profiles