Software defect prediction (SDP) is essential for early identification of software defects for any IT firm. Traditional machine learning models often lack the ability to capture contextual information from source code, relying solely on static features. In contrast, recent advancements in deep learning leverage abstract syntax tree (AST) to extract semantic features, potentially improving prediction ability. However, not all deep learning models perform equally well in SDP task. This article provides a critical evaluation of state-of-the-art AST-based deep learning approaches for SDP. Through extensive experiments conducted on ten Java-based open-source projects from the Promise repository, the performance of these models is assessed using precision, recall, and F-measure. The outcomes of the research offer valuable insights for researchers to select the most suitable AST-based deep learning approach for effective identification of the defects, thereby enhancing software quality assurance practices and minimizing defects in software products.