Intelligent Software Defect Prediction Using Multimodal Deep Learning Through the Integration of Source Code, Software Metrics and Historical Development Data

Software defect prediction has historically relied on a single view of source code — either handcrafted metrics, token sequences, or structural representations in isolation. Multimodal deep learning addresses the fundamental limitation that no single code view captures the full richness of software artifacts: their syntax, structure, semantics, history, and natural-language context. This report provides a comprehensive survey of multimodal deep learning approaches for software defect prediction, covering the major modalities (lexical/semantic, structural/AST, control-and-data-flow, metric-based, and natural language) and the fusion architectures that combine them — concatenation, attention-gating, cross-attention, and contrastive multi-view learning. We synthesize findings from over 50 recent publications (2020-2026), including GMCA-SDP cross-attention fusion, FusionVul multimodal vulnerability detection, hierarchical CNN fusion of AST/CFG/DDG, and emerging vision-language model applications. We benchmark performance across datasets, analyze fusion strategy trade-offs, address challenges in modality alignment and missing data, and map the trajectory toward unified multimodal foundation models for software quality assurance.