The Evolution of High-Level Synthesis for Machine Learning: A PRISMA-Based, Systematic, Bibliometric and Scientometric Review

High-Level Synthesis (HLS) has become an important methodology for implementing Machine Learning (ML) workloads on reconfigurable hardware; yet the development, technological characteristics, and research structure of HLS-for-ML have not been examined comprehensively across the literature. This study combines a PRISMA-based systematic literature review with bibliometric, scientometric, and systematic technological analyses to characterize the field and its development. Searches across Scopus, IEEE Xplore, Web of Science, and the ACM Digital Library retrieved 2911 records. Following deduplication, screening, full-text assessment, and a subsequent record-level eligibility and technological re-audit, 305 publications were retained in the corrected analytical corpus. Bibliometric and scientometric analyses examine publication growth, collaboration, citation structure, bibliographic coupling, keyword relationships, and thematic development, while a full-text-informed technological analysis characterizes ML workloads, Design Environments, hardware platforms, optimization strategies, evaluation approaches, and application contexts. The results show substantial growth in research activity, with 89.8% of the included publications appearing from 2019 onward, together with a distributed international research community and an identifiable intellectual structure connecting HLS, FPGA acceleration, and ML research. Technologically, the literature remains strongly FPGA-centered and CNN-oriented, while also encompassing emerging ML architectures, specialized ML-to-hardware frameworks, diverse optimization mechanisms, and multiple deployment and evaluation contexts. Cross-dimensional analysis shows that established techniques such as parallelism, pipelining, memory optimization, quantization, and dataflow coexist with workload-specific implementation strategies, including those used for Transformer and Graph Neural Network architectures. An evidence-derived multidimensional taxonomy organizes the field across six interacting dimensions. The review further identifies heterogeneous evaluation practices and limited benchmark commonality as challenges for cross-study comparison, while reproducibility, interoperability, and portability remain areas requiring dedicated investigation. Overall, HLS-for-ML is characterized by substantial research growth, persistent FPGA and HLS foundations, and a broad technological design space in which ML and hardware decisions interact across multiple abstraction levels.