Fake news detection has emerged as a critical challenge in the modern information ecosystem, where the rapid proliferation of misinformation threatens democratic processes, public health, and societal stability. Machine learning (ML)-based approaches have demonstrated significant promise in automatically identifying and classifying misleading information across diverse platforms. This study presents a comprehensive scientometric and systematic review of ML-based fake news detection research, drawing on 649 peer-reviewed articles indexed in the Web of Science database (1991--2023). Using bibliometric tools including R-Bibliometrix and VOSviewer, we systematically evaluate publication trends, collaborative networks, thematic evolution, and the relative effectiveness of key ML paradigms---including classical classifiers (SVM, Naive Bayes), deep learning architectures (LSTM, CNN), and transformer-based models (BERT, RoBERTa)---as reflected in the literature. Our analysis reveals a significant surge in research activity post-2016, driven by electoral misinformation and the COVID-19 infodemic, with transformer-based models increasingly dominating the field due to their superior contextual understanding. The findings identify critical research gaps, including limited cross-lingual detection capabilities and insufficient attention to multimodal fake news involving deepfakes. This review provides actionable insights for researchers and practitioners seeking to advance intelligent systems for misinformation detection.