Extracting pertinent information from unstructured medical texts is essential for advancing healthcare research, particularly in oncology. Medical documents, such as diagnostic reports, clinical notes and electronic health records (EHRs), contain valuable insights but pose challenges due to the complexity of the language. Transfer learning techniques, particularly pre-trained models like BERT (Bidirectional Encoder Representations from Transformers) and its variants such as BioBERT, RoBERTa, ClinicalBERT, and CancerBERT, have significantly improved tasks like named entity recognition (NER), information extraction and relation extraction. In this review we analyze 20 studies in the last 5 years (2020–2024) focused on transfer learning techniques in medical NLP, BERT and its variants for biomedical text extraction, NER and relation extraction in oncology, domain-specific applications of BERT in cancer treatment, highlighting advances in precision oncology. Although BERT models confronted with several challenges that can impact their performance and applicability in various contexts like dataset quality and availability, computational resource requirements and ethical considerations. Emerging trends include sector-specific pretrained model, integration of multimodal learning, generative AI and automation, ethical AI practice, and improvement of computational efficiency.