<p>Oncology research has witnessed an increase in the past decades. Therefore, the scientist found it difficult to stay up to date with recent advancements in the field. One option is to minimize the time and effort required to find the right article and discover the true knowledge based on the cancer type. The variations in writing style could complicate the searching process and put more pressure on the researchers.&#xa0;This study aims to build a cancer article topic classification model using a new embedding confusion technique. We used three embedding algorithm techniques: Lexical, Semantic, and Deep Contextual embedding. This approach enables the classification model to comprehend both the general meaning of language and the specific medical context necessary for interpreting biomedical texts.&#xa0;To accomplish that, we used TF-IDF with dimensionality reduction to derive the lexical features from the text. Then, we utilized BioBERT with named entity recognition to extract the primary biomedical entities, which were then converted into a semantic embedding vector. Also, we used the PubMedBERT model to generate deep contextual embedding vectors. Three embedding vectors have been fused using Canonical Correlation Analysis. Then we build different machine learning algorithms based on the fused embedding vectors. We compared our proposed approach to popular topic classification models that use BERTTopic and LDA algorithms. Also, we performed an ablation study to investigate the impact of different fuesion of embedding vectors. Finally, the results of the different models have been statistically tested using the Wilcoxon signed-rank test.&#xa0;The obtained evaluation and statistical results of the ablation study confirmed that using three fused embedding vectors produced better efficiency than using two or one embedding vectors, irrespective of the machine learning used. Also, the proposed model outperformed benchmark models such as BERTopic combined with SVM and TF-IDF with Random Forest.&#xa0;The study shows that the embedding fusion process made for a specific field with supervised learning makes a reliable system for classifying biomedical texts. Using many representations shows richer contexts and meanings in cancer research. This method can be applied to other medical fields and help improve tools for finding information, and integrating knowledge in oncology. It also makes future clinical decisions easier and supports medical studies based on evidence by automating the organization of large biomedical literature.</p>