Medical document classification is one of the active research problems and the most challenging within the text classification domain. Medical datasets often contain massive feature sets where many features are considered irrelevant, redundant, and add noise, thus, reducing the classification performance. Therefore, to obtain a better accuracy of a classification model, it is crucial to choose a set of features (terms) that best discriminate between the classes of medical documents. This study proposes TextNetTopics, a novel approach that applies feature selection by considering Bag-of-topics (BOT) rather than the traditional approach, Bag-of-words (BOW). Thus our approach performs topic selections rather than words selection. TextNetTopics is based on the generic approach entitled G-S-M (Grouping, Scoring, and Modeling), developed by Yousef and his colleagues and used mainly in biological data. The proposed approach suggests scoring topics to select the top topics for training the classifier. This study applied TextNetTopics to textual data to respond to the CAMDA challenge. TextNetTopics outperforms various feature selection approaches while highly performing when applying the model to the validation data provided by the CAMDA. Additionally, we have applied our algorithm to different textual datasets.