Managing the vast amount of unstructured data we generate today would be unthinkable without automatic text classification. This tool has become indispensable for tasks as varied as sentiment analysis and combating fake news. However, a significant technical dilemma arises: despite the vast array of algorithms designed for these purposes, there is still no clear consensus on which is the most effective for each specific dataset. To address this issue, the present study compares the robustness of two models: Random Forest, which operates using a set of trees, and Logistic Regression, a linear model. The corpus used is Symptom2Disease, a dataset of 1,200 text records containing descriptions of symptoms in natural language, evenly distributed across 24 diagnostic categories, covering conditions such as psoriasis, diabetes, dengue, and malaria. For this purpose, a factorial experimental design was adopted, incorporating standard preprocessing, data augmentation to mitigate sample size limitations, and representation analysis via N-grams. In other words, it appears that Logistic Regression performs better than Random Forest, achieving a maximum F1 score of 0.9797 when used in conjunction with data augmentation and unigrams. For short medical texts, the linearity of the model and synthetic corpus enrichment prove to be a more efficient and accurate solution compared to non-linear models.