<![CDATA[The integration of structured clinical measurements and unstructured textual symptom descriptions poses persistent challenges for automated medical diagnosis, particularly due to feature heterogeneity and class imbalance in real-world outpatient data. This study proposes a feature importance–driven multimodal machine learning framework for multi-class medical diagnosis classification that jointly models numerical clinical attributes and free-text symptom narratives within a unified pipeline. Beyond overall performance comparison, the proposed approach systematically examines the interaction between model architecture and feature importance through three controlled configurations: base, strong-feature, and weak-feature settings. Six supervised learning algorithms are evaluated using stratified five-fold cross-validation and imbalance-aware metrics. The results show that feature importance–driven modeling yields strongly model-dependent performance characteristics. Bagging-based tree ensembles benefit most from strong-feature selection, with the Extra Trees classifier achieving the best overall performance, reaching a macro-averaged F1-score of 0.821, compared to 0.811 in the base configuration and 0.642 when only weak features are retained. Conversely, margin-based classifiers rely on distributed feature representations. The kernel-based Support Vector Machine performs poorly under strong-feature selection (F1-score 0.450) but achieves a substantially higher F1-score of 0.795 and a macro-averaged recall of 0.808 under the weak-feature configuration. Linear SVM demonstrates stable behavior across configurations, maintaining a macro-averaged F1-score between 0.798 and 0.810, while attaining the highest overall recall of 0.819. These findings indicate that feature importance should be treated as a model-aware analytical tool for aligning feature selection strategies with the inductive bias of the learning algorithm, supporting robust and clinically meaningful diagnostic classification.]]>
