Leveraging ISMOTE-KPCA-STACKING Algorithm for Enhanced Vascular Vertigo/Dizziness Diagnosis and Clinical Decision Support

Vascular vertigo/dizziness is a complex clinical syndrome involving multiple disciplines and specialties, such as neurology and psychiatry. Due to the intricate etiology and the similarity between causes and symptoms, traditional diagnostic methods based on clinical symptoms and signs are often inaccurate. This study aims to establish an effective and accurate intelligent diagnostic method for vascular dizziness to address this issue. Initially, we collected patients’ medical history and biochemical indicators as research indices. To tackle the sample imbalance issue in clinical data, we employed an improved SMOTE(ISMOTE) algorithm to generate minority class data. The enhancement of the ISMOTE algorithm lies in its ability to more effectively identify and generate minority class samples in sparse regions, resolving the issue of traditional SMOTE algorithms potentially neglecting sparse areas when generating synthetic samples. Subsequently, we utilized the Pearson correlation coefficient for feature correlation analysis, screening and analyzing the original features, and identified 13 feature indices. To further improve model performance and simplify the computational process, we applied the KPCA algorithm to these indices for dimensionality reduction, ultimately obtaining three comprehensive feature indices. Finally, we constructed a Stacking ensemble algorithm model comprising base models (including KNN, RF, Naive Bayes, SVM, GBDT, and XGBoost). To optimize the overall model performance, we introduced a fully connected cascade neural network as a meta-layer model and employed grid search and the Levenberg-Marquardt (LM) algorithm to optimize the base models and meta-layer model, respectively. This enabled the Stacking ensemble algorithm better to learn the correlations among predictions from each base model, enhancing the model’s generalization ability. Experimental results demonstrate that the proposed ISMOTE-KPCA-STACKING model exhibits significant advantages in diagnosing vascular vertigo/dizziness, outperforming single base models in multiple evaluation metrics. Furthermore, the model excels in handling imbalanced data and feature selection, providing an effective method for accurately diagnosing vascular vertigo/dizziness.