Enhancing Classification of High-Dimensional Neurocognitive Disorder Data via Neural Network-based Feature Selection with Static Attention and Noise Robustness

Objectives: To develop a high-dimensional medical dataset feature selection method for neurocognitive disorders. To combine a static attention-based score with noise robustness analysis that can optimize the classification of diseases. Method: A Neural Network feature selection method is proposed, integrating mutual information, standard deviation, variance threshold and correlation-based static attention scores that were fused by perturbation-based stability under Gaussian noises. The pipeline consisted of preprocessing the data, estimating its score, powerful feature selection, the SMOTE-ENN resampling technique, and training a variety of models using three open datasets PD Speech, Darwin, and dyslexia. Gold-standard methods were used as the benchmark of performance. Findings: The proposed work improved the state-of-the-art and reached up to 98% accuracy (Darwin), 98.89% (Parkinson) and 97.64% (Dyslexia) with significant increases of F1-score and ROC-AUC. Hybrid static attention and perturbation learned more stable, more discriminative features, and allowed strong classification in class-imbalanced high-dimensional cases. Novelty: This work presented an integrated approach of consideration of feature ranking based on the combination of both the properties of fixed attention and noise resistance, unreported before on these datasets. Class balancing was achieved by the application of SMOTE-ENN post-selection and the empirical data confirmed that they significantly outperformed the existing filter, wrapper, and metaheuristic methods. Keywords: Neural network, Feature selection, Static attention, Noise robustness, Class imbalance, SMOTE-ENN, High-dimensional data, Neurocognitive disorder, Disease classification