Interpretable ADHD Classification using Hybrid CatBoost-MLP with Early Fusion of Structural MRI and Clinical Variables Using SHAP-Guided Ensemble Methods

Classification of Attention-Deficit/Hyperactivity Disorder (ADHD) using structural MRI remains limited by inter-subject anatomical variability, site-specific imaging protocols, and the inherently heterogeneous clinical profiles associated with the disorder. In this study, a multimodal classification framework was constructed by integrating mid-slice axial T1-weighted MRI features with standardized clinical variables, including demographic and cognitive measures. Image-based representations were extracted through a compact convolutional neural network, while clinical variables were processed using a dense encoder. The resulting latent features were concatenated and further expanded through second-order polynomial interactions and statistical descriptors to enrich the feature space. To ensure interpretability and reduce dimensionality, SHapley Additive exPlanations (SHAP) were applied to an XGBoost classifier, selecting the most informative 200 features from the fused set. Class imbalance was addressed using BorderlineSMOTE, and robustness was assessed by evaluating noise-augmented variants of the input data. Six traditional classifiers were benchmarked, and a stacked ensemble combining multilayer perceptron and CatBoost with an XGBoost meta-learner was developed. Evaluation was conducted using 25 repetitions of stratified five-fold cross-validation, yielding 600 validation runs in total. The hybrid model (mlp+cat) achieved the highest performance, with a 87.04% of mean accuracy, area under the ROC curve 0.9409%, 0.862% of F1-score, 81.48% of sensitivity, and 92.59% of specificity. Visualizations using UMAP and PCA-UMAP indicated improved separability in the fused feature space, while confusion matrix and ROC analysis confirmed balanced classification. The framework demonstrates improved diagnostic reliability through early fusion of multimodal data, incorporation of polynomial expansion, SHAP-guided feature reduction, and interpretable ensemble learning. These findings support the potential of integrative approaches for enhancing ADHD classification in heterogeneous clinical cohorts.