Caption-based Alzheimer's and mild cognitive impairment diagnosis at baseline using hybrid and transformer-based models

This study explores a novel caption-based diagnostic framework for early Alzheimer's Disease (AD) and Mild Cognitive Impairment (MCI) screening using baseline clinical and cognitive data from the TADPOLE challenge. We compare classical classifiers (SVM, Gradient Boosting), hybrid neural networks (MLP, MLP+LSTM), and Transformer-based models (custom and T5) for both classification accuracy and the generation of human-readable diagnostic captions. Results show that while single-task MLPs achieve the highest classification performance (balanced accuracy = 0.739), Transformer models excel in caption quality (BLEU = 0.54, METEOR = 0.78), enhancing interpretability. Our findings highlight the trade-offs between accuracy and explainability, establishing a baseline for intrinsically interpretable AI in dementia care. This captioning approach holds promise for transparent clinical decision support and paves the way for future longitudinal modeling and integration of multimodal biomarkers.