Mitigating semantic reconstruction bias and methodological contamination in dementia detection using attention-gated linguistic fusion

Dementia and other cognitive impairments like Alzheimer's disease (AD) are major public health problems worldwide. Early detection is crucial for timely intervention and improved patient outcomes. Natural Language Processing (NLP) has shown significant potential for detecting dementia by identifying cognitive decline. However, existing approaches may lack clinical interpretability, which can reduce reliability. This study proposes an Attention-Gated Hybrid Network (AGH-Net) that integrates handcrafted linguistic biomarkers and contextual semantic representations from DeBERTa-v3 embeddings to improve the robustness of dementia classification. The Pitt Corpus dataset is used to evaluate the model with a participant-level validation strategy to reduce subject-level data leakage and ensure unbiased performance estimation. The proposed AGH-Net achieved 80.18% accuracy and a high clinical sensitivity of 83.87%. A shadow analysis quantifies a 4.44% performance inflation from data leakage, supporting the leakage-controlled result as a more reliable performance estimate. The model achieves transparency through Explainable Artificial Intelligence (XAI) methods. External evaluation on ADReSS-IS2020 and the Lu corpus further examined cross-corpus transfer, with performance varying across datasets. We further validate the classification results against ground-truth clinical Mini-Mental State Examination (MMSE) scores. The analysis indicates that the LLM can reduce the patient's pathological speech characteristics by normalising linguistically informative errors. Because general-purpose LLMs prioritise semantic meaning over linguistic noise, they may suppress linguistic irregularities that contain clinically informative signals of dementia. These findings indicate that integrating linguistic features with contextual semantic representation provides a reliable, interpretable framework for non-invasive automated dementia screening.