Explainable Feature Attribution in Heterogeneous Clinical-to-Text Transformer Language Models for Alzheimer’s Progression Diagnosis

Alzheimer’s disease (AD) progression diagnosis remains challenging due to heterogeneous clinical variables, limited cross-cohort validation, and the lack of transparent learning systems suitable for clinical use. Although large language models (LLMs) have shown strong potential, their use on structured clinical assessments remains underexplored, particularly with respect to encoder–decoder comparisons, out-of-distribution (OOD) generalization, and explainability. This paper proposes a heterogeneous clinical-to-text framework that converts structured cognitive, demographic, and clinical variables into standardized natural-language patient profiles. The framework enables a unified comparison of encoder-based transformers, decoder-only LLMs, and baseline machine learning models under the same input representation. Experiments are conducted on in-distribution (IID) and OOD cohorts for binary and multi-class AD progression diagnosis. The results show that fine-tuned encoder models consistently outperform prompt-based decoder LLMs, with ModernBERT achieving the strongest predictive performance and cross-cohort robustness. Additional evaluations using tabular/text baselines, fusion strategies, calibration, uncertainty estimation, subgroup analysis, and domain sensitivity confirm the reliability of the proposed representation. Attribution-based explainability using Gradient-SHAP, Integrated Gradients, and Gradient-Activation further demonstrates stable and clinically meaningful cognitive feature contributions, with inter-method agreement supporting explanation consistency across cohorts and diagnostic tasks. Overall, clinical-to-text fine-tuning, particularly with ModernBERT, provides an accurate, robust, and interpretable framework for AD progression diagnosis, offering guidance toward trustworthy AI systems for neurodegenerative disease assessment.