Automated Language Analysis in Neurodegenerative Diseases: Application of Artificial Intelligence Models (AILAN)

Neurodegenerative disorders pose significant diagnostic challenges, particularly in differentiating conditions with overlapping clinical symptoms and neuroanatomy. Although language impairment is a hallmark symptom across several forms of dementia, existing diagnostic tools are often costly, time-consuming, and lack cultural sensitivity, particularly for the Italian language. Artificial Intelligence (AI) offers a promising, non-invasive approach to extracting objective acoustic, prosodic, and linguistic markers from speech data.

The primary objective of this study is to characterize the language profiles of patients diagnosed with AD, PPA, bvFTD, DLB, and PD compared to healthy controls (HC). Specific objectives include: 1) Identifying distinct linguistic and speech features that act as surrogate biomarkers for early and differential diagnosis; 2) Using machine learning to optimize classification algorithms; and 3) Implementing Explainable AI (XAI) to ensure the interpretability of diagnostic outputs, thereby facilitating clinical trust and adoption of AI-driven tools in the diagnostic process.

This is a prospective, experimental, multicenter study. A total of 168 participants (24 per group: AD, PPA, bvFTD, DLB, PD, and HC) will be enrolled. Participants will undergo a 1.5-hour remote neuropsychological and logopedic evaluation via a secure, HIPAA-compliant videoconferencing platform. The assessment includes spontaneous speech tasks (e.g., image descriptions) and standardized language and cognitive tests. Audio data will be recorded and processed using an AI pipeline, including automatic speech recognition for transcription and various Natural Language Processing (NLP) techniques-such as BERT, RoBERTa, and UmBERTo-to extract complex linguistic and semantic features.

Data analysis will involve training machine learning classifiers (e.g., Support Vector Machines, Random Forests) to distinguish between diagnostic groups, utilizing nested cross-validation to ensure robustness. XAI techniques, including SHAP (Shapley Additive Explanations) and LIME (Local Interpretable Model-agnostic Explanations), will be applied to enhance the transparency of the model's decisions, allowing clinicians to interpret the pathological language patterns driving the classification. The study aims to establish the first structured Italian corpus for the linguistic characterization of neurodegenerative disorders, providing a foundation for future cross-linguistic analysis.