This repository contains the DementiaNet dataset and research on multimodal detection of dementia from spontaneous celebrity interviews using natural language processing (NLP) and audio signal analysis.
The full research paper detailing the methodology, experiments, and results is available in this repository: Article_Dementia_NLP.pdf
Neurodegenerative disorders such as Alzheimer's disease and related dementias represent a growing public health challenge, where early diagnosis remains essential to slow progression and improve patient outcomes. Traditional diagnostic methods, although reliable, are often invasive, costly, and not widely accessible. In this context, speech-based markers and linguistic analysis are emerging as promising, non-invasive tools for scalable early detection.
This work investigates the automatic identification of dementia from spontaneous audio interviews, using a heterogeneous dataset characterized by strong thematic variability. We evaluate unimodal models relying separately on acoustic descriptors from audio signals and textual features from transcripts, employing traditional machine learning approaches. In addition, we explore the potential of large language models through diverse prompting strategies to detect subtle linguistic markers. Our study highlights two central challenges: handling the variability inherent to context-independent data and addressing the limited dataset size through optimized representations of audio and text.
Keywords—NLP, LLM, FewShot, ZeroShot, Audio, Machine learning, Deep Learning, Transformers
The implementation code for the models and methods described in the research paper will be available soon after some final cleaning and organization.