96. USING MACHINE LEARNING MODELS TO DETECT EARLY ALZHEIMER’S DISEASE THROUGH SPEECH ANALYSIS

Introduction

There is an urgent need for novel approaches that may facilitate early detection of Alzheimer's disease and thus, create targets for effective intervention and management. Current diagnostic methods often rely on expensive and/or time-consuming procedures such as brain imaging and cognitive assessments. A novel approach proposes leveraging AI/ML models to detect AD early through the analysis of spontaneous speech and language use. This method holds the potential to advance the process of AD diagnosis by offering a non-invasive, cost-effective, and easily accessible screening tool that may identify subtle variations in linguistic (and by extension, neurocognitive) function that may not yet be identified by standard screening tools. Here, we explore the range of deep learning models that have been applied to language and cognition. We also compare their analytic approaches and available results, with a view to identifying which approach may translate most readily to clinical care.

Methods

We used a multi-faceted approach that included a literature review, brainstorming sessions with an interdisciplinary team and field experts, and targeted internet searches for relevant web-based resources. The focus of our search was to compile studies that explored the development and application of AI algorithms to identify subtle changes in speech patterns, linguistic features, and acoustic properties associated with the early stages of AD. We considered, but did not apply a traditional biomedical search algorithm, since the literature in this space is often found outside of the biomedical literature, and because this is an exploratory project. We noted that by analyzing extensive datasets of speech samples from both healthy individuals and those with AD, all the identified studies sought to establish robust predictive models for early detection. We further examined whether confounding variables present in current linguistic AD models, such as those arising from language barriers, are also present in trained deep learning models.

Results

Our investigation demonstrated the consistent application across the literature, of a multimodal system, encompassing both neural networks and traditional analysis models, which were fine-tuned for the early detection of Alzheimer's disease. Among these, the ADReSS dataset emerged as the most effective, with the ensemble method achieving the highest accuracy in predicting Alzheimer’s disease based on speech patterns. However, we noted a crucial limitation: the model’s training relied solely on English speech data. This restriction introduces bias and hinders generalizability. Languages exhibit distinct phonetic structures, accents, and rhythms, potentially causing a model trained exclusively on English to misinterpret speech from other languages. Furthermore, while deep neural networks excel at discerning complex patterns, their internal workings often remain opaque, making it challenging to ascertain the precise rationale behind specific predictions.

Conclusions

Our review identifies a notable body of literature that outlines a range of deep learning models that have already been applied to identifying cognitive changes through the use of language. With large language models gaining rapid popularity, there is a tremendous opportunity to gather data samples from natural language, and by pairing the right model with the right type of language data, powerful new screening tools may be developed. Our work points to two key areas for future prioritization: 1) developing models trained on diverse languages and 2) expanding existing datasets to encompass a wider range of linguistic variations, including various dialects and demographics. These advancements will contribute to more equitable and reliable speech-based Alzheimer's disease detection tools.