Using natural language processing and speech processing techniques to understand and predict cognitive decline

The National Institute on Aging awarded The Research Foundation For The State University of New York (operating as Stony Brook University) $129,288 on September 17, 2025, under the Aging Research program (CFDA 93.866) to develop natural language processing and speech processing techniques for analyzing occupational exposure data and predicting cognitive decline in World Trade Center responders.

The research develops and applies NLP methods to extract occupational exposure variables from free-text descriptions of work activities in the WTC responder cohort, a population at high risk for cognitive decline due to disaster-related neurotoxic exposures. In the current F99 phase, the NLP-extracted exposure variables will predict mild cognitive impairment incidence, cognitive trajectory, and neurodegenerative biomarker distributions, with mask usage as a potential mediator. In the planned K00 phase, advanced NLP and speech processing techniques will analyze interview transcripts and audio recordings to extract linguistic and acoustic features—including word-finding difficulties, reduced vocabulary, pitch, pauses, and vocal quality—using pre-trained models such as RoBERTa and WAV2Vec to predict future cognitive changes in domain-specific cognitive functions. Machine learning models will process these features to advance occupational cognition health research by reducing reliance on structured lists and manual categorization of occupational activity exposures.

Performance occurs in Stony Brook, New York. The award runs through September 16, 2027.

Abstract

Alzheimer's disease and related dementias are a significant public health concern, especially for disaster responders and construction workers exposed to neurotoxic hazards. Although valuable data about occupational exposure and cognitive health can be obtained from open-ended questions, analyzing such unstructured text and audio data is often time-consuming and underutilized. Recent advancements in natural language processing (NLP) and speech processing (SP) offer new opportunities to efficiently extract this information and predict cognitive outcomes. This proposal aims to leverage NLP and SP tools to analyze unstructured free text and audio data from the world trade center (WTC) responder cohort, a population at high risk for cognitive decline due to disaster- related neurotoxic exposures. I developed and validated a linguistic tool (NLP method) to extract occupational exposure variables-termed WTC exposure to response activities (wera) -from free text descriptions of work activities. In the F99 phase, these wera variables will be used to predict mild cognitive impairment (MCI) incidence, cognitive trajectory, and neurodegenerative biomarker distributions, with mask usage as a potential mediator. This NLP method will advance research in occupational cognition health by reducing reliance on structured lists and manual categorization of occupational activity exposures. In the K00 phase, i will expand the research by utilizing advanced NLP and SP techniques to analyze interview transcripts and audio recordings. By extracting both linguistic and acoustic features-such as word-finding difficulties, reduced vocabulary, pitch, pauses, and vocal quality-through pre-trained models like roberta and WAV2VEC, i aim to predict future cognitive changes in domain-specific cognitive functions. These features will be processed using machine learning models (e.g., random forest regressor, support vector regression, and neural networks) to predict cognitive decline over time. This approach offers a non-invasive, scalable, and cost-effective method for early detection of cognitive impairment, potentially benefiting other at-risk populations, including veterans and older adults exposed to neurotoxic hazards.