Drexel Researchers Use Large Language Models To Predict Dementia From Spontaneous Speech

Recent technical developments have opened several gates for applying artificial intelligence (AI), especially in the healthcare industry. One such area of biomedical research that could significantly benefit from AI involves Alzheimer’s disease (AD). Alzheimer’s disease is a brain condition that is currently incurable. The current method of diagnosing AD is quite arduous and time-consuming, entailing a thorough assessment of medical history and a lengthy list of physical and neurological evaluations and testing. Speech often is a key indicator in detecting early signs of such neurodegenerative diseases since almost 60-80% of dementia patients suffer from language impairment. Researchers have extensively worked on using natural language processing (NLP) to identify early AD predictions. However, the use of large language models, such as OpenAI’s GPT-3, to assist in the early detection of dementia is still an uncharted area.

Working on this front, a research team from the School of Biomedical Engineering, Science, and Health Systems at Drexel University recently showed that OpenAI’s GPT-3 could successfully recognize early stages of dementia using spontaneous speech with almost 80% accuracy. The team’s research is the most recent of its kind, demonstrating the value of natural language processing in Alzheimer’s early detection and raising the possibility that language impairment may serve as a precursor to neurodegenerative illnesses.

The team concentrated on developing algorithms that could recognize minor cues like hesitation, grammatical and pronunciation errors, and even forgetting word meanings. These frequently aided medical professionals in the preliminary stages of determining whether a patient should go through a comprehensive checkup or not. Other frequently used tests for the early diagnosis of Alzheimer’s disease focus on auditory characteristics such as pausing, articulation, and vocal quality.

OpenAI’s GPT-3 employs a deep learning algorithm that emphasizes how words are used and how language is created, enabling it to respond to any language-related tasks more eerily than any other deep learning system. It is also a promising candidate for identifying the subtle speech characteristics that might predict the onset of dementia due to its remarkable performance on “zero-data learning” (ability to answer questions that would typically require consulting external knowledge sources), as well as its systemic approach to language analysis. The model was trained using a large dataset that was enhanced with speech pattern-related data needed to identify potential dementia patients.

The researchers created a vector representation from the text that accurately captures the essence of the input speech by utilizing the extensive semantic knowledge included in the GPT-3 model. Then, using only speech data, the vector representations were used to estimate the subject’s cognitive assessment score and identify people with AD from the healthy population. The researchers concluded that vector representation significantly surpasses the traditional acoustic feature-based approach and produces results on par with the current best-performing models.

The team’s commendable performance indicates great potential for developing fully deployable AI-driven tools for early dementia diagnosis, which could assist people in receiving the necessary care and treatment right from the start. Future work by the researchers involves creating a simple and user-friendly web application that could be used as a pre-screening tool for dementia at home or a clinic. If the initial proof-of-concept tests well, the application could be a game changer for early screening and risk assessment before a clinical diagnosis.

Check out the Paper and Reference Article. All Credit For This Research Goes To Researchers on This Project. Also, don’t forget to join our Reddit page and discord channel, where we share the latest AI research news, cool AI projects, and more.