A new peer-reviewed study published in PLOS Digital Health demonstrates how OpenAI’s GPT-3 program predicts early stages of dementia from spontaneous speech with a high degree of accuracy.
“To our knowledge, this is the first application of GPT-3 to predicting dementia from speech,” wrote professor Hualou Liang, Ph.D., and co-author Felix Agbavor at Drexel’s School of Biomedical Engineering, Science and Health Systems.
The most common type of dementia is Alzheimer’s disease, a neurodegenerative disease that affects an estimated 47 million people worldwide, according to the Alzheimer’s Association. By 2030 this figure is expected to grow to 76 million globally, according to the same source.
There are 5.8 million Americans with Alzheimer’s disease, of which two-thirds are women, according to a report by AARP and the Women’s Alzheimer’s Movement (WAM). The number is expected to nearly triple to 16 million Americans by 2050, according to the Harvard NeuroDiscovery Center at Harvard Medical School.
In the early stage of Alzheimer’s disease, a common symptom is short-term memory loss. The neuropsychiatric symptoms associated with Alzheimer’s disease may include agitation, distrust in others, disinhibition, depression, social withdrawal, psychosis, wandering, apathy, and delusions. The symptoms of Alzheimer’s disease vary depending on the stage of the disease.
Over time, the ability to perform various functions is impaired, affecting motivation, focus, executive functioning, decision-making, problem-solving, judgment, and the ability to multitask.
In the later stages of the disease, people with AD forget how to perform basic daily tasks and are eventually dependent on caregivers for survival. There is no known cure for Alzheimer’s disease, so early detection gives the patient valuable time to seek support, plan arrangements, and find treatment to manage the symptoms.
According to the researchers, language impairment is a noteworthy biomarker for neurodegenerative disorders as artificial intelligence (AI) natural language processing (NLP) has been used to predict Alzheimer’s disease (AD) in its early stages using speech. Using speech as a biomarker is a non-invasive, inexpensive, and fast method for clinical screening.
Prior work from other researchers used acoustic features from audio and linguistic features obtained from either speech transcripts or written texts through AI natural language processing techniques. The Drexel researchers point out that this feature-based approach is difficult to generalize as it is dependent on hand-crafted transformations and domain-specific information that makes it less flexible and adaptable for other stages and disease types.
“In this work, we show for the first time that GPT-3 can be utilized to predict dementia from spontaneous speech,” wrote the researchers. “Specifically, we leverage the vast semantic knowledge encoded in the GPT-3 model to generate text embedding, a vector representation of the transcribed text from speech, that captures the semantic meaning of the input.”
Venture-backed OpenAI based in San Francisco, California created GPT-3. OpenAI was founded in 2015 by Elon Musk (SpaceX, Tesla, Inc., and Twitter’s CEO), Y Combinator’s Sam Altman, Greg Brockman (Stripe’s former CTO), and others with backers that include serial entrepreneur Peter Thiel, LinkedIn co-founder Reid Hoffman, and more. Musk resigned from the board in 2018 to avoid any potential conflicts of interest with Tesla.
GPT-3 is a third-generation General Pretrained Transformer (GPT) model that consists of AI deep learning models named Davinci, Curie, Babbage, and Ada, that are capable of understanding and generating natural language. Davinci is ideal for comprehending the intent of text, determining cause and effect, and summarization for a specific audience. Curie is for language translation, tasks that involve analyzing complicated texts for sentiment classification, summarization, and complex classification. Babbage is for moderately complex classification tasks and search classification. Ada is a fast model suitable for keywords, address correction, parsing text, and simple classification tasks.
For this proof-of-concept, the Drexel researchers converted voice to text using wav2vec 2.0, a pretrained state-of-the-art model for automatic speech recognition. They trained the program with transcripts from speech recordings.
“Our results demonstrate that the text embedding, generated by GPT-3, can be reliably used to not only detect individuals with AD from healthy controls but also infer the subject’s cognitive testing score, both solely based on speech data,” the researchers reported. “We further show that text embedding outperforms the conventional acoustic feature-based approach and even performs competitively with fine-tuned models. These results, all together, suggest that GPT-3 based text embedding is a promising approach for AD assessment and has the potential to improve early diagnosis of dementia.”
Copyright © 2022 Cami Rosso All rights reserved.