The artificial intelligence that can write essays and pass tests can also help identify dementia.
Researchers at Drexel University in Philadelphia used the AI behind ChatGPT (which has grabbed headlines for writing believable term papers and passing bar exams) to analyze speech, and the system correctly identified Alzheimer’s patients 80% of the time, according to the study published in the journal PLOS Digital Health.
The researchers used GPT-3, the language model that drives ChatGPT, to analyze audio clips of people describing a picture in a standard test for dementia.
Alzheimer’s patients often repeated themselves, strayed from describing the picture’s contents, didn’t finish thoughts, and referred to objects vaguely as a “thing” or “something.”
“GPT-3 is able to capture such a subtle difference reflected in the text,” says study author Hualou Liang, PhD, professor of biomedical engineering at Drexel.
The software analyzed text transcribed (also by software) from 10-second recordings of healthy adults and Alzheimer’s patients. The text trained the GPT-3 model to identify the subtle differences between regular language and speech from someone experiencing cognitive decline.
The GPT-3 machine learning models understand passages of text by converting words into mathematical representations called “embeddings.” The embeddings are multi-dimensional signals, which allow the AI to identify subtle differences and similarities that even experienced doctors may miss. GPT-3 compares the text passages by measuring the distance between those signals in the embeddings.
Because GPT-3 only analyzes written text, the process bypasses the pauses and other sounds in spoken language that aren’t words. In this case, that turned out to be an advantage: The GPT-3 analysis outperformed some machine learning models developed by other laboratories that included those sounds.
Other studies, however, have found that the “ahs” and “ums” in speech can be important in revealing Alzheimer’s. A 2021 study that encoded those pauses allowed a machine learning model to detect Alzheimer’s disease with 90% accuracy, and a separate study conducted in Slovenia that combined text and acoustic features achieved an accuracy of 94%.
“The best combination tends to be combining both types of features together,” says Frank Rudzicz, PhD, associate professor of computer science at the University of Toronto. “There’s a lot of information in the words and structure of the transcripts, but also in our tone of voice.”
Using Voice to Spot Alzheimer’s
More and more researchers are looking at voice as a biomarker, a way to detect various diseases including Alzheimer’s.
Worldwide, Alzheimer’s cases are successfully detected just 48% of the time, according to estimates by the World Health Organization. Higher-income countries achieve a 54% diagnostic rate, while low- and middle-income countries are only identifying 24% of Alzheimer’s cases.
Researchers in this field hope to close that gap by developing a tool that can detect Alzheimer’s early — when the effects may be too subtle for a physician to notice. “There is no cure for Alzheimer’s disease yet, but there are life changes that can delay some of its effects, so early diagnosis is still important,” says Rudzicz, who co-founded a speech analytics mobile app called Winterlight. “These kinds of technologies could also be applied to other disorders, including Parkinson’s, depression, and so on.”
Doctors could eventually use a device or computer program to test a patient’s cognitive abilities in their office. Brain scans or other clinical tests could then confirm the Alzheimer’s diagnosis.
Another application might use smart devices like Alexa and Siri to monitor your regular conversations (with your consent) and alert you if it notices any worrying word fumbles. It may even detect other psychological problems like depression and stress.
“The analysis could be done in a privacy-preserving manner once the system is fully functional,” says Liang. “As such, it could make an immediate and significant impact on mitigating the dementia problem in the older adult community.”
Sources
Hualou Liang, PhD, professor of biomedical engineering, Drexel University, Philadelphia.
Frank Rudzicz, PhD, associate professor of computer science, University of Toronto.
Saturnino Luz, PhD, medical informatics researcher, University of Edinburgh.
PLOS Digital Health: “Predicting dementia from spontaneous speech using large language models.”
Frontiers in Computer Science: “Pauses for Detection of Alzheimer’s Disease,” “Temporal Integration of Text Transcripts and Acoustic Features for Alzheimer’s Diagnosis Based on Spontaneous Speech.”
Source: Read Full Article
