ChatGPT Could Become Alzheimer's Screening Tool, Study Suggests

ChatGPT Could Become Alzheimer's Screening Tool, Study Suggests unknown

The now-famous artificial intelligence (AI) program, ChatGPT, has made waves in recent months because of its ability to construct human-like text responses. A new report suggests the tool might also be able to use its language-analysis skills to help diagnose people with early signs of Alzheimer’s disease.

In a new study published in PLOS Digital Health, investigators from Drexel University in Philadelphia tested the AI program ChatGPT to see whether it was able to pick up subtle language cues that are associated with Alzheimer’s.

ChatGPT is a language model developed by the research company OpenAI. The initials in its name stand for General Pretrained Transformer. The program was trained on vast amounts of text and information, allowing it to answer questions and construct text in an easily readable format that mirrors the grammar and syntax of human writing. A preview version of the software was made available to the public in November.

While ChatGPT has generated attention for its ability to generate text that sometimes seems indistinguishable from human-written text, the new study evaluates ChatGPT’s ability to do the inverse: inspect transcriptions of human speech to identify speech patterns and usage that seem unnatural.

Hualou Liang, Ph.D., a Drexel biomedical engineering professor, and Feli Agvaor, a doctoral student, explained that one of the ways clinicians can identify early onset Alzheimer’s is by analyzing speech. Roughly 60-80% of people with Alzheimer’s have language impairment symptoms.

“The most commonly used tests for early detection of Alzheimer's look at acoustic features, such as pausing, articulation and vocal quality, in addition to tests of cognition,” Liang said, in a press release. “But we believe the improvement of natural language processing programs provide another path to support early identification of Alzheimer's.”

The investigators used transcripts of speech from people with Alzheimer’s to create a characteristic profile of Alzheimer’s speech that is sensitive to the subtle differences in how people with the disease construct sentences, use words, and sometimes forget the meaning of words. The authors then used that profile to train ChatGPT-3 to screen text samples for signs of Alzheimer’s. The goal was for the program to distinguish transcripts of speech from people with Alzheimer’s from speech by healthy controls.

Liang and Agbavor found ChatGPT was able to accurately predict Alzheimer’s about 80% of the time using the training profile created for the study. It outperformed two other natural language processing programs that were analyzed, they said.

The investigators next decided to see whether this type of analysis could be used to predict patients’ performance on other Alzheimer’s assessments. They asked ChatGPT to predict patients’ scores on the Mini-Mental State Exam (MMSE), a commonly used set of questions that are used to assess the severity of patients’ dementia. They found ChatGPT was about 20% more accurate at predicting MMSE scores than conventional assessments based on acoustic features.

Liang said the research suggests ChatGPT could be a promising diagnostic tool. He and Agbavor also highlighted a handful of important caveats. First, they noted that speaking patterns vary across geographic regions and across demographic groups, so it would be important that a diagnostic tool be based on data from all around the world and be assessed for bias. Patient privacy would need to be protected. And there needs to be more trust in AI before it can be widely used for healthcare applications, particularly since machine learning programs can make predictions using ultra-complex “reasoning” that even its human developers cannot explain.

Still, Liang said the study shows that if these problems can be solved, tools like ChatGPT could be a meaningful addition to the healthcare syste.

“Our proof-of-concept shows that this could be a simple, accessible and adequately sensitive tool for community-based testing," Liang said. "This could be very useful for early screening and risk assessment before a clinical diagnosis."