Generative AI is a term that is now echoed in every sphere. But did you know the capabilities of generative AI are not limited to creating photographs, artworks and writings?  

Jakob Nikolas Kather, Narmin Ghaffari Laleh, Sebastian Foersch & Daniel Truhn has put forward the research that generative AI models can create synthetic images very close to or indistinguishable from real images.  

According to the study, generative models like generative adversarial networks (GANs) and variational auto-encoders (VAEs) can create synthetic images in an unconditional way or conditional to a defined set of classes. Both approaches require dedicated training to a specific domain, for example, x-ray images. 

In the medical application of AI, generative models have been used in radiology, histopathology and endoscopy. 

Detecting Alzheimer’s 

AI algorithms behind ChatGPT have drawn attention for their ability to generate human-like responses to some of the most creative queries. 

In this scenario, a researcher from Drexel University’s School of Biomedical Engineering, Science and Health Systems recently demonstrated that Open AI’s GPT-3 program could identify clues from spontaneous speech that are 80% accurate in predicting early stages of dementia.  

Since there is still no cure for the disease, diagnosing it early can give patients more options for therapeutics and support. The current practice for diagnosis involves medical history review and a lengthy set of physical and neurological evaluations and tests. 

 Hualou Liang, PhD, a professor in Drexel’s School of Biomedical Engineering, Science and Health Systems and a co-author of the research, believes that the improvement of NLP programs provides another path to support early identification of Alzheimer’s.  

Testing the theory 

To prove the theory, two tests were conducted by the researchers:  

Test 1: The program captured meaningful characteristics of the word use, sentence structure and meaning from the text to produce what researchers call an “embedding”. They used the embedding to re-train the program, turning it into an Alzheimer’s screening machine.  

Result: Running two of the top NLP programs through the same paces, the group found that GPT-3 performed better than both in terms of accurately identifying Alzheimer’s examples, identifying non-Alzheimer’s examples and with fewer missed cases than both programs.  

Test 2: The second test used GPT-3’s textual analysis to predict the score of various patients from the dataset on a common test for predicting the severity of dementia, called the Mini-Mental State Exam (MMSE). 

Results: The text embedding generated by GPT-3 can be reliably used to not only detect individuals with Alzheimer’s Disease from healthy controls but also to infer the subjects cognitive testing scores, both solely based on speech data.  

Generative models in medicine 

Synthetic data generated by such AI systems are looked at with hopeful eyes by the medical community as a promising approach for data augmentation, data sharing and explainability in medical AI.  

Presently in medical AI, the systems focus on a ‘narrow niche’ with a single type of data and are validated thoroughly in this particular niche. Although these narrow AI systems have shown performance on par with experts and are very valuable in their specific domain, they fail to generalize or adapt to slight changes in the inputs.  

However, according to the research by Jakob Nikolas Kather, Narmin Ghaffari Laleh, Sebastian Foersch & Daniel Truhn, an important shortcoming of generative AI in medicine is the limited scope of such specialized systems, which often need to be laboriously trained to generate images in a single narrow domain. In their opinion, training on vast amounts of unselected text-image pairs scraped from the internet also conveys useful medical knowledge.