Opportunities for Generative Artificial Intelligence in Medical Research

AI Efficiency in Medical Data Analysis

In early trials of generative artificial intelligence in medical research, experts from the University of California, San Francisco, and Wayne State University found that AI can process large volumes of medical data much faster than traditional teams of computer science specialists. In some cases, the results achieved with AI even surpassed those obtained by human teams, who typically needed months to analyze similar datasets.

To objectively compare effectiveness, researchers assigned identical tasks to different groups: some relied solely on human expertise, while others used AI tools. The task was to predict preterm births based on data from over 1,000 pregnant women. Even less experienced researchers, with the help of AI, were able to quickly develop working predictive models. The system generated code in just minutes, whereas experienced programmers would usually spend hours or days on the same task. The advantage of AI lay in its ability to create analytical code from brief and specific prompts. However, only half of the tested AI systems managed to produce usable code, but the successful tools did not require large specialist teams for management.

Thanks to the high speed of AI, junior researchers were able to complete experiments, verify results, and submit them to scientific journals within just a few months. Such AI tools can help eliminate one of the main bottlenecks in data science—building analytical pipelines.

The Importance of Preterm Birth Research

Accelerating data analysis can foster the development of diagnostic tools for identifying preterm births, which are a leading cause of newborn mortality and long-term health issues in children. In the United States, about 1,000 babies are born prematurely every day.

The causes of preterm birth are still not fully understood. To identify possible risk factors, microbiome data was collected from around 1,200 pregnant women across nine separate studies. Such analysis is only possible with open data sharing and collaboration among different research groups.

However, processing such large and complex datasets proved to be a significant challenge. To address this, researchers turned to the global DREAM crowdsourcing competition, where over 100 teams worldwide developed machine learning models to identify patterns associated with preterm birth. Most groups completed their work in three months, but consolidating the results and publishing them took nearly two years.

Testing AI on Pregnancy and Microbiome Data

To test whether generative AI could shorten analysis timelines, research teams tasked eight AI systems with independently generating algorithms on the same datasets used in the DREAM challenges, without direct human programming. The AI chatbots received carefully crafted natural language instructions and analyzed medical data in a manner similar to the original competition participants.

The tasks included analyzing vaginal microbiome data to detect signs of preterm birth and studying blood or placental samples to estimate gestational age. Estimating gestational age is often imprecise, yet it determines the type of medical care women receive. Inaccurate estimates complicate birth preparations.

As a result, only four out of eight AI tools created models comparable in effectiveness to those produced by human teams, and in some cases, AI even delivered better results. The entire process—from starting the work to submitting the article—took just six months.

Prospects and Limitations of AI Use

AI systems require careful oversight, as they can produce erroneous results, and human expertise remains essential. However, thanks to the rapid processing of large medical datasets, generative AI allows researchers to spend less time debugging code and more time interpreting results and formulating meaningful scientific questions. This opens up opportunities for specialists with limited data science experience to focus on solving important biomedical problems without the need for large collaborations or extensive technical implementation.