New AI Model PodGPT Blends Research and Podcasts for Smarter Health Answers

Kolachalama Lab Launches PodGPT to Revolutionize Health and Science

By Maeve Smillie

In an age where medical misinformation runs rampant and science communication often falls flat, a new artificial intelligence tool promises to bridge the gap between expert knowledge and everyday understanding.

Introducing PodGPT: a tool trained not just on written material like textbooks and journal articles, but also on real conversations. This medical dialogue is sourced from science and medicine podcasts where experts talk through everything from heart disease to public health. By learning from the way people naturally talk about complex topics, PodGPT delivers answers that feel more relatable and easier to understand.

“By integrating spoken content, we aim to enhance our model’s understanding of conversational language,” said Vijaya B. Kolachalama, Associate Professor of Medicine and Computer Science, Hariri Institute Faculty Affiliate, and corresponding author of the study published in npj Biomedical Innovations. “This is special because it uses real conversations, like expert interviews, instead of just written material, helping it better understand how people actually talk about science in real life.”

PodGPT is the latest development in the expanding field of generative AI, where large language models (LLMs) like OpenAI’s GPT-4 are already transforming how we search for information, study, and make decisions. Yet, until now, most LLMs have been trained almost exclusively on text like books, articles, and scientific papers. That leaves out a rich, often more relatable source of knowledge: spoken conversation.

“We wanted to go beyond the written word,” said Kolachama. “By integrating spoken content, we aim to enhance our model’s understanding of conversational language and extend its application to more specialized contexts within STEMM [science, technology, engineering, mathematics and medicine] disciplines.”

To create PodGPT, the team compiled over 3,700 hours of publicly available podcasts and used advanced software to convert the speech into text. From there, they trained a language model and tested it on a variety of STEMM subjects, including biology, math, and medicine, sometimes in multiple languages. The results showed a clear boost in the model’s ability to understand questions and deliver precise, relevant answers.

But the bigger story is what this means for everyday people.

Too often, scientific and medical information feels out of reach. PodGPT helps close that gap and makes it easier for someone to understand a new diagnosis, research a health concern, or learn about public health issues without needing a medical degree. It’s also a step toward breaking language barriers, and can help people around the world access critical information in ways that feel natural and culturally familiar.

“This opens the door to using all kinds of multimedia options, including video or audio lectures or interviews, to build smarter, more human-like technology,” Kolachalama said. “It also shows promise in making science more accessible across many languages, helping people everywhere learn and stay informed.”

The researchers see potential applications in areas like Alzheimer’s disease, mental health, cancer, and global health education. The ultimate goal? Giving people not just answers, but answers they can actually use, grounded in expertise, but delivered to people in a way that resonates.

The study is published in npj Biomedical Innovations. 

This story has been adapted from Boston University’s Chobanian & Avedisian School of Medicine.