Challenges with LLMs in Disease Diagnosis and Treatment: An Interview with Michael N. Liebman (Co-Founder of IPQ Analytics) Posted on : Jun 12 - 2026

Large Language Models (LLMs) are creating new opportunities in healthcare, but important challenges around accuracy, bias, explainability, and patient safety remain.

I recently interviewed Michael N. Liebman , Managing Director and Co-Founder of IPQ Analytics, to discuss the challenges of using LLMs in disease diagnosis and treatment.

1.Can you tell us about your background and your work in healthcare analytics and precision medicine?

I have a mixed academic/industry background, having gone back and forth several times after my graduate training in theoretical chemistry and protein crystallography. I was Global Head of Computational Biology and Genomics for Roche Pharmaceuticals, Executive Director of a US Dept of Defense sponsored Breast Cancer Center, and have academic appointments at Drexel College of Medicine, Univ Massachusetts (Lowell) and Fudan School of Medicine (Shanghai) I have also recently launched a nonprofit, Woven, focused on women’s health research

2. LLMs are being rapidly explored in clinical settings—what are the biggest challenges in using them for disease diagnosis and treatment today?

LLM’s operate as a “black box” in that they are not transparent as to what they have been trained on….and there can both be great variability in quality of research publications, etc as well as the reality that information/data/terminology evolves over time and that is rarely if ever considered when seeking information from an LLM, hence the basis for our research project.

3. What are the key risks healthcare organizations should consider when deploying LLMs in diagnosis workflows (e.g., hallucinations, bias, liability, data quality)?

LLM’s continue to become more conversant and convincing about the quality and completeness of their responses to questions, yet do both hallucinate and also can provide variable answers when a specific question is repeated.

4. What are the most important takeaways attendees can expect from your session on “Challenges with LLMs in Disease Diagnosis and Treatment”?

While LLM’s can provide great support for exploring and summarizing the literature, etc, we need to more deeply focus on establishing the correct question to present for analysis and understand the current constraints on the answers that will be provided.

5. Looking ahead, what needs to happen—technically, clinically, or regulatorily—for LLMs to become truly trusted in healthcare

There must be greater transparency as to what the basis of literature, etc was used in training or improve the query process to recognize how great variability in quality of research publications, etc as well as the reality that information/data/terminology evolves over time may be addressed

Key Takeaways

LLMs have tremendous potential to improve healthcare, but their successful adoption requires rigorous validation, high-quality data, strong governance, and human oversight. As Michael highlights, AI should augment healthcare professionals—not replace clinical expertise.

Bio

Michael N. Liebman is a computational biology, translational medicine, and digital health leader with experience across academia, biotech, and pharma, including leadership roles at Roche, Wyeth, and the University of Pennsylvania Cancer Center.

He is currently Managing Director of IPQ Analytics and advises on disease modeling, AI-driven risk detection, healthcare strategy, and quantum computing.

His work focuses on computational models for clinical decision-making, pharmaceutical risk-benefit analysis, and women’s health, including cardiovascular disease, multiple sclerosis, breast cancer, pregnancy-related disorders, and health disparities.

Topic : Challenges with LLM's in disease diagnosis and treatment

Abstract : 

The use of Large Language Models, LLM’s, is increasing rapidly in medicine, both by patients and physicians.  The most common uses focus on diagnosis and treatment and are being used to influence clinical decision support, but foundational medical research evolves rapidly impacting data and interpretations on a daily basis.  

Take home lessons:

1. LLM’s are extremely sensitive to the data upon which they are trained and this is very critical in a rapidly evolving field

2. Gaps and conflicts may appear and disappear over the window selected to provide training data

3. While simply repeating a query is known to produce different results, complex queries can produce even greater variability, e.g. “diagnosis” vs “treatment” vs “diagnosis and treatment”

Want to learn how healthcare leaders are turning AI from hype into real-world impact? Join our AI in Healthcare Virtual Conference on June 25 and hear from top experts, innovators, and practitioners shaping the future of healthcare with Generative AI, Agentic AI, and Clinical AI.

Register today and gain actionable insights, emerging trends, and practical strategies you can apply immediately. https://www.eventbrite.com/e/global-ai-in-healthcare-virtual-conference-tickets-1988295787223