The Beat

How Limina Is Making Sensitive Health Data Safe for AI

Episode Summary

In this episode, host Sandy Vance sits down with Patricia Thaine, co-founder and chair of Limina (formerly known as Private AI), for a fascinating conversation about one of the most underappreciated bottlenecks in healthcare AI adoption: the privacy of unstructured data. With a background in natural language processing and privacy research, Patricia built the company from the ground up to solve a problem most organizations did not even know they had. Today, her platform helps health systems, research organizations, and payers de-identify everything from clinical notes to ambient listening data so they can train models, share data for research, and move their AI initiatives forward without putting patient privacy at risk. If your AI initiative is stalled because of privacy concerns, this episode is exactly what you need to hear.

Episode Notes

In this episode, host Sandy Vance sits down with Patricia Thaine, co-founder and chair of Limina (formerly known as Private AI), for a fascinating conversation about one of the most underappreciated bottlenecks in healthcare AI adoption: the privacy of unstructured data. 

With a background in natural language processing and privacy research, Patricia built the company from the ground up to solve a problem most organizations did not even know they had. Today, her platform helps health systems, research organizations, and payers de-identify everything from clinical notes to ambient listening data so they can train models, share data for research, and move their AI initiatives forward without putting patient privacy at risk. 

If your AI initiative is stalled because of privacy concerns, this episode is exactly what you need to hear.

In this episode, they talk about:

• 80 to 90% of healthcare data is unstructured, and most organizations have no idea what sensitive information is hiding in it
• Cloud providers require you to send your data outside your environment, and that alone is a dealbreaker for many health systems
• De-identification is not just about removing names; quasi-identifiers like age ranges, locations, and diagnoses all factor into re-identification risk
• The goal is to keep re-identification risk below 0.04%, not just strip out obvious fields
• Training AI models on real PHI creates a memorization problem where the model can regurgitate patient information in production
• Providence Health has used Limina since the early days to train patient and physician-facing chatbots safely
• A mature privacy-to-AI operating model requires statisticians, product teams, IT, governance, and legal all at the table
• LIMINA rebranded from Private AI because the old name kept attracting requests for on-premise LLMs, which is not what they do

A Little About Patricia:

Patricia Thaine is the Co-Founder & Chairwoman of Private AI, a Microsoft-backed startup that raised their Series A led by the BDC. Private AI won the Privacy Innovation Award at PICCASO 2024, was named a 2023 Technology Pioneer by the World Economic Forum, and was a Gartner Cool Vendor. Patricia is also the host of The Data Frontier podcast and was on Maclean’s magazine Power List 2024 for being one of the top 100 Canadians shaping the country.