The Future of AI in Healthcare

The 2024 HiPaaS Healthcare AI Conference in Burlingame, CA, brought together leaders in healthcare technology to discuss the latest advancements and challenges in artificial intelligence for the healthcare industry. The Healthcare AI panel included Nitin Shingate, Chief Technology Officer at GoodRx; Sandeep Deokule, Founder and CEO of HiPaaS Inc.; and our own Zoltan Gombosi, Senior VP of Engineering at Certify. In this blog, hear from Zoltan Gombosi as he discusses generative AI's current applications, future possibilities, and crucial considerations for responsible implementation within the healthcare industry.

Moderator: Why do we need AI in healthcare?

Zoltan: On the provider data management side, a major challenge is data disparity. Currently, credentialing, licensing, and enrollment processes require pulling data from numerous sources, determining which information belongs to which provider, verifying its accuracy, and ultimately deciding if a provider is in good standing. We currently aggregate data from over 500 sources, but that represents only about 10% of the estimated 5,000+ provider data sources in the US. Managing this volume of data necessitates entity resolution—identifying if an entity is the same across multiple sources, essentially de-duplication. AI is ideally suited for this. Entity resolution is a known problem in AI, with companies working on solutions, some more successfully than others. We're applying AI in this area and have found that while traditional machine learning algorithms have been used for de-duplication, large language models (LLMs) offer significant advantages. LLMs excel at entity resolution and, crucially, offer explainability—you can ask an LLM why it made a specific decision, which isn't possible with traditional neural networks without significant additional effort. This opens up new techniques and opportunities.

Another use case is summarization. At the end of the credentialing process, findings like sanctions or malpractice claims need to be summarized for a credentialing summary. This is currently a manual process. There's no reason for manual typing when summarization is a core strength of generative AI. We see substantial opportunities here.

Moderator: There are valid concerns about AI safety, regulation, and governance, as well as ensuring AI assists rather than replaces providers. What processes, guidelines, and best practices do you have in place to ensure safety protocols are followed?

Zoltan: There are established governance and security processes around AI. We're not reinventing the wheel. If you ask a large language model like Gemini about compliance requirements for AI, it generates a reasonable list covering governance, privacy, ethics, and other key areas. For me, a few things stand out. When using AI—and often we’re specifically discussing generative AI—we are in the business of accuracy and trust. We want clients to trust our platform and its decisions and recommendations. Generative AI can "hallucinate," generating incorrect or fabricated information. This necessitates having a human in the loop. We can't let AI make decisions autonomously without checks and balances to maintain trust and ensure accuracy. We have these checks in place, and our development process incorporates feedback from clients regarding their specific requirements for AI usage.

Generative AI raises significant data security questions. There have been reports of models leaking information provided in user prompts, not just from their original training data. This is a crucial consideration and key reason why we deploy private models rather relying on publicly available services, like ChatGPT.

Moderator: What was your transition into healthcare?

Zoltan: My first role after graduate school was in healthcare, working for a company developing electronic medical records. They were building a system to analyze data from anesthesiology machines during operations. I was present in the operating room during software rollouts, which solidified my decision not to pursue a medical career—I barely stayed conscious! This sparked my interest in healthcare data. After working in other industries and then transitioning into master data management (MDM), I returned to healthcare data. Life sciences companies were early adopters of MDM technology, and provider data was a key focus. When I sought a new opportunity, I was drawn back to healthcare where I could work on technologically interesting projects with real-world value.

Moderator: What is your vision for the future of healthcare, and what role do you see your team and company playing?

Zoltan: It's difficult to predict, especially given the rapid emergence of generative AI. We're in a phase of rapid development, and there will undoubtedly be new innovations we can't foresee. Currently, we're focused on applying AI techniques to existing provider relationship management (PRM) areas to improve efficiency, accuracy, and reliability. We started with summarization, a natural fit for generative AI. While chatbots are common now, I believe the next evolution will be AI providing more substantial value beyond simple interactions. This aligns with the broader trend of focusing on value and outcomes in healthcare. At the end of the day, I don't want to do AI just to do AI. I want to do AI because it brings some significant value to our clients and improves outcomes for the healthcare industry as a whole.

Audience: I heard that the average time to hack an LLM is five minutes. Do you think it will be possible to use LLMs with sensitive company data, beyond just regular ChatGPT use?

Zoltan: That's a critical question that needs to be addressed by those developing LLMs, like OpenAI and Google. We're primarily consumers of these models. As mentioned by others on this panel, deploying LLMs for sensitive applications like healthcare requires placing them behind a firewall, ideally as private LLMs. They absolutely should not be directly connected to the outside world. They should be integrated with your applications using pre-trained or purpose-built models. We use Vertex AI on GCP, which allows us to deploy our own models privately, leveraging the security of the underlying platform. Cloud providers like AWS and Azure offer similar capabilities. Protecting the data we handle is our responsibility. There's also an education component. We need to help Clients understand the real risks versus perceived concerns. Conversations like this are crucial for fostering that understanding.

Conclusion

This panel discussion at the HiPaaS 2024 Healthcare AI Conference underscored the significant impact AI is poised to have on the healthcare industry. Zoltan Gombosi believes that practical applications of generative AI can benefit the healthcare industry. While challenges remain, particularly surrounding data privacy and the need for human oversight, the potential for AI to drive efficiency, improve accuracy, and ultimately enhance patient outcomes is undeniable. As the technology continues to mature, collaboration and open dialogue, as discussed at the conference, will be crucial for navigating the evolving landscape of generative AI in healthcare and ensuring its beneficial impact on the industry.

You can watch the entire interview on YouTube.