Basic artificial intelligence models are rapidly evolving across the healthcare ecosystem. System integration plays an indispensable role in ensuring the use of generative AI results in safety, security and reliability.
Furthermore, effectively and responsibly integrating a domain-specific AI model with the broader healthcare system is a critical part of ensuring a trusted AI environment.
Srini Iyer is vice president and chief technology officer at Leidos Health. At the HIMSS24 Global Conference and Exhibition in March in Orlando, Leidos and Google will address the ongoing challenge of achieving trust and security with genAI by demonstrating their collaboration on the Medical Pathways Language Model 2 (MedPaLM2), highlighting use cases to demonstrate the criticality of designing of confidence in genAI to maximize benefits for healthcare organizations.
We spoke with Iyer to get a preview of his HIMSS24 educational session titled “The Impact of Domain-Specific Models on Health AI.”
Q. What is the overarching focus of your session? Why is this important to healthcare IT leaders in hospitals and healthcare systems today?
A. Generative AI models represent a huge change in the field of AI. Specifically, the impact of AI on healthcare highlights the benefits and potential of using AI models trained on medical data for various healthcare tasks. This session highlights the potential of domain-specific AI to revolutionize healthcare by delivering more accurate, efficient and cost-effective care.
According to the Gartner Healthcare Provider Research Panel, June 2023A majority of respondents (85%) believe AI big language models will have a significant to disruptive impact on healthcare, while 14% rate it as a moderate impact.
There are several use cases of interest to healthcare IT leaders in hospitals and healthcare systems. Chief among these are automated data analysis, automatic document generation, and EHR search and summary. They should be interested in this topic for several reasons:
- Improved accuracy and relevance. Domain-specific healthcare AI models, such as Med-PaLM 2, are trained on large amounts of medical data, allowing them to understand and answer complex medical questions with greater accuracy and relevance compared to generic AI models.
- Better patient outcomes. More accurate analysis of medical data can lead to faster diagnoses and better treatment plans.
- Streamlined workflows and administrative tasks. AI can automate routine tasks, allowing healthcare providers to focus on important patient care.
- Increased efficiency. Domain-specific AI models require less data and training time than traditional AI models, making them more scalable and cost-effective to implement. This can be especially beneficial for smaller hospitals and healthcare systems with limited resources.
In the coming years, more than half of the generative AI models used by companies will be domain-specific, up from 1% today. Domain-specific AI can act as a valuable assistant for healthcare professionals, giving them instant access to relevant medical information and insights, ultimately improving decision-making and patient care.
Q. What is one of the most important lessons you want HIMSS24 session participants to walk away with?
A. In a short period of a few months, Leidos with a small team developed a successful Med-PaLM 2 Proof of Concept to validate trustworthy genAI in healthcare, demonstrating how trust and security can be seamlessly integrated into genAI systems to maximize benefits for healthcare organizations.
We selected a use case that focuses on the three most important needs of healthcare executives. Medical professionals play a critical role in delivering quality care, but their time is often challenged by administrative tasks such as completing complex medical reports.
Agencies like the VA, SSA, and CMS require detailed documentation, but generating reports places a significant burden on physicians, impacting both efficiency and accuracy. The private healthcare sector also faces the same challenges.
We got better responses and our accuracy improved when we used vector storage. These are ideal for generative AI applications because they allow finding relationships between unstructured data points and help LLMs remember these relationships over time.
There were challenges we encountered and addressed while working on this project:
- Length and complexity. Reports can be extensive and require navigation through complicated sections and fields, which takes a lot of time and attention.
- Information overload. Physicians may need to consult several sources and references to accurately complete these reports, which often adds to the time burden.
- High error potential. The sheer volume of information and complexity of the sections can increase the risk of errors, potentially impacting patient care and reimbursement.
Q. What is another lesson you would like session participants to take away with?
A. People with AI skills are hard to find and often expensive. Building generative AI capabilities within a company is a journey, not a destination. We have been able to provide our teams with hands-on experience to learn skills related to developing, training and deploying models.
We were able to collect many lessons learned along the way. AI platforms and tools are still maturing; Applying these evolving tools to support your specific use case requires experimentation, deep knowledge, and patience.
We had early access to some of these domain-specific models and knew that documentation for these rapidly evolving tools was limited. Our developers had to collaborate with product teams and go through an iterative process to determine the right path forward.
Access to good data is crucial for the success of these healthcare projects. This can be a major challenge for healthcare IT as we deal with PII/PHI and HIPAA compliance. This limits access to real-world data, meaning we have to rely on synthetic data or anonymized data.
As an early adopter of implementing foundational models in healthcare, we are cautiously optimistic that we can address some of our critical healthcare challenges to improve patient safety, resulting in better outcomes for our patients.
The session, “The Impact of Domain-Specific Models on Health AI,” is scheduled for March 12 from 3:00 PM to 4:00 PM in Room W208C at HIMSS24 in Orlando. More information and registration.
Follow Bill’s HIT coverage on LinkedIn: Bill Siwicki
Email him: bsiwicki@himss.org
Healthcare IT News is a HIMSS Media publication.