Blogs

Can you tell us about Triomics and your role there?

I am Sebastian Rhodes and I lead strategy and operations at Triomics. Our mission is to align the latest advancements in generative AI with oncology. At the core is ONCO LLM—an oncology-focused large language model that's a constellation of models designed to efficiently and cost effectively extract key data elements from unstructured data sources.

In oncology, roughly 80% of key data elements needed for care and research decisions exist in unstructured documents. There's a high burden on physicians, clinical research staff, and nurses to manually review patient records to synthesize key insights. We've built an engine to automate that manual chart review process, and on top of it, we've built several downstream use cases.

What are the key use cases you've developed?

The core use case is automated screening of patients for clinical trials and clinical research—matching patients to trials in cancer care. These inclusion and exclusion criteria are extremely complex with tight time windows.

We've created a platform that can automatically screen a patient's record against all ongoing clinical research at an institution. At an institution with 600-700 active studies, it wouldn't be possible for nurses to screen every patient manually.

We also have pre-charting and summarization capabilities to help physicians quickly understand new patients who might have 20 years of medical history, currently working across academic NCI-designated research centers, community oncology, pharma, and a subset of data providers.

This shift from trial-to-patient to patient-to-trial matching seems significant. Why is that such a fundamental unlock?

The patient-to-trial model did not exist before generative AI. One physician might see 10-20 patients a day– with hundreds of trials, that's thousands of patient-trial pairs to evaluate.

Even at the largest academic centers focused on research, they weren't measuring what percentage of patients were getting screened for clinical research. We heard anecdotally it was only 20-30%. Our core metric with partners is that every single patient who has an appointment gets screened for clinical research, and if they're a good potential match, that insight is provided to the physician and care team before the appointment.

Instead of being reactive or expecting patients to know about research opportunities, we can detect these matches ahead of time and make that appointment more useful.

What steps has Triomics taken to drive transparency and trust with clinical partners?

We do extensive validation early on. We've published in journals like Nature Digital Medicine with multiple academic research partners, showing 95% accuracy in our platform for identifying patients for eligible studies.

In terms of model architecture, when we work with an institution, we either bring the model directly into their environment with no external data transfer, or if they're comfortable with our HIPAA-compliant, SOC 2 Type 2 cloud, their environment is completely partitioned. When we start working with a customer, each institution gets their own version of ONCO LLM.

All learnings and refinements from user interactions are contained for each customer. None of the patient data or user interactions are shared across institutions.

How did Triomics become involved with CHAI?

As a technology company, we recognized there are things we don't have expertise on that we need to learn from provider colleagues who see patients day-to-day. At the same time, we can't expect physicians to know everything about generative AI the way we do as a technology company. We joined CHAI because we wanted to provide back those learnings and be part of the team that helps set reliable roadmaps for deployment going forward.

How do you envision CHAI resources helping organizations like yours accelerate AI adoption?

It's been a roller coaster with larger academic research centers and health systems as they actively try to define their own AI guidelines and policies while deploying technology like ours. You can iterate on requirements, and three weeks later there's a whole new list.

We're in the first few innings of enterprise AI deployment at scale. I think the best way to narrow in on what works is to share experiences across institutions, vendors, providers, community and academic settings—instead of everybody working in silos on the same problem. In healthcare research, if someone developed a novel therapy or surgical method, it would be shared everywhere. In the world of AI, we haven't seen that as much yet, but CHAI is critical to scaling adoption and getting to the best place faster.

Where do you hope to see the most progress for AI in clinical trials?

We targeted trial enrollment because it's the biggest area AI can address today. When patients are asked if they'd be open to enrolling in a study targeted to their condition, upwards of 70% say yes. Yet only around 5-7% of patients enroll in clinical oncology research.

While some of that gap exists because of social determinants of health and geography, a huge factor is identifying eligible patients at the right time. As we move toward personalized medicine, the complexity of inclusion/exclusion criteria is only increasing. That problem is getting bigger, not smaller, and it is a really apt use case for generative AI.

What are you most excited about with Triomics being part of CHAI?

The opportunity to create a forum where learnings in technology are shared in a structured, meaningful way to help accelerate all of our adoptions—not only vendor adoptions but hospitals and physicians coming up the curve on generative AI.

CHAI is the main forum right now bringing together a broad collective to have these conversations and get learnings down on paper. The healthcare industry has been very poor about tech adoption overall, but there's a big opportunity with generative AI. We'll only see the fruits of that labor if we double down on sharing technology insights the same way we share insights when we have new therapies or surgical methods.