Making Patient Data Usable: Q&A with Noga Leviner

Noga Leviner
CEO
Picnic Health

The life sciences industry runs on data. Every step of the drug development process is built around collecting as much data as possible. However, this can often lead to issues, as researchers often end up with mountains of data that isn’t always in usable formats. PicnicHealth CEO Noga Leviner spoke with Pharmaceutical Executive about how AI is solving this problem and creating new opportunities.

Pharmaceutical Executive: Is AI discovering new ways to organize massive data sets?
Noga Leviner: AI is changing how biopharma companies clean, organize, structure, and abstract information at a much faster speed and greater scale than a person or team of people ever could. We can now gather and curate data more efficiently and complete manual, tedious tasks easier to unlock new insights.

But AI isn’t a silver bullet. The key to any successful AI implementation is also having robust quality controls and clinical experts in place to validate and verify data accuracy and relevance.

Algorithms are able to retrieve specific points of data from massive sets with incredible speed. How is this impacting research/development timelines?

A big challenge in research is that data lives in silos, making it difficult to collect real-world data and evidence to evaluate the ongoing safety and efficacy of therapeutics. AI captures and organizes data much faster and helps find insights, patterns, and signals that may have been hard to see before.

Post-marketing safety studies, for example, are incredibly burdensome in terms of the time and cost it takes to collect data and extract insights. There is significant potential for AI to easily capture data from more sources and transform this process so researchers can better understand the safety of treatments and more effectively monitor and improve patient outcomes.

PE: How is AI able to take data sets in different, potentially incompatible formats, and present a unified story?
Leviner: The newest generation of AI, large language models (LLMs), is very good at synthesizing information from diverse, incompatible data formats. Once an LLM is trained on enough examples, it demonstrates remarkable flexibility in applying its knowledge to new and different datasets, just like a human would.

LLMs are proving to be especially valuable when working with medical records. When exposed to a wide variety of examples from lots of different care settings, EMR systems, specialists, and demographics, for example, LLMs dramatically improve their accuracy and performance –and continue to get better with hundreds of millions of clinical annotated examples.

Once again, a big key is having the infrastructure, both systems and people, in place to identify instances where AI cannot confidently match or interpret different formats, and course-correct. AI is great at handling the harder, more tedious tasks, but human intervention should be used to ensure the quality and accuracy of the story being told.

PE: What impact is AI having on the collection of RWE?
Leviner: AI is gathering and organizing information faster than ever before to help researchers make more informed, evidence-based decisions, and achieve complete, accurate, and traceable data that is fit-for-use when generating real-world evidence (RWE) and getting a complete picture of patient experiences.

But it’s important to have guardrails and systems around AI. Off-the-shelf AI by itself is not going to be trustworthy, reliable, or fine-tuned for research purposes. You need an AI that is purpose-built for a task, like interpreting medical records, and has visibility and plausibility checks in place.

It is critical to have an AI that not only has medical knowledge, but also has been domain-trained on patient records to make sure it’s producing high-quality, meaningful output for research.