We’re Overcomplicating Healthcare AI — Here’s How We Achieve Impact Beyond The Hype

The early days of any foundational technology are characterized by overhype and skepticism. When the Internet first hit the scene in the 1990s, enthusiasts were awed by the potential to search for nearly anything online – even if nascent webpages weren’t entirely useful – and imagined opportunities for a more connected world. Skeptics saw the Internet as a fad that would never take over traditional means of communication or research. As time went on, there was a boom and a bust – but eventually, Internet applications turned from cool novelties to practical solutions to real problems. We’ve seen comparable cycles play out with mobile phones and social media. Today, we’re at a similar juncture for artificial intelligence (AI) in healthcare.

When most people think about AI in healthcare, they envision futuristic applications like “AI doctors” helping to diagnose cancer and AI agents that automatically scan your body for cardiovascular events before they happen. While these innovations show the promise of AI for the future, most have yet to be deployed and validated at scale. If we want to find useful AI applications in healthcare today, we need to look further upstream to where it can be transformative in solving long-standing, complex problems.

Clinical research is one such field that is ripe for AI’s impact. Today, data management remains a significant challenge, and clinical researchers rely on manual processes to organize and analyze mass amounts of health data. While in most areas of our lives we’re accustomed to typing queries into a search engine and receiving rapid answers, most clinical researchers are still stuck using spreadsheets to store, exchange, access, and analyze data. As the quantity and complexity of multimodal health data has grown with the rise of omics, the tools we used to use for analysis are not working. Outdated processes lead to research delays and stall innovation.

This is where AI comes in – not with hype, but with a real solution to clinical researchers’ data management woes. AI systems – particularly large language models (LLMs) – can help researchers organize data into ready-to-use formats and empower them to interact with their data more efficiently.  

Let’s dive a bit deeper into how these technologies are already enabling scientific breakthroughs in clinical research.

One area where we’re seeing AI add value today is in accelerating data harmonization. In its source form, clinical data is very messy. Researchers typically spend months parsing through digital silos and disparate data formats to harmonize information into a common, research-ready structure. With the advent of LLMs, researchers can speed up this process by 100 times through AI-powered automation. LLMs harmonize mountains of data from registries, research databases, and electronic medical records (EMRs) into a controlled vocabulary that’s primed for analysis. Downstream, researchers can forgo time-consuming data wrangling and get straight to the  science. 

Another high-value application is in extracting structured data from previously unusable “unstructured” data, such as the free-text notes from an EMR. Because this information – which includes important details on a patient’s medication usage, for example – isn’t organized into a structured format, it’s often excluded from research databases. LLMs can map this text from clinical notes to more structured fields, answering questions like “did this patient receive radiotherapy” or “did this person have a cancer recurrence.” This works with images as well. For example, we use AI-powered computer vision to extract metadata from histopathology images like cell counts, cell types, staining, and more. Once it’s structured, all of this data can then be made available for research – accelerating the time to insight. 

AI-powered user interfaces are also radically improving the accessibility of data, creating a new paradigm for interactions between researchers and datasets. Rather than needing to learn the technical intricacies of databases – or working with data specialists who know how to write code – AI-powered interfaces allow researchers to “chat with data” using natural language. Rather than writing a line of code, researchers can simply type, “Find me all patients that have had breast cancer of pathological stage II or higher and that have pathogenic variants on the genes associated with Lynch’s Syndrome,” into the AI-powered user interface. The LLM understands the request and converts it into a query on the harmonized dataset, returning an answer in seconds instead of weeks. Not only does this accelerate time to insight, it democratizes access to data, opening the door for even more medical experts to get involved in research – especially those without coding expertise. 

I believe hype and skepticism are both good when it comes to AI. These points of view will push the boundaries of what’s possible while reeling in the less effective use cases. But it’s important that we focus attention on the areas where AI is already having a significant impact. These are the practical applications that will allow the best and brightest researchers to focus on science rather than data munging. Let’s get to work on those.

Photo: rudall30, Getty Images

This post appears through the MedCity Influencers program. Anyone can publish their perspective on business and innovation in healthcare on MedCity News through MedCity Influencers. Click here to find out how.