Healthcare AI Is Failing And the Fix Isn’t More Data

 

If Netflix recommends a bad documentary, you scroll past it. Maybe you roll your eyes. Life goes on.

But what happens when an AI model misses something critical in a clinical trial, such as when a group of patients stops responding to a drug that looked promising in earlier phases?

The consequences aren’t an annoying recommendation. They’re measured in human lives.

That comparison, offered by Dr. Abhishek Jha on the SaaS That App podcast, perfectly captures the tension at the heart of AI in life sciences. Dr. Jha is the Co-founder and CEO of Elucidata, a company that’s spent the last decade building AI-driven data platforms for drug discovery and precision medicine.

His core message is one the broader tech world needs to hear: the AI playbook that works for consumer tech doesn’t translate to healthcare. Not even close.

 

The Data Gap Nobody Talks About

When most people think about AI, they think about scale. Facebook has 3.5 billion daily logins. Amazon tracks billions of product interactions. Google indexes the entire internet. These companies built their AI empires on oceans of data, and the models got better as the oceans got deeper.

Now compare that to healthcare. The richest, most well-funded clinical trial in history might include tens of thousands of patients. Many rare diseases and small cancers have far fewer data points than that. As Dr. Jha puts it, the scale is “off by orders of magnitude, not even close.”

This is a fundamental constraint that changes what kind of AI you can build and how you should build it. Traditional machine learning leans hard on massive datasets. In life sciences, that luxury simply doesn’t exist.

 

Why “Out of Distribution” Is an Existential Problem

In machine learning, there’s a well-established concept called “out-of-distribution” detection. Basically, if your model encounters data that looks different from what it was trained on, its predictions become unreliable.

In consumer tech, that means a weird movie recommendation. In a self-driving car, it could mean failing to detect an obstacle. In drug discovery, it might mean missing the reason a promising therapy suddenly stops working in a subset of patients.

Dr. Jha’s team at Elucidata has made this challenge central to their work. They’ve built benchmarking systems that evaluate how models perform as task complexity increases, and what they’ve found is sobering: many models that look impressive in published papers collapse when tested against real-world proprietary data. The performance drop-off is dramatic, and it explains a lot about why AI adoption in life sciences has been slower than the hype suggests.

 

Flipping the Script: Data-Centric AI

Elucidata’s answer to all of this is what they call data-centric AI, an approach championed more broadly by Andrew Ng of Stanford and Coursera fame. Instead of obsessing over model architecture and parameter counts, data-centric AI focuses on the quality, preparation, and context of the data going into models.

In practice, this means curating and linking multiple types of data for the same small patient population, cleaning and structuring that data carefully, and building what Dr. Jha calls data priors that improve model performance without changing the model itself. His team recently presented research showing consistent improvements across five different model architectures, all achieved by improving the data.

 

The Business Model Had to Change Too

The technical challenges ripple into how Elucidata runs as a business. The classic SaaS model didn’t fit. Each customer’s use case requires enough customization that a one-size-fits-all product just doesn’t work. Instead, they’ve embraced a managed services model that combines platform access with hands-on expertise. Some engagements wrap up in three months; others have run for six years.

Dr. Jha and his team tried traditional SaaS pricing. It didn’t stick. They were “trying to force-fit a model that was not natural to our offering.” When your AI solution requires deep domain context, the pricing model needs to reflect that reality.

 

The People Problem Underneath the Tech Problem

Perhaps the most striking takeaway from the conversation is Dr. Jha’s insistence that the hardest challenges in this space aren’t technical; they’re human. Leadership teams are told by their boards to “do something about AI” without clear guidance on what that means.

The discourse online swings between apocalyptic doom and dismissive skepticism. Scientists and engineers chronically over-index on technology and under-invest in educating stakeholders, building trust, and solving distribution problems.

After a decade of building in this space. Dr. Jha’s advice to aspiring founders is equal parts encouraging and sobering: if you feel genuinely compelled to do this work, go for it. But understand that the toll is real. No Excel model of projected returns will justify the journey on its own. It has to come from something deeper.

And maybe that’s the most important thing separating AI in life sciences from AI everywhere else. The stakes demand it.

 

Dr. Abhishek Jha’s Background

Dr. Abhishek Jha is the Co-Founder and CEO of Elucidata, a biotech company focused on AI-driven data solutions for life sciences R&D and drug discovery. He has over 20 years of experience at the intersection of life sciences, data science, and machine learning. Abhishek co-founded Elucidata in 2015, launching Polly as a cloud-based SaaS platform that harmonizes multi-omics, clinical, and real-world data into ML-ready formats, speeding data preparation by 10x with high accuracy via LLM-powered curation.

Listen to the new episode on: