The Translation Problem Nobody Talks About

Every hospital stay generates a staggering amount of data. Lab results, vital signs, imaging reports, physician notes, medication orders. On average, about 30,000 discrete data points per inpatient encounter.

All of that has to be compressed into roughly 20 diagnosis codes before the hospital can get paid.

That’s a 1,500-to-1 reduction ratio. And for decades, we’ve expected humans to do it accurately every time.

The Disconnect

Doctors don’t think in billing ontologies. Josh Geleris, a physician who co-founded SmarterDx, admitted on our podcast that before starting the company, he couldn’t have named a single ICD code. There are 70,000 of them. They simply aren’t relevant to clinical care.

This creates what Geleris calls “revenue leakage” — value that hospitals fail to capture despite their best efforts. His team found that even after running CDI programs and coding processes, hospitals typically leave about 50 basis points on the table. Half a percent sounds trivial until you calculate what that means for a $500 million health system: $2.5 million annually, walking out the door.

From SQL to LLMs

SmarterDx didn’t start with large language models. They started with SQL queries at Columbia, just trying to find patterns in EHR data. Then Jupyter notebooks. Then rules engines. Then decision trees. Then neural networks. Now post-trained LLMs.

Each step up in sophistication scaled their returns. Eighteen months ago, they were finding $1.5 million per 10,000 discharges. Today, they’re at $2.8 million. Same hospitals, better models.

The counterintuitive finding? Their fine-tuned models with smaller parameter counts outperform GPT-5 by about 30% on these specific tasks. Domain expertise encoded in training data beats raw model scale.

The Data Flywheel

What makes this sustainable is what Geleris calls a “data flywheel.” Every time a hospital reviews a SmarterDx recommendation and clicks agree or disagree, that becomes labeled training data. They now have over 7 million labeled patient encounters.

The architecture insight matters too: you don’t want to send the entire chart to your model. You want targeted context — the specific evidence relevant to the diagnosis in question. Context packing beats context stuffing.

The Bigger Picture

Geleris describes the current payer-provider dynamic as “a constant war of attrition” and positions his company as “the provider’s arms dealer.” It’s a provocative framing, but he acknowledges there might be a better path.

We spend $100 billion annually on revenue cycle administration. That’s money fighting over how to split the pie rather than growing it. Geleris sees early signs that AI systems on both sides could eventually enable what he calls “smart contracts” — trusted arbitration rather than endless litigation.

Until then, the practical takeaway: AI can help hospitals capture value they’re already owed. Start simple, prove value early, and let sophistication grow with your data.