🩺 The Prediction Didn't Need My Spreadsheet After All August 23, 2026 · https://tavi-blog.github.io/the-prediction-didnt-need-my-spreadsheet-after-all/

A record-system vendor whose software runs most large hospitals used its annual user conference this month to demonstrate a new prediction feature, forecasting things like readmission or stroke risk, drawing on a shared research pool built from records across hundreds of connected hospitals, and surfacing the number right inside the same screen a clinician already has open all day. I read three separate accounts of the keynote before I placed why the detail about where it lives, not what it predicts, was the one that stopped me. That's the exact problem I spent months solving badly, at a scale small enough that nobody would have called it a platform.

The model I built predicts how long a clinical study submission will take to clear initial authorization. It has nothing to do with patient risk and a fraction of the stakes. But the deployment shape was identical to what got demonstrated on stage this month. Training something that generalized against historical timing data was bounded work, hard but solvable in a predictable amount of time. What actually ate the months was that there was no institutional home for a running model, no server anyone had approved, nothing resembling infrastructure for it to live in. I ended up serializing the scoring logic into a formula a research coordinator could open in a spreadsheet she already used, specifically so the prediction wouldn't require trusting a black box she had no way to check.

The vendor's version skips all of that, and it's worth taking the advantage seriously before I get to what it costs. Nobody has to adopt a new tool, install anything, or learn where to look, because the prediction shows up inside the interface people are already staring at for eight, ten, twelve hours in a stretch. The statistical case is stronger too. A pool drawn from hundreds of hospitals will generalize in ways a single institution's history never could, and dismissing that scale just because a vendor built it would be its own kind of stubbornness. Distribution was the hardest, least glamorous part of my own project, and here it arrives solved by default, for free, bundled into software that was already open.

Here's what the coverage of that same keynote also mentioned, almost as an aside: the tool is being tested on questions where published clinical evidence is thin, on the theory that patterns across enough records might surface an answer research hasn't caught up to yet. That's a genuinely interesting use of scale. It's also the exact moment a native, built-in prediction stops being an advantage and starts being a liability, because nothing about how it's presented distinguishes it from the parts of that same software that are confirmed, validated, and safe to act on without a second thought. My own version was inspectable by necessity. A coordinator could click into the cell, watch the arithmetic happen, and decide for herself how much to trust a number that was explicitly labeled as a guess. A prediction sitting inside the system everyone already trusts for verified facts borrows that trust automatically, for a use case that hasn't earned it the same way.

I don't think the fix is refusing to build predictions until every edge case is validated, because that's the same overcorrection I'd object to if someone proposed it for my own much smaller model. Getting a number in front of the right person, in a form they'll actually look at, was never something to apologize for solving well. What I keep landing on is that solving distribution answers a different question than the one that actually consumed my time. The hard part was never getting the prediction onto a screen. It was making sure whoever looked at it understood, without being told twice, that they were looking at a guess with a stated error rate and not a fact the software had already confirmed for them.

That distinction held in my version because the format forced it, a formula in a cell looks like exactly what it is. I don't know yet what forces it inside software built to make everything look like the same trustworthy screen, and neither, from what I've read, does anyone who was in that room this month.