Book a demo

LastPatient Research · for research teams

Coming later

Causal estimates, from data built for them.

LastPatient Research is being built to run causal and Bayesian analysis on defined visit data. It estimates what would have happened under a different decision, for patients like this one, with a range showing how certain the estimate is.

Talk to our research team See how it works
Asking what would have happenedA search bar cycles through research questions, such as whether physical therapy first changes outcomes for meniscus tears. For each, a matching group of patients lights up, two recovery curves show how function changes over time with each choice, and an estimate of the difference narrows as more patients are counted, ending in a plain verdict such as leans PT first or no clear difference. All values are illustrative.EVERY VISIT, DEFINEDPatients like this oneSAME PATIENT, TWO PATHSFUNCTION ↑Today’s visitMONTHS AFTER →HOW SURE WE ARE

Why LastPatient Research

Research data dies in the note.

Written to bill

The note is written to get the visit paid. Researchers re-read it later and guess at the rest.

Every question starts over

Each new question goes back to the charts. 10,000 charts turn into 340 usable cases, and one answer takes 18–36 months.

Illustrative funnel, not a measured result.

Centers can’t pool

Each site records the same visit its own way. So a multi-center study starts with re-abstraction, reading charts by hand to pull the data back out.

The data

What causal methods need, and what the data layer records.

Most reports tell you what happened. Causal AI estimates what would have happened if you’d chosen differently. That estimate is only as good as the data under it, and it needs five things.

What causal methods need, and how the LastPatient data layer records it
What causal methods needHow the data layer records it
Facts recorded before the decision, timestampedSo the estimate only uses what was known at the visit, not what happened after.Each finding from the visit is stored as its own defined variable, with a record of where it came from and when.
The decision itself, as a defined variableSo the treatment chosen is a fact, not a guess made later from a billing code.The treatment chosen at the visit is stored as its own variable, in the same record as the findings.
Outcomes at 6 weeks, 3 months and 1 yearSo every comparison has an end point measured the same way.LastPatient Outcomes will collect PROMs, short standard questionnaires on pain and function, tied to the exact procedure and pre-op baseline.Coming next
One standard across every siteSo patients from different centers can be compared without re-reading charts.Centers using the same data structure record the same defined variables, in the same way.
Facts confirmed by the physicianSo the inputs are what the physician saw, not what a model inferred.The physician reviews the values and signs off before the note is closed. Each value keeps who confirmed it.

Causal AI needs to know why you made the call. The EMR records what you billed. LastPatient records what you saw.

Methods

The methods, in plain terms.

Causal AI is a family of statistical methods that estimate the effect of a decision, not just its correlation with an outcome. These are the methods the research module is being built around. Each gives an estimate with a range, never proof.

Coming later

The road not taken

Counterfactual estimation

Estimates the outcome a patient would likely have had under the treatment they didn’t get. It compares them with patients who looked the same at the visit (same tear, same age, same goals) but were treated differently, across thousands of patients and dozens of factors.

Start from what surgeons know

Bayesian updating

Start from what surgeons already know, update with each new result, and report how sure we are. The starting point is called a prior. Each new patient’s result moves it, so evidence adds up instead of starting from zero with every study.

The trial you’d never get funded

Target trial emulation

Write down the trial you’d want to run: who’s eligible, what each group gets, when follow-up starts. Then emulate it with visit data already recorded, with the same rules a real trial would follow.

Who actually benefits

Heterogeneous treatment effects

An average effect can hide groups that do much better or much worse. These methods estimate how the effect of a treatment changes with age, tear size, diabetes and other factors, and when it matters most.

How sure are we

Sensitivity analysis

Any estimate from real-world data could be bent by a factor nobody recorded. Sensitivity analysis asks how big a hidden factor would need to be to flip the answer. A small number means be careful. A large one means the result is hard to explain away.

Planned to sit beside every estimate, not in a footnote.

A worked example

Patients like this one: PT first, or an injection?

Newer methods estimate, patient by patient, how one treatment would compare with another, with a calibrated range of certainty.1 They only work on consistent, well-defined data. Here is what that could look like for one patient.

Illustrative, not a result

This patient, at the visit

  • Age58
  • Dxknee osteoarthritis
  • Severitymoderate
  • Pain6 / 10
  • Goalhike again

Compared with patients who looked the same at the visit but got the other treatment.

The decision

Estimated knee score at 1 year (0–100, higher is better)

  • PT first

    physical therapy, 8 weeks

    71range 64–77

  • Injection first

    steroid injection

    65range 57–72

Estimated difference, PT first minus injection

+6 pointsrange −2 to +13

The range crosses zero, so the honest reading is “PT first probably does better for this patient, but it isn’t certain.”

Dot: the estimate. Bar: a 90% range. If the method is well calibrated, the true value lands inside that range 9 times in 10. Fictional patient and values, not a result. Evidence for your conversation, your call.Illustrative example, not a result. A fictional patient, 58, with moderate knee osteoarthritis and pain of 6 out of 10, who wants to hike again, is compared with patients who looked the same at the visit but got the other treatment. One decision splits into two paths. PT first: estimated 1-year knee score 71, range 64 to 77. Injection first: estimated score 65, range 57 to 72. The estimated difference is 6 points in favor of PT first, with a range from minus 2 to plus 13, which crosses zero, so the benefit is likely but not certain.
Coming later

The research module isn’t built yet. This example shows what it’s designed to do.

Footnotes

  1. Example: Singh et al., “CausalGuard: Conformal Inference under Graph Uncertainty,” arXiv:2605.21928 (2026). A preprint tested on benchmark datasets; cited as an example of the method, not a model LastPatient runs. ↩

How it works

It starts with the data layer.

Every visit is stored as defined variables, and the note, the codes and the analysis all come from that same record. LastPatient works alongside the EMR. Research is built on that record, not on notes read back after the fact.

See how the data layer works

The return

Research without the chart review.

1 search

per new question

Model estimate. The alternative is 18–36 months of chart review.

Every center

one research database

Variables are defined the same way at every center, so the data pools without re-abstraction.

~18 mo

from first use to national publication

Early-user result, one surgeon.

Built for multi-center research

One standard, every center.

Same variables at every site

Variables are defined the same way at each center, so data pools without re-abstracting charts, and it lines up with the registries the hospital already reports to.

Outcomes attached to the visit

With LastPatient Outcomes, patient-reported outcomes, the scores patients give their own pain and function, will be tied to the exact operation and its baseline.

Coming next

Governance stays with the hospital

Data use agreements and IRB review still apply. We take away the prep work. The oversight stays where it is.

Private by design

Transcription happens on the device, and we work under a signed BAA. SOC 2 Type I is planned for 2027.

Running a multi-center study? Start with the visit.

Bring the next protocol. Our research team will walk through it.

Talk to our research team