Can’t recover what was never captured
The details that go missing aren’t random. Pile on more records and you pile on more bias.
How it works · The data layer
Most systems store a note and try to pull data out of it later. LastPatient’s own database structure defines the variables first, AI fills them from everything that happens in the visit, and the note is written from that record.
The structure
LastPatient’s proprietary database structure defines every variable and the kind of value it holds. The AI is given that structure and has to place what happens in the visit into those variables, nowhere else. Each value is weighed against the structure itself, the patient’s history and the context of the visit.
External rotation
30°
The database structure
Consistent: A whole number of degrees, in range
The patient’s history
Consistent: Up from 20° at the last visit
The visit context
Consistent: Matches the exam macro and what was heard
Saved to the record
Laterality
?
The database structure
Consistent: Must be left or right
The patient’s history
Unclear: Both shoulders treated before
The visit context
Unclear: Not said, typed or dictated
Ask at checkout
Coming laterMost important question first.
Coming laterAnything below the confidence threshold will go to the agentic checkout, where the agent asks for what it needs, most important question first.
Context
A follow-up after rotator cuff repair needs different things from a follow-up after an injection. The visit, the procedure, the timing and the patient’s history tell LastPatient which variables matter. The variables themselves come from the database structure.
Context
One patient’s record
Defined variables
Diagnosis
Exam
Imaging
Plan
One record
Every output is a view of the same record, so the note, the codes and the research data can’t drift apart.
Sources
One record
Defined variables
Each visit, stored as variables defined by the proprietary database structure.
Outputs
From the room to the record
Provenance
Each value keeps a record of its source and who confirmed it, so a value in a note, a code or a study can be traced back to the visit. That record gets richer as the platform grows.
Why structure at the visit
The details that go missing aren’t random. Pile on more records and you pile on more bias.
A value confirmed by the physician is ground truth. One pulled out of a note later is a model’s guess.
A new study or a new question doesn’t mean reprocessing every record.
Add your own variables with macros (dot phrases) and your own text, so the detail nobody planned a field for still gets saved.
Where it’s going
During onboarding, LastPatient will learn the patterns you repeat. When a visit matches one, the variables collapse to your own text.
A patient is 2 weeks after a left total knee replacement, and you say “standard 2-week post-op”. The variables collapse to your own text:
After the visit, the agent will ask only for what it still needs, most important question first. A question that’s still open, like “undetermined, awaiting CT”, will be saved as data, not treated as an error.
We’ll walk through one of your own visits.