Longitudinal clinical review
AI medical record summarization has to preserve what changed.
A short summary is not enough when medications, diagnoses, test results, and follow-up actions change across encounters. A useful review helps a reader reconstruct the patient's changing state and verify the evidence behind each important claim.
The review problem
Five questions a longitudinal review must answer
- 01What happened, and when?
- 02What changed since the earlier encounter?
- 03Which findings, treatments, and problems are related?
- 04Which claims are supported, conflicting, historical, or uncertain?
- 05What follow-up or conditional action remains undocumented?
Why context gets lost
Generic summaries can be fluent while flattening the history.
Common failure patterns include showing a discontinued medication beside an active order without making the state clear, repeating a copied-forward problem-list label after a later report challenges it, combining events from different encounters without preserving their dates, describing a relationship without showing the supporting source, or turning missing documentation into an unsupported recommendation.
Changing state
Active, historical, superseded, held, and undocumented states need different labels.
Source evidence
A reviewer should be able to inspect the document, page, date, and evidence state behind a material claim.
Contradictions
A later test or specialist record can challenge an earlier assertion. A smooth paragraph should not hide that conflict.
CareNote's research direction
A patient-specific temporal graph is the approach CareNote is testing.
CareNote turns fragmented clinical-record PDFs into a patient-specific temporal graph, then produces a longitudinal synthesis or temporal answer with links to the source record and page. The intended review surface separates source-record assertions from generated synthesis, relationships from clinical conclusions, and undocumented follow-up from recommendations.
CareNote is research-stage. It does not diagnose, recommend treatment, determine what a patient is actually taking, or replace clinical judgment. The current implementation works from uploaded PDFs rather than a direct EHR integration.
A fair first evaluation
Compare graph-assisted review with ordinary document review.
Use a fixed synthetic or appropriately governed de-identified record set. Ask the system and the reviewer to reconstruct state at a specified time, identify material changes and dates, trace claims to source pages, find contradictions without silently resolving them, identify unresolved follow-up, and mark what is supported, ambiguous, conflicting, or undocumented.
Record missed events, unsupported additions, wrong citations, false resolution of contradictions, and false confidence. Positive reactions are not clinical validation.
Questions people ask
What is longitudinal chart review?
Longitudinal chart review is the reconstruction of a patient's clinical course across records created at different times, by different providers, and often for different purposes. The task includes chronology, changing status, relationships among events, source verification, contradictions, and unresolved work.
How is this different from AI medical record summarization?
AI medical record summarization usually describes the output format: a generated account of the record. Longitudinal chart review describes the reasoning task: reconstructing changing state and checking the evidence across time. A useful summarizer can support that task, but should not hide dates, source pages, uncertainty, or contradictions.
Can citations make an AI medical summary safe?
No. Citations make a claim easier to inspect, but they do not prove that the system interpreted the source correctly or included every important event. Source evidence, explicit uncertainty, and human review remain necessary.
Is CareNote clinically validated?
No. CareNote is a research-stage prototype with no external pilot proof or clinical outcome evidence. It is exploring whether a patient-specific temporal graph can help a knowledgeable reviewer organize and verify fragmented records.
Continue the research
Read the technical architecture, inspect the public proof, or see how CareNote is framing the search for evaluators. The research question is open to critique.