Limits & review guidance
Suggestions are a starting point.
NoteCoder is experimental. It can miss a condition, propose an unrelated code or give a high score to an unsupported description. Human review is part of the workflow, not an optional final step.
Matching can miss wording
Candidate discovery relies on packaged phrases, official aliases and proposed relationships. It is not exhaustive understanding of every possible clinical expression. A condition without a matching anchor may never be sent to the model.
Highlights are not evidence
A highlight explains how a candidate was found. It may describe another person, a past episode, a suspected condition or a negated diagnosis. Related matches and proposed details are especially important to verify in context.
Scores are not probabilities
A score reflects the model's answer under the selected question and model. It is not a calibrated clinical probability. A cutoff only filters the screen; it does not establish a diagnosis, coding validity or absence of disease.
Specific codes need specific support
Check every required part of the code description: site, stage, side, severity, subtype, timing and complications. The presence of two conditions does not by itself establish a causal relationship between them.
HCC scope is a mapping filter
Medicare CMS-HCC V28 and ACA HHS-HCC 2026 are separate models. Both shows their union without duplicating ICD codes. A code absent from the selected mapping can still represent a diagnosis in the note.
ACA categories may involve age/sex edit conditions and different scoring-population rules. Displayed tags do not decide patient eligibility. This interface does not calculate demographic coefficients, hierarchy effects or a patient risk score.
Data and model boundaries
The code catalog is pinned to the FY2026 April 1 ICD-10-CM update. It should not be assumed to include later releases. The ACA source supports January–September 2026 coverage; its FY2027 mapping column is empty. New October–December 2026 codes and mappings are not included. Some legacy phrase mappings are retained, and additions do not certify the entire dictionary as correct.
The model checks only the discovered candidates. Model choice, prompt wording and the review method can change suggestions. Limited development testing does not establish general clinical accuracy.
Browser support, memory and long-note limits can prevent analysis. Notes are not silently shortened to fit: a limit error requires your attention. The model download and local cache can use substantial storage.