Medmio says its medical coding AI hit 98.1% accuracy on unseen charts

Medmio released a public benchmark for its CodeSight AI medical coding engine and says the system matched a vascular practice’s billed level 98.1% of the time on charts it had never seen. The company also published the underlying evidence so the results can be independently checked.

Why it matters: - Medical practices are under pressure to automate billing because experienced coders are in short supply. - Medmio is positioning CodeSight as a reviewable alternative that can assign codes with confidence scores and send uncertain charts to a human coder. - The company is also making a broader claim about transparency in healthcare AI by publishing the full evidence behind the benchmark.

What happened: - Medmio released the Medmio EM-945 benchmark as a public measure of how accurately CodeSight reads clinical notes and assigns billing codes. - On charts the system had never seen, CodeSight matched the exact billing level a vascular practice actually billed 98.1% of the time. - The result carried a 95% confidence interval of 96.2% to 99.1%. - The benchmark was measured without custom training on the practice’s data. - Any chart below Medmio’s confidence threshold was routed to a human coder for review.

The details: - The benchmark was built from 945 billed-and-paid charts. - The test set was locked before the result was measured. - Medmio scored the system against the codes the practice actually billed. - The company published confidence intervals, a calibration analysis and the complete accuracy-coverage curve. - CodeSight is the coding engine inside Medmio’s broader platform, which also includes voice dictation, patient intake and mobile charge capture. - Medmio says CodeSight combines an ensemble of large language models with deterministic, policy-driven rules to keep outputs accurate and consistent. - Medmio says CodeSight integrates with major EHRs via HL7 standards. - The full report is available as a CodeSight Benchmark. - Medmio’s platform also includes autonomous medical coding, voice dictation, patient intake and mobile charge capture.

Between the lines: - The benchmark is limited to office visits at one practice, so the result is a narrow test rather than a broad industry-level validation. - Medmio is using the release to signal that it wants outside verification, not just vendor-controlled claims. - The human-review fallback suggests the company is trying to balance automation with billing risk management.

What's next: - Medmio says it will publish future results the same way. - The company plans to test larger datasets and a broader, more complex range of codes. - Medmio says it wants openness to be a standard for its own technology and, by extension, for healthcare AI more broadly. - CEO Levan Sulakvelidze said the company sees room for higher quality performance as CodeSight is applied to more practices.

The bottom line: - Medmio is not just claiming high coding accuracy; it is also publishing the evidence package behind the claim in a way that invites scrutiny.