Medmio says new ultrasound coding model hit 99.2% accuracy

Medmio reported that its CodeSight model matched billed codes on diagnostic ultrasound reports 99.2% of the time in its first ultrasound benchmark. The result tops the company’s prior 98.1% score on office-visit charts and underscores how the startup is positioning public benchmarks as proof points for medical coding automation.

Why it matters: - Medical coding errors can trigger claim denials, rework, delayed revenue and patient billing issues. - A higher-accuracy model on diagnostic ultrasound suggests Medmio is expanding its coding system beyond office visits into more technical records. - The benchmark gives physician practices a public data point on whether AI can handle real billed claims at scale.

What happened: - Medmio said its newest CodeSight model coded diagnostic ultrasound reports correctly 99.2% of the time. - The benchmark covered 1,287 studies, with 254 held-out studies scored once at the end. - The result is the highest accuracy Medmio has published. - The company said the same model had not previously been tested on ultrasound notes. - Medmio published the full ultrasound benchmark online at the ultrasound benchmark.

The details: - The 99.2% score exceeded Medmio’s 98.1% result from July, when CodeSight was benchmarked on office-visit charts. - Office-visit charts are doctors’ written accounts of patient visits. - Ultrasound reports are technical imaging records that include protocol, findings and laterality. - Both tasks involve medical coding, which converts clinical notes into standardized diagnosis and procedure codes used for payment. - CodeSight combines several large language models with deterministic, policy-based rules. - Medmio’s platform also includes voice dictation, patient intake and mobile charge capture. - The benchmark followed the same protocol used in July: held-out studies were selected before development, scored once at the end and evaluated against a configuration fixed in advance. - Medmio said the full method and limitations are published with the result. - A manuscript is in preparation. - The July office-visit benchmark remains available in full.

Between the lines: - Medmio is using repeatable public benchmarks to frame product updates as measurable progress, not just feature launches. - The company is also signaling that ultrasound may be an easier step than surgical operative notes, which Medmio described as a substantially harder coding task. - President and CEO Levan Sulakvelidze said the company intends to benchmark every new model against real billed claims and publish the results either way.

What's next: - Medmio said a third benchmark on surgical operative notes is underway. - The company said it will publish that benchmark under the same testing terms. - Medmio also said future model upgrades will be measured in public against real billed claims.

The bottom line: - Medmio’s first ultrasound test suggests its coding engine can extend beyond office notes while maintaining very high accuracy, but the company still needs to prove that performance on harder, higher-stakes records.