A medium-sized breast imaging facility with a single interpreting physician was managing a high-volume workflow. At the time, the facility was also dealing with issues involving its dictation system, the kind of operational friction that adds pressure to an already busy environment.
A group of mammography cases came through for interpretation. The physician dictated the reports. They were transcribed, signed, and released through the normal internal process. Nothing flagged. Nothing held. The reports moved forward.
Approximately four hours after the reports were signed and released, one of them reached the Mammologix inputting and coding team. The team didn't enter the assessment code and move on. They read the report.
The body of the report described a screening mammogram with no suspicious findings. The narrative was clear: normal result, no acute findings, routine follow-up. But the imaging assessment assigned to that same report was suspicious. And the recommendation was for an immediate biopsy.
The report's narrative said the patient was fine. The assessment code said she needed a biopsy immediately. One of them was wrong.
The patient lay letter had not gone out, and it never would have. Mammologix reviews every report for internal consistency before any patient communication is generated. An inconsistency like this one does not proceed to a letter. It gets flagged immediately and held for resolution. The communication workflow does not move forward until the discrepancy is resolved. This is not a quality check applied occasionally or to high-risk cases only. It is applied to every report, every time.
Mammologix notified the facility immediately. Because the report had already been finalized and released, an addendum was issued and the referring physician was notified. The facility reviewed the original dictation, identified the source of the mismatch, and corrected the record. The patient never received any communication. The error was resolved entirely before it reached her.
Without that review, the patient would have received a letter telling her she had a suspicious finding requiring urgent biopsy, based on a report whose actual clinical findings described nothing of the kind. She would have had no way of knowing the letter was wrong. Her physician, reached by a panicked call, would have been reading from a report that had already been corrected but whose original error had already reached the patient.
A peer-reviewed study of 213,977 speech-recognition-generated radiology reports from 147 radiologists found that 9.7% of reports contained errors, and 1.9% contained material errors believed to potentially alter the interpretation of the report. (Ringler, Goss, Bartholmai. Journal of Innovation in Health Informatics, 2015/2017.)
This is not a story about an unusual event at an unusual facility. Speech recognition dictation errors in radiology are a documented, peer-reviewed finding in the literature. At 1.9% material error rate, a facility interpreting 5,000 mammograms a year could see approximately 95 reports per year with errors significant enough to potentially alter clinical interpretation. Not every one of those errors will involve a mismatch as stark as this case. But some will, and in a fully automated communication workflow, those reports become letters before anyone has read them.
The Mammologix review process is not a pass-through. It looks at why the patient presented, what examination was performed, what the report narrative describes, what the impression states, what follow-up is recommended, and whether the assigned imaging assessment is consistent with all of that. Those elements are reviewed together. A code-in, letter-out model cannot do this, because the inconsistency in this case existed inside a single signed report, invisible to any process that reads only the assessment field.
Patient communication is the last step in the clinical chain, and the step that reaches the patient directly. That is exactly the wrong place to find out something was wrong earlier in the process.