Decision support

How Mira manages risk in its predictive decision support

Mira's EMR includes features that use AI models to help clinicians: drafting visit notes, proposing chart updates, orders and codes from a signed note, answering questions about a chart, summarizing a patient before a visit or after surgery, reading implant labels, and turning dictation into structured entries. Mira does not train these models. It uses large language models from Anthropic and Google under business associate agreements, and those vendors do not train on our customers' data.

Every one of these features produces a draft, a proposal or a summary. Nothing reaches a patient's record, an order, a prescription or a bill until a clinician reviews and confirms it. Each proposal shows where in the chart or note it came from.

Before release, and before any change to a model or its instructions, we evaluate each feature for validity, reliability, robustness, fairness, intelligibility, safety, security and privacy, comparing the new version against the current one on test charts. We review clinician edits, declines and feedback at least monthly. Patient data reach a model only from the patient's own chart, under the same access rules a person follows, and only to the extent the feature needs.

Users can see the source information for each feature in the EMR: who developed it, what it is for, its known limitations, and how it is validated and monitored. Where we have not yet measured something, such as performance broken down by patient group, the EMR says so, and where information is not available for review it shows "Information not available for review". The same information is available for the rule-based alerts Mira supplies. Practices can record their own information about each feature as well.

Questions about Mira's decision support: Mira Technologies, Inc., at support@mirahealth.care.