The audio stream is processed in real-time during conversations and automatically deleted as soon as the audio is transcribed by Scribe. This ensures that the audio, which is highly sensitive, is not retained any longer than necessary to fulfil the necessary purpose.
Outputs generated by Scribe, such as transcriptions, clinical notes, and summaries, are retained as identifiable personal data for a period of 30 days for healthcare professional users to access.
During this 30-day period, a small number of outputs are extracted for clinical review, in line with statutory obligations relating to medical devices. The number of outputs selected for this process is derived from a statistically-robust methodology to ensure a meaningful analysis of error rates. In addition to these sampled outputs, any outputs identified by clinicians as erroneous or otherwise concerning are also included for review. Please see here for more information about how these outputs are processed for clinical safety purposes.
After this 30-day period, the majority of outputs are permanently and securely deleted. A small number, again derived from a statistically-robust methodology, are retained to support testing of any changes to the underlying AI model’s system prompt resulting from clinical review or other identified improvements. These outputs are:
- selected using a statistically-robust methodology to ensure meaningful regression testing can occur;
- minimised using Named Entity Recognition techniques to redact both structured and free-text patient data, resulting in outputs that align with the ICO’s definition of “effectively anonymised” (ensuring that for this specific use case, it is no longer considered personal data); and
- support statistically-valid post-market performance monitoring, detection of rare safety-relevant errors, and compliance with ISO 14971, ISO/TR 20416, and DCB0129 obligations.
Regardless of the steps taken to minimise personal data in the outputs, they continue to be treated with the same technical and organisational measures as if they were fully identifiable, further reducing the likelihood of unintended re-identification of individuals.
Accurx have also taken into account peer-reviewed research and other published guidance relating to data minimisation and re-identification risk, such as this.
.png){kind=small}
