What is Patchs AI?
Patchs AI is an optional feature of Patchs that automates parts of the triage and workflow when processing patient requests.
The aim of Patchs AI is to reduce GP practice workload and improve patient safety.
Patchs AI has been available since January 2021, and all GP practices that use it have found it both useful and safe.
There are different Patchs AI modules, which do different things:
- Urgency AI flags patient requests as urgent or emergency as soon as they are submitted.
- Signpost AI signposts patients away to self-care, NHS 111, or emergency services as appropriate.
- Topic AI works out the clinical topics of a patient request and can be used to ask patients further questions before they submit their request.
- Assign AI assigns requests to a ‘Clinical’ inbox if they need input from a clinician.
- Face-to-Face (F2F) AI highlights requests that may need an in-person consultation so they can be booked in sooner.
How does Patchs AI work?
Patchs AI using Natural Language Processing (NLP) and both deep and shallow machine learning to analyse free text responses from patients.
Patchs AI has been ‘taught’ how to do this by the triage decisions made by GP practice staff using Patchs each time they process real life patient requests (this is called supervised learning), and through other unsupervised learning techniques.
In addition to the free text written by patients, Patchs AI also analyses the type of request chosen by the patient (e.g. health problem vs admin), and their age and sex to make its decisions.
Patchs is not a keyword search or pre-programmed algorithm, though for certain AI modules we sometimes use algorithmic approaches to improve performance.
Patchs AI has learned from the experience of hundreds of Patchs users across the country who have triaged thousands of real patient requests (see: the wisdom of the Crowd). And it's learning more every day - each time a user makes a triage decision they are teaching Patchs how to triage similar patient requests in future.
Because Patchs AI makes clinical decisions, we primarily use triage decisions made by clinicians to teach Patchs AI rather than triage decisions from non-clinical staff (although we do use triage decisions made by non-clinical staff when they don’t need clinical input). Therefore when a request needs clinical input, GP practices must assign requests to clinicians within Patchs so they can add triage decisions.
Patchs AI can therefore deal with real-world patient health problems, spelling mistakes, and uncertainty - all of which are essential for General Practice. However, just like the humans that teach it, Patchs AI isn’t perfect. It isn’t 100% accurate – nothing is.
How is Patchs AI regulated?
Patchs AI is considered a Class I Medical Device because it offers ‘triage and signposting of next steps based on filters by severity and probability of a match’, but without ‘direct diagnosis’.
It is registered with the Medicines and Healthcare products Regulatory Agency (MHRA) and is UKCA marked. This means we undertake rigorous clinical evaluations and safety assessments before releasing PATCHS AI modules to customers, and conduct ongoing post-market surveillance after they are released. This process includes multiple stages of testing including on historic patient requests, newly collected patient requests, interviews with staff and patients, and direct comparisons with GPs.
Because Patchs AI is used in the NHS, it is also governed by DCB0129: Clinical Risk Management: its Application in the Manufacture of Health IT Systems. This means it regularly undergoes clinical risk assessment and has its own Clinical Safety Case Report, which should be read and understood by GP practices before it is used.
How should Patchs AI be used by GP practices?
Patchs AI is meant to assist, not replace, human decision-making. It is an optional feature of Patchs and is used by GP practices at their own risk.
We recommend that GP practices become familiar with using Patchs without AI first so they fully understand how Patchs works.
If GP practices use Patchs AI before this point they may become reliant on it, and because Patchs AI isn’t 100% accurate, the risk with being reliant on Patchs AI is that they will not have processes in place for the cases where Patchs AI is incorrect.
As a guide we recommend you become familiar using Patchs without AI at levels of incoming requests recommended in this article before switching on Patchs AI, i.e. between 12 requests per 1000 patients on your practice list per week and 12 requests per 1000 patients on your practice list on an average working day.
We would recommend using Patchs at these levels for at least three months - though for some practices this may be longer.
Practices can start using Patchs AI before this point, though they do so at their own risk.
Because it isn’t 100% accurate, staff should still read, respond, and triage all patient messages as soon possible when using Patchs AI .
To conduct post-market surveillance of Patchs AI as required by medical device regulations, we monitor triage decisions made by staff to ensure that it is safe.
Because Patchs AI makes clinical decisions, we primarily use triage decisions made by clinicians to do this rather than non-clinical staff (although we do use triage decisions made by non-clinical staff when they don’t need clinical input). Therefore, in order to use Patchs AI, when a request needs clinical input, GP practices must assign requests to clinicians within Patchs so they can add triage decisions.
If GP practices do not do this, we cannot monitor the safety of Patchs AI and they will be unable to use Patchs AI.
How does Patchs AI work with Patchs Translate and Telephone Assistant?
Patchs is primarily meant to be used in English. However, GP practices can choose to use Patchs Translate.
This feature allows patients to submit requests in non-English languages that are then automatically translated into English by Patchs.
Patchs AI has primarily learned from requests written by patients in English, so non-English language requests are first translated to English then analysed by Patchs AI. This may reduce the accuracy of Patchs AI.
To make translated patient requests quicker and easier to identify, there is a language icon next to them in the Patchs inbox.
Patchs Telephone Assistant allows patients to use the telephone to submit Patchs requests by speaking to an automated voice assistant.
Their voice message is then automatically transcribed into a written message, which can be analysed by Urgency AI, Signpost AI, Assign AI, and Face-to-Face (F2F) AI.
Patchs AI has primarily learned from requests written by patients, not spoken by them. Therefore, Patchs AI may be less accurate when analysing messages submitted via Telephone Assistant.
Updating Patchs AI
We constantly monitor the performance of Patchs AI using triage decisions made by GP practice staff.
If GP practice staff change a decision made by Patchs AI, we consider Patchs AI to have been incorrect. As we collect more data, we use it to re-train Patchs AI to keep it up to date.
Periodically, we release these new versions of Patchs AI - but only once we are satisfied they perform better than older versions.
What work are we doing on the use of AI in general?