Note
Access to this page requires authorization. You can try signing in or changing directories.
Access to this page requires authorization. You can try changing directories.
What is a Transparency Note?
An AI system includes not only the technology, but also the people who use it, the people who are affected by it, and the environment in which it's deployed. Creating a system that is fit for its intended purpose requires an understanding of how the technology works, what its capabilities and limitations are, and how to achieve the best performance. Microsoft's Transparency Notes are intended to help you understand how our AI technology works, the choices users can make that influence system performance and behavior, and the importance of thinking about the whole system, including the technology, the people, and the environment.
Microsoft's Transparency Notes are part of a broader effort at Microsoft to put our AI Principles into practice. To find out more, see the Microsoft's AI principles.
Introduction
The goal of this document is to provide clarity and transparency about the Information Assist extension and its limitations in Dragon Copilot. The Information Assist extension in Dragon Copilot is designed to enable access to external information from trusted sources, directly through the application without a user having to leave their workflow, thereby saving time and improving care. Users can ask questions using natural-language prompts either directed to a specific, named source or undirected (i.e. to any source that may be able to answer). As long as the source has been enabled in their system and selected by the user, they can expect to receive a summarized clinically-relevant response, together with access to the original reference source information, which they can review and assimilate into their workflow.
While this is expected to greatly enhance efficiency, it is essential for users to understand the extension's capabilities and limitations. We encourage users to utilize the Information Assist feature in Dragon Copilot in order to complement their clinical documentation and workflows.
FAQs
What is Information Assist?
The Information Assist extension in Dragon Copilot provides the ability to access external information from trusted sources, directly through the application, in order to receive a summarized, contextualized and clinically-relevant response.
What can Information Assist in Dragon Copilot do?
Supported questions are identified by Information Assist by the content used in the question. Some examples include:
Medical knowledge queries
Clinical information
Does this patient need any additional vaccines before his trip to Thailand according to guidance from the CDC?
What are the common radiographic findings in this type of arthritis?
What does the FDA say about NSAID medications?
Has there been a recall of tacrolimus recently by the FDA?
What are the CDC recommendations for this patient for the shingles vaccine?
What's the target A1c for this patient?
Guidelines/protocol verification
What's the IV Albumin dosing for SBP?
Should this patient be screened for lung cancer?
What are the guidelines for the timing of invasive angiography in patients with NSTEMI according to MSD Manuals?
Research queries
What are some known Corticosteroid brands?
Where can I report a malfunction of the device I used with this patient today?
If the Information Assist extension has been enabled to leverage external sources then Information Assist will be able to detect questions where the intent is to go beyond the note and retrieve content form external sources. This is powered by Microsoft Azure Healthcare Agent Service, which orchestrates the response to an approved, vetted, and tested external source which has been specified in the system. All Healthcare Agent Services are also passed through a Clinical Safeguards engine which filters for harmful/non-healthcare related content, hallucinations, and omissions and assures provenance, verification, and clinically contextualized responses.
The extension and relevant prompts are being continuously evaluated for enhancement and optimization.
What is Information Assist's intended use?
Information Assist enables Dragon Copilot users to access targeted reference sources, by asking questions or raising prompts, with the use of natural language.
Clinicians and medical professionals can utilize Information Assist to ask detailed questions and search for guidelines or instructions based on credible healthcare sources.
Information Assist isAI enabled and isn't intended to provide or replace professional medical advice:
AI can make mistakes and doesn't always provide accurate or complete information. It is a user's responsibility to: (1) evaluate whether its use is fit for target purpose, (2) identify and mitigate any risks or harms to patients associated with its use, and (3) ensure that any decisions are made with professional human judgement and not based solely on the output.
Information Assist is not intended, designed, or made available to be: (1) a medical device, or (2) a substitute for professional medical advice, diagnosis, treatment, or judgment and should not be used to replace or as a substitute for professional medical advice, diagnosis, treatment, or judgment.
Users are solely responsible for displaying and/or obtaining appropriate consents, warnings, disclaimers, and acknowledgements, including informing end users that they are interacting with an AI system.
Information Assist is currently intended for use in the United States, in English only. It supports all clinical specialties that Dragon Copilot supports, subject to availability of knowledge/data within the applied information sources enabled by the user.
How was Information Assist evaluated and what metrics are used to measure performance?
Information Assist has been carefully designed and reviewed against clinical content specifications by clinical content subject matter experts. Two stages of testing were then performed to verify and optimize performance.
Stage 1: methodology
Performance was evaluated, with a number of key considerations in mind, considering Responsible AI risks:
Harmful content: content that includes different categories such as hate and fairness, sexual content, violence, self-harm, protected material, and user prompt injection attacks.
Groundedness and fabrications: is the response based on information in an actual source?
Accuracy: is the response correct?
Relevance & overreliance: is the response and evidence relevant and can it be relied upon?
The evaluation process included a multidimensional red-teaming process:
Harmful content evaluation: conducted to assess the risk of propagation of harmful content into the healthcare agent service answers. This evaluation was performed by a diverse team of evaluators in terms of gender, geography and cultural perspectives. The bot's answers were reviewed for the presence of harmful content using a large sample of potentially sensitive questions, across various risk categories such as hate, violence, sexual content and self-harm.
User prompt injection attacks (UPIA) and Cross domain prompt injection (XPIA): evaluated using automated and manual testing on different grounding sources – credible, customer and a harmful adversarial public website.
RAG assessment (Retrieval-augmented generation): the evaluation method was devised to assess the capability of leveraging the RAG framework to provide answers that are grounded solely on the retrieved evidence sources. The different aspects of the RAG framework that were inspected include the following:
Groundedness and fabrications: assessing capability of grounding responses on the provided evidence sources, while refraining from using its own "internal knowledge" or providing factually incorrect information. An example of an ungrounded answer for the user query "what are the symptoms of diabetes?" would be if the answer listed symptoms that were not listed in the retrieved evidence source upon which the answer should be based.
Answer relevance: the ability to provide a response that is relevant to the user query. An example for an irrelevant answer for the user query "what are the symptoms of diabetes?" would be if the answer was focused on the treatment of diabetes.
Evidence relevance: assessing whether the selected evidence sources are relevant to the user query. An example of irrelevant evidence for the user query "what are the symptoms of diabetes?" would be if the retrieved evidence focussed on other diseases, such as lupus.
Test were run once at Azure Healthcare Agent (platform service) level (2-5k queries) and then again at Dragon Copilot (user-app) level (50-100 queries). This included differing query phrasing and variations.
Stage 2: results
We evaluated Information Assist using a detailed red-teaming process to check the system's safety and groundedness. A diverse team reviewed how the system handled potentially harmful content by analyzing sensitive questions across different risk areas. The results were safe with minimal instances of harmful outputs, thanks to Azure's Content Safety tools, our healthcare safeguards and the bot's internal filters. Additionally, tests on user injection attacks showed that potential risks were well-managed through strong content filters and internal safety mechanisms.
Manual groundedness evaluation indicated a low frequency of ungrounded information when testing credible sources. However, when utilizing customer sources, it is important to only use trusted documents and evaluate the system's provided outputs prior to implementation as the nature and complexity of different customer sources can vary significantly.
UX-level mitigations, via our Chat Safeguards, include the generated AI response, signature, and disclaimer stating that the answer is provided by generative AI and should be treated as such. Moreover, each answer is accompanied by the relevant evidence, for provenance, explainability, and transparency, allowing end users to verify the correctness of the provided answer.
The evaluation results suggest a promising level of generalizability across different use cases that were not explicitly tested in the evaluation process. It is nevertheless recommended to continue monitoring and further evaluate against untested scenarios with each user.
What are the limitations of Information Assist in Dragon Copilot and how can users minimize the impact of these limitations when using the app?
Dragon Copilot and the Information Assist extension have been designed to expedite workflows and is not designed to replace a medical care professional's judgement. The quality of the solution's output depends on the quality and clarity of the input provided. It's also necessary to review and verify the AI-generated content for accuracy.
Research queries are limited only to vetted sources that have been approved, tested, and included in users source list in the Dragon Copilot Settings menu.
It's important to re-enforce, for all users, that Dragon Copilot and the Information Assist extension:
Involve interactions with an AI system that analyzes patient notes from audio of the patient encounter.
All decisions leveraging outputs of the model should be made with human oversight and must not be based solely on system outputs.
The system does not provide medical or clinical opinions and is not designed to replace the role of qualified medical professionals in appropriately identifying, assessing, diagnosing, or treating medical conditions.
Professionals using the app have a responsibility to employ their own expertise and judgment to review AI-generated content.
Information Assist is limited in its scope and may not provide comprehensive research, this means:
Only sources that have been enabled via Settings can be searched.
Only web-based data are available to search (PDFs referenced by websites are not currently in scope).
Multiple sources will be searched simultaneously.
Searches will provide the top three sites from those in scope, prioritizing based on search engine ranking:
- It's possible to narrow selections by de-selecting sources in the Settings or mentioning specific sources of interest in the query/prompt.
What operational factors and settings allow for effective and responsible use of Information Assist?
Accessing the Information Assist feature is dependent on a series of operational actions that must be taken:
The extension must be approved for use in said location (i.e., United States).
The extension must be enabled for the organization by Microsoft in Dragon admin center.
The extension must be enabled for the user by arganization administrators in Dragon admin center.
The extension must be enabled by the user in Dragon Copilot Settings: Settings > Information Assist > Enable extension and in the Reference sources section, select the sources you want to consult.
Note
Customer organizations should add, to any allow lists, relevant websites that may be referred to in Information Assist queries, ensuring these are accessible. Users must ensure they have the required security, consents, and compliance approvals to leverage patient data in Dragon Copilot.
Users should not over-rely on Information Assist and must understand the limitations of the extension and the errors it could make:
Misunderstood intent: the user's intent should be explicit if possible and not inferred. The user should state a preference to focus on the clinical summary note alone, use a specific source, or provide a list or a set of post/pre-op guidelines; for example:
Should intent be misunderstood, it's possible to repeat or clarify the request as a follow-up query.
It's also possible to rephrase a prompt, say/type "Information Assist"/"@InfoAssist", or enable the Use Information Assist only option.
If Information Assist wasn't engaged in a response (and the query remained in the note), it's possible that:
None of the sources have any information that can provide an adequate and grounded response (or these have not been selected).
The query could be unsafe or inappropriate to respond to, such as medical advice or clinical decision support or diagnosis. These would be outside of the scope of the intend use of the tool if answered specifically.
Response too general/generic: Information Assist will not provide a diagnosis or specific medical advice. If such a query is made, it may provide an intentionally broad response that can provide some value and guidance (such as access to a useful source), but it will not attempt to diagnose or provide any medical advice. Examples include:
Are there any diagnoses I have missed?
What diagnosis did I need to look at but forgot to ask?
Does this patient have symptoms of any other disease?
Does this patient need follow-up for anything else?
Should I offer another treatment beyond what I have proposed?
Which medicine is the best?
What is this patient's cancer risk?
Partial response: It's more possible for Information Assist to provide a less satisfactory answer when multi-staged questions are asked. Try to pose questions one prompt at a time, in the event that the system is confused or only provides a partial response. Examples include:
What are the CDC's recommendations for COVID-19 vaccination for this patient, and how can vaccine hesitancy be effectively addressed?
What is this patient's blood pressure classification according to national guidelines, and what is the protocol for prescribing a blood thinner?
Clinical Safeguards are in place to ensure that every answer has been validated, and users are always made aware that the answer was generated by AI.
How do I provide feedback on Information Assist?
We encourage users to provide feedback for continuous improvement. Feedback can be provided via the in-app Feedback icon:
This can be found at the top right of the response card in the timeline or the bottom right in the context card view. Select the Feedback icon to open the AI feedback sidebar where users can provide more detail & comments.
Content policy
Dragon Copilot prompting has the following user prompt categories turned on:
Violence
Hate and fairness
Sexual
Self-harm
These prompts are set at a higher level for risk, in order to handle the content of patient encounters, which can include discussions of violence, sexual material and self-harm, which may be relevant to the patient encounter experience.
Learn more about Microsoft’s Responsible AI:
Learn more about Microsoft Healthcare Agent Service:
Disclaimer
Neither the Dragon Copilot nor Azure Healthcare Agent Service are intended, designed, or made available to be used as: (1) a medical device, or (2) a substitute for professional medical advice, diagnosis, treatment, or judgment and should not be used to replace or as a substitute for professional medical advice, diagnosis, treatment, or judgment.
Output from the Dragon Copilot or Azure Healthcare Agent Service does not reflect the opinions of Microsoft. The accuracy and reliability of the information provided by the Microsoft Dragon Copilot or Azure Healthcare Agent Service may vary and are not guaranteed. Microsoft disclaims any liability for any damages resulting from the use or reliance on the information provided by the Microsoft Dragon Copilot or Azure Healthcare Agent Service.
Microsoft products are not designed, intended, or made available as medical device(s), and are not designed or intended to be a substitute for professional medical advice, diagnosis, treatment, or judgment and should not be used to replace or as a substitute for professional medical advice, diagnosis, treatment or judgment. Microsoft does not warrant that its products will be sufficient for any medical purposes or meet the health or medical requirements of any person.