In my last post, I mentioned how HCD can ensure that questionnaires are only used when they are necessary; for example, data that is typically collected as unstructured (e.g., clinical notes) would not be easily collected using FHIR and may cause users additional burden. One comment on this post noted that data quality goes beyond questionnaires and I completely agree. In that post I mentioned how AI could potentially improve this and here I want to expand on that with a case study, showing how HCD can help solve the real problem of transforming unstructured clinical notes into FHIR-friendly data that can be easily shared and accessed.
The vast majority of medical providers and hospitals use an electronic health record (EHR). According to the Office of the National Coordinator for Health Information Technology, as of 2021, 78% of office-based physicians and 96% of non-federal acute care hospitals (96%) had adopted a certified EHR (1). Yet, in many clinical workflows, there is still a lot of unstructured data that is produced with every clinical encounter. These include clinical notes, images, emails, PDF reports, etc - basically any information that is not structured in a form or spreadsheet or directly entered in fields in the EHR. Currently FHIR relies on structured data typically in discrete fields such as a phone number that follows the (XXX) XXX-XXXX format. So how can FHIR be used for these common but not FHIR-friendly components of a patient’s medical record?
This is where HCD and AI can both lend a hand. Imagine a practice needs to comply with MIPS Promoting Interoperability standards (2) and ensure their clinical notes are adequately captured so that when their EHR API connects to the large hospital EHR, those patient records are all correctly shared. The FHIR API exchange may be happening seamlessly for all of the structured data in both EHRs, but the unstructured data may be left behind. Currently, for unstructured data, someone might need to manually extract information from clinical notes to place the information in the individual discrete fields for the data elements required for a FHIR form. This is not only a duplication of efforts, but time consuming, fraught with potential risk or bias as an individual is deciding what is important to extract, not to mention tedious and soul crushing as a task.
There are a few areas where HCD can help solve this problem.
- The first is in understanding the current clinical workflow of the provider and the types of unstructured data that are being produced, and at what frequency. Does the provider transcribe his or her notes after hours? Does the provider write notes between visits? These are important questions that can help identify if there is an option for Structured Data Capture to replace any of the content that is currently unstructured and fit into current workflows.
- The second is in understanding how to implement AI within the clinical workflow to provide an assist. EPIC Systems, the biggest EHR vendor, has a tool called AI Charting that records conversations between patients and providers and uses generative AI to draft notes. In one example of how this tool was implemented, the authors cite numerous HCD methods including examining provider preferences and workflows, and measuring patient experience when providers use this tool. They found that it saved providers time but also improved patient experience as providers now made more eye contact and were able to more actively listen to patients (1).
- The third is in leveraging AI into structured data by looking for ways for tools like AI Charting to use Natural Language Processing (NLPs) and trained Large Language Models (LLMs) to create structured data. In this case, the generated draft notes can be extracted into discrete fields. However, to be useful these tools need to be designed to extract information in a way that makes clinical sense, and they need to be evaluated by users to ensure that they are performing as desired. HCD can help here by co-designing with users to define what to extract and how to transform that data in a way that doesn’t lose meaning as well as engaging users in evaluating outputs and refining the tool.
- The last is in ensuring the use of AI tools are ethically responsible and accessible. The potential benefits of tools like AI and LLMs to transform data like clinical notes into meaningful structured data that FHIR can fully leverage are significant but there are considerable ethical and privacy concerns. Here HCD can ensure that these tools are accessible, they collect only the necessary information, and that patients and providers are fully aware of how the data are being used and provide consent in this process.
If we consider our example of a practice needing to meet MIPS Promoting Interoperability standards but struggling with what to do with their clinical notes, how might they accomplish this? If they are one of EPICs many users they can use the AI Charting tool and an extraction tool like Nabla to summarize their notes, extract structured data, and input the data into the EHR to be shared with the FHIR API. HCD guidance can be shared with the practice to help them figure out how implementing these tools may influence workflow, how to evaluate what the tools are producing, and how to communicate with their patients about the use of these tools.
Similar to my previous post, in making the case for why FHIR needs HCD, we should examine the risks of not including HCD in implementing a solution to a problem like the one presented here. I can easily see how implementing these tools without thinking about the humans that interact with them can lead to lack of uptake, workarounds, and poor data quality. Imagine if your practice rolls this out without adequate patient information and you end up with a PR nightmare headline of “[practice] is giving your patient data to AI”? If you include HCD early we can ensure the design and implementation of these tools are being used in an ethical and legal way and can effectively communicate the risks to patients and providers.
In comments on my last post the question was asked of whether HCD needs FHIR (as opposed to my original question of whether FHIR needs HCD). I think of this relationship as mutually beneficial. I’d be remiss if I didn’t note that HCD practitioners have to grapple with questions of whether AI will replace us. I agree with the author of that article who notes that “...if your focus is on understanding people and designing usable interfaces, there will always be work for you.” Even in system-to-system communication like FHIR there are people still involved and even AI “conversations” still behave like human interactions. Just my two cents (RIP the penny).
While I’m not new to HCD, I am very interested in learning from folks who are much more knowledgeable in how FHIR is currently being used and how it will be used in the future. I know there are other examples where HCD can amplify the value of FHIR and I’d love to hear more about these in the comments. Please let me know what you think - I’d love to keep this conversation going as part of the FHIR Department.
References: