The Doctor Will See You Now, But AI May Be Listening In

In the era of the electronic health record (EHR), patients have become used to getting a ping on their smartphone alerting them that a visit summary is now available to review. For patients, it’s a chance to see into the mind of the physician, to review the provider’s recommendations and to figure out how to spell that five-syllable medical term they uttered.

However, physicians have begun to have that same experience in recent years. Artificial intelligence (AI) has gained the ability to make sense of complex interactions, such as patient-provider visits.

In April, medical technology company Augmedix released Augmedix Go, a fully automated, generative AI-powered ambient medical documentation mobile app. The app is designed to assist with medical documentation in the emergency room (ER). Physicians and other clinicians wear a Bluetooth microphone connected to their smartphones. In turn, the app captures the patient-clinician conversation and generates medical notes. The clinician then reviews the notes and makes any needed edits or additions before finalizing it.

The company partnered with Tennessee-based HCA Healthcare to pilot the technology at four of the company’s hospitals. The pilot found that patients consented to allow the ambient AI technology in 99% of cases, and the company said it incorporated insights from the pilot into the release version of the software.

Veronica L. Cassese Klasko, M.H.Sc., M.M.S., a physician assistant at HCA’s Florida Trinity Hospital, said the technology decreased the amount of time she had to spend in front of her computer. “I was able to essentially eliminate charting after my shift was over, which makes a huge difference,” she said.

Yet the use of the technology also raises a host of technical, ethical and business questions, and chief among them is: What if AI makes a mistake?

Indeed, two days before Augmedix announced its app’s release, researchers from Oregon Health and Science University (OHSU) reported the results of a study that was published in the Journal of Medical Internet Research on the use of ChatGPT in generating medical notes based on simulated patient-provider conversations. The OHSU researchers found that the software had an average of 23.6 errors per clinical case, though most were errors of omission; only 3.2% of errors involved introducing incorrect facts. In short, the authors concluded that ChatGPT was not ready for use in a real-world ER setting.

Augmedix Go is more sophisticated than ChatGPT, a publicly available tool that has been used in a wide variety of settings. Augmedix says its AI was built specifically for medical uses. The company notes that it uses proprietary large language models (LLMs) to power its technology, along with industry-specific tools such as Google Cloud’s MedLM models. In addition, Augmedix Go, which is compliant with the privacy and security standards of the Health Insurance Portability and Accountability Act (HIPAA), requires notes to be reviewed by providers before being finalized. Aside from accuracy and privacy concerns, though, one major question is whether payers will be willing to pay claims based on documentation that was largely generated by AI.

Manny Krakaris, MBA

Managed Healthcare Executive reached out to Manny Krakaris, MBA, Augmedix’s CEO, to pose those and other questions. Below is an edited transcript of his responses.

This type of technology could be used in all healthcare settings. Are there specific reasons the technology is uniquely useful in an
ER setting?

HCA’s decision to start with implementing the program in an ER setting stems from several factors. There is a significant industry-wide gap in medical documentation solutions in ER settings. The ER workflow demands technology capable of handling complex conversations and recording numerous nonsequential interactions with multiple clinicians, all within the backdrop of a typically noisy environment. Prior to our recent launch of Augmedix Go, there was no solution that was built for the acute care setting of an ER — only ambient AI products with basic LLMs built for ambulatory care settings like primary care. Any clinician will tell you that emergency medicine and primary care are worlds apart, so we set out to solve the unique documentation-related challenges and pain points that organizations and ER clinicians face. Filling the gap in the ER was urgent for HCA Healthcare.

What type of review process is in place after AI generates these notes? For instance, is it always reviewed by the physician/provider who saw the patient?

Transparency builds trust. That concept is at the core of everything that we’re putting into Augmedix Go. ... The app makes it easy for users to see the relationship between input and output, which makes reviewing and editing notes more straightforward. At the end of the day, the clinician is always in control of what goes into the final note. AI is there to help process information faster than a human can, and to capture things in the moment that a clinician sometimes can’t document until hours or even a day later. Human memory is not perfect, and when you add the cognitive load that doctors and nurses have to carry around in the ER, it’s easy for details to get missed.

For Augmedix Go, the notes are reviewed and signed off by the clinician who saw the patient. The Augmedix platform provides a high degree of transparency in the note-creation process. Users can see the transcript being created in real time [and] have access to the clinical data page, which allows them to easily confirm that all the information captured is accurate. For example, they can verify that the correct medications and tests are included. In this way, they can ensure accuracy without relying on their memory. They also have full editing and in-line dictation capabilities directly from the app, or they can send draft notes to their EHR and edit there. No other ambient medical documentation solution offers this level of transparency into the data
capture process.

[Although] Augmedix Go is a pure AI solution that relies solely on the clinician for review, we offer other products that blend AI with human-in-the-loop support that helps with all the other steps in the charting process, like pulling in the right templates and macros [and] thoroughly reviewing notes before sending them to the EHR for the clinician’s review and signature.

To facilitate review and build trust after a patient visit, we highlight the key elements within the transcript that our technology considers relevant to the medical note. Then we put the key data from the patient visit, along with the patient history, into an easy-to-follow format that can be seamlessly exported into our customers’ data lakes.

We recognize that different health systems and clinicians have varying needs, [which] is why our solutions are designed to be flexible to fit different workflows and comfort levels of using AI technology.

In the pilot, you found that patients had a very high acceptance rate for the product. What about other stakeholders? Has there been any pushback from payers?

The pain points we’re helping healthcare provider organizations solve impact their relationships with payers and the burdensome process of getting properly reimbursed. Payers are looking for specific criteria within a note to justify reimbursement for whatever services were provided by the clinician. In the ER, such criteria are presented in the MDM [medical decision-making] section of the note. That’s where we’ve focused much of our research and development — not because payers asked us to but because the back and forth with payers is a major pain point for our customers. We’re helping organizations more accurately document all the MDM details to enable proper coding and reimbursement. Over time, healthcare organizations should see improved reimbursement levels that reflect services that were actually rendered during the patient encounter. The supporting detail contained within the MDM is tied to the ultimate source of truth: the transcript/recording. Having well-supported claims should help providers and payers by reducing the amount of back-and-forth communication that occurs today as a result of claim ambiguities.

Does Augmedix use patient visits to train its AI? If so, how does the company mitigate privacy concerns?

Augmedix uses deidentified notes, deidentified transcripts, and structured data from our proprietary notebuilder tool to train our proprietary [natural language processing tool]. We mitigate any privacy issues by using only these transcripts and notes, and we follow all HIPAA compliance rules and regulations.

What about the malpractice angle? If AI makes an error in a medical record and it is not caught by any human reviewer, who is at fault from a liability standpoint?

Augmedix emphasizes the importance of review and final sign-off by the clinician. Ultimately, any AI tool — no matter how sophisticated — is designed to enhance workflows and support clinicians, not replace them. It is paramount that clinicians leverage AI responsibly and review all notes to ensure their accuracy.

The significant value that Augmedix brings is substantially reducing the time spent reviewing notes through a highly sophisticated technology that delivers accurate medical note drafts. However, it does not altogether eliminate the time spent on documentation. No technology does that, nor should it.