Could ChatGPT Unlock Answers from a Patient's Medical Record

Imagine you are sitting in the room with your doctor and you are asked if you have had a lab test or procedure completed in the past. You know you had it done - maybe even at that same health system - but your doctor has spent at least 5 minutes staring into the computer screen and simply can’t find it.

Now imagine you are the doctor. You have 30-40 patients on your schedule today and this scenario plays out over and over and over again. Why is it so hard to find the information you need?

---

How search is shifting away from 10 blue links

10 blue links refers to the results you get when you type in a search at Google. Sometimes you want to read a few different opinions on a topic but other times you are just looking to find the answer to a specific question. Howeer, most search results don’t give you the answer, instead you must click through 10 blue links to try and find it. That’s where ChatGPT, Gemini, Perplexity and similar tools are different. Type in a question and it’s goal is to give you the answer. It’s not always perfect, but when it does work it can provide answers in seconds that would have taken much longer to find, if you found them at all.

Searching the Medical Record

Doctors traditionally have searched a patient’s medical record in two ways.

The first is to simply go through different sections of the chart looking for the information needed. If its a lab test, the doctor might start by going to the lab results section and filtering through to see if they find the results. Sometimes the data is buried with a specific patient note, in which case the doctor may try to identify a couple notes they think are mostly like to have the information and skim through them.

The second approach is to search the medical record. This might work similarly to using Google to search the internet, but sometimes you have to separately search the various aspects within the EHR (search in the labs section, search in procedure notes, search the patient portal communications, etc).

---

Chatting with a Patient’s Medical Record

Last month, I was invited to give a Keynote lecture at a National Institutes of Health (NIH) workshop on the potential role of using artificial intelligence with the electronic health record (EHR). The talk focused on the four areas that clinicians spend most of their time on while using the EHR:

1. Chart review (33%)

2. Documentation (24%)

3. Order entry (17%)

4. Communicating with patients

Today, I’ll focus on chart review and specifically how to find the information you need quickly. To illustrate what chatting with a patient’s medical record might look like, I found an example of a publicly available de-identified patient note. UNC’s School of Medicine has 9 de-identified notes available here. I chose the 2nd one (HP2). Here’s an image of the what the 10 page note looks like from a zoomed out view:

I took the the note in PDF format, uploaded it to ChatGPT 4.0 and starting asking questions. Wouldn’t it be great if I could get a short, succinct summary of this long note?

Not bad. ChatGPT was able to summarize the patient’s demographic information and main medical problems. It interpreted why she came to the emergency room and then how it compared to symptoms from past episodes. Next, it summarized what the initial treatment plan entailed and what was outlined for continued management.

Ok, now lets ask it to find specific information from within the note.

It tells me that the note does not describe a prior colonoscopy. It is able to differentiate that a colonoscopy is discussed but that there is no evidence it has been completed. This saves a lot of time because the in order to be sure something isn’t there, a doctor would have to read the entire note.

Here’s another specific question:

This time ChatGPT gives me a bullet list included what I asked for (past surgeries) as well as when they occurred.

Let’s try one more. We often ask patients if they have any allergies. This information should be documented in the allergies section of the EHR but many times it is not or the information is incomplete. So the type of reaction is often buried within a note somewhere.

Pretty good. It lists the allergies in bulleted form, along with what it caused. In order to do this, that information had to be present within the note. But imagine this note only had allergies but a different note had what the allergies caused. A ChatGPT-like interface might be able to piece these bits of information together so a clinician doesn’t have to read through multiple notes. Lets take it one step further, what if an AI-bot was designed to pull this information and then put it into the right place in the EHR - that could address the documentation challenges described earlier. Of course, we might want a patient and/or doctor to verify they are correct before they go into the chart, but if that information was queued up in an easy way to review and confirm that would make it much easier.

---

Barriers to enabling this at scale

If you are not convinced yet there is potential, I urge you to read the 10-page note and see how long it takes you to find the same information. Then imagine doing that for 30-40 patients in a row. The potential impact on clinician workflow and patient outcomes is significant. But several barriers need to be addressed first to enable this type of technology at scale:

1. Response accuracy and bias. LLMs or large language models like those used by ChatGPT are evolving rapidly. Google has developed one specifically for the medical domain. Most of these models work by predicting the next word in the response based on a algorithm that is weighted using training data. Sometimes the models hallucinate - they generate a response that is not true. Over time, these models will get better but the potential to generate false information is a patient safety issue that will hinder broader adoption. Model bias is another key issue that must be evaluated and addressed.

2. Data security. Patient information is sensitive. A chat function would need to ensure that patient data is not being mixed with other dataset. Instead, they will likely need a secure cloud environment to store information so it can be queried without mixing it to other environments. At first, this may be costly to setup. Over time, the price may come down or companies may spin up to do this specifically. We’ve already seen Amazon develop cloud environments for other health care use cases and maybe this could be applied here.

3. Workflow Integration. To work, these capabilities would need to exist within the medical record. This could be done by the EHR companies integrating it or through an EHR application that could be accessed from within the medical record in the same way you access an app from your smartphone. There is evidence of this occurring already for other use cases. Here’s an article from Epic on how its using generative AI to draft responses to patient portal communications.

---

Frequently Asked Questions

Let’s take this one step further. I asked ChatGPT to give me 3 questions on this topic that would be likely asked by a reader but were not answered by my post. I then asked ChatGPT to give me the answers to the questions. Here they are below. See what you think.

1. How do patients perceive the use of ChatGPT and artificial intelligence in managing their medical records?

Patients' perceptions of using ChatGPT and AI in healthcare can vary widely based on several factors, including their familiarity with technology, concerns about privacy and data security, and their experiences with healthcare systems. Some patients may appreciate the efficiency and personalized care that AI can offer, such as quicker access to medical information, more accurate diagnoses, and tailored treatment plans. However, others might have concerns about the privacy of their medical records, fearing unauthorized access or misuse of their sensitive health data. Additionally, there may be apprehension about the depersonalization of care, with AI possibly reducing the human interaction between patients and healthcare providers. Overall, patient perception could improve with increased transparency about how AI is used, assurances about data security, and evidence of improved care outcomes.

2. What are the specific ethical considerations involved in using AI like ChatGPT within EHR systems, especially regarding patient consent and data anonymization?

Ethical considerations in using AI like ChatGPT within EHR systems include ensuring patient autonomy, privacy, and confidentiality. Obtaining informed consent is crucial, where patients are made aware of how AI will use their medical records and the benefits and risks associated with it. This involves transparently communicating the scope of AI's role in their care, including how it may analyze their medical data, the security measures in place to protect their information, and their rights to opt-out or restrict how their data is used.

Data anonymization is another critical aspect, ensuring that patient data used for training AI models does not compromise individual privacy. Ethical use of AI in healthcare also necessitates addressing potential biases in AI algorithms that could lead to disparities in patient care. Ensuring that AI systems are trained on diverse datasets and regularly audited for bias is essential for equitable healthcare outcomes.

3. How does the integration of AI like ChatGPT into EHR systems impact the training and skill set requirements for healthcare professionals?

The integration of AI and technologies like ChatGPT into EHR systems is likely to significantly impact the training and skill sets required for healthcare professionals. Clinicians will need to become proficient in using these technologies, which includes understanding how to interact with AI systems, interpret their outputs, and integrate AI-generated insights into patient care decisions. This requires a foundational understanding of data science and AI principles, as well as training on the ethical considerations and limitations of AI in healthcare.

Medical education curricula may need to evolve to include courses on digital health, AI, and data privacy, equipping future healthcare professionals with the skills needed to navigate the intersection of technology and healthcare effectively. Additionally, ongoing professional development programs will be essential to ensure that current healthcare providers can update their skills and knowledge as AI technologies evolve. This shift emphasizes a multidisciplinary approach to healthcare, where clinicians work alongside data scientists, engineers, and ethicists to provide patient-centered care in the age of AI.