When I first heard about OpenAI’s ChatGPT Health, I felt a familiar itch.
Since being diagnosed with a malignant brain tumor 18 years ago, at age 29, I’ve developed a deep curiosity about my own health. That curiosity has driven me to enroll in numerous studies, connect my health records to the NIH All of Us research program, and even donate my brain tissue for research-grade genomic sequencing.
Over time, this curiosity became a career. Today, I’m part of the OpenNotes team at Beth Israel Deaconess Medical Center, where we study how sharing clinical notes improves care and enhances trust between clinicians and patients.
I use every tool at my disposal to make sense of the messiness of living with brain cancer. And starting in August 2024, after learning it was possible for non-programmers like me to create our own tools within ChatGPT, I did what many tech-curious patients would do: I built my own custom GPT.
I called it Kaiser GPT, and it was trained on an enormous 4,839-page PDF of my medical records during the eight years when I received care at Kaiser Permanente in Northern California. Kaiser GPT is a time capsule of a younger me. When my new neuropsychiatrist at the University of California San Francisco asked when I last had a cognitive evaluation, I typed the question into Kaiser GPT and it responded in seconds: Aug. 6, 2013.
It was right.
This is something I can’t easily do in any of the MyChart portals I use today. I can’t search across my entire record for a specific phrase. I can’t ask it to summarize eight years of notes from my neuro-oncologists. But I can with Kaiser GPT.
So when I was granted early access to ChatGPT Health recently, I knew I had to try it.
But to be honest, as I wrote in a blog post about my first 72 hours with ChatGPT Health, it was a little disappointing, especially compared with Kaiser GPT.
To test the setup, I connected ChatGPT Health with my two active patient portals at UCSF and UC Davis Health, as well as the Kaiser system where I am no longer an active user. ChatGPT Health also offers the option to connect with consumer apps such as Apple Health, AllTrails, and MyFitnessPal.
Once connected, I prompted ChatGPT Health to “conduct a qualitative analysis of all the data” it had access to. I was expecting to see three patient portals and my Apple Health data, but more than half of the time, something was missing. I’d prompt again, naming the missing system, and suddenly ChatGPT Health would respond, “Oh yeah, that’s here too.” It was like it forgot what it had until I reminded it.
None of my medical records connected successfully at first, although the Apple Health data synced flawlessly.
Apparently, I wasn’t the only person experiencing issues. After my blog post received attention on LinkedIn, an executive at b.Well Connected Health reached out to suggest disconnecting and reconnecting my records. b.Well is a health technology company that enables health consumers (i.e., the general public) to create their own longitudinal health records. They’ve been entrusted by major health systems, including the Department of Veterans Affairs, to be their FHIR API of choice. (Fast Healthcare Interoperability Resources Application Programming Interfaces are a standardized way for different health systems and apps to securely exchange medical data like labs, medications, or visit notes so information can move between platforms without manual copying or faxing.) bWell’s partnership with OpenAI was announced on Jan. 7.
After following the advice from b.Well, my health system connections appeared within a couple of days.
I’ve since run my tests again.
Prompt: “Compare and contrast the differences between my brain surgeries in October 2021 and May 2022 in 1 sentence.”
Output: “In short: the 2021 surgery treated broader non-enhancing progression, while the 2022 surgery was a focused resection of an anterior-margin recurrence (both reported as uncomplicated, maximal-safe resections).”
Not too shabby.
What makes ChatGPT Health promising is that, in theory, unlike my static Kaiser GPT, it can access recent data from across all health systems with which b.Well has partnered. And yes, like all LLMs, I expect it to get confused. But I’m not too worried about so-called hallucinations, in which it makes up information. I catch the errors because I’m the one with brain cancer. And I’ve read every single one of my clinical notes over the last 18 years. I know when something is off.
Since leaving Kaiser in 2017, my care has only gotten more complex. I’ve had two additional brain surgeries, radiation, and I now coordinate care across UCSF (neuro-oncology) and UC Davis Health (primary care, sleep medicine). Referrals come and go, and no single system holds my entire story — except my brain. And now, partially, in these GPTs.
Some of my research, clinician, and patient advocacy colleagues have raised their eyebrows at my willingness to sync my medical records to ChatGPT Health. As my boss often reminds me, “If you’re getting something for free, you are the product.”
That’s fair. But I’ve been blogging publicly about my medical history since 2008. My records have already been exposed — by my choice — for years. Someone commented on LinkedIn, “So you just gave away 4,800 pages of your data for free … can you explain the logic?” My response was simple: “Some people do things out of curiosity, without financial motivation. I am (often) one of those people.”
I don’t recommend this approach for everyone. Before you sync your health data with any AI tool, think critically about what you’re doing, who you’re trusting, and the long-term implications. In my case, I synced first and researched later. But I was relieved to discover that large public systems like the VA already depend on these same FHIR APIs.
While I have a high level of health literacy, I never earned a bachelor’s degree. I don’t think people should have to go through 18 years of brain cancer to understand their own health records. My hope is that models like ChatGPT Health can help regular people become more curious about their health, learn to ask informed questions, and feel more prepared for their medical visits.
Liz Salmi is a patient-turned-researcher with the OpenNotes Lab at Beth Israel Deaconess Medical Center and a cognitive science student working to bridge lived experience and health research.
link