As a Nurse With Cancer, I Saw the Value of Patient-Facing Materials

— Offering useful info is only becoming more complex in the age of AI

When I was diagnosed with head and neck cancer 4 years ago, my radiation oncologist pulled out a copy of the National Comprehensive Cancer Network's (NCCN) treatment recommendations, which oncologists consult to decide what treatment to give patients. He showed me my stage and what it called for: 10 rounds of chemotherapy and 35 sessions of radiation. Then he handed me the document and sent me home.

As a clinical guideline, it was flawless: accurate, comprehensive, and evidence-based. It was also written by clinicians, for clinicians. I am an emergency and critical care nurse, and I study how health systems evaluate artificial intelligence (AI)-generated patient information. I have also built hundreds of treatment order sets in electronic health records for clinicians to use. My care team knew me that way, as a clinician. But that day, I was not one. I felt like just a patient.

A 2024 study in BMC Primary Careopens in a new tab or window of doctors treating fellow doctors found that clinicians changed how they communicated when the patient was a colleague, including making assumptions about what the patient already knew. Being legible to them as a medical professional made me illegible as a patient. But I was a patient, and I had just been told I had cancer.

What I remember most sharply is the not knowing. But my questions weren't about whether it would work, because I was promised a cure; they weren't about when to go to the emergency department, because that was a call I could make in my sleep. What I did not know was what was normal. Was swallowing supposed to feel like my throat was full of broken glass, or was something else wrong? Would my hearing recover? How would I learn to eat again after the feeding tube came out? They were all versions of the same question: what is happening to me, and what happens next?

Instead, what I found in the guidelines was general advice about whom to ask for help, rather than answers to the questions I was actually asking.

I later downloaded the patient-facing NCCN guidelinesopens in a new tab or window for my diagnosis. For radiation burn, its advice was to ask my healthcare provider for the best treatment. I was using Aquaphor on my neck, per the guidelines, including before my radiation sessions. Only later did I find another documentopens in a new tab or window and learn that I had been using it at the wrong time.

I wound up depending on a Facebook groupopens in a new tab or window for people with my cancer, taking suggestions from anonymous strangers, because the papers in my hand did not answer my questions and the patient portal was static -- a place to retrieve documents rather than a place to ask anything.

That was 4 years ago. Today, patients who are up at 2 a.m. trying to get answers to their specific health questions can turn to AI, which can quickly pull together information scattered across medical sources and explain it in language tailored to their questions. That can be enormously useful when you are sick, exhausted, and trying to understand what is happening. But AI can also be wrong, incomplete, or confidently misleading. It is not a substitute for the clinicians caring for you.

Meanwhile, health systems are increasingly using AI to write and rewrite information for patients. As they do so, they need to test this material not only for medical accuracy, but for whether patients find it complete, clear, and helpful.

All patient instructions should meet that standard. But with AI, getting the standard right becomes more urgent: it can improve information for millions of patientsopens in a new tab or window or reproduce the same gaps millions of times.

Epic, one of the largest electronic health record platforms, has diagnosed the problem correctly and now sells a patient-facing virtual assistant called Emmieopens in a new tab or window. One feature rewrites the after-visit summary, which is handed to the patient on their way out, replacing, in Epic's words, "terminology that's appropriate only for hospital staff" with plain language. The NCCN guidelines page on side effects in my chemotherapy binder warned of "ototoxicity manifested by tinnitus." It meant the chemotherapy could damage my hearing and cause ringing in my ears, but the warning was written for whoever was monitoring me rather than for me. The same is true of "NPO," "STAT," and "PRN," which appear on discharge paperwork every day and may mean everything to a nurse but nothing to a patient: nothing by mouth; right now; and as needed.

Rewriting for the patient is crucial. A randomized trial of 128 patientsopens in a new tab or window in Germany compared standard physician-written discharge summaries, conventionally designed for communication between providers, with AI-generated versions that rewrote the same information for patients. Patients who received the patient-oriented version scored 9.6 points higher on patient activation, a standard measure of how prepared someone is to manage their own care.

Before anyone in your health system hires humans or contracts for AI tools to write and rewrite patient education, they need to ask three questions: Does this material work for the patients we actually serve, across language, age, and social circumstance? Did anyone check it for more than just medical accuracy? And did actual patients test whether they could understand and act on it?

Ask before the information reaches the rest of the patients in your hospital. The people best positioned to say whether a discharge instruction works are the patients holding it at 2 a.m., with no one to call. Until we ask them, we are grading our own homework.

The views in the piece are the author's alone and do not represent NIH, AIM-AHEAD, Ochsner Health, or Xavier University of Louisiana.