As a physician, I often find myself answering friends’ and family members’ medical questions. “Should I get this looked at?” and “What does this medical term mean?” are some of the most common.
When I answer these questions, I reflect on the fact that not everyone has a physician or nurse in their family or social circles. After a visit to the emergency department, for example, most people don’t have someone to whom they can turn to help them make sense of their discharge instructions.
And that’s unfortunate, because the discharge process can be confusing and disorienting. In medical school, we're taught that discharge is "the last chance to get it right." In this blog post, I’d like to talk about how new technology — specifically artificial intelligence (AI) — can help both patients and clinicians make the most of this final touchpoint in a hospital visit.
Helping patients when they need it most
Every year, new physician residents learn to recognize when a patient may be at risk for adverse avoidable outcomes during care transitions. Modern medical education does a good job highlighting issues that can arise between care teams, including discharge from the hospital — a time when patients are particularly vulnerable.
However, physicians and care teams are stretched to their limits. Transitions of care remain challenging and potentially dangerous times for patients, who not infrequently leave the hospital and arrive home with a sense of uncertainty about what comes next. The problem is exacerbated by impersonal, one-size-fits-all discharge instructions, as well as a sense of urgency to turn the bed over to the next patient.
It is well known that lack of communication is a primary driver of medical malpractice litigation. When combined with rapid patient turnover and understaffing, patients can be left without a sense of complete understanding of their care
But recent advancements in AI now offer the promise of delivering a more truly patient-centered experience through easy-to-understand clinical communication, such as personalized discharge instructions. Although the responsibility ultimately rests with the physician, AI-based tools can augment education and communication with patients, allowing them to better understand their care and make genuinely informed medical decisions.
Large language models (LLMs) such as the kind utilized in ChatGPT offer scalable avenues to simplify the complex and elevate hyper-relevant information in mere seconds. Today’s LLMs are quite good at summarizing and simplifying highly complex language, summarizing, and spotting patterns among all the “noise” in long and complex documents.
For example, the discharge instructions may include terms like “supraventricular tachycardia,” “moderate sedation,” “etomidate,” and “electrical cardioversion.” This sort of medical jargon may be lost on anyone that doesn’t have advanced medical credentials.
With LLMs, this jargon can be converted into something like, “Your heart was beating rapidly due to an abnormal electrical pattern. The emergency department staff converted your heartbeat to a normal rhythm using electricity and you were observed until your sedation medication wore off. Please do not drive today and return to the ED immediately if your symptoms occur again.”
Such tools are rapidly evolving and show promise to become integral parts of patient care, particularly in areas like discharge processes, where safety and quality depend on clear communication between patients and care teams.
While developing language-based tools requires careful and systematic evaluation of safety, these technologies hold potential to strengthen communication and continuity of care at vital points where gaps tend to be prevalent. Here are three examples of how LLMs can help with today’s discharge processes:
1. Simplifying clinical notes and elevating relevant information
Upon leaving the hospital, patients need to know what to do once they get home. But discharge instructions are often intimidating and vague, making it difficult to find the most actionable information. As a physician, clinical documentation is essential for providing an accurate picture of a patient's medical history, diagnoses, and treatment plans. So in that sense, the medical terminology, abbreviations, and other technical language found in medical notes are the result of clinicians trying to be as specific as possible.
As a result, patients themselves have not been able to utilize the content of clinical notes to reference their assessment and treatment plan. Although physicians must make every effort to communicate the discharge plan clearly, such information can be complex and nuanced, and even the most important elements easy to forget. This is true now more than ever, as physicians have less time with patients than in the past.
LLMs not only can easily simplify medical terms ("edema" → “swelling”) and decipher abbreviations (“FAST” → “Focused Assessment with Sonography for Trauma”), but they can also be utilized to pull forward the most relevant information that a patient may need to know upon discharge, such as symptoms that could warrant a return to the ED, relevant precautions about new prescriptions, and anticipatory guidance for newly diagnosed medical conditions.
2. Realizing the full potential of imaging reports
Imaging reports for tests like CT scans, X-rays, MRIs, and ultrasounds are often filled with technical and anatomical terms and complex medical jargon. AI-powered translation of radiology reports can empower patients to better understand the core findings. What’s more, LLMs can elevate and highlight incidental discoveries such as precancerous lesions and other abnormalities, which might otherwise go unnoticed, making AI a potentially life-saving tool. The rates of clinically relevant incidental findings has been reported to be variable, but may be as high as 30% to 40% of radiology studies. Rates of notifying patients of such findings, however, has been reported to be as low as 27%, with the proportion of patients who both received notification of the findings and a follow up plan to be as low as 7%.
LLMs can be used to highlight such findings for the physician and aftercare teams, as well as generating personalized notifications for the patient to follow up on important findings that require additional workup — an important step for safer care delivery.
3. Pre-populating discharge instructions
Traditionally, discharge instructions issued to patients are standardized and lack customization, offering limited insights into the patient's clinical findings and treatment plan. Even trained medical staff can find it challenging to discern a patient's admitting diagnosis or hospital course from these instructions, due to their lack of detail. As a result, patients often leave the hospital with paperwork that appears promising, but lacks easy-to-understand descriptions of what occurred in the hospital or plan of care. By processing clinical notes and imaging reports, LLMs can pre-populate discharge instructions with personalized and relevant information that’s customized to each patient's encounter.
AI-assisted discharge processes can also ensure that patients are able to arrange follow-up with the appropriate medical specialists and become well-connected into the healthcare systems, which can facilitate growth objectives
The future AI for health care delivery
As language-based AI technologies mature and their safety is evaluated, it’s reasonable to expect that their function will extend to interactive applications. For example, LLMs could be used to answer basic questions from patients and caregivers, facilitate inquiries about post-discharge status (i.e., automated check-ins from the hospital to the patient), or enable any number of interactions that enhance the patient experience.
These applications will need to be designed thoughtfully and will require thorough investigation of safety and continued quality assessment. Of course, there’s still much work to be done. AI developers must work with medical professionals to continuously assess their safety and accuracy. And I’m fortunate to be part of that effort.
Find out how Vital’s Discharge Summary feature uses large language models to provide patients with important discharge instructions that are clear, easily understandable, and actionable.