LLM Discharge Summaries Improve Patient Activation

Hospital discharge is a sensitive transition from inpatient care to self-management at home, when patients need clear information about diagnoses, medicines and follow-up. Patient activation covers the knowledge, skills and confidence needed to manage health after leaving hospital. A randomised controlled trial published in The Lancet Digital Health assessed whether patient-oriented discharge summaries generated by a large language model could improve activation compared with standard discharge summaries. The work took place in the Department of Internal Medicine at Helios University Hospital Wuppertal in Germany. Adults scheduled for discharge received either routine discharge information or a patient-oriented version generated from an anonymised standard discharge summary using GPT-4o. Patients receiving the generated summaries had higher activation at discharge than those receiving standard documents.

Plain Language for Discharge Information
Participants were adults scheduled for discharge after treatment in the internal medicine department. They needed German proficiency and informed consent, including agreement for cloud-based processing through a large language model. Patients were excluded when rare diseases, serious previously undisclosed diagnoses or severe impairments made participation unsuitable. After screening and consent, 128 patients entered random allocation, with equal numbers assigned to standard discharge information or the generated patient-oriented summary.

For the intervention group, physicians first prepared the standard discharge summary as part of routine care. A duplicate version was then limited to key sections, including diagnosis, medication and recommendations, before identifying details were removed. GPT-4o generated a patient-facing version through Microsoft Azure OpenAI Service. The model was instructed to simplify the existing content without adding clinical information or changing facts. Physicians reviewed each output before release, reinserted essential dates and telephone numbers and corrected inaccuracies when needed. No critical errors were flagged in the initial generated outputs. The approach focused on language simplification rather than adding visual tools or structured teaching methods.

Must Read: Chatbot Health Use Reveals Personal and Practical Needs

Activation Increased at Discharge
Patient activation was measured with the German version of the Patient Activation Measure, which assesses confidence and ability to manage health. All participants completed the primary assessment. The enrolled group had a median age in the mid-60s, and baseline characteristics were broadly similar across the control and intervention groups. The intervention group recorded higher activation scores at discharge than the control group, with a statistically significant difference between groups.

Activation levels also shifted in favour of the generated summaries. More patients receiving the patient-oriented summaries reached the highest activation level, while fewer remained at a lower readiness level compared with patients receiving standard summaries. Health literacy did not differ significantly between groups. Scores were slightly higher among patients receiving the generated summaries, but the difference was not statistically significant. The proportions of patients with inadequate, problematic and adequate health literacy were also similar. The health literacy measure covered broader ability to find, understand, evaluate and apply health-related information, rather than only understanding discharge instructions. This distinction matters because the generated summaries addressed discharge communication directly, while health literacy reflects a wider set of capabilities.

Patients Rated the Summaries More Favourably
Patients assessed their discharge information across helpfulness, comprehensibility, empathy, trust and overall satisfaction. Ratings favoured the generated patient-oriented summaries in most areas. Patients receiving these summaries rated them as more helpful, easier to understand and more empathetic than patients receiving standard discharge summaries. Overall satisfaction was also higher. Trust remained similar between groups, suggesting that plainer language did not reduce confidence in the discharge information.

The generated summaries differed from standard discharge summaries in audience and style. Standard summaries mainly support communication between healthcare professionals and often use dense clinical terminology. The generated patient-oriented summaries used clearer explanations, more accessible wording and a structure aimed at the patient’s needs after discharge. The workflow still included physician oversight, and patients in the intervention group also received the standard discharge summary required under routine care.

Several implementation issues remain. The single German hospital setting and internal medicine focus limit wider generalisation. Patients with lower language proficiency, cognitive impairment or functional limitations were not included, although simplified discharge information could be relevant for such groups. Longer-term outcomes such as readmissions, medicine adherence and later health service use were not measured.

Large language model-generated patient-oriented discharge summaries improved patient activation at hospital discharge and were rated more positively than standard summaries in several quality dimensions. Health literacy did not improve significantly, and longer-term clinical effects were not assessed. The workflow relied on existing discharge content, model-generated plain language and physician checking before release. Wider use would require dependable quality assurance, strong privacy safeguards, smooth integration into clinical systems and clearer regulatory pathways. Further evaluation across broader patient groups and care settings is needed before routine adoption can be considered.

Source: The Lancet Digital Health

Image Credit: iStock