ChatGPT: the future of discharge summaries?

ChatGPT (Open AI, San Francisco, CA, USA) has taken the world by storm.1Open AI ChatGPT: Optimizing language models for dialogue. OpenAI, 2022https://openai.com/blog/chatgpt/Date accessed: December 27, 2022Google Scholar Released to the public in November, 2022, ChatGPT is based on artificial intelligence (AI) technology and trained on data from the internet written by humans, including conversations. This AI-powered chatbot has vast capabilities ranging from poem composition, essay writing, solving coding issues, and explanation of complex concepts including “how can we fix the UK National Health Service?” Some consider that ChatGPT has advanced the online search to the next level;2Grant N Mets C A new chat bot is a code red for Google's search business. The New York Times, Dec 21, 2022https://www.nytimes.com/2022/12/21/technology/ai-chatgpt-google-search.htmlDate accessed: February 1, 2023Google Scholar the program offers rapid and in-depth understanding of complex matters and generates custom responses in a conversational manner to the exact question asked, recalling its own previous responses. Although still only a research preview, ChatGPT has shown impressive performance across multiple use cases. One potential area of ChatGPT's application could be to generate discharge summaries. Composition of high-quality discharge summaries containing the requisite information can be time consuming, with the burden usually falling on junior doctors. Therefore, when balanced against clinical commitments, discharge summaries are often left under-prioritised resulting in delays in patients' discharges or insufficient discharge summaries, which not only places pressure on an already stretched junior doctor workforce, but also presents potential patient safety issues in the transition of care from secondary to primary care.3Allen J Hutchinson AM Brown R Livingston PM Quality care outcomes following transitional care interventions for older people from hospital to home: a systematic review.BMC Health Serv Res. 2014; 14: 346Crossref PubMed Scopus (127) Google Scholar ChatGPT allows doctors to input a brief of the specific information to include, concepts to elaborate on, and guidance to explain, outputting a formal discharge summary in a matter of seconds (panel). Discharge summaries are an obvious choice for this technology because of their largely standardised format; this standardisation has previously been recognised through the creation of templates—eg, from the Royal College of Physicians.4Royal College of PhysiciansImproving discharge summaries—learning resource materials. RCP, London2021https://www.rcplondon.ac.uk/guidelines-policy/improving-discharge-summaries-learning-resource-materialsDate accessed: December 27, 2022Google Scholar Automation of this process could therefore ease the work burden upon junior doctors allowing more time for the delivery of patient care and the seeking of training opportunities in often roles centered on service provision. ChatGPT could also improve the quality of the discharge summaries themselves, with previous literature showing discharge summaries that are written through traditional means often missing details.5Earnshaw CH Pedersen A Evans J Cross T Gaillemin O Vilches-Moraga A Improving the quality of discharge summaries through a direct feedback system.Future Healthc J. 2020; 7: 149-154Crossref PubMed Google ScholarPanelChatGPT's response to a simple discharge summary requestPrompt“Write me a discharge summary for John Doe, a 70-year-old man who has had a total hip replacement 5 days ago and now can leave hospital to go home.”Response *Although this summary is likely to be accurate for most patients, the additional information that was not included in the prompt was automatically generated by ChatGPT, highlighting the need for manual checking by clinical staff before finalisation. *Although this summary is likely to be accurate for most patients, the additional information that was not included in the prompt was automatically generated by ChatGPT, highlighting the need for manual checking by clinical staff before finalisation.“Patient: John DoeAge: 70 years oldDate of admission: dateDate of discharge: dateReason for Admission: Total hip replacement surgerySummary of Hospital Course:Mr Doe was admitted to the hospital on date for a total hip replacement surgery. The surgery was successful and Mr Doe has made good progress in his recovery. He has been compliant with his physical therapy and pain management regimen, and has demonstrated good mobility and strength in his affected leg. He is able to perform activities of daily living independently and has demonstrated good understanding of his post-operative care instructions.Recommendations for follow-up care:Mr Doe is advised to continue with his physical therapy as prescribed by his physical therapist. He should also continue to take his pain medication as directed and follow a healthy diet and exercise routine to promote healing and maintain good overall health. He should follow up with his orthopedic surgeon in 1–2 weeks for a post-operative check-up.Overall, Mr Doe is in good condition and is ready for discharge from the hospital. Signed,physician's name” Prompt “Write me a discharge summary for John Doe, a 70-year-old man who has had a total hip replacement 5 days ago and now can leave hospital to go home.” Response *Although this summary is likely to be accurate for most patients, the additional information that was not included in the prompt was automatically generated by ChatGPT, highlighting the need for manual checking by clinical staff before finalisation. *Although this summary is likely to be accurate for most patients, the additional information that was not included in the prompt was automatically generated by ChatGPT, highlighting the need for manual checking by clinical staff before finalisation. “Patient: John Doe Age: 70 years old Date of admission: date Date of discharge: date Reason for Admission: Total hip replacement surgery Summary of Hospital Course: Mr Doe was admitted to the hospital on date for a total hip replacement surgery. The surgery was successful and Mr Doe has made good progress in his recovery. He has been compliant with his physical therapy and pain management regimen, and has demonstrated good mobility and strength in his affected leg. He is able to perform activities of daily living independently and has demonstrated good understanding of his post-operative care instructions. Recommendations for follow-up care: Mr Doe is advised to continue with his physical therapy as prescribed by his physical therapist. He should also continue to take his pain medication as directed and follow a healthy diet and exercise routine to promote healing and maintain good overall health. He should follow up with his orthopedic surgeon in 1–2 weeks for a post-operative check-up. Overall, Mr Doe is in good condition and is ready for discharge from the hospital. Signed, physician's name” Although ChatGPT offers significant promise, deployment into clinical practice will require multiple barriers to be overcome. First, ChatGPT depends on the data it receives and therefore still necessitates salient information to be manually inputted. Future iterations might allow automatic scraping of data from the patient's electronic health record without the need for manual inputting of a brief. This practice, however, presents significant issues with respect to data governance. Previous use of AI within health care has been associated with failings that significantly damage patients' trust6Powles J Hodson H Google DeepMind and healthcare in an age of algorithms.Health Technol. 2017; 7: 351-367Crossref PubMed Scopus (187) Google Scholar and therefore care must be taken around data storage and access. Acceptability of this technology, from patients in particular, must be considered. The concern that automation might lead to depersonalisation of care could lead to resistance to this technology. Pilot trial data is therefore necessary to gather stakeholder views and ultimately show improvements in efficiency and quality of processes. Finally, the consequences of technology failure must be considered. Although asking trivial questions to a chatbot at home might have few consequences, incorrect or inadequate information concerning medication changes or follow-up can greatly affect patient care. Therefore, although ChatGPT can be considered as a first step in the writing process, manual checking of its output by a doctor will be required before completion. ChatGPT demonstrates the power of this next generation of chatbots powered by large language models. Although the use of chatbots within health care is by no means novel, having previously been applied to areas such as triage,7Baker A Perov Y Middleton K et al.A comparison of artificial intelligence and human doctors for the purpose of triage and diagnosis.Front Artif Intell. 2020; 3543405Crossref PubMed Scopus (16) Google Scholar this next generation aims to advance the field significantly through its generative capacities. Health care is an industry that has significant scope for automation, especially in areas of documentation. We should be reactive to these advances in technology and explore potential use cases and feasibility of using chatbots such as ChatGPT across diverse areas in health care. Proactive early adoption of this technology at a central level might also reduce the risk of issues in the future; parallels can be drawn with interoperability issues due to differing electronic health records adopted at a local level.8Centre For Public ImpactThe electronic health records system in the UK.https://www.centreforpublicimpact.org/case-study/electronic-health-records-system-uk/Date accessed: December 27, 2022Google Scholar The question for the future will be how, not if, we adopt this technology. We declare no competing interests. Funding and infrastructural support was provided by the UK National Institute for Health and Care Research (NIHR) Imperial Biomedical Research Centre. KL is supported by a NIHR Academic Clinical Fellowship.