AI and Healthcare: ChatGPT, GPT-4 and Med-PaLM 2
In mid-March, Google unveiled Med-PalM2, its new medical AI, which already achieves an 85% accuracy rate [2]. Additionally, Open Health Stack [3][11] was introduced, a set of components for creating digital health apps that can operate offline and rely on data for healthcare workers in low-resource contexts [4].
Furthermore, several works have been published using ChatGPT and GPT-4 to address various issues within the field of medicine and specialized sectors. This article provides an overview of the current landscape, from news to the state of the art in research papers, while also highlighting the potential applications of ChatGPT in improving physician-patient communication, decision-making, and data management. Although the revolutionary potential of AI in medicine is emphasized, it is concluded that issues of accuracy, integrity, and ethical concerns need to be addressed before its implementation in clinical research and medical practice.
(Originally published in Spanish for Linkedin at: 3 Jun, 2023.)
Index:
- The Importance of multidisciplinary work in AI and Medicine
- The Current Evidence of ChatGPT in Medical Research: Challenges and Opportunities
- Conclusions
- Bibliographic references
1. The Importance of multidisciplinary work in AI and Medicine
This summer, I spent several days with my family, and my brother brought along several books and lent me “Mastery” by Robert Green [5], a book that discusses how to achieve mastery and shares various biographies of different masters.
One of the stories that I particularly enjoyed was that of Yoku Matsuoka. After completing a postgraduate program in robotics and enrolling in a robotics master’s program at MIT, she became involved in a project to design a large-scale robot and decided to focus solely on developing its hands. She started by constructing a model that closely replicated the human hand, pondering, “What makes something alive and organically complex?” This led her to delve into the study of neuroscience, human psychology, and evolution.
Thus, multidisciplinary work eventually becomes a necessity [12] in various projects and sets professionals apart. This is also relevant in the fields of AI and medicine, where IBM defines it as:
“Artificial intelligence in medicine is the use of machine learning models to search medical data and uncover insights that can improve health outcomes and patient experiences. […] AI algorithms and other AI-driven applications are used to assist medical professionals in clinical environments and ongoing research.” [1]
We can recall the example of Dendral [6], an expert system for inferring molecular structures, which later led to Mycin [7], a system used for identifying bacteria that cause severe infections and recommending appropriate antibiotics, and CADUCEUS [8], a system for diagnosing internal medicine diseases (covering up to 1,000 different diseases).
Today, we continue to witness innovations like Med-Palm2, Google’s new AI designed to assist healthcare workers. In its latest version, it was tested with expert-level medical questions and achieved a score of 85%. Although it has achieved state-of-the-art performance, this model still has work to do before it can be implemented in the real world.
Additionally, Google is working on adding more features, such as synthesizing information based on medical images like X-rays and mammograms, in order to improve patient outcomes [13][14][15].
Furthermore, in March, Open Health Stack was announced: a set of open-source development components created under an interoperable data standard [10], following the Fast Healthcare Interoperability Standards (HL7 FHIR ©). Teams are already working with Open Health Stack, such as Intellisoft Consulting in Kenya, which is developing a maternal health app called “Mama’s Hub” to support volunteers and pregnant women in rural communities.
On the other side of the spectrum, research involving ChatGPT and GPT-4 has been continuously expanding in recent months, encompassing various medical contexts, environments, and issues.
2. The Current Evidence of ChatGPT in Medical Research: Challenges and Opportunities
One recurring functionality that has been investigated is ChatGPT’s ability to translate or convert medical reports into common language. Lyu Q. et al. [16] demonstrate that ChatGPT can successfully translate radiology reports into plain language with a score of 4.27/5. Furthermore, when using GPT-4, the quality of the translated reports into plain language increased significantly.
Komorowski, M. et al. [17] explores the applications and limitations of ChatGPT from the perspective of an intensivist, highlighting the improvement of communication between doctors and patients, as well as supporting doctors in making diagnoses and decisions. The following table presents a more detailed overview of the different applications and limitations:
Lu, Y. et al. [22] follows the previous line of research and highlights the potential uses of ChatGPT in intensive care medicine. They emphasize the support in decision-making and improving communication between patients and doctors, but also emphasize ChatGPT’s ability to handle and analyze data:
- Data extraction and preprocessing
- Data annotation and labeling
- Data quality evaluation
- Data anonymization and de-identification
- Predictive analysis and modeling
They also explore different applications to assist healthcare professionals in enhancing their knowledge, education, and training.
However, Ruksakulpiwat, S. et al. [18] conduct a systematic review of the current evidence on ChatGPT in medical research, where they conclude that it has the potential to revolutionize this field in various ways. However, its accuracy, integrity, and ethical concerns need to be improved before implementation in clinical research and medical practice (a conclusion shared by several authors in different disciplines).
Does this signal the need to include artificial intelligence (or other related branches such as optimization or computer science) as a necessary field in healthcare careers? Not exactly, but Mendiola, M. S. [19] recommends staying up-to-date with this dynamic landscape by taking courses [23] and analyzing if curriculum updates are necessary (for example, the author mentions that UNAM introduced Biomedical Informatics as a mandatory subject in the medical surgeon career). Additionally, accompanying the development of AI with an emphasis on emotional intelligence.
Ultimately, the adoption of AI cannot occur without a strategic plan from governments for education and massive implementation. Florez, O. [20] emphasizes this, citing the case of Peru, but it is something that should be replicated in various Latin American countries, starting with the need to collect relevant data and incentivize AI research.
3. Conclusions
In conclusion, the importance of multidisciplinary work in various projects, including those related to artificial intelligence (AI) in medicine, is highlighted. The use of machine learning models and AI algorithms can help improve health outcomes and patient experiences, as well as support medical professionals in clinical settings and ongoing research. AI has proven its utility in different areas, such as molecular structure inference, disease diagnosis, and medical report translation.
However, it is acknowledged that the development of AI in medicine still faces challenges. The need to improve accuracy, integrity, and address ethical issues before implementation in clinical research and medical practice is mentioned. Furthermore, the question of including artificial intelligence as a necessary component in healthcare careers is raised. Although there is no clear consensus on this matter, staying updated with advances in the field is suggested.
Acknowledgments: The author acknowledges that this article was partially generated by ChatGPT (powered by OpenAI’s GPT-4 language model; http://openai.com). Editing and research were conducted by the author.
4. Bibliographic references
[1] IBM. ¿Qué es la inteligencia artificial en la medicina?. https://www.ibm.com/mx-es/topics/artificial-intelligence-medicine (Accessed 30 May, 2023).
[2] Domingo E. (2023, May 18). Med-PalM2, nueva IA médica de Google, ya acierta en el 85% de diagnósticos. https://www.redaccionmedica.com/secciones/tecnologia/med-palm2-nueva-ia-medica-de-google-ya-acierta-en-el-85-de-diagnosticos-3099 (Accessed 30 May, 2023).
[3] Cacho J.M. (2023, May 15). Med-PalM 2 y GPT-4 comparten escenario ¿coincidencia? no creo. https://medium.com/@CuraeSalud/med-palm-2-y-gpt-4-comparten-escenario-coincidencia-no-creo-18123bbcbd60 (Accessed 30 May, 2023).
[4] Google for Developers. Open Health Stack. https://developers.google.com/open-health-stack (Accessed 30 May, 2023).
[5] Green, R. (2012). Maestría (Mastery).
[6] Feigenbaum, E. A., & Buchanan, B. G. (1993). Dendral and meta-Dendral. Artificial Intelligence, 59, 233–240.
[7] Shortliffe, E. H. (1974). MYCIN: a rule-based computer program for advising physicians regarding antimicrobial therapy selection. Stanford Univ Calif Dept of Computer Science.
[8] Ayangbekun, O. J., & Bankole, F. O. (2014). An expert system for diagnosis of blood disorder. Int J Comput Appl, 100(7), 975–8887.
[9] Google Research (2023). Med-PaLM. https://sites.research.google/med-palm/ (Accessed 31 May, 2023).
[10] Hersch, F. (2023, Mar 14). New tools to help developers build better health apps. https://blog.google/technology/health/health-developer-tool-thecheckup/ (Accessed 31 May, 2023).
[11] Mehta, I. (2023, Mar 14). Google introduces Open Health Stack for developers. https://techcrunch.com/2023/03/14/google-introduces-open-health-stack-for-developers/ (Accessed 31 May, 2023).
[12] Widner, K. (2023, May 30). 5 myths about medical AI, debunked. https://blog.google/technology/health/5-myths-about-medical-ai-debunked/ (Accessed 31 May, 2023).
[13] PASTOR, J. (2023, May 16). Google lanza PaLM 2 con un objetivo claro: lograr cambiar las tornas y ganarle la batalla a ChatGPT y GPT. https://www.xataka.com/robotica-e-ia/google-lanza-palm-2-objetivo-claro-lograr-cambiar-tornas-ganarle-batalla-a-chatgpt-gpt (Accessed 31 May, 2023).
[14] Narváez, A. (2023, May 10). ¿Qué es Med-PALM 2, la IA de Google dedicada al cuidado de la salud?. https://www.unotv.com/salud/med-palm-2-que-es-inteligencia-artificial-ia-google-cuidado-de-salud/ (Accessed 31 May, 2023).
[15] Google [@Google]. (2023, May 10). Med-PaLM 2 can help answer questions and summarize insights from a variety of dense medical texts. We’re working to add more capabilities to Med-PaLM 2, so that it can synthesize information from medical imaging like mammograms or help radiologists interpret results. Twitter. https://twitter.com/Google/status/1656347311671214098 (Accessed 31 May, 2023).
[16] Lyu, Q., Tan, J., Zapadka, M. E., Ponnatapura, J., Niu, C., Myers, K. J., … & Whitlow, C. T. (2023). Translating radiology reports into plain language using ChatGPT and GPT-4 with prompt learning: results, limitations, and potential. Visual Computing for Industry, Biomedicine, and Art, 6(1), 9.
[17] Komorowski, M., del Pilar Arias López, M., & Chang, A. C. (2023). How could ChatGPT impact my practice as an intensivist? An overview of potential applications, risks and limitations. Intensive Care Medicine, 1–4.
[18] Ruksakulpiwat, S., Kumar, A., & Ajibade, A. (2023). Using ChatGPT in Medical Research: Current Status and Future Directions. Journal of Multidisciplinary Healthcare, 1513–1520.
[19] Mendiola, M. S. (2023). ChatGPT y educación médica: ¿estrella fugaz tecnológica o cambio disruptivo?. Investigación en Educación Médica, 12(46), 5–10.
[20] Florez, O. (22 May, 2023). “Esta medicina llamada Inteligencia Artificial”, por Omar Flórez | OPINIÓN. El Comercio. (Accessed 22 May, 2023).
[21] Rumbo-Prieto, J. M. (2023). ¿ Es ChatGPT una inteligencia artificial fiable para aconsejar al profesional sobre lesiones cutáneas?. ENFERMERÍA DERMATOLÓGICA, 17(48), 6–7.
[22] Lu, Y., Wu, H., Qi, S., & Cheng, K. (2023). Artificial Intelligence in Intensive Care Medicine: Toward a ChatGPT/GPT-4 Way?. Annals of Biomedical Engineering, 1–6.
[23] Grabb, D. (2023). ChatGPT in Medical Education: a Paradigm Shift or a Dangerous Tool?. Academic Psychiatry, 1–2.
