Using AI Applications to Enhance the Healthcare Quality
Before the pandemic, there was a significant need for efficiency and improved quality of care within the healthcare system, which is a health and economic issue in all countries. Now, in the aftermath of the pandemic, reforms within the healthcare system are needed more than ever. Testing, diagnosing, and treating patients to improve the quality of care is a question that has proven to be very complex. There are hopes that technologies such as AI can be an answer to the problem. The question is: How can AI applications be used to improve the quality of healthcare?
Using AI as an Effector Arm
Doctors Ezekiel J. Emanuel and Robert M. Watcher [1] examines the American healthcare system and argues that the problem is not with analyzing large amounts of data, but rather the issue lies in changing the behavior patterns of patients and medical doctors. They point out that doctors’ behavior patterns account for 80% of the American healthcare costs. This includes the prescribing of tests, treatments, and medications. At the same time, they note that patients’ lifestyle habits such as diet and exercise account for over 50% of all chronic diseases.
These behavior patterns are meaningful because both physicians and patients can change them. Emanuel and Watcher conclude that this gives us an indication that AI applications focusing solely on data analysis will not improve the quality of care, thereby missing the main problem, which is meaningful behavior change. They explain the concept of “effector arm,” meaning that AI applications should reflect the transfer from scientific evidence to being used in healthcare. Here, technology can play an important role, they argue, since AI can be an “effector arm” for changing meaningful behavior patterns.
Guiding Physicians Through AI Wayfinding
In contrast to changing behavior patterns, Julia Adler-Milstein et al. [2] emphasize the importance of guidance. They point out that AI has the potential to assist human intellect by reducing increased diagnostics and thus related human errors. However, AI has not achieved this yet, and they highlight the importance of AI moving from predicting labels to guiding doctors. They mention that there are many examples of AI solving well-known isolated diagnostic problems. One example of this is image analysis that can predict diabetic retinopathy and whether a chest X-ray shows pneumothorax.
They write that these problems are predicted by AI applications through a marker representing the end part of the diagnostic process. Thus, the applications overlook the whole that a physician must navigate through, thereby losing the physicians trust. They argue that AI applications should instead focus on the process, which they call “wayfinding.” They explain that “wayfinding” involves orienting oneself in the current position, making route choices, confirming the route choice, and delivering a result.
For a physician, this process looks similar. By gathering information such as symptoms, having a diagnostic hypothesis, and eliciting new data based on the diagnostic hypothesis such as test results and physical examination. Since this can continue to ultimately reduce uncertainties that exist and then the final goal becomes clearer for setting a final diagnosis. They conclude that this is a challenge for AI, but the technology could be of great help. Healthcare already uses computer interpretation, ECG, and AI applications should then be available for physicians so that they know where in the diagnostic process they are.
An AI Chatbot By Google Shows Promising Results
In recent efforts by Google Research and DeepMind, their large language model (LLM) has been interpreting medical anamnesis by actors that were trained to portrait medical conditions. Behind the research, Tu, T., et al. [3] point out that the dialogue between a physician and the patient is “fundamental to effective and compassionate care.” For a physician, the data is collected through dialogue with the patient, together with after following tests. Tu, T., et al. [3] introduced AMIE (Articulate Medical Intelligence Explorer), an LLM based AI system optimized for clinical history-taking and diagnostic dialogue.
AMIE is a medical AI designed for diagnosing through conversations, developed using real and simulated medical dialogues, and datasets on medical reasoning and summarization. It was trained using a unique approach that includes self-play simulations and feedback to improve its ability in different medical areas. This training involves two main steps: first, AMIE practices simulated conversations with an AI patient, adjusting based on feedback. Second, it uses these refined dialogues for further improvements. AMIE employs a reasoning strategy during actual conversations to provide accurate responses.
In a study comparing AMIE with real primary care physicians using simulated patients, AMIE was found to perform better or equally well in most assessments [3]. This is a step in the right direction for delivering world-class healthcare for everyone, and LLMs can be one of the amplifiers in delivering large-scale healthcare. However, the results should be met with caution and further research has to be done in the field.
Conclusion
Healthcare is comprehensive, and there are many ways for AI applications to improve the quality of care, which is highlighted by the mentioned authors. What distinguishes them is the approach to how AI applications can improve the quality of care. Emmanuel and Watcher emphasize behavior patterns in primarily patients that can be changed rather than diagnostics. They see the opportunity for AI to apply scientific evidence to something practical. Adler-Milstein et al. also note the possibilities with AI applications as a guide for physicians to make crucial decisions through diagnostics. Meanwhile, Tu, T., et al. have developed an innovative chatbot that provides hands-on research on how AI can assist physicians in the fundamental part, which is communication between the patient and caregiver. This breakthrough suggests promising possibilities for improving patient care and a more empathetic healthcare environment.
Author
Filip Dimitrijević is Chairman of the Board at KTH AI Society and, a student in Computer Science at the KTH Royal Institute of Technology. You can reach him on LinkedIn or by email at filip@kthais.com.
Reference
[1] Emanuel, E. & Watcher, R. (2019), “Artificial Intelligence in Health Care: Will the Value Match the Hype?” in: Journal of the American Medical Association.
doi:10.1001/jama.2019.4914
[2] Adler-Milstein, J. & Chen, J. & Dhaliwal, G. (2021), “Next-Generation Artificial Intelligence for Diagnosis: From Predicting Diagnostic Labels to “Wayfinding”,” in: Journal of the American Medical Association.
doi:10.1001/jama.2021.22396
[3] Tu, T, et al. (2024), “Towards Conversational Diagnostic AI?” in: Arxiv.org.
doi:10.48550/arXiv.2401.05654
