Devising health improvement plans acceptable to patients based on their individual characteristics and diverse preferences is critical in leveraging medical results toward the improvement of patients’ daily lives. However, it is difficult for clinicians to provide such services to a large number of patients. In collaboration with Kyowa Hakko Bio and Hirosaki University, a research group consisting of Professor Yasushi Okuno, special lecturer Ryosuke Kojima, and graduate student Kazuki Nakamura, all from the Graduate School of Medicine, Kyoto University, succeeded in the development of an AI model that proposes optimal and effective health improvement plans for individuals based on personal examination data by combining machine learning, a type of AI technology, with hierarchical Bayesian modeling. The results of this research was listed in Nature Communications.
Proposals for personalized health improvement plans currently rely primarily on the experience of clinicians. A system that supports clinicians in making these plans in a data-driven manner has been desired for some time. To aid the medical and healthcare field, prediction models using AI / machine learning technology have been created and it has become possible to provide diagnostic support and future disease predictions based on comprehensive patient information. While such machine learning models are capable of high-performance predictions, there is a problem in that the prediction process can be considered a "black box". Therefore, even if the AI / machine learning model predicts that the risk of developing a disease is high for an individual’s health, it would not be possible to illustrate a concrete plan regarding the kind of health improvement action that should be taken. The recommended health improvement plan should be set up so that it is effective, and the plan itself is relatively "easy to carry out" for the patient. For example, an improvement plan that shows a combination of test values that cannot realistically be achieved by an individual is not feasible in practice, and an improvement plan that involves a combination of test values that is actually present in the health examination data is considered preferable. Furthermore, to improve health, it is not advisable to restrict individuals across the board by mandating exercise, drinking restrictions, and changes in eating habits. Patients often find methods that remain efficient despite only targeting a few habits or lifestyle aspects more acceptable.
The research group developed an AI model to propose a more effective health improvement plan while considering "ease of implementation." To assess "feasibility," they studied the patterns of real data distribution using a hierarchical Bayesian model. The use of this model in addition to the usual machine learning model allows for the proposal of a health improvement plan that combines realistically actionable values. The developed AI model was applied to the screening of big data acquired through the Iwaki Health Promotion Project at Hirosaki University COI. Researchers evaluated whether it would be possible to formulate an effective health improvement plan for subjects at risk of hypertension or chronic kidney disease (CKD).
As a result, it was confirmed that the developed AI model could formulate individual health improvement plans according to an individual’s health condition. It was also demonstrated that the developed AI model improvement plan was easier to implement compared to other plans, while still obtaining equivalent improvement effects. The current study developed a new instrument to propose an "easy-to-perform" health improvement plan for patients and validated it using existing data. In the future, positive verification of the effectiveness of the developed AI model will be necessary for its practical application.
"Currently, a variety of AI models are being developed worldwide. However, even though AI models have high-performance predictions, the reasons for such predictions are often opaque; they are considered a black box. In particular, the black box problem of AI models is a major barrier to their practical application in human life-threatening medical healthcare. This study solves the black box problem of AI models by presenting optimal plans to improve prediction results as well as the reasons for such predictions. It is our hope that by applying this AI model to a variety of diseases, it will lead to more effective treatment and health promotion," Professor Okuno said.
This article has been translated by JST with permission from The Science News Ltd.(https://sci-news.co.jp/). Unauthorized reproduction of the article and photographs is prohibited.