The Necessity of Prescribing the Optimal Oral Hypoglycemic Medicines for Each Patient
Today, all over the world, the number of diabetic patients is increasing. Diabetes is a disease in which blood glucose levels become too high. Prolonged high blood glucose levels cause various complications. According to the National Health and Nutrition Survey Japan, 2016* by the Ministry of Health, Labour, and Welfare, the number of individuals strongly suspected of having diabetes is estimated to exceed 10 million (including untreated cases). This situation must be addressed urgently.
To treat diabetes, it is vitally important to provide appropriate medical treatment by medication, such as oral hypoglycemic medicines or insulin preparation. In particular, blood glucose levels should be controlled continuously to keep "HbA1c levels below 7.0%," the target for preventing complications recommended by the Japan Diabetes Society. HbA1c is short for "hemoglobin A1c," which refers to the ratio of hemoglobin bound to glucose in the blood; in diabetes treatment, this is used as an indicator of blood glucose.
In long-term treatment, it becomes difficult to keep this indicator at a low level. Because the diabetic patient's condition tends to become more complex due to intercurrent illnesses and other factors during long-term treatment, many different oral hypoglycemic medicines must be combined appropriately according to the patient's condition. However, at present no method has been established for determining the optimal prescriptions in consideration of the sequence and timing of medication.
* The results of the National Health and Nutrition Survey Japan, 2016(Japanese)
AI to Learn the Medical Data of Approximately 5,000 Patients and Created Learning Models to Predict Treatment Effects
Since February 2019, Fujitsu and Fujitsu Hokuriku Systems have been conducting joint research with a research group led by Professor Hirofumi Onishi (Chief of the Medical Information Division) of Sapporo Medical University to use AI to optimize prescription of oral hypoglycemic medicines in diabetes treatment.
A dataset will be created from a massive amount of medical data—such as medical records, test values, and prescription data—from approximately 5,000 diabetic patients examined at Sapporo Medical University Hospital in a format that strips personally identifiable information. AI will then learned the dataset and learning models will be generated to predict the treatment effects of medication.
From Sapporo Medical University, experts in medicine and medical information, including clinicians, are participating in the joint research. Fujitsu, which is well-versed in safe, secure handling of clinical data, such as patients' prescription data and test values, and Fujitsu Hokuriku Systems, which has full-time AI engineers, are also participating.
Key Points: Technology to Create a "Highly Accurate Dataset" for AI to Learn and Creation of Learning Models
This joint research has two technical key points. One is technology to create a "highly accurate dataset" for AI to learn. Collecting and processing clinical data to prepare data for analysis requires ingenuity. For example, diabetic patients each visit the hospital at intervals that may vary and undergo tests that may vary each time. Thus, missing data and variance become issues. While considering these issues, the research team first determines "which part of changes in HbA1c levels should be looked at" for each patient. Next, the team extracts "medicines used and test values that indicate the patient's condition." This process enables the team to produce data relevant to the answers it wishes to obtain, and such data is used to create a dataset for machine learning.
The other key point is the "creation of learning models using AI technology." For the aforementioned dataset, the team then explores an appropriate learning method. Also, by incorporating what experts think is important, the team has AI learn the relationships between factors, such as medicines and test values. Based on such AI learning, the team creates learning models to predict the effects of treatment.
Technical key points of the joint research
Toward the Development of "Personalized Treatment" of Diabetes
At present, diabetes treatments vary greatly from patient to patient. Some patients can control their HbA1c levels effectively with one type of medicine, while others require many different medicines. For this reason, "personalized treatment," such as selecting the personally optimal medicine or combination of medicines, is important. The physical burden on patients will be reduced if this joint research can enable the creation of models to predict the effects of oral hypoglycemic medicines suitable for individual diabetic patients and to help provide the optimal prescriptions.
In view of specifically what kind of learning models can be created, the joint research experimentally creates and evaluates learning models repeatedly, aiming to generate successful results. Thereafter, the team plans to take steps toward practical use of such learning models. For example, it will explore ways to help physicians understand and make use of AI's outputs regarding diabetic medicines in order to help them choose the optimal medication for each patient. We can sufficiently anticipate that this know-how of understanding data will be applied to treatment of other diseases in the future.
Fujitsu will continue to make full-scale efforts in the joint research and produce innovative value for society to bring about a prosperous future for people's lives.