- Data-driven approach to support pharmacotherapy selection
Hitachi, Ltd. announced the development of a machine learning-based outcome prediction and comparison technology that predicts with high accuracy the success and failure of various medication options for patients with type-2 diabetes mellitus (T2DM), in collaboration with University of Utah Health (“U of U Health”). The technology analyzes electronic medical records and creates a prediction model which calculates the probability of different medications attaining a target value for hemoglobin A1c (HbA1c)*1 after 90 days of commencing treatment, thus potentially helping medical practitioners choose the most effective medication option. The high accuracy*2 of the technology in predicting the effectiveness of various medications was confirmed using simulation on past records of patients with diabetes.
Currently, the importance of value-based healthcare, which aims to realize both better outcomes while reducing healthcare costs, is receiving attention worldwide. In the United States, the number of patients with diabetes has risen to about 23 million, and one in four people over the age of 65 is diagnosed with T2DM*3. Treatment of T2DM spans from several months to several years or more, during which time the drug(s) selected and the dosage(s) need to be adjusted based on the patient’s condition. This significant variability in pharmacotherapy regimens often leads to a trial-and-error approach in drug selection.
Hitachi has been working on developing various measures to address diabetes using IT, such as life-style modification support and diabetes prevention services *4. In this research collaboration, the knowledge and experience of Hitachi, and doctors, pharmacists and biomedical informaticists at the U of U Health, were used to develop technology to predict the probability of achieving a treatment target with each medication. In developing the technology, data from approximately 6,800 patients were analyzed chronologically from various aspects such as drug category, dosage, treatment period, weight, trend in test results, etc. Machine learning techniques were then applied to the resulting information to build a prediction model for HbA1c, generating patient-specific predictions of the effectiveness of alternate treatment options. Using the resulting model, it is possible to predict and compare the effectiveness of treatment at 90-days, a common period in the United States between clinic visits for patients whose diabetes therapy is being adjusted. As a result, the technology has the potential to support the selection of the most effective medication depending on the patient’s background and condition. The technology was verified using data from another 2,200 patient files by simulation, and the highly accurate nature of the predictions was confirmed.
Hitachi will continue this collaborative research with U of U Health to realize the practical application of this technology and its development of healthcare informatics technologies that support medical practitioners and patients with improved healthcare outcomes*5. Additionally, part of these results will be presented at the IEEE-NIH Special Topic Conference on Healthcare Innovation and Point-of-Care Technologies to be held from 6th to 8th November 2017, in Bethesda, Maryland, U.S.A.