"The disease risk prediction AI service" inputs one year's worth of health examination data (hereinafter referred to as "health examination data"), and by six years from now, diabetes, hypertension, dyslipidemia, renal dysfunction, liver dysfunction. And it is a service that predicts the possibility of developing six lifestyle-related diseases called obesity.
One of the diseases that can be predicted by this service is diabetes. The AI model we have developed predicts diabetes with an over 90% accuracy rate. This high level of accuracy has been made possible by the broad-ranging AI technologies that Toshiba has accrued in its over 50 years of AI research, together with massive amounts of medical examination data and medical expense statement data for Toshiba Group employees, collected over the course of multiple years. This involves the handling of employees' precious personal information, so we use Toshiba's information security technologies such as data anonymization to appropriately manage and handle the data.
Toshiba's AI researchers have a tradition of placing great importance on field expertise and knowledge when utilizing data, and they value dialog with people working in the field. AI researchers, occupational health physicians, public health nurses, and other professionals have collaborated and discussed how this data should be used and the direction to be taken by the AI that is developed.
Through these efforts, Toshiba has created an AI prediction model based on machine learning and used it to perform analysis of the patterns shared by people who develop lifestyle-related diseases, their associated traits, and more. When creating an AI model, the quality of the data used for learning is extremely important. Machine learning requires a huge amount of learning data, so the quality of the data has a particularly large impact on the predictive accuracy of the AI model.
For this service, we used medical examination data and medical expense statement data as learning data. Changes in result values in the medical examination data and information regarding medications in the medical expense statement data were used to determine whether people had lifestyle-related diseases. The AI model then studied lifestyle habits and how medical examination values in the medical examination data changed over time.
Toshiba has a large number of employees and a low attrition rate. This means that we have a large amount of data covering multiple years for each employee. We were able to use this data in machine learning to develop an AI model that produces highly accurate predictions.
We also used the "random survival forest" method, a statistical machine learning method, and applied our own unique optimization technologies to achieve a high level of predictive accuracy. We used the AI model to predict who had a high risk of developing diabetes within the next three years. The results that were produced had a high predictive performance, with an AUC* of 0.96.
* AUC (Area Under the Curve): An indicator of prediction quality. AUC is a number between 0.0 and 1.0. The closer to 1.0, the greater the predictive accuracy. An AUC of 0.5 is the equivalent of random prediction.
Not only does the model have high predictive accuracy, it is also exceptionally generalizable, effective even when used with unknown data. The same AI model was applied to a different group of samples, separate from those of Toshiba's health insurance union, and achieved an AUC of 0.94. These results were presented both at the Scientific Meeting of the Japan Society of Ningen Dock* and at the World Congress on Ningen Dock**.
* 60th Scientific Meeting of the Japan Society of Ningen Dock (2019) "Development of an Algorithm for Predicting the Onset of Diabetes through AI Analysis of Health Examination Data"
** Joint Meeting of the 27th International Health Evaluation and Promotion Association & 4th World Congress on Ningen Dock (2020) "Prediction of diabetes, dyslipidemia, hypertension, liver function & renal function using AI"
Our solutions for preventing complications use data such as medical examination and medical expense statement data to identify individuals at high risk of suffering diabetes complications and to warn them of this danger. We are also conducting research in conjunction with a university with the aim of using Toshiba AI technologies to prevent increases in the severity of diabetic nephropathy. Our AI model is being used to divide diabetic nephropathy patients into multiple categories and develop optimized preventative methods that are stratified and systematized for each individual risk. These preventative methods assist with providing patients with guidance regarding lifestyle habits, tailored to their individual risks and symptoms. We believe that this can help them enjoy a greater QOL.