NEC, RIKEN, and Nippon Medical School, in collaboration with several university hospitals, have been conducting research on the integration of electronic medical records and AI technology in the medical field. They have constructed a multimodal AI that analyzes medical big data for prostate cancer from multiple perspectives. Through this multimodal AI, they found that the pattern of predictive factors captured by AI differs depending on the number of years between surgery and recurrence. The three parties will promote collaboration for the practical application of this multimodal AI with the aim of optimizing treatment planning and early prostate cancer detection.
As medicine becomes highly specialized, there is a need for tools that analyze medical big data from multiple perspectives. However, many conventional medical AI systems target single test data, and there exists a problem wherein making an integrated judgment using multiple test data is not possible. In the current study by the three parties, a multimodal AI was constructed to enable simultaneous analysis of multiple sets of test data.
Based on these results, the three parties will combine NEC's platform technology that integrates various data based on electronic medical records, multimodal AI that utilizes wide-range image analysis technology and feature selection developed by RIKEN, and highly reliable verification data from doctors at multiple university hospitals including Nippon Medical School. They have decided to aim for the practical application of a medical AI system that analyzes various medical data from multiple perspectives.
If this medical AI system is put into practical use, it is expected to optimize treatment plans, detect diseases at an early stage, and safely manage data, thereby reducing medical costs by shortening treatment periods and reducing the workload and efficiency of medical personnel.
In this study, multimodal AI analysis was conducted on prostate cancer, which is one of the most common cancers among Japanese men, using preoperative electronic medical record data and pathological biopsy images. Results showed that the patterns of predictors captured by the AI differed depending on the number of years from surgery to recurrence. This suggests that the mechanism of recurrence may differ depending on the number of years before cancer recurrence.
The accuracy of predicting recurrence up to five years after surgery was improved by approximately 10% as compared with the existing method, namely, the prostate cancer prognosis prediction model (Kattan Nomogram). This improved accuracy was accomplished by 1) improving dimensionality reduction, which is a method of reducing multidimensional information into fewer dimensions while preserving its meaning and which applies machine learning techniques also used for generative AI, and by 2) multidimensional optimization of predictors captured by the AI.
In the future, the target data will be further expanded and verified for practical use. Some of the results of this research were introduced at the 5th Annual Meeting of Japanese Association for Medical Artificial Intelligence JMAI2023) held in Tokyo from June 17 to 18.
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.