Hitachi and U.S. Medical Institution Jointly Develop AI for the Medical Field

NAONOI managed the joint project between Hitachi and Partners HealthCare Pivot Labs. During my undergraduate and graduate years, I was engaged in basic research on numeric operations in supercomputers, and I chose to work at Hitachi because I wanted to work for a company that was doing research and development (R&D) on supercomputers. Since joining the company, I have been involved in putting high-performance computing (HPC) technology into practical use. Although I was assigned to the R&D department, I have often been involved in relatively flexible projects. I have had experiences in various businesses through projects on finance, and oil and gas, and on this occasion, I was put in charge of an informatics (information sciences) project in the healthcare field.

SHIBAHARAAlthough I studied mathematics at university, I went to graduate school for information sciences because I wanted to apply math to real-world problems. I conducted research on a type of image processing called computer vision, and upon completing a Ph.D. course in information sciences, I joined Hitachi. For a few years I conducted research on electron microscope image processing techniques. However, with the advent of deep learning around 2012, AI became mainstream and I felt as though the age of processing on an ad hoc basis had come to an end. Seeing this as an opportunity, I decided that I wanted to conduct research on AI in a new field, joined a doctoral program in medicine, and researched AI that could be used for medicine. Although I had thought that AI could be used widely in the medical field, my supervisor who was also a physician said, "If it cannot explain why it has made a certain decision, AI cannot be used for medicine." With this in mind, my doctoral degree focused on the development of a method by which an AI could explain the reasoning behind the results it predicted, and the application of this method to epidemiological research. Currently, I conduct research by tackling new problems with researchers from different fields.

TAKATAI was involved in researching the application of machine learning methods in bioinformatics during my years as a student. However, I really wanted to study overseas, and upon obtaining a Master's degree in bioinformatics, I joined a Ph.D. program at a university in the United States. I switched to a Master's program while I was there and ended up with two Masters degrees. The thing that triggered my connection to Hitachi was a career forum of Japanese companies held in Boston. The person in charge of personnel strongly encouraged me to get an interview. However, I had a full schedule at the career forum so at first I refused to give up time to interview with Hitachi. However, they insisted that I have an interview at lunchtime, and I ended up joining Hitachi because of this. At the outset, I was involved in research on data management platforms such as databases and big data processing. Now I conduct research based in the U.S. at Hitachi America.

NAONOAround the 2000s, the fields where IT could be applied began to expand, and so did the research areas. In the midst of this, there was an increase in projects where we would collaborate and co-create solutions with clients. It was a research style where we would use a project team at Hitachi to make new discoveries with client data; the collaboration with Partners HealthCare Pivot Labs is one example of the research projects we conducted in this manner. It was necessary to have experts in machine learning and data processing, which is what prompted Shibahara and Takata to both join the project. This project was fairly agile, and small corrections could be made to the theme of the project and its members throughout its course. I inherited the project in FY2017, and we have continued R&D on it ever since.

SHIBAHARAThe project with Partners HealthCare Pivot Labs required people who were familiar with analyzing big data and using AI in healthcare. In conducting projects with data analysis, I believe the most difficult part is assembling experts from each field as members on the team. At Hitachi, we were already running projects on developing AI for medicine.

TAKATASince I am with Hitachi America, I report to a completely different set of people than the healthcare project team members. With all three of us belonging to different departments, the project activities had to be negotiated with each of our supervisors.

NAONOI also have my hands in multiple special research projects. You could say that our style is something like the "xx production committees" you see in movies, which are agile and produce results by gathering the necessary resources.

NAONOOur collaborator in the project, Partners HealthCare Pivot Labs, is a division of Partners HealthCare, an integrated medical center based in Boston that is comprised of 14 hospitals and community health centers. In addition to having two academic medical centers and multiple community and specialty hospitals, it is also a principal teaching affiliate of Harvard Medical School. The value that data carries is important in big data and AI R&D, and it was important that we were able to collaboratively create with a medical institution that has state-of-the-art technology and data.

SHIBAHARAPartners HealthCare sees 1.5 million patients annually and has 20 years' worth of medical records, all stored on patients' electronic health records (EHRs). Both the quantity and the quality of this data is particularly high: even physicians' consultation notes are clearly recorded and archived. Through our collaboration with Partners HealthCare Pivot Labs, Hitachi had the opportunity to work with a uniquely robust set of de-identified patient data. That we conducted AI R&D on medical data linked with hospitals that are teaching affiliates of Harvard University provides credibility to our project. As a result, we believe further investment is necessary for the medical information management field in Japan.

NAONOIn the U.S., medical institutions have been required to adapt to health insurance reforms introduced by former president Barack Obama's Affordable Care Act, commonly referred to as "Obamacare." One of these changes was the enactment of the Hospital Readmission Reduction Program (HRRP), which places penalties on readmissions within 30 days of being discharged from a hospital. If a hospital's readmission rate is higher than the U.S. average, 2% to 3% of revenue can be penalized. This is a significant issue for hospitals, because if a hospital has revenue of 10 billion dollars, it could be fined 200 to 300 million dollars, an enormous amount. Because of this, there is an incentive among hospitals to develop new ways to improve patient outcomes and therefore reduce readmission rates.

Partners HealthCare had already been working on implementing several initiatives to reduce readmission rates. Although a major electronic health record company and others had provided solutions, they were not producing results. In the midst of this, an American employee who worked at Hitachi's headquarters and had connections at Partners HealthCare suggested the two companies collaborate on a solution.

SHIBAHARAFollowing the completion of the first phase of the project in December 2017, we issued a press release announcing that we had jointly developed a way to predict the readmission risk of heart failure patients with high accuracy using AI, and that this AI-based program could potentially reduce readmission rates by more than half when compared to conventional methods. As a result, medical costs per patient could be reduced by 800,000 yen a year.

Up until this point, there had been a significant barrier when using machine learning – particularly in deep learning – for medicine: because deep learning cannot indicate the reasoning used to make a prediction, physicians cannot take informed action on those predictions. This is, simply knowing a patient is at risk is not sufficient: Physicians need to know why a patient is at risk in order to effectively take action.

At Hitachi, we possessed AI technology that could indicate reasons for decisions based on research that I conducted in the medical field. This relates to how physicians make decisions on diagnoses and treatment methods from statistical processing on a daily basis. From the 1950s onwards, physicians have been using statistical processing called logistic regression to consider probabilities and factors for the onset of illnesses. In fact, the formula used for logistic regression is the same as a formula called perceptron in the field of AI, and its only difference was its purpose.

For example, physicians have selected some variables from the clinical laboratory values (e.g., white blood cells) to construct a logistic regression model. My research enables the construction of algorithms and the selection of variables to be conducted by machine learning: the AI generates logistic regression models compatible with the data physicians have compiled over the last several decades. This enables physicians to determine which variables have what influences on the model presented by the AI based on statistical processing. In other words, physicians are now able to confirm the reasoning of the model presented by the AI, aiding new awareness.

In the collaboration with Partners HealthCare Pivot Labs, we utilized this method to create a formula that predicts readmission risk of heart failure patients. First, we developed an AI that was based on medical guidelines and information from a subset of Partners HealthCare's hospitalization and consultation records. Next, the AI created a model that conducts risk prediction by only selecting data elements with high impact on risk prediction, such as age, comorbidities, prescription information, and the number of days a patient has spent in the hospital. When patient admission information is processed by the AI, the risk prediction algorithm calculates the probability of readmission for that patient.

TAKATAThis outcome – the development of an AI that could predict and explain readmission risk – was published in a press release distributed in December 2018. While Dr. Shibahara continued to research AI models that predicted readmission risk in Japan, I was processing Partners HealthCare's data in the U.S. To comply with Partners HealthCare's security protocol, the data could not be taken outside the U.S., so Dr. Shibahara often visited to continue developing the AI using the dataset I processed.

SHIBAHARAAlthough the AI model initially achieved a high accuracy in the first phase of the collaboration, we saw a decline in accuracy during the subsequent validation phase. At first, we were unable to determine what could have caused this decline.

TAKATAHowever, one possibility was considered, which was a situation that occurred outside of Hitachi's purview: the electronic medical record vendor utilized by Partners had changed between phases. As a consequence of this, the format in which some data were stored also changed, rendering some of this data undigestable by the model, which of course would have an impact on the accuracy of the model's predictions. Although we were able to address the biggest source of data change (laboratory results) by mapping the new data labels to old data labels, the accuracy improved, but not quite to back what was achieved in the first phase. It is possible there were other data changes that we were unable to detect. Realizing that this situation is a risk which could occur in other systems as well, we decided to develop an AI that could accommodate such unforeseeable changes in data structure, and create a system that enables one to track the source of these problems immediately.

This led to the development of the "explainable data management technology for AI prediction results." The components which predict readmission risk with AI include the original training data, calculated risk features, and the most current version of model. Predictions of readmission rates are made in percentage values by applying the model on test samples. By doing this, we developed a data management technology that can search for the data source that brings the calculated features and the model from the predicted risk factors obtained with AI. By providing a dashboard where data to be used by the AI are centrally controlled and made visible, we were able to easily gather supporting evidence for the AI's prediction results. This made it easy to confirm results by returning to the data source, making the AI even more reliable.

For example, one of the features calculated by the heart failure readmission risk prediction AI that ranked high was the term "milk." If one only looks at the predicted results, it appears as if consuming milk increases the readmission risk of heart failure patients, but its real meaning was unknown. Upon using the explainable data management technology, we found a record in the data source, on which the AI based its prediction, referring to "milk of magnesia prescribed for constipation." This led to the realization that it was not milk that was the problem, but rather the prescription of milk of magnesia (typically prescribed for constipation) that was in some way related to readmission risk for heart failure patients.

NAONOExplanation data management technology is not something that we envisioned as part of the original project. It is something which emerged from what we learned about the nature of EHRs during the first and second phases of the project, and how it can affect prediction accuracy.

TAKATAIn reality, we are somewhat fortunate that the EHR data management system changed and revealed the impact this has on prediction accuracy, which is something we might not have learned otherwise. Because of this, we were able to produce a new technological development: a data management platform for AI. This technology can be applied not only for readmission risk prediction, but also for a wide range of other AI technologies. Because the process of data collection, processing and model creation is ongoing in AI development, it would be likely to become individualistic and confuse the details of the processed data. We believe that this data management technology can holistically manage even the reasoning of matters, and could become a solution in various fields.

NAONOWhen I first had a discussion with Partners HealthCare Pivot Labs in May 2017, we talked about what kind of AI we wanted to create. The idea that we settled on was "Human Power Enhancement AI." We agreed that AI would not replace physicians, but that AI is necessary to enhance the power of physicians. Although the joint project began with the goal of helping hospitals avoid paying fines related to HRRP, we found that in the end, physicians care about what they can do to help individual patients. We believe that the AI developed by Hitachi contributes to this desire to help patients regardless of medical condition or readmission risk.

SHIBAHARAAlthough the technology behind AI has advanced significantly, we feel as though AI does not even come close to human physicians. Physicians are able to reach diagnoses immediately after just looking at a few images from MRI scans. I was once told that physicians are not simply looking at images, but that they also overlay anatomic images with the actual state of the disease in front of them when conducting surgery in order to conduct proper diagnoses. Acquiring the various data that physicians have obtained and organizing them for use by AI is extraordinarily difficult. As a researcher, I am even hesitant to call the current state of AI, AI. In order to create "true AI" that could solve real problems in the medical field, I feel strongly that it is necessary that physicians, AI, and data processing experts would have to collaborate as professionals and design AI as a system.

NAONOThe two turning points of the project were when we came up with the idea to create Human Power Enhancement AI through discussions with Partners HealthCare Pivot Labs, our creative collaborators, and when we learned the impacts of the EHR transition, which revealed a gap in terms of the reproducibility and generalizability of AI predictive performance. Each of these led to the development of the readmission risk prediction AI and the explanation data management technology. Although managing the project was difficult, I believe that Shibahara and Takata worked together to break through two large barriers to success.

SHIBAHARAFor me, research is like a hobby. I believe this might be true for most informatics researchers, and we always check our ideas against programs immediately. On the other hand, I also enjoy playing the viola in an orchestra that plays gaming music. In research, even if you do not obtain the results you hoped for, you gain knowledge that will help you in the future. However, with music, no matter how much you practice, sometimes it just does not turn out well. I believe that going back and forth between research and playing instruments helps both my research and my music.

TAKATAI enjoy many hobbies such as traveling, flamenco, and playing the guitar. I sometimes find new sides to my research when I show off the results of my research to my friends who have no relation to science, such as my flamenco and guitar friends. I hope I am able to value the inspiration that rises up when I do things apart from my research.

SHIBAHARAIf you only do research, ideas often stop flowing. Just as collaborating with other researchers is important, I believe that researchers themselves should try out many different things.