The application of artificial intelligence (AI) to the medical field is a theme with great potential, but the stakeholders who play an important role in the research and development face individual challenges. Although physicians and other medical professionals (medical institutions) are well versed in medical issues and problems (needs) and possess abundant medical data and ideas, they are unable to initiate specific research and development due to a lack of knowledge on how to utilize AI technology and networks with IT vendors. On the other hand, IT vendors with AI technologies are also interested in applying them to the growing medical field, but have few networks with medical professionals (medical institutions), making them difficult to obtain medical needs and medical data. Furthermore, they lack experience in pharmaceutical administration and regulations, including the Pharmaceutical and Medical Device Act, making practical application of AI technologies not easy. In addition, pharmaceutical and healthtech companies, at the exit of the R&D, that wish to commercialize medical applications of AI often find it difficult to handle everything from research to business development on their own, both in terms of time and resources. It is important to have a framework in which medical professionals (medical institutions) with issues, IT vendors with AI technology, and exit pharmaceutical/healthtech companies can collaborate from the outset to promote development. To this end, open innovation in the division of labor among different fields is important, and data scientists, AI researchers, and pharmaceutical experts need to work together with physicians.
Currently, we are working on projects such as respiratory function diagnosis, support for maintenance hemodialysis, support for diabetes treatment, assessment of impaired swallowing function, pathology of breast cancer, prediction of arrhythmia and heart failure in patients with cardiac implantable devices, and prediction of blood clots in artificial heart.
Patients with chronic kidney failure undergo maintenance hemodialysis three times a week for life to remove water and waste products to replace the crippled kidneys. The number of patients exceeds 330,000 and the medical cost exceeds 1 trillion yen. Insufficient dehydration causes heart failure, hypertension, and other cardiopulmonary dysfunctions, while excessive dehydration causes hypotension during dialysis, resulting in adverse events such as bad mood and loss of consciousness. Typically, a dialysis hospital manages several dozen patients with one physician, several nurses, and clinical engineers. Human resources at dialysis hospitals are inadequate. The occurrence of adverse events during dialysis consumes scarce human resources and adversely affects the life expectancy of patients, making it an important medical issue.
To achieve safe and secure hemodialysis, our laboratory has been developing software as a medical device (SaMD) equipped with AI that predicts appropriate target amount of total water withdrawal and decrease in blood pressure during dialysis, in collaboration with university hospitals, private dialysis clinics, NEC and NEC Solution Innovator. By learning from the experience (tacit knowledge) of dialysis specialists (experienced nephrologists) (about 2,800 patients with 725,619 dialysis sessions), the SaMD can predict the target amount of total water withdrawal on the day of dialysis with an error of about 1 glass of water, and can also predict the occurrence of decrease in blood pressure (below 20 mmHg) before dialysis starts with an accuracy of AUC 0.91.
Dialysis treatment is a highly specialized field, and it is difficult for non-specialists to set the same volume of water removal as experienced dialysis specialists. However, the number of dialysis specialists is not sufficient, and non-specialists are often engaged in dialysis treatment in rural areas or at night, and in many dialysis facilities, non-specialists, experienced nurses, and clinical engineers assist in setting the water volume under the direction of a few dialysis specialists. This SaMD can reduce the burden of medical personnel involved in dialysis treatment with fewer human resources, and enables safe and secure dialysis treatment.
The number of diabetes patients is considered to exceed 10 million in Japan, and the treatment options have been rapidly expanding with the development of various therapeutic agents for diabete. Insulin injection therapy is necessary to strictly control blood glucose levels in diabetes and prevent diabetic complications. However, the safe dose range of insulin is narrow, and the optimal type and dosage must be carefully selected for each patient, as overdosing results in hypoglycemia. On the other hand, since diabetologists account for less than 2% of all physicians and are geographically unevenly distributed, diabetes patients currently do not always see diabetologists as their primary care physicians, but rather often see a non-diabetologists.
Our laboratory, in collaboration with NEC and NEC Solution Innovator, has been developing software as a medical device (SaMD) that mimics a diabetologist's treatment and provides optimal rapid-acting insulin (morning, afternoon, and evening) and long-acting insulin (before bedtime) dosage units based on blood sugar levels. This SaMD can be utilized in clinical settings to assist non-diabetologists to perform diabetologist-level insulin therapy. Utilizing Skill Acquisition Learning (SAiL) based on deep learning, DM-SAiL was developed and customized in collaboration with physicians and AI researchers to optimize the prediction of diabetic insulin dosage. Development based on patient data admitted to Tohoku University Hospital has been completed, and the AI can now predict insulin dosage with an error margin of about 2 units from the dosage prescribed by experienced diabetologists.
Oral function declines with age, and if such oral frailty is left untreated, it can lead to many physical and social impairments such as eating disorders and dysarthria, as well as generalized muscle weakness (frailty), so early diagnosis and appropriate treatment are important. In an aging society, dysphagia, a form of impaired oral function, is on the rise, and it has been reported that approximately 70% of pneumonia, a leading cause of death among the elderly, is caused by aspiration. Early detection of impaired swallowing function and therapeutic intervention such as rehabilitation are important to prevent aspiration pneumonia, but currently there are only swallowing evaluation methods that place a heavy burden on patients are available, such as swallowing endoscopy and swallowing fluoroscopy. Since the organs used in swallowing and speaking have many parts in common, such as the tongue, oral cavity, and pharynx, we focus on the possibility of evaluating swallowing function from speech or voice, and have been developing software as a medical device (SaMD) that can evaluate swallowing dysfunction from speech data during conversation.
In collaboration with several departments of Tohoku University (Department of Otorhinolaryngology, Dentistry, and Rehabilitation Medicine, School of Biomedical Engineering) and NEC, our laboratory has been analyzing the voice data using an AI engine (time series model-free analysis) specialized for analyzing time series data of all frequencies of speaking sounds of patients visiting the Tohoku University Hospital Treatment Center for Swallowing Dysfunction. At present, we have developed the AI that establishes the baseline (gender, age, individual differences, etc.) of the voices of healthy people, can detect differences between healthy people and patients, and can diagnose an impairment of swallowing function. We will further develop the AI for practical use by learning from medical data of elderly patients with impaired swallowing function. This SaMD can facilitate early diagnosis of patients with impaired swallowing function who could suffer from aspiration pneumonia in the future.
The World Health Organization (WHO) considers respiratory diseases as important non-communicable diseases (NCDs) in addition to cancer, diabetes, and cardiovascular diseases. Typical respiratory diseases include chronic obstructive pulmonary disease (COPD) and asthma. COPD is also highlighted as an important disease in the revision of “Health Japan 21” by the Ministry of Health, Labour and Welfare, and is referred to as a “lifestyle disease of the lung”. However, the prevalence, morbidity, and mortality rates of COPD and other respiratory diseases are not clear due to the lack of widespread use of diagnostic tests for respiratory function. Spirometry is the most important test for respiratory diseases and respiratory function, but it is not widely used. In addition to requiring the effortful cooperation of subjects (patients) (effort breathing), spirometry diagnosis requires determining whether the test is performed correctly and interpreting the output results (flow volume curve). The development of a system that allows non-respiratory specialists to easily interpret the results of spirometry is considered an important medical issue for the diagnosis and early treatment of respiratory diseases.
In collaboration with Kyoto University and NEC Solution Innovator, Ltd., our laboratory has been developing an AI algorithm to determine respiratory diseases and errors during the spirometry using information obtained from spirometry test results (flow volume curves).
The development of software as a medical device (SaMD) that automatically analyzes spirometry results will enable non-specialists to interpret tests, and can facilitate early diagnosis and intervention in respiratory diseases.
・AI for Pathology of Breast Cancer
Breast cancer is the most common cancer among Japanese women, and one (1) in 11 Japanese women is considered to suffer from breast cancer in their lifetime. When breast cancer is suspected by lumps or imaging diagnosis, the definite diagnosis relies on the pathology, which requires an experienced pathologist. Our laboratory had been developing an AI to detect breast cancer lesions from pathological images. Currently in the exploratory research phase, the detection model is classified into three classes (benign, non-invasive cancer, and invasive cancer) or two classes (benign and malignant) and has achieved diagnostic accuracy at 88.3% and 90.5% respectively (published in the scientific journal "Journal of Pathology Informatics"). In the future, we plan to develop AI diagnosis using "intraoperative rapid pathology specimens" in breast cancer.
・Software as a Medical Device (SaMD) for Prediction of Arrhythmia and Heart Failure in Patients with Implantable Cardiac Devices
Implantable cardiac electrical devices such as implantable cardioverter defibrillators (ICDs) and cardiac resynchronization therapy pacemakers (CRT-P) are widely used in heart failure patients. These implantable cardiac electrical devices enable the remote monitoring of ever-changing biological information over time from the patient's home. Our laboratory has been developing an AI that predicts the onset of heart failure and fatal arrhythmias in advance by utilizing the telemonitoring information of patients with cardiac implantable electrical devices.