【Dialogue】 Exploring Data Science for the Future of Digital Healthcare

I graduated from university in 1964, the year of the previous Tokyo Olympics. That summer, I entered a graduate program in the United States, and in the summer of 1971, I returned to Japan and joined a research institute at a Japanese company. While in the U.S., I worked under Professor Satoshi Watanabe at the University of Hawaii, who had transitioned from theoretical physics to pattern recognition. I assisted with his work and received a salary for it¹⁾. Although I earned my degree in theoretical physics, much of my work focused on pattern recognition. One of my labmates in the late 1960s was C. A. Kulikowski, who obtained his degree by applying pattern recognition to thyroid disorders and later joined the Artificial Intelligence Department at Rutgers University. My work involved a variety of topics, including machine learning techniques, applied research, and the development of imaging devices using laser light.

Upon returning to Japan in the summer of 1971, I joined Hitachi’s newly established information systems research institute. There, I became involved in a project to enable computers to perform precise differential diagnoses of heart diseases, collaborating with Dr. Kiyoshi Machii (then at Mitsui Memorial Hospital, later a professor at Toho University Medical School) and young researchers at the institute. This effort involved transferring the experiential knowledge of human specialists to computers to automate the process of enhancing diagnostic accuracy—what we now call an expert system. This work took place in the early 1970s. Additionally, we explored the use of R. Bellman’s dynamic programming to optimize drug administration based on data from gout patients accumulated at Toranomon Hospital. Incidentally, the term “expert system” was coined by Kulikowski. After moving to Rutgers University, he developed a system using expert knowledge for glaucoma diagnosis, which later became part of Japan’s so-called Fifth Generation Computer initiative.

In 1976, I joined the newly established Tokyo Metropolitan Institute of Medical Science. There, I aimed to create an environment capable of handling all types of medical data, including documents, numerical data, waveforms, images, 3D images, and maps—a platform-like system in today’s terms²⁾. By the late 1970s, I had established a Medical Informatics group focused on developing statistical packages, though we did not engage in clinical diagnostic research. However, I became acquainted with researchers like E. H. Shortliffe, who developed MYCIN, an interactive system supporting antibiotic usage at Stanford, and P. Szolovits, who conducted clinical application research at MIT’s Artificial Intelligence Laboratory. These young researchers, in their early 30s at the time, formed the foundation of medical AI groups in the U.S., which eventually led to the creation of AMIA⁴⁾. This movement resonated strongly with Dr. Shigeyoshi Kaihara, then an assistant professor at the University of Tokyo Hospital’s Information Processing Department. Dr. Kaihara, with some support from me, leveraged a series of lectures by these young U.S. researchers to establish the Japan Association for Medical Informatics in 1983. However, our research group has since had limited involvement in clinical diagnostic research.

In the 1980s, while leading a research group in Bioinformatics, I also contributed to establishing a Clinical Epidemiology group. In 1981, we launched the precursor to today’s CBI (Chem-Bio Informatics) Society, based on work that transferred the core elements of the NIH and EPA’s Chemical Information System to our institution. From 1983, we introduced experimental equipment such as clean benches and optical microscopes into our computational facilities and began conducting experiments, including observing embryogenesis using *C. elegans* and performing compound screenings. This involvement connected me to various projects and trends of the 1980s, such as the Ministry of International Trade and Industry’s (now METI) Fifth-Generation Computer Project, molecular and bio devices (and subsequently bio-computers), and neural network initiatives, some of which I supported.

The Fifth-Generation Computer Project is well known for its goal of creating an implementation environment for expert systems. However, I advocated a “semi-empirical methodology,” suggesting that “directly inputting the knowledge of specialists into a computer cannot yield reliable results. Such knowledge systems should serve as a starting point for improvement through learning from real-world data in practical settings.”

Today’s CBI Society is highly focused on drug discovery, but its initial goal was to integrate systems for compound databases and computational chemistry. Toward the late 1980s, I moved to the National Institute of Health Sciences, where I worked on establishing the newly accessible internet infrastructure, launching projects like the “Global Information Network on Chemicals (GINC)” in collaboration with international organizations such as WHO and UNEP, and addressing emerging issues like microplastic debris in oceans.

Since my retirement in 2001, I have stepped away from active research. Reflecting on that era, the period when I was engaged in research marked a time when Japan’s science, technology, and corporate activities (economy) were truly on the rise. Some researchers at institute meetings even declared that “there’s nothing to learn from America.”

Indeed, Japan was vibrant in the 1980s. However, some argue that the U.S. launched efforts to curb Japan’s rise, starting with the 1985 Plaza Accord. This is often seen as the beginning of Japan’s international stagnation. Domestically, the introduction of the consumption tax in the late 1990s is viewed as a self-inflicted setback. Yet, rather than saying Japan declined during the Heisei era, it might be more accurate to say that other countries advanced. Singapore and China, in particular, made remarkable progress, largely driven by their response to the opening of the internet in 1994.

In 1992, as a WHO consultant, I spent a week in Singapore visiting agencies involved in chemical management, and I was struck by the experience. Singapore was ambitiously aiming to establish an electronic government by 2000. Around the same time, at a WHO conference in Malaysia, I was surprised to see participants from various countries, including China, Hong Kong, Canada, and Fiji, all of whom were ethnically Chinese.

Inspired by the introduction of internet and WWW environments to my research institute and the CBI Society, I felt compelled to write a book to highlight the opportunities and warn of the challenges of this era. This resulted in my book *The Third Opening of Japan: The Shock of the Internet* (Nijiman Kaminuma, Kinokuniya, 1994). Unfortunately, it did not receive the attention I had hoped for. I believe the theme of “The Third Opening of Japan” diluted the urgency of the message that Japan needed to address the internet wave immediately. The goal should have been not an “opening” but a “reformation.”

Subsequently, the nations and companies that embraced the internet wave experienced significant growth. In this sense, the 1990s was a major turning point, and the defining keyword or zeitgeist of the era was undoubtedly the response to the (opened) internet.

It’s a field where generalizations are challenging, but in areas I’ve been somewhat involved with—such as natural sciences, health and medical care, drug discovery, and environmental science research and development—there are frameworks that can be made more accessible.

For instance, in the context of medical diagnosis, doctors constantly make decisions. The foundational knowledge, information, and materials they rely on are diverse, including language (descriptions, natural language), waveforms, images (photographs), and observations of movements. Today, many of these elements can be handled digitally, though not completely. At a basic level, they encompass language, waveforms, diagrams, and images, including videos. This is a classification by data format.

Another classification pertains to data associated with domain-specific knowledge. For example, molecular spectra and structural data related to molecular interactions are vital to chemists; genomic sequences matter to geneticists and biologists; and images of malignant skin tumors are tied to the expertise of pathologists. Chemists may see proteins as large molecules, while biologists or informaticians regard genes as nucleotide sequences (codes).

There is also a cognitive purpose-based classification, including deduction (reasoning), induction, creativity (inspiration), and planning techniques for goal achievement. For instance, the first data discrimination (classification) technique I worked on in the 1960s, which would now be called an SVM (Support Vector Machine), was related to quadratic programming methods. However, its use was primarily in pattern classification and inductive reasoning (inference or conjecture). In other words, cognitive computational techniques (AI) are intricately nested within these domains.

My classification is experiential, and I am not aware of others’ perspectives. Given the current AI boom, I think there’s a need to explain these data science techniques and their classifications in simple general terms. However, to be honest, it’s a formidable challenge.

It’s not just AI—there isn’t an established academic framework for informatics, like there is for physics. This seems to be related to challenges in education and human resource development. The unconventional physicist Richard Feynman once delivered a lecture on the theme of “computation,” which was later published as a book⁶⁾. It’s a fascinating book that connects to quantum computing, but in it, he says something along the lines of, “Programming and computational techniques are not science but art.”

In the United States, there are departments or faculties of Computer Science, where Ph.D. degrees are awarded. While the situation may be different now, in Japan, creating an excellent program alone was often insufficient to earn a doctorate. This may be one of the factors contributing to the shortage of talent in this field in Japan.

The concepts of “health” and “illness” are not scientific but rather convenient constructs. During the 1980s, when biology was in the spotlight, Dr. Ken Ohno of the City of Hope, a renowned Japanese researcher based in the U.S., discussed genes that humans could modify. His proposals have become realistic issues with advancements in genome editing technology. It’s often said that pharmaceutical companies “create diseases,” as defining diseases allows for the development of new drugs⁷⁾. From a data perspective, this may relate to biomarker discovery. While individual indicators are easier for doctors to understand, in reality, composite indicators are sometimes necessary.

I’ve proposed the concept of “Health Metrics.” Intervention at the early warning stage is the most effective, as outlined in a roadmap by former NIH Director Elias Zerhouni. While I advocate for interventions at the tertiary prevention stage, there’s growing recognition in the U.S. of the importance of maintaining health through lifestyle innovations in next-generation healthcare⁸⁾. The key question is whether aging should be considered a disease to be treated. Recently, there has even been research using AI to measure healthy life expectancy⁹⁾.

I’m not sure if my perspective is accurate, but about a decade ago, there was widespread discussion about whether the Human Genome Project, which concluded successfully over a decade prior, had significantly contributed to medicine or drug development. Francis Collins, who led the international team for the Human Genome Project at NIH, became the NIH Director in 2011. Despite budgetary constraints, he established NCATS (National Center for Advancing Translational Sciences) under NIH, appointing his former subordinate, Christopher Austin, as its director. NCATS not only promoted collaborative projects with big pharmaceutical companies but also launched the BD2K (Big Data to Knowledge) initiative, which aims to extract knowledge from big data. Additionally, NIH created a position for a leader (P. E. Bourne) to oversee digital initiatives, such as AI and data science. Collins also advocated for Precision Medicine, initially targeting cancer (Precision Oncology) and later focusing on PGx (Pharmacogenomics) to promote the proper use of medication¹⁰⁾. This reflects his determination to fulfill the promises he made as a leader of the genome project.

Transforming genetic data into breakthrough drugs is still in its infancy and will take time, but I believe progress is steady.

I don’t know the details, but reading about the success of CAR-T therapy¹¹⁾ gives me hope. However, Japan seems to be significantly lagging behind. While Japan appears to have a strong tradition of excellent immunological research, the issue seems to lie in translating that research into drug development.

This is a very personal impression, but based on my experience since the 1970s working on immunology (such as hepatitis B and complement systems) from a computational application standpoint, I’ve found immunology to be extremely challenging to handle from a systems (engineering) perspective. This difficulty arises because the subjects—cells and molecules (like complements)—continuously change their states. I believe this issue is also relevant to regenerative medicine. It relates to the challenge of “cell fate and control,” which suggests that computational applications require some innovative approaches. However, this very complexity makes it an appealing and worthwhile area to tackle from the perspective of computational and information techniques.

That said, addressing this challenge requires teams composed of biomedical experts and younger generations specializing in computational and information techniques. One example, based on online information, is the European (German-French) LifeTime Initiative¹²⁾.

In any case, cancer, immunology, and research, as well as clinical applications, appear to be advancing at an accelerated pace. However, I feel that adequate frameworks to fully process and utilize the data generated are not yet in place. Perhaps it’s necessary to start by focusing on specific targets, such as cancer or autoimmune diseases, and systematically collecting clinical data related to them.

Many of Japan’s research and development plans lack strategic depth and often prioritize visually appealing, simplified ideas. At some point, it became mandatory to include illustrative “concept diagrams” in grant applications for scientific research funding. For example, even in robotics research, “human-like qualities” tend to be highly valued. I call this phenomenon the “Astro Boy Syndrome.”

In AI, I believe there should be increased funding for socially critical issues, such as nuclear decommissioning, fraud prevention (e.g., “It’s me” scams), and understanding and addressing dementia. Furthermore, I believe the government should support the development of cancer treatment drugs as a response to the aftereffects of nuclear exposure. Innovation doesn’t emerge from merely aiming for it; it arises from tackling the challenges that need to be addressed.

Regarding biobanks and cohort studies, the 10th issue of *Drug Discovery Plaza* discusses baseline studies. What’s currently needed is a numerical understanding of normal states relative to disease states. Additionally, there needs to be more recognition of the importance of clinical record management techniques and databases that can generate disease-specific knowledge from real-world data (RWD).

For instance, the U.S. company Flatiron Health was acquired by a major pharmaceutical firm for a substantial amount, likely due to the value of its techniques for recording clinical (support) data. Analyzing such data is expected to yield useful insights into cancer care. A similar approach could be applied to clinical records and database construction for autoimmune diseases like rheumatoid arthritis. These can be considered “heuristic databases.” The same principle applies to biobanks.

Developing such systems in Japan is extremely challenging, but analyzing why this is the case could reveal new areas for innovation.

That’s an excellent idea. Japan’s insurance system operates on the assumption of inherent goodwill. The internet of around 1994 was also supported by volunteers operating on similar goodwill principles. Both of these systems now seem to be faltering, which is why I believe we need to explore new models.

Perhaps self-serving, but the nonprofit organization we established, the Institute for Cyber Bond (ICA), aims to address these two changes. Specifically, we are promoting “participatory healthcare.” In simple terms, this means encouraging patients and citizens to make smart use of existing services while taking active steps to manage their health through measures like exercise, diet, sleep, and environmental adjustments without necessarily relying on prescriptions.

In Japan, there is still a prevailing ethos of “patients should just follow instructions,” reflecting a paternalistic approach among doctors. While such paternalism has also been criticized in other developed nations, there is growing recognition in the U.S. and Europe of the importance of patient-centric care. In Japan, companies like Takeda Pharmaceutical have started adopting a “Patient First” approach. In contrast, we advocate for empowering citizens and patients to take proactive roles in maintaining their health and addressing illnesses.

Lifestyle diseases are a prime example where ensuring behavior changes among citizens and patients is crucial. Practical methods to achieve these changes are at the heart of what we should be implementing.

I was involved in founding what is now the CBI Society around 1980 and continued to engage in its operations until 2010. However, I am not certain how much of an impact this activity has had in the field of drug discovery. At that time, pharmaceutical companies were highly secretive. Nonetheless, a recent success story has been reported in detail by a group from Taiho Pharmaceutical involved in the development of Lonsurf¹³⁾. It seems that data science tools are now recognized as indispensable at various stages of drug development rather than being tied to the development of specific drugs.

I find training human resources for ICT utilization extremely challenging. The fundamental issue lies in a lack of understanding of specialization. Informatics itself has not been systematically organized, and there is no comprehensive overview of how techniques that support or replace human thinking and cognitive functions relate to other fields such as natural sciences, engineering, and medicine.

I foresee that current data science and AI will evolve into a new academic domain, independent of mathematics, natural sciences, and engineering, which could be called “quantitative thought studies.” This field will include applications in healthcare and drug discovery.

One of the core challenges in understanding informatics or computational techniques, including AI and machine learning, is the rapid pace of technological advancements and the increasing breadth and depth of their applications. Moreover, acquiring these skills tends to favor younger generations while being more challenging for older individuals, akin to gymnastics or swimming, where early involvement offers significant advantages. Consequently, a gap widens between developers and users of data science technologies in fields requiring years of expertise.

This generational gap raises issues of how to harmonize the roles, job functions, and career advancement of younger professionals. Their work often does not align with traditional evaluation criteria for researchers in science and engineering. This misalignment is because platforms for presenting results, such as conferences, journals, and research communities, have not kept pace with their work. It is not uncommon for data scientists to be appreciated for their contributions but not adequately compensated or recognized in terms of position or salary.

In the private sector, hierarchical structures like subcontracting often result in poor recognition and treatment of IT specialists who perform the actual work. While this issue resembles the treatment of skilled experimental technicians, it differs in that computational techniques directly relate to human cognitive tasks like reasoning and decision-making. These techniques may lead to “creative destruction,” altering the design of work or research itself.

Given the rapid pace of ICT advancement, professionals focusing on computational techniques need dedicated environments to specialize. Becoming an expert takes time, and young professionals should ideally develop their skills by tackling real-world problems. Creating an environment that facilitates this approach is crucial.

In Japan, particularly in healthcare, what is lacking are strategies and the logistics mindset. From a strategic perspective, NIH declared its commitment to transforming into a data science-based organization¹⁴⁾, with “continuous knowledge generation and utilization” as its goal. Data science is inextricably linked to knowledge generation. In the U.S., the National Library of Medicine (NLM), whose director in the 1980s, Donald Lindberg, was both a physician and an AI researcher, has become a foundational institution for genomic medicine¹⁵⁾. Japan lacks such a foundational institution. Additionally, the application of AI techniques requires a glossary of technical terms. Japan has few researchers in natural language processing related to healthcare, compounded by challenges with the limited user base of the Japanese language.

To promote data science in Japan, a robust support infrastructure is essential. Training human resources must also be viewed from this perspective. In drug discovery, as NCATS suggests, it’s vital to restructure workflows, particularly pipelines, in pharmaceutical companies and reassess each component from a data science perspective¹⁶⁾.