Designing Clinical AI: From Technical Possibility to Meaningful Care

A conversation with Florent Pollet — Biomedical Informatics Researcher at Columbia University — on bridging the gap between machine learning and patient care.

By Louise Servoin · 2026-04-30 · 12 min read

## Florent Pollet – PhD Student in Biomedical Informatics at Columbia DBMI (NYC)

Expert in Data Science and Machine Learning applied to Healthcare. Trained in engineering, computer science, and applied mathematics, Florent works at the intersection of machine learning and biomedical informatics. Based in New York City.

Florent Pollet's path has been shaped by a movement from understanding the world to wanting to improve it. Trained in engineering and computer science at Mines Paris PSL and ENS Paris-Saclay, he was first drawn to the power of models, systems, and computation to make sense of complexity. But over time, this technical foundation found its most meaningful direction in biomedical informatics at Columbia University, a field where abstract tools meet one of society's most concrete and consequential challenges: health.

For Florent, healthcare is not just another domain for AI application. It is one of the most complex spaces in which technology must prove its value, because it involves biology, uncertainty, institutions, ethics, and human judgment. Machines may become faster, more capable, and even more natural in the way they communicate. But the essential question remains human: how do we move from what is technically possible to what is clinically meaningful, responsibly implemented, and genuinely useful?

We spoke with Florent about the role of AI and data science in shaping the future of healthcare education, the importance of connecting technical possibility with real-world clinical practice, and how biomedical informatics can help transform information into better decisions, better systems, and more personalized care.

## The Hybridization of Worlds

Florent, your profile is striking. You come from a background in top-tier French engineering and applied mathematics but now operate at the heart of Clinical Informatics in New York. What does this technical perspective bring to biomedical informatics?

It is fundamentally a matter of translation, but also of learning to reverse the usual direction of technical thinking. Engineering teaches you how to move from a well-defined problem toward an efficient solution. Clinical informatics reminds you that, in healthcare, the problem is rarely well-defined at the beginning. You first have to understand the clinical workflow, the constraints, the uncertainty, and the human context before deciding what kind of tool should be built.

But once the problem is properly defined, technical expertise becomes essential. This is where engineering, computer science, and applied mathematics bring real value: in formalizing the problem, choosing the right modeling strategy, designing robust systems, and evaluating whether the solution actually works. The goal is not to apply technology for its own sake, but to use technical tools with precision once we understand what needs to be solved.

When I work on machine learning models using electronic health records, I do not see the data as a clean abstraction. Behind every variable, timestamp, and missing value, there is a clinical workflow: a nurse under pressure, a sensor that may be unreliable, a physician making decisions with incomplete information, or a documentation process shaped by the realities of the hospital. Understanding this context is the first step, because it tells us what problem we are really trying to solve.

That is why clinical data is never neutral. It carries the biases, constraints, and rhythms of the system that produced it. In healthcare AI, the challenge is therefore not only to build models that perform well on standard metrics, but to define the right objectives in the first place. We need metrics that reflect clinical reality and tell us whether a tool is actually useful, safe, and actionable in practice.

For me, this is where engineering and clinical informatics meet. Clinical informatics helps us start from the right place: the clinical problem, the workflow, and the human context. Then technical expertise becomes essential to translate that understanding into robust models, reliable systems, and meaningful evaluations. The goal is not to optimize models in isolation, but to build technology that corresponds to the world in which it will be used.

"The goal is not to apply technology for its own sake, but to use technical tools with precision once we understand what needs to be solved." — Florent Pollet

## AI and the Evolution of Education

The link between education and AI is a burning topic. How is AI changing the way we train future researchers and practitioners?

AI is forcing us to rethink what it means to be trained as a researcher or practitioner. The question is not simply whether students should memorize less and synthesize more. The deeper question is: what forms of understanding remain essential when some layers of work can be automated or delegated?

We have seen this before in other technological revolutions. When programming moved from assembly to higher-level languages, we did not expect everyone to reason constantly at the machine-code level. But we also did not want engineers to lose all understanding of what happens underneath. AI creates a similar challenge. It can help us navigate literature, generate code, test ideas, and move faster across abstractions. But speed is not the same as understanding.

So the critical skill is not only knowing how to use AI. It is knowing when to use it, what to delegate, and what must still be practiced directly. In education, this means students need to struggle with the fundamentals before outsourcing them. They need to write, code, read papers, formulate hypotheses, and make mistakes themselves before AI becomes a co-pilot. Otherwise, the tool risks becoming a shortcut around learning rather than a support for deeper thinking.

For future researchers and practitioners, the highest-order skill may not be execution, but ideation: the capacity to frame meaningful problems, ask the right questions, and understand which challenges are worth solving. AI can accelerate many layers of work, but humans remain closest to the complexity of the real world: the clinical constraints, ethical tensions, institutional realities, and lived experiences that make a question important. This is why metacognition becomes central: we need to understand not only the answers AI gives us, but the assumptions behind our questions and the consequences of acting on them. AI can help us move up the ladder of abstraction, but education must ensure we still understand what is being abstracted away.

## Data at the Patient's Bedside

Your research at Columbia focuses on Biomedical Informatics. What is the greatest current challenge in moving data science out of the lab and to the patient's bedside?

I think the challenge is not only to move data science from the lab to the bedside, but to move it there responsibly.

Technically, interoperability remains one of the central problems of biomedical informatics. This has been true since the beginning of the field. We have made real progress, but healthcare data is still fragmented across electronic health records, imaging systems, genomic platforms, devices, institutions, and workflows. Building useful AI requires more than collecting data; it requires making these systems communicate in a way that preserves clinical meaning.

I experienced this very concretely when working on data harmonization before training a foundation model. Even for one modality, I was surprised by the diversity of units, formats, outliers, missing values, and local conventions in the data. What may look like a simple preprocessing step from the outside is actually a major part of the scientific and clinical work: understanding what the data really means before asking a model to learn from it.

Regulatorily, the key issue is responsibility. As AI systems become more integrated into clinical decision-making, we need to rethink how responsibility is defined and shared. If a physician relies on an AI recommendation, who is accountable for the outcome? The clinician, the hospital, the developer, the institution that deployed the tool, or the system that validated it? These questions become increasingly important as physicians may reasonably come to rely on these tools more and more.

Socially, the challenge is trust. Explainable AI is often presented as the solution, and it is useful, but I think it is limited. An explanation is not automatically meaningful just because it is technically interpretable. Trust comes from knowing that a system has been evaluated in the right context, that it works for the right patients, that its limitations are understood, and that it fits into clinical reality.

So for me, the greatest challenge is building the full ecosystem around AI in healthcare: interoperable data systems, clear responsibility frameworks, and forms of trust that go beyond simply opening the black box.

Editor's note: Explainable AI (XAI) is a set of processes and methods that allows human users to comprehend and trust the results and output created by machine learning algorithms. In traditional AI (often called "Black Box" models), an algorithm might provide a highly accurate prediction—such as a medical diagnosis or a loan approval—but it cannot explain why it reached that conclusion. XAI aims to bridge this gap by making the internal logic of the model transparent.

## The Future: Augmented Expertise

In your view, how will AI change the roles of researchers and physicians in the coming years?

In the short term, I do not think we will see one general AI system running the hospital. What we will see, and are already beginning to see, is the integration of more specific intelligent tools: AI scribes, automated responses, decision-support systems, diagnostic assistance, and workflow optimization systems. A good example is Columbia DBMI's CONCERN Early Warning System, which uses patterns in nursing documentation to help detect patient deterioration earlier. What is interesting about this type of system is that it does not replace clinical expertise; it tries to make visible and actionable some of the subtle signals that nurses already perceive in their daily work.

The challenge is to design these tools so they are not only powerful, but scalable, efficient, and responsible. Healthcare already operates under major constraints: time, money, staffing, energy, infrastructure, and regulation. So the goal is not simply to add more technology, but to build systems that genuinely improve the way care and research are delivered.

For physicians, I think the future is not about becoming less human, but more human. As some repetitive or administrative tasks are automated, clinicians should have more space to focus on what remains deeply human: listening, empathy, contextual judgment, and understanding the limits of any tool in a complex clinical environment. A good physician of tomorrow will not blindly rely on AI; they will know how to use it, question it, and integrate it into patient care responsibly.

For researchers, the transformation is slightly different. AI will accelerate discovery by helping connect fields that used to be separated: biology, medicine, computer science, statistics, engineering, and ethics. The researcher of tomorrow will need to navigate across these boundaries, ask meaningful questions, and create synergies between disciplines.

So I do not see AI as replacing physicians or researchers. I see it as forcing us to rethink their roles; and hopefully to build a healthcare system that is more intelligent, more efficient, and also more human.

## The Golden Thread

We often see a gap between what technology can do and what professionals actually adopt. How do we ensure AI becomes a 'co-pilot' that elevates human judgment rather than a tool that encourages intellectual passivity?

First, we need the right environment: time to think, psychological safety, and a genuine desire to contribute. Productivity may be a valuable by-product, but it should not be the primary goal. The incentives around AI should reward judgment, responsibility, creativity, and meaningful contribution; not simply speed or output.

Second, we need to design AI systems that are reliable, understandable, and adapted to professional contexts. The goal is not an opaque "super-intelligence" that replaces human reasoning, but a partner with whom we can co-think. This also requires developing our own metacognition: knowing when to trust the system, when to question it, and when to rely on human experience. Humility must go both ways.

Third, knowledge remains necessary for judgment. AI should not make us passive because information is easily accessible. Instead, it should help us prioritize what knowledge is worth memorizing, synthesize complex material more effectively, and learn how to search, evaluate, and connect ideas with greater depth.

Tags: AI in Healthcare, Biomedical Informatics, Data Science, Columbia University, Machine Learning, Digital Health