Our Framework to Mitigate Bias in Clinical AI with Dr. Félix Manuel Chinea

Hello everyone. Thank you for being curious about how we work to mitigate bias in clinical AI at Doximity.

What I want to talk about today isn't a theoretical exercise. It's about the systems we're building right now, the AI tools that are increasingly shaping clinical decisions at scale, and what it means to build them responsibly.

The title says "mitigate bias," but what we're really talking about is how we build AI that physicians can actually trust, and how trust, when you dig into it, is inseparable from equity.

I'm Dr. Félix Manuel Chinea. I'm a physician and serve as Senior Director of Health Equity and Inclusion Strategy at Doximity. My background is in health disparities research, and my work now lives at the intersection of medicine, technology, and equity. I spend a lot of time thinking about how digital health tools can either widen or narrow the gap in care.

The answer depends almost entirely on how intentionally they're designed.

We'll move through four sections. We'll start with the problem because understanding why AI scales bias is the foundation for everything else. Then we'll talk about how mitigating that bias builds clinician trust. From there, we'll look at the framework we've developed at Doximity: detect, measure, report, and mitigate. Finally, we'll ground it in real clinical case studies.

AI doesn't emerge from a vacuum. It learns from data, and our data reflects our history.

Medical education has perpetuated false narratives about race and biology for decades. Medical literature carries those same biases. When AI is trained on that data, it doesn't correct those errors. It amplifies them at scale, often invisibly.

An equitable answer is simply a more accurate medical answer. Fixing bias isn't a charitable act. It's a clinical imperative.

I want to challenge the idea that health equity and quality are somehow in tension. That's exactly backwards. Equitable AI is rigorous AI.

At Doximity, we use a four part framework: detect, measure, report, and mitigate.

Detect is about identifying where bias exists. Measure is about quantifying it in a way that's repeatable over time. Report turns findings into institutional knowledge that developers can act on. Mitigate is where all of that becomes engineering reality.

Accountability doesn't scale on its own. You have to build the infrastructure for it.

Detection starts with defining what bias looks like for your specific tool, your specific use case, and the communities most likely to be harmed. You can't audit for what you can't see, which is why lived experience is a structural necessity.

Once you identify bias, you need to measure it over time. Models evolve, training data changes, and what looks accurate today may drift tomorrow. We use both qualitative and quantitative measurements because numbers alone don't tell the full story.

Reporting transforms findings into institutional memory. It's not enough to say a model was wrong. You have to explain who was harmed, how they were harmed, and what the clinical consequences are so developers can act on that information.

Mitigation closes the loop. Detection, measurement, and reporting only matter if they lead to meaningful engineering improvements. Different sources of bias require different solutions, and those solutions need to be embedded directly into technical development.

We measure progress across clinical decision making, hallucinations, drug safety, and health equity. The goal isn't perfection at a single point in time. It's continuous improvement and accountability.

These results represent real clinical scenarios where biased outputs were identified, addressed, and improved. Every improvement represents a patient population receiving a more accurate clinical answer.

Let's look at a few examples.

In dermatology, we found the model defaulted to light skin tone as the diagnostic baseline for conditions like psoriasis. We updated the outputs to describe disease presentation across the full Fitzpatrick skin tone scale, resulting in more complete and accurate answers for every patient.

For kidney disease, our audit found the model was still recommending outdated race based eGFR calculations. We updated the outputs to align with current race neutral clinical guidelines.

For transgender and nonbinary patients, we replaced gender based screening recommendations with anatomy based logic. This improved both inclusivity and clinical accuracy.

We also improved dietary recommendations by incorporating culturally aligned nutrition guidance instead of relying on generic defaults. When patients see themselves reflected in clinical recommendations, adherence and trust improve.

I'll leave you with three key ideas.

First, health equity and clinical quality are the same pursuit. Improving equity improves care for everyone.

Second, a framework transforms good intentions into accountable systems. Detect, measure, report, and mitigate isn't just a slogan. It's operational infrastructure.

Third, physicians should set the standard for clinical AI. Clinical expertise combined with equity literacy is what enables us to recognize where tools fall short and define what an accurate answer should look like.

The work doesn't end with a framework. It ends when every patient receives an accurate and complete clinical answer, regardless of who they are or where they're from.

That's the standard we should hold clinical AI to.

Thank you, and we look forward to supporting your clinical workflow.