The Human Variable: What AI Still Misses in Medicine

Last month, I watched a radiologist friend review chest X-rays on her phone during lunch. She swiped through cases with the fluid efficiency of someone scrolling social media, pausing only when an AI flag appeared in the corner of the screen. "Pneumonia probability: 87%," it read. She zoomed in, nodded, moved on. The whole interaction took twelve seconds. Later, she told me the patient had refused treatment because he believed his illness was a spiritual matter requiring prayer, not antibiotics. The algorithm had been right about the lungs. It knew nothing about the man.

This gap between diagnostic precision and human understanding is what haunted my recent conversation with Dr. Arvin Faundo, a hematopathologist and stem cell therapy expert whose work sits at the convergence of cutting-edge medicine and deeply human complexity. We tend to celebrate AI's pattern-recognition prowess in healthcare—the algorithms that spot tumors earlier than any human eye, that predict sepsis before vital signs crash. But Dr. Faundo reminded me of something I'd been ignoring in my own enthusiasm for medical AI: the datasets training these systems have a cultural address, and most patients don't live there.

The evidence is stark when you examine it. Western medical databases—the foundation for most healthcare AI—are built predominantly from North American and European patient populations. When these algorithms deploy in Southeast Asia, they carry embedded assumptions about disease presentation, body mass indices, genetic markers, and treatment responses that simply don't translate. Dr. Faundo described cases where diagnostic thresholds calibrated for Western patients flagged false positives in Filipino populations, or missed patterns of illness that manifest differently across genetic backgrounds. The AI wasn't wrong, exactly. It was answering questions nobody in that examination room was asking.

This matters because medicine isn't just applied biology—it's applied biology filtered through culture, belief, economics, and trust. I've spent the past year interviewing leaders across the AI landscape, and healthcare consistently emerges as the sector where stakes are highest and context is most maddeningly specific. An algorithm can learn that certain imaging patterns correlate with malignancy. It cannot learn why a patient might view Western medicine with suspicion born from colonial medical experiments, or why family consensus might trump individual patient autonomy in collectivist cultures. These aren't edge cases. They're the texture of real medical practice, especially outside the Western clinical settings where most AI training data originates.

What struck me most in our conversation was Dr. Faundo's insistence that patient beliefs belong in the same analytical frame as lab values. Not as obstacles to overcome, but as legitimate variables in treatment efficacy. A technically perfect treatment plan that ignores a patient's spiritual framework or cultural context isn't perfect—it's likely to fail. This is where AI's current architecture shows its limits. Machine learning excels at finding patterns in the data it sees. It cannot account for the data that was never collected: the grandmother's insistence on traditional remedies, the economic reality that makes adherence impossible, the historical trauma that makes clinical settings feel unsafe.

I'm not arguing for abandoning AI in medicine. The potential is too significant, the early successes too promising. What I am arguing for is a fundamental expansion of what we consider "medical data." If an algorithm is going to recommend treatment protocols, it needs training sets that include cultural context, economic constraints, and belief systems as formal variables—not background noise to be filtered out. This requires something the tech industry finds uncomfortable: admitting that Western clinical data isn't universal, that "bias" isn't just a technical problem to be engineered away, but a structural feature of how medical knowledge has been created and validated.

Dr. Faundo's work in stem cell therapy offers a glimpse of what culturally-informed medical AI might look like. When treatment protocols must be adapted for patient populations with different genetic backgrounds, different environmental exposures, and different healthcare expectations, you're forced to build flexibility into the system from the start. You cannot simply port over a Western protocol and expect equivalent outcomes. The same principle applies to AI: you cannot port over Western-trained algorithms and expect them to serve diverse populations equitably. The medicine must be adapted. So must the machine learning.

For leaders navigating this space—whether you're developing healthcare AI, setting hospital policy, or funding medical innovation—the implications are clear. First, diversify your training data geographically and culturally, not as an afterthought but as a core requirement for accuracy. Second, build clinical decision support systems that flag cultural and contextual variables the same way they flag abnormal lab values. Third, invest in interfaces that help clinicians understand not just what the AI recommends, but which assumptions undergird that recommendation and whether those assumptions apply to the patient in front of them. The goal isn't to make AI more "sensitive." It's to make it more accurate for more people.

What keeps me up at night isn't that AI will replace doctors. It's that we'll deploy AI systems that appear to work brilliantly in aggregate while systematically failing individual patients whose lives don't match the training data. My radiologist friend's algorithm was 87% confident about pneumonia. It was 0% aware that the patient it was assessing had reasons for his choices that no chest X-ray could reveal. The future of medical AI isn't just about better algorithms. It's about remembering that every percentage point represents a person, and people come with context that no dataset has fully captured.

We can build AI that augments medical judgment without flattening human complexity—but only if we accept that the patient's story is data too.# The Human Variable: What AI Still Misses in Medicine

Last month, I watched a radiologist friend review chest X-rays on her phone during lunch. She swiped through cases with the fluid efficiency of someone scrolling social media, pausing only when an AI flag appeared in the corner of the screen. "Pneumonia probability: 87%," it read. She zoomed in, nodded, moved on. The whole interaction took twelve seconds. Later, she told me the patient had refused treatment because he believed his illness was a spiritual matter requiring prayer, not antibiotics. The algorithm had been right about the lungs. It knew nothing about the man.

This gap between diagnostic precision and human understanding is what haunted my recent conversation with Dr. Arvin Faundo, a hematopathologist and stem cell therapy expert whose work sits at the convergence of cutting-edge medicine and deeply human complexity. We tend to celebrate AI's pattern-recognition prowess in healthcare—the algorithms that spot tumors earlier than any human eye, that predict sepsis before vital signs crash. But Dr. Faundo reminded me of something I'd been ignoring in my own enthusiasm for medical AI: the datasets training these systems have a cultural address, and most patients don't live there.

The evidence is stark when you examine it. Western medical databases—the foundation for most healthcare AI—are built predominantly from North American and European patient populations. When these algorithms deploy in Southeast Asia, they carry embedded assumptions about disease presentation, body mass indices, genetic markers, and treatment responses that simply don't translate. Dr. Faundo described cases where diagnostic thresholds calibrated for Western patients flagged false positives in Filipino populations, or missed patterns of illness that manifest differently across genetic backgrounds. The AI wasn't wrong, exactly. It was answering questions nobody in that examination room was asking.

This matters because medicine isn't just applied biology—it's applied biology filtered through culture, belief, economics, and trust. I've spent the past year interviewing leaders across the AI landscape, and healthcare consistently emerges as the sector where stakes are highest and context is most maddeningly specific. An algorithm can learn that certain imaging patterns correlate with malignancy. It cannot learn why a patient might view Western medicine with suspicion born from colonial medical experiments, or why family consensus might trump individual patient autonomy in collectivist cultures. These aren't edge cases. They're the texture of real medical practice, especially outside the Western clinical settings where most AI training data originates.

What struck me most in our conversation was Dr. Faundo's insistence that patient beliefs belong in the same analytical frame as lab values. Not as obstacles to overcome, but as legitimate variables in treatment efficacy. A technically perfect treatment plan that ignores a patient's spiritual framework or cultural context isn't perfect—it's likely to fail. This is where AI's current architecture shows its limits. Machine learning excels at finding patterns in the data it sees. It cannot account for the data that was never collected: the grandmother's insistence on traditional remedies, the economic reality that makes adherence impossible, the historical trauma that makes clinical settings feel unsafe.

I'm not arguing for abandoning AI in medicine. The potential is too significant, the early successes too promising. What I am arguing for is a fundamental expansion of what we consider "medical data." If an algorithm is going to recommend treatment protocols, it needs training sets that include cultural context, economic constraints, and belief systems as formal variables—not background noise to be filtered out. This requires something the tech industry finds uncomfortable: admitting that Western clinical data isn't universal, that "bias" isn't just a technical problem to be engineered away, but a structural feature of how medical knowledge has been created and validated.

Dr. Faundo's work in stem cell therapy offers a glimpse of what culturally-informed medical AI might look like. When treatment protocols must be adapted for patient populations with different genetic backgrounds, different environmental exposures, and different healthcare expectations, you're forced to build flexibility into the system from the start. You cannot simply port over a Western protocol and expect equivalent outcomes. The same principle applies to AI: you cannot port over Western-trained algorithms and expect them to serve diverse populations equitably. The medicine must be adapted. So must the machine learning.

For leaders navigating this space—whether you're developing healthcare AI, setting hospital policy, or funding medical innovation—the implications are clear. First, diversify your training data geographically and culturally, not as an afterthought but as a core requirement for accuracy. Second, build clinical decision support systems that flag cultural and contextual variables the same way they flag abnormal lab values. Third, invest in interfaces that help clinicians understand not just what the AI recommends, but which assumptions undergird that recommendation and whether those assumptions apply to the patient in front of them. The goal isn't to make AI more "sensitive." It's to make it more accurate for more people.

What keeps me up at night isn't that AI will replace doctors. It's that we'll deploy AI systems that appear to work brilliantly in aggregate while systematically failing individual patients whose lives don't match the training data. My radiologist friend's algorithm was 87% confident about pneumonia. It was 0% aware that the patient it was assessing had reasons for his choices that no chest X-ray could reveal. The future of medical AI isn't just about better algorithms. It's about remembering that every percentage point represents a person, and people come with context that no dataset has fully captured.

We can build AI that augments medical judgment without flattening human complexity—but only if we accept that the patient's story is data too.