Flipping the Script on 'Deep Fake' Technology
We hear a lot about the dark side of artificial intelligence, especially "deep fake" technology. It’s often in the news for its potential to mislead and deceive. But as a researcher and clinician, I found myself wondering: what if we could flip the script? What if we could harness the very same technology for social good?
That question was on my mind when I confronted a major challenge in my field: the timely detection of cognitive impairment and dementia. As a clinician, I see the profound impact that a diagnosis can have, opening doors to treatment, support, and crucial planning for patients and their families. Yet, an invisible health crisis is unfolding in our aging world. Up to 80%, and in some reports 90%, of individuals with cognitive impairment never receive a formal diagnosis. They are, in effect, missing from our healthcare systems.
A Clinical Dilemma: To Screen or Not to Screen?
The problem is, the medical literature has been confusing on how to find them. On one hand, influential bodies like the US Preventive Services Task Force have advised against universal cognitive screening for all older adults, citing insufficient evidence. On the other hand, clinical intuition and a growing body of evidence tell us we must do something. The idea of "case-finding"—or targeted screening—in high-risk groups makes perfect sense to clinicians on the ground. But this raises another question that has been left unanswered: who exactly is "high-risk"? The term has been frustratingly uncertain.
This apparent conflict left policymakers and healthcare providers in a bind. How do you find the undiagnosed without screening everyone, which is costly and resource-intensive? And how do you decide which "high-risk" group to focus on?
Building a 'Flight Simulator' for Public Health
That’s when the two threads of my thinking came together. What if we could use the technology behind deep fakes—a deep learning model called a Generative Adversarial Network (GAN)—to create a 'digital twin' of an entire population? To explain our approach and what we found, we created this short animated video that summarizes our entire journey.
As the video shows, our journey began by building this digital twin. It wasn't as simple as pressing a button. A huge part of our 'behind the scenes' work involved a painstaking validation process to ensure our virtual population was a high-fidelity replica of the real one. We had to be sure that the complex relationships between age, health conditions, education, and cognitive function were all accurately preserved. Think of it as building a flight simulator, but for public health policy. We built a virtual replica of Singapore's older adult population—all 753,905 of them—to safely "test drive" dozens of different screening strategies without any real-world cost or risk.
Finding the Sweet Spot: A Clear Answer Emerges
Once our digital twin was ready, we started running the simulations. What if we only screened people over 85? Or those with known hypertension? Or what if we focused on individuals who were already worried about their memory?
The results were illuminating. As our infographic shows, without active case-finding, the vast majority of people with cognitive impairment remain in the dark. After analyzing the results of dozens of strategies, one hit the sweet spot. We found that the most effective way to balance diagnostic reach and cost was to screen adults aged 75 and over who also expressed worries about their cognitive decline.
This single strategy successfully reduced the undiagnosed rate to below 50% at a remarkably low cost of just SGD$25 (about USD$19) for every new person we identified. It’s a simple, two-step approach that can be easily implemented in primary care: first ask about age and worries, then proceed with a cognitive test if needed.
A Practical Roadmap for a Global Challenge
But we knew that healthcare doesn't operate in a one-size-fits-all world. A strategy that’s feasible in a high-resource system might be impossible in a low-resource one. Our simulations also revealed that the "best" strategy changes depending on a health system's budget, or its "willingness-to-pay" for a new diagnosis.
This led us to develop a practical, tiered framework that policymakers anywhere can adapt.
In settings with very limited resources, it’s most efficient to start with highly targeted groups, like those aged 85 and over. As more resources become available, programs can expand to our "optimal balance" strategy, and eventually to even broader groups like those with known hypertension. This provides a flexible, evidence-based roadmap for tackling undiagnosed dementia globally.
The Road Ahead
This study was about answering a pragmatic, operational question: how can we efficiently find the missing millions? By using AI for good, we've provided a data-driven answer. The next step, of course, is to evaluate the long-term benefits and harms for the individuals identified through these strategies. But for now, we hope our work provides a clear, adaptable blueprint for health systems around the world to begin closing the diagnostic gap for one of the most significant health challenges of our time.
Read the full paper at Nature Health.