Through massive strands of data, Dr. Yulin Hswen trains AI to listen for the answers to very human health concerns: What do these symptoms mean? Where will I receive good care? Am I at risk for a chronic disease?
Hswen develops AI methods that convert online human expressions about emotional and physical needs — such as posts in social media forums — into population-scale evidence. By integrating digital communication, clinical data and medical imaging, her research helps uncover hidden patterns in human health, disease and healthcare, providing new insights that can improve clinical care, public health, and disease prevention.
Hswen is one of five leading experts in the first faculty cohort at UMD’s Artificial Intelligence Interdisciplinary Institute at Maryland (AIM), with academic homes in the School of Public Health’s Department of Epidemiology and Biostatistics and at AIM in the College of Computer, Mathematical and Natural Sciences.
Hswen spoke with Fid Thompson in this abridged interview, sharing her views on how ethical use of AI could change health for the better.
What is “social listening” and why is it important?
Social listening uses artificial intelligence to transform millions of everyday digital interactions into meaningful public health insights. It analyzes how people express their health experiences through social media, online communities, images, videos and other digital content to better understand the environments that drive health and societal change.
Every day, people generate vast amounts of such unstructured digital data that reflect their health experiences, symptoms, healthcare encounters and perceptions. Individually these data are anecdotal, but collectively they provide an unprecedented view of population health. Artificial intelligence allows us to analyze these data at scale, producing actionable evidence that can improve healthcare and public health surveillance.
How does this work in your research?
My lab develops AI methods that transform millions of unstructured digital interactions into structured, analyzable data, enabling us to detect patterns in health, disease and healthcare that would otherwise remain hidden.
One example is a study I led using AI to analyze nearly two million patient-experience posts from more than 1.3 million Twitter (now X) users across the United States. Results revealed healthcare disparities experienced by LGBTQ patients that would have been historically difficult to capture through traditional surveys and clinical data.
In a more recent study using AI, we analyzed patient-generated discussions from a large online health community to identify the dominant concerns, treatment preferences and emotional experiences surrounding the only FDA-approved therapy for interstitial cystitis, a debilitating chronic bladder condition without a clear cause. Our study provided authentic symptoms and side-effects of treatments, adding crucial information to what isn’t always available in clinical trials.
For me, every individual story is important. Gathered together, these human experiences can help identify meaningful population-level signals, offering evidence that can guide clinical practice, health policy, and population health. In this way, I believe AI gives us an opportunity to make healthcare more responsive to the people.
Every individual story is important. Gathered together, these human experiences can help identify meaningful population-level signals.
Can you share an example of your current research?
Perimenopause and menopause affect half the world's population and yet remain understudied and often poorly documented in traditional health data. So we decided to use AI to analyze large-scale Reddit discussions to better understand people’s lived experiences of menopause.
As public figures such as Halle Berry and Penélope Cruz raise our cultural awareness of menopause, this study offers evidence from organic real-world experiences that I hope will help improve clinical care and guide future research.
In a dearth of research about menopause, it is exciting to apply a social listening approach to see what we can learn when we listen outside of the usual channels. Find out more about the study in JAMA Network Open here.
What are the challenges you see using AI in population health?
Like any transformative technology, AI presents tremendous opportunities and important challenges. The quality of an AI system ultimately depends on the quality and representativeness of the data used to develop it. If training data are incomplete, inaccurate or systematically unrepresentative of the populations the model will serve, those limitations can be reflected in the model's performance.
Many existing datasets do not fully capture the diversity and complexity of real-world patient populations. Menopause is a great example of this, where a normal aging process experienced by almost half the global population, has been disproportionately underresearched.
The result of omissions like this is that AI models may not perform equally well across different clinical settings or populations. Building reliable AI requires high-quality, on-the-ground, representative data and rigorous validation to ensure models are accurate, robust and generalizable.
As AI becomes increasingly integrated into healthcare, it is important to evaluate not only how models perform, but also how they are developed, validated and ultimately deployed. In 2025, I created AI-Y — short for AI, Why? — a framework that helps researchers and developers systematically consider key questions throughout the AI development process. A checklist encourages users to think about issues such as data quality, model validation, clinical applicability, performance across different populations and how models may behave when applied in new settings or geographic regions.
In general, I worry less about AI replacing humans than about humans gradually outsourcing their thinking to AI. As these systems become more capable and persuasive, there's a real risk of human cognitive complacency — accepting recommendations without sufficient scrutiny or critical evaluation. The goal should be an artificial intelligence that augments human expertise while encouraging, not replacing, independent judgment.
The goal should be an artificial intelligence that augments human expertise while encouraging, not replacing, independent judgment.
Can AI help prevent disease?
AI allows us to move beyond simply describing disease toward predicting it. By analyzing millions of medical images alongside other sources of health data, AI can uncover subtle biological patterns that are difficult or impossible to recognize through conventional approaches. These advances have the potential to detect disease earlier, guide preventive care and improve long-term health outcomes.
My research is increasingly focused on using multimodal AI to understand how diseases develop, long before symptoms appear. For example, we are combining retinal images collected during routine eye examinations with electronic health records, genetic information, brain imaging, blood biomarkers and other health data to identify early biological signatures of Alzheimer's disease and other neurodegenerative disorders. Predictive diagnosis can be greatly enhanced with AI.
Ultimately, my lab’s goal is to leverage multimodal data and AI methods to enable earlier disease detection, improve prevention and support more social, empathetic and connected care.