Scaling AI in Healthcare: Insights from Dr. Alvin Liu on Real-World Implementation and Governance

From Ophthalmology to AI Leadership

Dr. Liu’s foray into AI began in the late 2010s during his clinical training, sparked by groundbreaking studies—particularly one from Google—that demonstrated the ability of AI models to predict cardiovascular risk factors from retinal images. This superhuman diagnostic capability was a turning point for him. As a retina specialist immersed in an image-rich field, Dr. Liu recognized the untapped potential of deep learning to transform how clinicians interpret complex visual data.

At Johns Hopkins, Dr. Liu leads the Gills AI Center—the first endowed AI initiative at the Johns Hopkins School of Medicine—while also maintaining an active clinical practice. He contributes across four pillars: AI development, implementation, governance, and scientific innovation, giving him a panoramic view of the opportunities and challenges in healthcare AI.

Autonomous AI in Primary Care: A Case Study in Diabetic Retinopathy Screening

One of the most compelling examples Dr. Liu shared was the deployment of an FDA-approved autonomous AI system to detect diabetic retinopathy in primary care settings. This system was the first of its kind to be approved for autonomous clinical use, and Johns Hopkins began implementing it in 2020.

Traditionally, patients needed to see a separate specialist to complete an annual retinal screening—an extra step that often led to missed appointments and lower screening rates. The AI system allows primary care physicians to take retinal images in their office, with AI analyzing them in real time. Patients receive immediate results, and only those with positive screenings are referred to an ophthalmologist.

The outcomes have been striking. Johns Hopkins observed a marked improvement in screening adherence, especially among underserved populations such as African Americans and Medicaid recipients. These results, published in Nature Digital Medicine, underscore how AI can help close gaps in preventive care—if implemented thoughtfully.

Generative AI for Revenue Cycle: From Clinical to Operational AI

AI’s impact at Johns Hopkins isn’t limited to the clinic. Dr. Liu described a pilot project using generative AI for revenue cycle management, specifically prior authorization. This is a high-friction area in healthcare, involving extensive paperwork and delays in care.

By leveraging large language models (LLMs), Johns Hopkins automated prior authorization workflows, reducing the time required and handling unstructured data far more effectively than traditional robotic process automation (RPA) methods. These results illustrate how AI can unlock value beyond clinical domains by streamlining healthcare operations and improving provider efficiency.

Startups and the Reality of Healthcare AI

Drawing from his experience working with numerous startups, Dr. Liu offered candid advice to AI entrepreneurs: understand reimbursement from day one. “I think one of the common mistakes that startup companies make in the healthcare AI space is not considering or not understanding their reimbursement issue from day one,” Dr. Liu added. Many startups make the mistake of focusing on building a great product without planning for how it will be paid for—especially in a field as complex and regulated as healthcare. 

He emphasized that FDA approval alone isn’t enough. Startups must also determine whether existing CPT codes apply to their solution, and if not, navigate the lengthy and uncertain process of obtaining new ones. Beyond regulatory hurdles, they must build business models that reflect the real-world economics of health systems.

Startups often underestimate the cost of this journey—$3 to $5 million for FDA approval is typical—and many don’t budget appropriately. Dr. Liu’s message was clear: clinical AI solutions need sound financial strategies as much as innovative technology.

Creating Enterprise-Ready AI: The Johns Hopkins Governance Model

To manage the influx of AI tools and ensure responsible adoption, Johns Hopkins established a robust AI governance framework. Dr. Liu is part of an eight-member enterprise leadership team that evaluates all AI-related initiatives across the health system.

This governance model is built around seven core principles: fairness, transparency, accountability, ethical data use, safety, evidence-based effectiveness, and sustainability. Any AI vendor seeking to partner with Johns Hopkins must complete a standardized intake process, provide detailed documentation on their tool’s safety, ROI, and evidence base, and undergo a rigorous review process.

The system categorizes tools based on their use case—clinical, operational, or imaging—and advances each proposal through specialized review committees. This ensures that tools align with Johns Hopkins’ mission, technical infrastructure, and patient care goals before they are deployed at scale.

This governance model could serve as a blueprint for other integrated health systems navigating a crowded and often chaotic AI vendor landscape.

Looking Ahead: Omics, Risk Prediction, and Scaling Innovation

Dr. Liu also shared his excitement about the emerging field of AI-driven “omics,” particularly using retinal biomarkers to predict systemic health conditions such as cardiovascular disease, kidney damage, and dementia. AI-enabled retinal screening programs in community settings could identify at-risk individuals years before symptoms emerge.

However, he was quick to point out that identifying risk is only part of the equation. Health systems must also build the care pathways to ensure those flagged by AI are connected to the appropriate subspecialists and receive timely follow-up care. Without that, the potential of predictive AI will remain unrealized.

A Call for Collaboration: Startups, VCs, and Health Systems

In his closing remarks, Dr. Liu highlighted a growing but still insufficient level of collaboration between AI startups, venture capitalists, and integrated health systems. Startups drive innovation and speed—but they often lack the domain knowledge and infrastructure to scale safely. Health systems, on the other hand, deliver the majority of care but tend to move slowly due to regulatory and operational constraints.

Bridging this gap, he argued, is essential for sustainable AI deployment. Startups need to understand the realities of clinical practice and reimbursement. Health systems need to improve agility and decision-making. And investors need to align their expectations with the long, complex arc of healthcare innovation.

Dr. Liu hopes to see more structured partnerships where these groups work together to solve real problems, share risk, and scale proven solutions responsibly. He believes that such collaboration is essential for delivering long-term value—and ultimately, for improving health outcomes.

AI is Here to Stay

As Dr. Liu puts it, “The train has left the station.” AI is already reshaping healthcare, and the focus must now shift to responsible scaling, thoughtful implementation, and real-world results. “ I think the vast majority of people will agree that AI will change medicine and society as we know it,” he adds. 

Whether through autonomous diagnostic tools, generative AI for operational efficiency, or predictive omics models, the future of healthcare will be defined by our ability to integrate AI into the fabric of care—ethically, equitably, and effectively.

This episode is a powerful reminder of what it really takes to turn promising AI into real-world results. For health systems, startups, and investors, Dr. Liu’s insights highlight why successful innovation depends as much on execution as on technology.