AI in Healthcare: Promise, Peril and the Path Forward

As artificial intelligence revolutionizes industries worldwide, healthcare stands at a critical inflection point. The potential for AI to transform medical diagnosis, treatment planning, and patient care is immense — but so are the challenges of implementation, concerns about bias, and questions about safety. Few understand this complex landscape better than Gordon Gao, Professor at Johns Hopkins Carey Business School and co-director of the Center for Digital Health and Artificial Intelligence (CDHAI), who will be teaching in the CKGSB-Johns Hopkins program on AI-Driven Healthcare Innovation beginning May 5.

“I think at this moment, what’s really holding healthcare executives back is uncertainty,” says Gao. “They’re not sure whether AI is useful — that’s the value perspective. And they’re not sure how safe it is to use — that’s the safety perspective.”

The Value Question: Can AI Deliver?

Healthcare leaders face difficult decisions about which AI applications merit investment. While marketing materials from vendors promise revolutionary advances, reality is often more complex. Gao cautions against taking AI’s capabilities at face value, pointing out that many models excel in laboratory settings but struggle in real-world clinical environments.

The terminology itself can be misleading. “In the AI literature, ‘predict’ most often means ‘classify’ rather than forecasting what will happen in the future,” Gao explains. “This leads to a lot of misunderstanding when business leaders or healthcare organizations are pitched by vendors.”

Consider medical imaging, where AI has made remarkable strides. According to FDA approvals of AI and machine learning-enabled medical devices, radiology and cardiovascular imaging applications dominate the field. What makes these applications successful is their well-defined scope, standardized data formats, and clear performance metrics.

“All these image-based AI applications use images to do prediction or classification,” notes Gao. “The job is very well defined, and the information set is limited to the image itself. You have sufficient information, clean data, and consistent, standardized formats. You don’t see missing data in the image. It’s actually a very good task for AI.” Contrast this with less successful implementations, like chatbots designed to respond to patient inquiries. “This didn’t work well,” notes Gao, “because there’s too much risk in advising patients to ‘do this, do that’ using a chatbot. The questions patients ask are so diverse, going beyond what the bot was trained to handle.”

An example of successful AI implementation comes from Tissue Analytics (now part of Net Health), a Baltimore-based company that uses AI to measure wound healing. “Previously, clinicians would place a ruler on a wound to measure it, but the boundary is never clear-cut — it’s subjective,” Gao explains. “Now, we can take pictures with a smartphone from different angles, and AI constructs a 3D model that even measures wound depth, which was previously almost impossible to determine non-invasively.”

The CGM-LSM Model: A Breakthrough in Diabetes Management

One promising frontier in healthcare AI is diabetes management. In a recent paper, Gao and colleagues introduced CGM-LSM, a large sensor model for continuous glucose monitoring that dramatically improves prediction accuracy.

“While previous studies of artificial intelligence in diabetes management focus on long-term risk, research on near-future glucose prediction remains limited but important as it enables timely diabetes self-management,” the researchers note. The CGM-LSM model was pretrained on nearly 16 million glucose records from 592 diabetes patients to predict glucose levels up to two hours in advance.

The results were remarkable: “CGM-LSM achieved exceptional performance, with an rMSE of 29.81 mg/dL for type 1 diabetes patients and 23.49 mg/dL for type 2 diabetes patients in a two-hour prediction horizon,” the paper reports. When tested against a standard benchmark dataset, it achieved a one-hour prediction error of just 15.64 mg/dL—half the error rate of previous best models. This approach shows how foundation models, similar to those powering large language models, can be applied to sensor data to unlock new capabilities in healthcare.

The Safety Imperative: First, Do No Harm

Beyond questions of value, healthcare leaders must consider whether AI applications might inadvertently cause harm. While the FDA rigorously tests drugs for safety before evaluating efficacy, the evaluation of AI tools often prioritizes performance over safety. “I think here, we are too rushed to the value part and we ignore the safety part,” Gao cautions. “How do we ensure that AI is not giving misleading answers? How do we verify that? How do we put guardrails on AI use?”

The challenge extends beyond technical performance to unintended consequences. For example, AI tools that generate clinical notes for physicians might improve efficiency but could affect how new doctors learn the art of clinical documentation. “We need to learn a new way to work with AI,” notes Gao. “It’s like having a new assistant in your team. How do you change your routine and workflow to best integrate that?”

The Bias Challenge: Ensuring Equitable AI

Perhaps no issue is more critical in healthcare AI than bias and fairness. As Gao explains, today’s AI systems learn from massive datasets that may encode historical biases in healthcare delivery. “The large language models are basically projecting our human language and images into a high-dimensional space, ingesting all the knowledge we have,” Gao notes. “We have bias in human society, historical bias. Maybe the bias has been corrected now, but AI is learning everything from all the way back.”

These biases often operate implicitly rather than intentionally. In a paper co-authored by Gao, “Addressing Algorithmic Bias and the Perpetuation of Health Inequities: An AI Bias Aware Framework,” the researchers present a four-step analytical process for building and deploying fair AI algorithms in healthcare:

1. Problem scope: Ensuring the right problem is being addressed
2. Data collection: Addressing bias in training data
3. Model building: Carefully designing algorithms to avoid amplifying bias
4. Decision-making: Implementing AI recommendations in an equitable manner

The paper emphasizes that “to the degree that ML algorithms are trained on data that may reflect existing biases in diagnosis, treatment, and provision of services to marginalized populations, they pose the danger of automating and further exacerbating that bias through subsequent learning cycles.”

A striking example comes from a widely used algorithm for predicting which patients would benefit from additional care support. Researchers discovered the algorithm was using healthcare costs as a proxy for health needs — seemingly reasonable, but with unfair consequences.

“Using healthcare costs as a risk measure, the system would give high-risk patients more resources, therefore they’d utilize more resources,” explains Gao. “In the end, the rich get richer. Nobody did that intentionally, but it had been used for multiple years affecting millions of patients.”

Harnessing Sensor Data for Precision Healthcare

Gao’s work extends beyond identifying problems to creating innovative solutions. His research on large sensor models (LSMs) demonstrates how massive datasets from wearable devices can be leveraged for precision healthcare.

“The superior performance of CGM-LSM demonstrates the feasibility of introducing foundation models from texts, images, and videos to human sensor records,” Gao’s research team notes. “Such data is becoming increasingly massive and ubiquitous given the widespread adoption of wearable devices.”

This approach represents a significant shift in how we think about medical data. Rather than treating each patient’s data in isolation, LSMs learn patterns across thousands of patients, enabling more accurate predictions even for individuals with limited personal health history. The implications extend far beyond diabetes management:

“With these data, various LSMs could be developed to advance the understanding and management of other diseases, like uncovering new knowledge about body mechanisms from heart rate, blood pressure, body temperature, weight, respiratory rate, and oxygen saturation,” the researchers explain. This could revolutionize preventive care by identifying subtle warning signs long before symptoms appear.

In another research direction, Gao has investigated how large language models (LLMs) often absorb and perpetuate biases in healthcare language. His team found that stigmatizing language in clinical notes — such as using terms like “drug abuse” instead of “substance use disorder” — not only affects patient care but also impairs AI performance. “We found that this kind of language will hurt certain people even more than others, then it actually leads to worse AI performance,” Gao explains. “We need to remove stigmatizing language.”

The Way Forward: Leadership for the AI Era

How can healthcare leaders navigate these complex challenges? Gao emphasizes the importance of understanding both the business challenge and the data landscape. “Leaders need to have a very good understanding of what AI can and cannot do, and identify the right scenario or use case,” he notes. “They should start with baby steps rather than ambitious, ambiguous projects. You need to understand your business challenge, but also the data — where it is, whether it’s labeled, how clean and complete it is.”

Organizational adaptation is equally crucial. When Kaiser Permanente nurses went on strike partly due to concerns about AI, it highlighted the fear many healthcare workers feel about technology displacing jobs. “How do you manage that and make everyone become an AI explorer and advocate to improve themselves?” Gao asks. “That’s a great challenge for leaders these days.”

Looking ahead, Gao sees transformative potential in how AI reshapes patient-provider relationships. First, by handling administrative tasks like clinical documentation, AI can free physicians to focus on patients rather than screens. “Doctors become practicing medicine and interacting with patients again, rather than becoming data input specialists,” he says.

Second, AI can help clinicians better understand patients by analyzing patterns in continuous monitoring data. “With sensors like continuous glucose monitors providing measurements every five minutes, AI can summarize patterns and help doctors understand, ‘This patient is not doing well with diet, so next time they come in, I should focus on that.'”

Finally, AI could shift incentives toward prevention rather than treatment. “If we can quantify a doctor’s value in maintaining health rather than just performing procedures, that could motivate clinicians to focus on long-term health and prevention,” Gao suggests.

Democratizing AI Skills

As healthcare systems worldwide grapple with these opportunities and challenges, Gao believes education will be crucial. “Nowadays is a great time to learn AI,” he says. “AI skills are no longer limited to computer science PhDs—they’re available to everyone.” With this democratization comes the opportunity for healthcare professionals at all levels to understand AI’s capabilities and limitations, helping to ensure that this powerful technology serves patients safely and equitably.

“I hope that we all have an open mindset towards AI,” Gao says. “That definitely goes a long way in our understanding of AI, in our use of AI, and in our attitude toward AI. It will be a subtle change, but it’s a different mindset.” For healthcare leaders navigating the AI revolution, that mindset shift — embracing the potential while acknowledging the risks — may be the most important step toward realizing AI’s promise to transform care delivery, improve outcomes, and make quality healthcare more accessible to all.