Though it has been years since I practiced as an interventional cardiologist, one moment remains etched in my memory. A middle-aged mother of two came to me, describing the “worst headache” of her life. Her words spilled out, a mixture of symptoms and urgent fears for her children as she grappled with the fragility of her own life. As her anxiety deepened with every passing second, my mind raced to pinpoint the cause. Was this a cluster headache, a searing, stabbing pain on one side of the head, excruciating but not life-threatening?? Or was this something far more dire, a thunderclap headache signaling a stroke?
Around the same time as this incident, I collaborated with researchers at Emory University and the University of Michigan Medical School to develop and validate a computer-based statistical algorithm. This algorithm was designed to identify potential stroke cases from hospital admission records, transforming clinical data into actionable risk assessments. Translated into a simple scoring system, it incorporated variables such as mental status, dizziness, slurred speech, weakness, and visual disturbances. Our goal was to embody this algorithm by integrating it into a physical device.
Today, the convergence of artificial intelligence and cloud computing has made that vision possible. The incredible advances in compute make it now possible to develop handheld devices capable of real-time stroke risk assessments, offering lifesaving insights precisely when they are needed most.
The low adoption of AI in healthcare
AI is poised to revolutionize healthcare at every stage of the treatment journey, from screening and diagnosis to therapeutic interventions. At its core, AI thrives on vast amounts of data, learning from patient records, medical images, sensor outputs, and clinical notes to generate insights that have the potential to elevate the quality of care. Managing these enormous datasets requires immense computational power, seamless accessibility, and robust security, all of which are capabilities enabled by cloud computing.
Today, the healthcare industry accounts for over a third of digital data generated worldwide. And yet, cloud adoption in healthcare remains just half of the overall industry.
Why is this so??
A 2024 NIH study might provide a clue.
The study highlighted explainability and transparency as critical hurdles to AI adoption in healthcare. The opacity of AI algorithms can erode trust among clinicians, who may hesitate to rely on machine learning outputs due to the lack of confidence in the generated results. In addition, their hesitation to adopt these new technologies was influenced by elements such as fairness, explainability, and potential for bias—qualities essential for the successful integration of AI into healthcare systems.
These concerns are understandable. The inner workings of neural networks and machine learning models can remain hidden and enigmatic, even to their creators. While the reliability of outputs persists as an area for improvement in AI, this need not be dispositive of the adoption of cutting-edge machine learning and generative AI technologies, even in the healthcare space. With billions of people around the world lacking access to adequate healthcare, it is a moral imperative to deploy AI to at least help lower the barriers that exist and ensure more equitable access to life-saving interventions.
Healthcare leaders, policymakers, governments, and the public must embrace a truth long understood by scientists: certainty is a luxury, and the inherent unpredictability of the world is an opportunity for mathematical precision to shine. By skillfully characterizing and managing this complexity, we unlock transformative possibilities. However, to harness the full potential of any new technology, we must embrace the nuances inherent in nonbinary decision-making, recognizing that uncertainty is not a barrier but a challenge to be met with ingenuity.
The upsides to such a mindset are profound.
Creating a more equitable world
Consider the breakthroughs in computational linguistics and machine learning that have enabled the restructuring of mRNA sequences, leading to lifesaving vaccines like those developed for COVID-19. Transformer-based architectures like those powering ChatGPT have been adapted to optimize mRNA sequences by treating genetic code like a language. These models analyze vast datasets of biological sequences to identify patterns, redundancies, and structural inefficiencies in mRNA transcripts.
One example is the use of deep learning models to predict secondary RNA structures, which affect how the mRNA folds inside cells. Poorly folded mRNA can degrade quickly or fail to produce the desired proteins efficiently. By applying NLP techniques such as sequence embedding (analogous to word embeddings in language models), researchers optimized the ordering of nucleotide triplets to improve both stability and translation efficiency while avoiding unintended immune responses.
While the acute phase of COVID-19 might have passed, there is another silent pandemic that plays out in the world day after day—the pandemic of limited or no access to care for the 4.5 billion people worldwide who remain underserved. According to the World Health Organization, in the Americas, one in three people continues to face barriers that make accessing health services difficult. These barriers include high treatment costs, lack of insurance, transportation difficulties, lack of time to seek care, long waiting times, and, in some cases, poor care experiences.
Take cardiovascular disease as an example to illustrate how inequities in access to care are helping exacerbate preventable consequences. Cardiovascular diseases have been the leading cause of death for more than three decades. According to another World Health Organization study, in the United States, nearly 40 percent of heart failure cases are diagnosed too late, often at the point of crisis. Globally, the statistics are grimmer still, with millions of individuals far from hospitals equipped with echocardiograms, MRI machines, or trained specialists.
AI is transforming this reality. Technologies like GE HealthCare’s VScan Air devices, equipped with Caption AI, enable real-time guidance for acquiring diagnostic-quality echocardiographic images. These tools empower even less experienced clinicians to acquire diagnostic-quality images, offering new possibilities in underserved regions where access remains one of the primary barriers to care.
Beyond specific devices, the rise of foundation models marks a paradigm shift in healthcare. These pretrained, large-scale AI systems can process complex multimodal data—text, images, and audio—to analyze diagnostic images, integrate patient information, and identify medical conditions that can help develop personalized treatment plans. Foundation models, adaptable across medical domains and even anatomies, will eventually streamline workflows, reduce costs, and empower clinicians in resource-limited areas to deliver accurate, timely care.
Yet the transformative power of these technologies, like any tool, comes with inherent complexities and uncertainties. To fully leverage AI’s potential and equip physicians to deliver precision care, we must embrace nonbinary decision-making, recognizing that certainty is neither a guarantee nor a necessity.
Embracing uncertainty
Americans rely on the Food and Drug Administration (FDA) to ensure the safety and effectiveness of their food and medical products. However, the approval process is based on studies conducted on a limited number of patients, meaning rare adverse effects may only emerge after a product reaches the market. To manage this uncertainty, the FDA continuously monitors medical products through post-market surveillance systems like the FDA Adverse Event Reporting System (FAERS) for drugs and biologics and the Manufacturer and User Device Experience (MAUDE) for medical devices. When I served as the Chief Informatics Officer for the FDA, we also used techniques like change point analysis (CPA) to detect shifts in data trends for post-market surveillance. Through CPA, we were able to identify safety concerns earlier in post-market surveillance, saving countless lives in the process.
Today, similar techniques to help navigate a world of Bayesian uncertainty can mitigate risks in the deployment of AI technologies by defining intended use and employing visual grounding. Beginning with a specific application—such as a tool for semiautomated segmentation of anatomical structures in ultrasound images—and gradually expanding its scope enables control, risk management, and compliance. Visual grounding techniques further enhance trust by linking AI outputs to interpretable visual data, such as annotated medical images. Combining these methods with strategies like ontology-based reasoning, explainability, and human oversight can help facilitate the safe and effective deployment of generative AI in healthcare.
Embracing uncertainty is a time-honored tradition in medical innovation. As AI converges with cloud computing, the next decade holds the potential to reshape healthcare more profoundly than the past century. By embracing uncertainty and taking bold action today, we can address one of humanity’s greatest inequities: the lack of quality healthcare for billions. It is only by embracing uncertainty that we can purge modern society of one of our greatest moral failings: having both the means and the technology to care for the least privileged of us, but all the same, choosing not to act.
