In 2022, Google DeepMind's AlphaFold system was awarded the Nobel Prize in Chemistry — the most prestigious recognition in science — for solving a problem that had stumped biology for half a century. The problem was protein folding: predicting the three-dimensional structure of a protein from its amino acid sequence. Before AlphaFold, mapping the structure of a single protein could take a team of researchers three to four years and hundreds of thousands of dollars. There are more than 200 million known proteins. At that pace, it would have taken humanity hundreds of thousands of years to map them all.
AlphaFold did it in months. The AlphaFold database now contains predicted structures for over 200 million proteins, freely available to any researcher on Earth.
The Nobel committee does not award its prizes for incremental progress. The recognition signals something the scientific community has been observing for several years: artificial intelligence has become a genuine force in medicine — not a supplement to existing methods, but a transformation of what those methods can do.
The Numbers Behind the Revolution
The scale of investment and deployment makes clear that this is not a pilot phase.
McKinsey & Company estimates that AI could increase US healthcare productivity by between 1.8% and 3.2% annually — equivalent to $150 to $260 billion in annual savings across the health system. Healthcare AI startups attracted over $7 billion in venture capital in 2024 alone. The global AI in medical imaging market, valued at $1.67 billion in 2025, is projected to reach $26.23 billion by 2034.
As of mid-2024, the US Food and Drug Administration had cleared approximately 950 AI and machine learning-enabled medical devices — nearly double the figure from two years earlier. Roughly 100 new AI medical device approvals are now issued each year, spanning radiology, cardiology, dermatology, and pathology. In December 2024, the FDA issued new finalised guidance specifically designed to streamline the approval of AI systems that continue to learn from new data — a significant regulatory update for a technology that was straining the boundaries of a framework designed for static devices.
These are not pilot programmes or research experiments. They are deployed clinical tools, actively affecting how patients are diagnosed and treated.
Reading the Invisible: AI in Medical Imaging
The earliest and most mature application of AI in clinical medicine is in medical imaging — and the results have been striking enough to shift the research consensus on what machines can do.
In 2020, a study published in Nature demonstrated that a deep learning system developed by Google Health could detect breast cancer in mammograms with greater accuracy than six radiologists working independently, while simultaneously reducing false positives and false negatives. The performance gains were consistent across two datasets — one from the UK and one from the US — suggesting the result was not a statistical artefact of one particular population or imaging protocol.
This pattern — AI performing at or above specialist level in specific image recognition tasks — has since been replicated across multiple imaging modalities. In dermatology, algorithms trained on dermoscopic images can classify skin lesions with accuracy comparable to board-certified dermatologists. In ophthalmology, AI systems can detect diabetic retinopathy and age-related macular degeneration from retinal scans with sensitivity that rivals specialist review. In radiology, AI tools can identify pulmonary nodules, intracranial haemorrhages, and spinal abnormalities in CT and MRI scans, and automatically prioritise urgent findings for human review.
The practical significance extends beyond accuracy alone. Radiologists in well-resourced hospitals review hundreds of scans per day; those in under-resourced hospitals or rural regions may be doing so with inadequate support, fatigue, and time pressure. An AI tool that provides a reliable second opinion — flagging findings for human verification rather than replacing human judgment — can meaningfully improve diagnostic reliability without requiring specialist deployment in every location. Indian health technology company Qure.ai has developed AI tools for chest X-ray and CT analysis specifically targeting high-volume, resource-constrained healthcare environments in low- and middle-income countries, bringing imaging AI to clinical settings where specialist radiologists are unavailable.
Predicting Disease Before It Happens
Beyond recognising existing pathology in imaging, AI systems are increasingly being used to predict acute medical events before they occur — giving clinicians a window for intervention that traditional monitoring cannot provide.
Sepsis — a life-threatening systemic response to infection — kills approximately 11 million people globally each year according to the World Health Organization. Its onset can be rapid and its early signs subtle, and the difference between a good outcome and a fatal one is frequently the speed of antibiotic administration. Machine learning models trained on electronic health record data — integrating vital signs, laboratory values, medication records, and clinical notes — are now able to identify patients at elevated sepsis risk hours before clinical deterioration becomes obvious to a human observer.
The University of Michigan Health System deployed an AI early-warning system that analyses over 100 patient variables in real time and alerts clinical teams when sepsis risk rises above a defined threshold. Trials demonstrated measurable reductions in mortality and shorter ICU stays in patients for whom the system triggered earlier intervention.
In cardiology, the results have been equally striking. A study published in the New England Journal of Medicine found that an AI system could identify patients with low ejection fraction — a sign of weakened heart muscle associated with heart failure — from a standard 12-lead ECG alone. Previously, identifying low ejection fraction required an echocardiogram. The AI finding means that a test available in virtually every clinical setting in the world can now screen for a condition that previously required specialised cardiac imaging. Separately, algorithms analysing ECGs can detect previously undiagnosed atrial fibrillation in patients who show no outward symptoms — identifying a major stroke risk factor before the stroke occurs.
Drug Discovery at a Fundamentally Different Speed
The most consequential long-term application of AI in medicine may be in pharmaceutical research — the process of discovering and developing new drugs, which has historically been one of the most expensive and slow processes in science.
Traditional drug discovery costs an average of $2.6 billion and over a decade to bring a single drug to market, according to PhRMA. The failure rate is extraordinary: only about 10% of drug candidates that enter Phase I clinical trials ultimately receive regulatory approval. Most candidates fail because of inadequate efficacy, unforeseen toxicity, or both — problems that are often predictable from molecular properties that were difficult to assess with pre-AI methods.
AlphaFold changed the foundational layer of this process. AlphaFold 3, released in May 2024, extended predictions beyond protein structures to protein-DNA and protein-molecule interactions — the kind of data that underpins drug binding research. The European Bioinformatics Institute makes the AlphaFold database freely available to researchers worldwide, providing structural data that previously would have required years of experimental work for every protein studied.
Building on this foundation, AI-driven companies are now doing what would have been considered science fiction a decade ago: designing entirely novel drug molecules. Insilico Medicine used generative AI to design a novel drug candidate for pulmonary fibrosis — a condition with limited treatment options — and advanced it from initial design to Phase II clinical trials in under four years, roughly half the conventional timeline. The company has described its approach as using AI to compress what would previously have been years of laboratory work into weeks of computational exploration.
Recursion Pharmaceuticals has built a drug discovery platform that generates millions of images of cells under thousands of experimental conditions per week, using machine learning to identify biological patterns associated with specific diseases and potential drug candidates. The company has partnered with major pharmaceutical manufacturers to apply this approach to their existing drug libraries and to novel target discovery.
The broader pharmaceutical industry is integrating AI at speed. Healthcare AI startups attracted over $7 billion in venture capital in 2024, with drug discovery receiving a significant share alongside clinical decision support and administrative automation.
Personalised Medicine: From Population Averages to Individual Treatment
One of medicine's longstanding limitations is what researchers call the "average patient" problem. Treatment protocols are designed for populations, not individuals. A chemotherapy regimen that produces remission in 60% of patients may cause serious harm in 40% — but without the ability to predict in advance which patient falls into which group, oncologists have had to make difficult decisions under uncertainty.
AI is beginning to change this. By integrating genomic data, medical history, lifestyle factors, biomarker measurements, and treatment response patterns from large patient cohorts, machine learning models can identify which patients are likely to respond well to specific treatments and which are likely to experience adverse effects.
The field of pharmacogenomics — studying how individual genetic variation affects drug response — has been dramatically accelerated by AI tools capable of identifying patterns in genomic data at a scale and speed that human researchers cannot approach. At institutions including Memorial Sloan Kettering Cancer Center, oncologists are using AI-driven genomic analysis to match patients with targeted therapies that exploit specific mutations in their tumours — moving away from broadly toxic chemotherapy toward precision interventions designed for the biology of a specific patient's cancer.
The long-term vision is a treatment approach in which every clinical decision — which drug, which dose, which timing — is informed by a model of the individual patient's biology rather than the characteristics of the average patient in a clinical trial conducted decades earlier. That vision is not yet clinical reality, but it is closer than it has ever been.
Barriers That Are Real and Not Going Away
The case for AI in healthcare is strong. So are the obstacles, and intellectual honesty requires taking both seriously.
Data privacy is a foundational challenge. AI systems require vast quantities of patient data to train effectively — and patient data is among the most sensitive personal information that exists. The EU's GDPR and the US HIPAA impose strict constraints on how medical data can be collected, shared, and used. The legal and ethical requirements for patient consent, data security, and re-identification prevention are not bureaucratic obstacles — they are protections for real people. Building AI systems that are both powerful and genuinely privacy-preserving is a hard technical and governance problem that has not been solved.
Algorithmic bias is a documented and serious concern. AI models trained predominantly on data from white Western populations have shown degraded performance on patients from underrepresented groups. A 2019 study published in Science found that a widely used commercial algorithm for identifying patients who needed additional healthcare resources was significantly less likely to flag Black patients than white patients with equivalent clinical need — not because the algorithm was designed to discriminate, but because it used healthcare cost as a proxy for need, and Black patients had historically received less care. The bias was structural and invisible until studied.
Ensuring that AI systems perform equitably across diverse populations is not a technical footnote. It is an ethical requirement for deployment — and in many cases, the data required to build equitably performing systems requires deliberate collection from underrepresented populations, which requires resources and time.
Regulatory lag is another genuine concern. The FDA's traditional approach to medical device approval — approving a specific algorithm at a point in time — is poorly suited to AI systems that continue to learn from new data in deployment. An algorithm approved in 2022 may be a substantially different system by 2025. The December 2024 FDA guidance on AI/ML medical devices represents a significant step toward addressing this, but implementation and enforcement are still developing.
Clinical adoption is slower than the technology. Clinicians who did not train with AI tools, who do not fully understand how they work, and who are appropriately sceptical of systems they cannot interpret need education, evidence, and time. Deploying AI tools without adequate clinical training is a genuine patient safety risk — an AI system that generates incorrect outputs can cause harm if clinicians accept its recommendations uncritically.
The Collaborative Future
The most evidence-based vision of AI in healthcare is not one in which algorithms replace physicians. It is one in which humans and AI systems work together in ways that exploit the specific strengths of each.
AI excels at processing large volumes of data without fatigue, identifying patterns across thousands of cases simultaneously, and maintaining consistent attention to specified criteria. Human clinicians excel at integrating contextual information that AI systems cannot access, communicating with and responding to patients as people, exercising ethical judgment in complex and ambiguous situations, and taking responsibility for decisions in ways that matter for accountability and trust.
A radiologist augmented by AI can review more scans, with greater accuracy, with less fatigue-induced error. A general practitioner with access to an AI-powered differential diagnosis tool can catch rare conditions that would otherwise go unrecognised for months. An oncologist using AI-driven genomic analysis can offer patients a precision of treatment matching that was not possible with any prior tool.
The technology is advancing faster than most health systems can absorb it. Training clinicians to understand, use, and critically evaluate AI outputs is as important as building the algorithms themselves. Investment in AI literacy across the medical profession is not optional. It is the condition under which the promise of AI in healthcare can be realised without the risks becoming its dominant story.
What do you think is the most important application of AI in healthcare — or the risk that concerns you most? Share your perspective in the comments below.