Medicine has long followed the same basic sequence: a patient becomes unwell, seeks help, receives a diagnosis and begins treatment. Artificial intelligence is beginning to disrupt that sequence at its earliest stage. The most consequential developments emerging from healthcare research are not about improving what happens after a patient falls ill. They are about identifying who will fall ill, and when.

The evidence arriving from clinical research is striking in its breadth. Researchers at Mayo Clinic have demonstrated that AI can detect signs of pancreatic cancer on routine CT scans up to three years before a conventional diagnosis would be reached. For a disease where prognosis is heavily tied to how early it is caught, that gap is clinically significant. Separately, AI models analysing patient data have shown the ability to predict cardiac arrest before visible deterioration occurs. A study combining proteomics and AI found it possible to estimate stroke and heart failure risk up to 15 years in advance. AI systems trained on data from wearable devices have identified infection risk in children undergoing cancer treatment before any clinical symptoms were present. Retinal imaging, long used to assess eye health, is now yielding predictions about osteoporosis risk. Routine MRI scans are being found to contain signals associated with future diabetes and cardiovascular disease that clinicians were not previously in a position to detect.

What connects these findings is not the sophistication of the technology alone. It is the realisation that the data required for earlier prediction has, in many cases, already been collected. The signals were present in existing scans and records. The question was whether any system could read them accurately enough to act on.

Predictive AI is also reaching into conditions that have historically been harder to manage because of their complexity and variability. Platforms that combine wearable device data with patient-reported outcomes are being developed to anticipate Vaso-Occlusive Crises in patients with Sickle Cell Disease, with the aim of enabling intervention before a crisis requires hospitalisation. In oncology, AI tools are being applied within bispecific antibody treatment pathways to identify patients at elevated risk of Cytokine Release Syndrome, a serious complication of some cancer therapies.

AI's role in treatment decisions is also expanding, though this draws less attention than its predictive applications. Researchers have used AI to identify multiple biologically distinct forms of Parkinson's disease, a finding that could support more targeted clinical trials and treatment approaches. Valar Labs received FDA Breakthrough Device Designation for an AI platform designed to predict outcomes and guide treatment selection in bladder cancer. Tools are being used to identify women at increased risk of postpartum depression during pregnancy. Pathology companies including Leica Biosystems, Indica Labs and Lunit have announced a collaboration on AI-powered tools to improve biomarker assessment in cancer care.

Generative AI is making separate inroads into clinical workflow. A study published in JAMA Network Open, involving more than 1,500 clinicians, found that AI documentation tools produced measurable reductions in administrative burden. A meta-analysis covering more than 28,000 participants found that conversational AI tools produced small to moderate improvements in mental health outcomes. In the United Kingdom, surveys suggest approximately one in seven people have already used AI tools as a substitute for contacting their GP, a figure that points to how quickly patient behaviour is shifting regardless of what health systems formally endorse.

Governance has not kept pace with adoption. Healthcare leaders continue to debate the appropriate conditions for deploying AI across health systems, and the Coalition for Health AI has released new implementation guidance in response. Polling from KFF found that roughly a third of American adults now use AI chatbots for health information, yet clinicians remain the most trusted source of medical advice. The two facts are not incompatible, but the gap between them is where most of the current regulatory and institutional debate is focused.

The direction of travel in healthcare AI is becoming clearer. Organisations that can accurately predict clinical deterioration before it occurs will be better positioned to intervene earlier, at lower cost and with better outcomes. The technology's most durable contribution may not arrive at the point of treatment. It may arrive years before a patient ever becomes one.