An AI tool built from personal medical records flagged a persistent calf cramp as a sign of deep vein thrombosis, sending its creator to the emergency room where ultrasound scans revealed four blood clots in his left leg. The warning came not from a doctor but from a software assistant trained on the man’s own health data – and it may have saved his life.
The cramp that led to a diagnosis
Gleb Tsipursky, a future-of-work consultant and author, had been suffering for five days from what he assumed was a stubborn muscle spasm in his left calf. The area was tender, swollen and growing worse. He visited his chiropractor, who treated it as a muscle problem. But the pain continued to intensify.
Tsipursky, who has a PhD and runs the consultancy Disaster Avoidance Experts, had created a custom AI health tool for himself based on his expertise in training companies to adopt AI effectively. The system had access to his medical records, medications, lab work and visit notes. When he described his symptoms, the tool flagged deep vein thrombosis – DVT – and pointed him to the diagnostic step that mattered: an ultrasound scan.
Deep vein thrombosis is a blood clot that forms in a deep vein, usually in the leg. According to the US Centers for Disease Control and Prevention, symptoms can include pain, swelling, warmth and discolouration of the skin, particularly when one leg is affected. The condition is often under-diagnosed. The National Heart, Lung, and Blood Institute warns that a pulmonary embolism – caused when a clot breaks off and travels to the lungs – can be fatal, especially if the clot is large or multiple clots are present. Common signs of a pulmonary embolism include sudden shortness of breath, sharp chest pain that worsens with deep breathing, a rapid or irregular heartbeat, lightheadedness, coughing (sometimes with blood) and clammy or discoloured skin. In some cases symptoms are mild or atypical, leading to delays in diagnosis.
Tsipursky called his primary care office, which advised him to make an appointment or go to urgent care. But neither of those options could provide an ultrasound scan. Following that path risked losing days before being referred to hospital anyway. So, guided by his AI tool, he went to the emergency room despite knowing he would wait hours. The ultrasound found four clots in his left leg.
After the diagnosis, the danger became personal. Tsipursky learned that his wife’s grandfather had died from a pulmonary embolism, as had the mother of one of her close friends. What had felt like a nagging cramp suddenly looked like the edge of a cliff.
The promise and peril of AI in healthcare
Tsipursky is clear that this is not a case of machines replacing doctors. The emergency room physicians did the indispensable work: ordering imaging, interpreting results, deciding whether to admit him, consulting specialists and sending him home on blood thinners when it was safe. The AI did not cure him; it helped him ask the right question in time.
The science is catching up with these stories. A study published in the journal Science, led by researchers affiliated with Harvard Medical School and Beth Israel Deaconess Medical Center, tested a large language model on clinical reasoning tasks using real emergency department cases. Science News reported that the model was more likely than physicians to include the correct diagnosis among the possible answers. The researchers suggested that AI is ready for rigorous clinical trials.
Other work is focusing specifically on DVT. A machine-learning algorithm called AutoDVT, developed by researchers from Oxford University, Imperial College and the University of Sheffield in collaboration with the company ThinkSono, has demonstrated the ability to diagnose DVT as accurately as traditional ultrasound scans. The system can also guide non-specialists in acquiring the correct images, potentially reducing waiting lists and avoiding unnecessary prescriptions of anticoagulant drugs. However, not all attempts have succeeded: one study found that an AI tool designed to help non-radiology specialists diagnose proximal DVT failed to meet accuracy targets and requires further optimisation.
The use of AI for health advice is already widespread in the UK, and the trend raises serious concerns. A report by The Guardian found that one in seven people in the UK are using AI chatbots for medical guidance instead of seeing a GP. A study by King’s College London put the figure at 15% of the public, with convenience, curiosity and long NHS waiting lists cited as reasons. The same Guardian investigation revealed that Google’s AI Overviews had provided misleading or incorrect health advice – including telling pancreatic cancer patients to avoid high-fat foods, the opposite of recommended guidance – which experts described as “dangerous.” A significant proportion of those who used AI for health advice reported that it discouraged them from seeking professional medical opinions or led them to decide against consulting a doctor, prompting concerns about an “unregulated AI healthcare system alongside the NHS.”
The regulatory landscape is evolving. AI in UK healthcare is currently governed by medical device regulations, with AI classified as software-as-a-medical-device or AI-as-a-medical-device. The Medicines and Healthcare products Regulatory Agency (MHRA) launched the National Commission into the Regulation of AI in Healthcare in September 2025, bringing together experts to advise on a new framework. Recommendations are expected in 2026. A consultation by the MHRA revealed that most respondents favour substantial reform, with key areas of concern including safety, performance standards, data governance and privacy. The UK is adopting a principle-based approach where individual regulators decide the way forward, contrasting with the EU’s single comprehensive framework. Among the questions the MHRA is grappling with are whether current regulations are adequate, how to ensure rapid access to safe AI medical devices, and who is accountable when something goes wrong.
Tools are also emerging to help patients navigate the system. “Pocket Patient Advocate” uses AI to explain medical terms, prepare for appointments and organise health records. Another platform, advocate.ai, records appointments, generates transcriptions and provides personalised questions. AI can also streamline administrative tasks such as financial clearance and prior authorisations.
Tsipursky’s lesson is not that everyone should outsource their health to software. It is that patient advocacy now has a new tool. A safe AI assistant, trained on an individual’s medical data, can help people gather their records, ask whether something urgent is being missed and push for the right diagnostic step. Medicine has always depended on second opinions. The next one may come from software, and the urgent task is to make sure it is accurate, accountable and used to save lives.