Dr Abdulbasit Sadiq

The UK's cyber security agency now says it is "highly likely" that AI-powered malware will be used in a ransomware attack on Britain's critical infrastructure by 2028. Hospitals are part of that infrastructure. We already know what one of these attacks does to a hospital. In June 2024, ransomware hit Synnovis, the pathology provider for two large London trusts. Blood testing was badly disrupted, and by August more than 11,000 outpatient appointments and planned procedures had been postponed. That attack didn't need AI. The NCSC expects AI to make attacks like it bigger and more efficient. Hospitals rehearse for fires and power cuts. I'd like them to rehearse a week without computers too.

Two big heart trials are presented in Chicago this afternoon. We already know both missed their main goal, and today we see the detail. Before the hot takes arrive, these are the four questions I ask of any negative trial. 1. Did the drug hit its target? If it never lowered the thing it was aimed at, the trial didn't test the idea at all. 2. Were these the right patients? A drug can't show benefit in people who were never at much risk from what it treats. 3. Was the trial big and long enough? A narrow confidence interval around "no effect" rules out a useful benefit. A wide one leaves the question open. 4. What were patients already taking? If background treatment was very good, there may have been little risk left for a new drug to remove. If a trial gets through all four and still shows nothing, I take the result seriously, even when I hoped for the opposite.

The US added 165,000 jobs in October. Economists expected about 60,000, and the unemployment rate fell to 4.1%. One number in the same report went the other way. The average spell of unemployment rose to 27.1 weeks, the longest outside the pandemic since 2016. About 2 million Americans have now been looking for work for more than six months. On a ward round, a good average doesn't mean every patient is doing well. You still check the sickest patient first. In this report, that means the long-term unemployed.

One of the most watched AI leaderboards now keeps most of its test hidden from the labs it ranks. Artificial Analysis rebuilt its ranking so that 60% of the score comes from private test sets or answers the AI labs never see. If a model has been trained on the questions, a high score may only show that it has seen them before. When the private tests went in, the gap between the best closed model and the best open-weight model grew from 3 points to 12. Medical licensing exams have worked this way for decades. Candidates don't see the real questions before the day. Any AI tool sold to a hospital should be tested the same way, on cases it has never seen.

A normal ECG does not rule out a heart attack. In the first hours, the tracing can look normal or nearly normal while heart muscle is already being damaged. That is why emergency teams check a blood test called troponin and usually repeat it an hour or a few hours later, along with the ECG. If you have chest pain and you are sent home after one normal ECG and no blood test, it is fair to ask about troponin.

Doctors and nurses: what's one part of your job you're sure AI won't do in the next 10 years?

Geoffrey Hinton wants AI approved the way we approve new drugs. I'm a doctor who tests AI. Here's what that would actually look like, phase by phase 🧵