Ten electrodes and a nurse, and the trace goes into the record [s1 c8]. Someone reads it for rhythm, and the rest of the recording sits there for the life of the file. The asset Imperial's group is working with is the unread part of a test that hospitals already run in volume, which is why an echocardiogram, with its sonographer, its equipment and its booked appointment, is the scarce thing and the ECG is not [8][11].
Teams tell themselves they are buying a model that beats a clinician, described by one of its own authors as "superhuman AI" [5]. In practice, the person who owns the echo list works down a queue in referral order, for a wait that Professor Fu Siong Ng of Imperial puts at several months [6]. What is being sold is detection accuracy; what is being used is a sorting function [9].
That distinction decides where the risk lands. At the ceilings reported, roughly 19% of heart failure cases and 10% of valve disease cases are not flagged [19], and in a triage system a low score is not a neutral output. It is a reason someone stays where they were in a queue for a disease that gets worse while the queue moves [7]. The consent form, the audit trail and the conversation in February about a patient scored low in October all belong to the hospital, not to the model.
The training design is the strongest part of the story and the least quoted. The models were built on 1.6 million Brazilian ECGs linked to patient records, plus several million more recordings from the US [12], which is about 24 times the population of the 67,000-patient trial itself [20]. The Brazilian linkage matters because the traces are tied to what happened next, so the model can learn patterns that preceded a later diagnosis rather than patterns that a clinician had already labelled at the time of the test [15]. Triage is a prediction about who will turn out to be ill, so outcome-linked data is the right shape of data for it.
The economics are unusual for diagnostics. That is the honest reason to look at this rather than the accuracy figures. Nothing new gets wheeled into a room. ECGs are among the cheapest and most widely used tests, and the spend is software applied to recordings already being collected [11]. That puts it in a different budget line from buying scanners, and it means the pilot can be small.
Two questions sort this kind of tool for any operator, and they form the grid. First: is the input already being collected as part of routine work, with no new staff time? Second: do you control the downstream capacity you are re-ordering? If the input is free and the bottleneck is controlled, that is genuine capacity relief. If the input is free but the bottleneck is not controlled, you end up with a better-sorted queue that moves at the same speed, plus a new record of who you ranked low. If the input is new and the bottleneck is controlled, it is a workflow project dressed up as a software purchase. If neither condition holds, it is a research result.
The number worth asking a vendor for is therefore not detection rate. It is median time from ECG to echo for flagged patients against unflagged, and the proportion of flagged patients whose scan confirmed something. Those are measurable in one clinic in a quarter. On the reported figures, a fifth of heart failure cases will not be in the flagged group [19], so a pilot that only counts what the model found will look better than the service does.
