The Devastating Impact of Artificial Intelligence (AI) on the Traditional Ways of Practising Imaging and Cardiology Are ‘imagers’ an endangered species, and should trainees rethink their careers?

Cardiac imaging is a fast-expanding field, with demand for certain imaging tests, such as CT coronary angiography (CTCA), growing at astronomical speeds. ^[12] This focuses investors’ attention and pressures to maximise profits are greater than in other fields of health care such as, say, cardiac rehabilitation, or palliative care in end-stage heart failure. Having said that, the impact of such ‘non-sexy’ interventions on quality of life is far higher than, for instance, prophylactic coronary calcium scoring coupled with other non-invasive cardiovascular screening tests.^[13]

Diagnostic tests are poorly ‘scalable’:^[14] there is a limited number of scans that can be done and interpreted per unit of time, determined by the physical laws of the Universe and by the availability of scanners and of trained radiologists (whereas, for instance, a 3-minute song can become a hit, be downloaded or streamed billions of times, and generate almost limitless revenue - it is highly scalable). It follows that in order to maximise profits, increasing the availability of imaging is a less effective strategy than cutting the costs associated with it. Together with the increasing potential profits at a time when demand for imaging is increasing, its poor scalability realises a perfect storm for the ‘rise of the machines’ and the elimination of human involvement in imaging.^[15]

Reviews on the topic of AI in cardiac imaging are generally upbeat and emphasize the ‘bright side’ of AI¹⁶, which is deemed to make cardiac imaging ‘thrive’^[17] and which is credited with benefits, for instance at the time of the COVID pandemic.^[18]

Image reporting by radiologists or cardiac ‘imagers’ is patternrecognition backed up by experience and judgement. AI can do that much faster, more accurately and cheaper than humans. While a human needs decades to train in order to deliver a productive professional life-span of 2 or 3 decades, AI trains itself in days or weeks using the whole accumulated experience of humanity in any given field. Throughout a doctor’s career, relatively high wages,^[19] sickness benefits, and pension plans need to be paid by the employer, who remains financially liable (through having to pay a pension) potentially for decades after the doctor has retired and has stopped generating any value for the health care system. Humans make mistakes, and their performance may vary unpredictably due to external factors over which they have no control, which causes increasing levels of litigation, with their added financial costs to the employers.^[20]

In comparison, AI is almost implausibly cheap,^[21] and – as a consequence of Moore’s law – (although we accept its continued validity is debated)^[22] it is likely to become progressively cheaper over time. Moreover, AI image interpretation is consistent, fast, and at least of similar accuracy to human reporting, even in studies performed from an ‘AI-hostile’ point of view^[23] leading to the opinion that ‘we should stop training radiologists now’.^[24] AI also makes radiomics possible, i.e., the extraction of quantitative data from an imaging dataset;^[25] early and already widely used applications include longitudinal strain in echo and T1 mapping in cardiac magnetic resonance (CMR), but the potential of this type of ‘deep-dive’ into data mining is much broader and not yet fully realised.

Finally, AI can identify patterns and structure where the human eye does not discern them, and, as such, can identify new cardiac phenotypes, e.g. in aortic stenosis^[26] or predict the risk of adverse outcomes from seemingly normal ECG tracings.^[27]

One of the most common indications for performing a transthoracic echocardiogram in hospital is the assessment of left ventricular ejection fraction.^[29] At present, at least in the UK, the process for this involves submitting to the local echo department a digital or paper-based request for a scan, taking the patient to the scanner when a slot becomes available for scanning, having a highly-specialised echo physiologist perform the scan over approximately 45min, and then generate a report which is sent to the requester on paper or digitally (depending on local set-up).

The advent of ‘pocket’ and smart-phone echo scanners^[30] demonstrated that useful, diagnostic quality images can be acquired by almost anyone after minimal training.^[31] Even better, AI can now guide (in real time!) a novice to acquire high-quality transthoracic echocardiograms, as demonstrated in some spectacular YouTube videos.^[32]

AI is already deployed on high-end, commercially-available scanners, found in most cardiology departments in the West, and is able to measure autonomously (w/o any human intervention) ventricular and atrial volumes and ejection fractions, as well as global longitudinal myocardial strain, a process that takes about half a minute and which requires a single 3D, 4-chamber acquisition from the apex.^[33]

Here is how we see the hospital echocardiography service, 10 years from now: There will be at most 2–3 people specialised in echo (with skills equivalent to those of any of the 8–10 highly specialised echocardiography physiologists in a contemporary UK department) in any average-sized cardiac department. They will be there primarily for quality control (QC) and for the occasional complex patients. Just like today’s nurses do echo ward rounds, scanning bladders in oliguric patients, there will be minimally-trained health-care assistants able to acquire apical 4-chambers view in 3D, using dedicated small scanners with on-board (or Bluetooth/ Wi-Fi accessible) AI, to guide the acquisition of echo images, doing echo ward rounds and acquiring apical 3D datasets potentially in all patients. In poorer settings they will acquire just 2D images under AI guidance, perhaps using pocket or smart phone-based scanners. Once a study is finalised, the datasets are uploaded on a central server where AI produces bi-atrial and bi-ventricular volumes and EF and (if deemed useful) some myocardial strain-based parameters.

The research to prove the next point is yet to be performed, but we think a single 3D colour flow mapping acquisition from the apex can resolve the assessment of the severity of valvular regurgitant lesions in most patients with regurgitation of the atrioventricular valves, and a parasternal one in those with aortic valve lesions. Patients with difficult images will be reviewed by the few remaining sonographers, with a view to consider either echo contrast studies or (more likely) immediate CMR (which will also be entirely automatized and with mostly autonomous reporting – current reporting software already uses AI to measure volumes and EF, and it is increasingly accurate).

We emphasize that the technology to implement such a model is already available; all that is needed is for the penny to drop and for someone to join the dots - the cost savings would be astounding. Through a combination of inertia, possible active resistance from various professional medical bodies (totally understandable, because this would ruin the huge industry of echo reporting, in fee-for-procedure healthcare systems) and simple, blissful ignorance, we continue to practice in the old ways, and to train new generations to acquire skills that would make them redundant within years from now.

We thought, until recently, that an obvious imaging niche where humans seemed protected was intra-procedural echo (mainly TOE, but also ICE). However, we are now less sanguine than we were just a few weeks ago. At 14:35hrs on 14/06/2023, on the Twitter feed (@GHCardiology) of the Glenfield Hospital (a major teaching and NHS institution in Leicester, in England) a post appeared describing how a nurse practitioner (identified only as ‘John’, and depicted with a mask covering the lower half of his face) had just completed a TAVR procedure as a first operator. The story was presented with a very positive spin, as a major breakthrough, and ‘a true transformation addressing NHS needs’. Within minutes, a huge backlash developed, to the point where the British Cardiac Intervention Society (BCIS) was forced to release an official statement to address the outcry. Trainees as well as senior, established consultants joined in an indignant chorus of opprobrium: such a development was perceived as putting patients at risk, flying in the face of training needs of structural fellows, and demeaning a complex procedure by devolving it to a non-physician grade. Clearly, if a NP does TAVR, there is no reason for which they (or echo physiologists) cannot do intra-procedural TOE to guide the small number of structural interventions still requiring it - in fact apocryphal evidence suggests this is already happening in some centres.

We think the stethoscope should be abandoned, and ubiquitously replaced with ‘point of care’ pocket ultrasound scanners (POCUS). Trainees should be taught to image all scannable parts of the body, as part of the physical exam, starting with their first clinical attachment, and POCUS should become the new stethoscope. In this way it may be possible to maintain the relevance of the doctor in the rapidly-shifting, unfamiliar landscape of contemporary medicine. Otherwise, our prediction is that we will be decimated and annihilated by the rise of AI, coupled with dwindling resources and increased demand from the ageing, sicker population.