The cardiologist in the age of artificial intelligence: what is left for us?

With the transition from unicellular to multicellular life, the coordination of increasingly complex organisms became a need. Certain higher organisms, namely, animals, developed neurons and eventually neuronal networks that allowed for rapid adaptation to the environment through electrical signal transduction. In evolution, the brain as the centre of the nervous system became most developed in vertebrates. In particular, primates became endowed with brains that allow them to master complex tasks, assemble in groups, communicate among each other, and, in the case of humans, speak. The vast and highly interconnected neuronal network in human brains equipped us with the most advanced intelligence known to date. For centuries, there was consensus that human-level intelligence can only arise from biological brains such as ours, the ‘crown of evolution’. The human brain is endowed with a seemingly miraculous network of around 80 billion neurons linked through 100 trillion synaptic connections. Since its development, the brain of humans and their ancestors invented increasingly sophisticated tools from the earliest stone tools to modern technologies. In sharp contrast to the slow and gradual changes occurring in biological organisms, technological change happened exponentially, with shorter and shorter intervals between one innovation and the next. After millions of years of biological and technological co-evolution, the introduction of machine learning algorithms has rapidly increased the capacity of computer systems to solve intellectual tasks (Figure 1).1 For example, on 10 February 1996, Deep Blue, a then supercomputer built by IBM, beat the reigning world chess champion Garry Kasparov for the first time. Similarly, the world’s leading Go player Lee Sedol was at first able to defeat the AlphaGo Algorithm, developed by Google DeepMind, but resigned only a few years later in view of the increasing sophistication of the programme. It has become evident that intelligence is a phenomenon generated by sufficiently complex signal transduction systems, be it biological or artificial in nature. Like human brains, today’s artificial neuronal networks learn from data they receive from their environment. Such neuronal networks work tirelessly with uninterrupted attention, exponentially increasing capabilities, and the ability to create copies of themselves. Their application in medicine has only just begun.2 Schematic graph showing the relative importance of physical (green) and intellectual work (purple) conducted by humans and their ancestors since the Old Stone Age. Indeed, although what is possible with artificial intelligence (AI) in medicine and cardiology even now is impressive,2,3 next-generation models will surpass what is possible today. Algorithms analyse images of any sort, be it the face of a patient with its wrinkles, hair structure, eyes and ear lobes or images of any organ obtained by ultrasound, computed tomography (CT), and magnetic resonance imaging (MRI). Digital records such as electrocardiogram (ECG) readings contain much more information than a human eye would appreciate on a paper printout.4 AI integrates seemingly endless amounts of data on demographics, symptoms, laboratory values, and imaging into the most likely diagnoses. The avatar ‘Emily’ from OpenAI’s ChatGPT already masters diagnostic skills at the level of an experienced cardiologist and not rarely beyond. Some algorithms even talk to the patient and obtain their medical history; one of them outperformed general practitioners in a randomized study by rapidly providing the most likely diagnoses based on a conversational interaction with clinical scenarios played by actors.5Natural language processing reads unstructured reports and is able to integrate that information into a diagnosis, for example, heart failure with preserved ejection fraction.6 Moreover, AI is increasingly used to provide precise predictions of future clinical events and outcome analyses,7,8 and more is to come as we summarize what has been achieved. No doubt, AI will transform cardiology as it is a specialty integrating vast amounts of data on patient history, symptoms, ECGs, laboratory values, and imaging of all sorts. How will it change the role of the physician in cardiology and medicine at large? Will it become doctor’s little helper, the co-pilot of the physician, or will it take over the plane on the journey to the future altogether? As we try to answer this question, let us recapitulate where AI algorithms surpass doctors already today. First, algorithms work tirelessly with consistent performance and at unprecedented speed. If appropriately developed according to current quality criteria9 and externally validated,10 they can provide reliable, highly integrated, and rapidly available information. Second, AI is able to learn and get better as we speak, integrating new information in real time. Third, algorithms see what we cannot see, for instance, when reading an ECG or a coronary CT.11 So, what is left for the physician? First, it seems likely that physicians will remain legally responsible for the foreseeable future. Therefore, there is an educational gap that needs to be closed; physicians have to learn to assess the quality of algorithms. Have they been derived from large and reliable data sets? Have they been evaluated externally? van Royen and associates have defined five quality criteria for AI algorithms that physicians should be aware of: first, complete reporting and reproducibility of results; second, clear intended clinical use; third, rigorous internal and external model validation; fourth, sufficient sample size for AI model development and validation; and lastly, openness of data and software.9 Registration agencies consider algorithms a device and evaluate and approve such tools based on quality criteria, providing further confidence in the clinical usefulness of an algorithm. No doubt, AI will soon have more profound effects on the medical community. As AI systems will be able to talk to patients, characterize their phenotype and ECG, and analyse imaging scans with unprecedented speed and precision, the number of physicians and allied professionals required for an effective clinical service will likely decrease. Indeed, echocardiography laboratories will need fewer medical and allied professional personnel and the interpretation of CT11 and MRI images far fewer cardiologists or radiologists. Likewise, fewer pathologists may be required to assess biopsy specimens histologically, skin checks will need fewer dermatologists, and clinical chemists may almost completely be replaced. In view of a global shortage of physicians and allied professionals, coupled with an increasing desire for shorter working hours, AI may help to cope with this discrepancy and to improve health care. On the other hand, surgeons and interventionalists are harder to replace. Certainly, there are robotic systems for ablation procedures12 and coronary interventions,13 but their use is complicated, time-consuming, and currently slower than experienced operators. The cardiologists or general practitioners in their clinic will benefit from AI as they will rapidly receive processed information even before seeing their patients including probabilities of different diagnoses and guideline-based recommendations for further examinations and treatment. Medical decisions and actions will become faster and more precise. Above all, physicians may have more time to speak to their patients and engage with their questions, fears, and hopes—in short, they may become better doctors. At present, AI models are tools without emotions. However, many models under development are being trained to act as if they were able to feel and do respond accordingly. Will AI eventually develop empathy? Empathy, the ability to understand and share feelings with one another, is a typical trait of specific animals with sufficiently advanced nervous systems, humans in particular. In 1990, Giacomo Rizzolatti and his team at the University of Parma discovered its neurobiological basis, i.e. the so-called mirror neurons.14 As the basis of empathy within biological neural network has been at least in part unravelled, it would be no surprise, if future AI models would be able to have and share real feelings—we will see. Lastly, creativity and innovation: AI algorithms are already able to write papers and reviews on what is available in the literature. While the majority of meaningful scientific problems are still solved by humans, this may change as more and more aspects of human intelligence are disentangled by neuroscience and incorporated in AI algorithms. As the philosopher Thomas S. Kuhn pointed out,15 there are two forms of scientific discovery, i.e. research along a paradigm and scientific revolutions. While most of what we are doing would be normal science, some discoveries were revolutionary in nature, e.g. in cardiology, William Harvey’s ‘De motu Cordis’,16 the description of the double-stranded DNA,17 the cholesterol hypothesis,18 and the open-artery hypothesis for acute myocardial infarction. Could algorithms spot the holes in our knowledge base, produce new hypotheses within an existing paradigm, or independently make fundamental scientific discoveries leading to new paradigms? We would not be surprised and agree with leading AI scientists that AI has the potential for both—leading to unprecedented progress and ending humanity. Luckily as for now, the future is still in human hands. None declared.