The author is an emergency physician. Full professor at the University of Montreal, he teaches, participates in research in emergency medicine and speaks frequently on health issues.
In recent years, artificial intelligence (AI) has developed at such an impressive rate that I, like many others, am quite convinced it will transform our healthcare. For the well-being of patients, everyone hopes so — although we are at the beginning of this type of manifestation.
I will take as an example a specific subject, the interpretation of the electrocardiogram (ECG), because I know it well. Not only are the advances in this area tangible, but some pretty amazing tools are now accessible.
Reading the ECG, printed on paper, depends on the doctor’s cognitive performance and experience. A complex reflection of cardiac electrical activity, this tracing makes it possible, in particular, to diagnose myocardial infarction, this acute blockage of a coronary artery that feeds the heart and which requires rapid detection.
In addition to the signs of a heart attack, the ECG provides information on a series of more or less serious diseases that affect our formidable pump. And as its interpretation constitutes a field of knowledge in constant evolution, the potential for improvement is therefore enormous.
However, if we use this centuries-old medical technology hundreds of times a day, in all hospitals and beyond, its teaching generally remains as insufficient as it is laborious. Despite everything, clinical answers must often be obtained without delay, well before the cardiologist’s final reading.
Or automated algorithms already integrated into contemporary ECG devices, which are not AI-based, provide part of the solution. But since their reliability is limited, they regularly miss several crucial anomalies.
Finally, even if well interpreted according to standardized and taught criteria that aim to recognize, for example, a heart attack on the ECG, the tracings themselves present imperfections, which can compromise the diagnosis by leading to a false negative (when a true disorder is not detected) . or on the contrary, a false positive (when an apparent problem on the ECG does not correspond to anything real).
The first case causes delays in treatment (here, reopening the artery by balloon or medication) that are harmful, while the second can lead to unnecessary interventions when these coronary arteries are not actually blocked. However, recently, AI already seems superior to doctors, and even specialists, when it comes to recognizing an acute heart attack.
This is because artificial intelligence does not work like the human brain, even the most talented one. Trained from a very large number of ECGs, the AI associates these strange electrical signals with possible diagnoses, based not only on the theory taught, but also on data that is sometimes too subtle for the human eye from the countless real cases used.
Looking at the speed of these developments, I can predict that AI ECG interpretation will replace mine within a few years, which will quickly open the door to much broader questions, the answers to which top experts are already investigating.
For example, in the case of errors, which occur despite solid AI performance, who will be responsible? The clinician, who depends on a tool superior to his own brain in this matter? The hospital, who implemented the technology? The legislator, who authorized its use? Or the manufacturer, who guaranteed their accuracy, even though the internal learning processes of these tools are poorly understood?
The issue of computer data security will also arise much more acutely than today, in particular due to the masses of clinical information used to train AI systems.
And, as with all emerging technologies, it remains to demonstrate the superiority of AI over human reading in a real-world situation, thus leading the implementation of AI, we hope, to improved care, but also, and above all, to favorable patient outcome.
Finally, a delicate issue will soon arise, namely that of maintaining competence in several areas. If, for example, a reliable AI reads ECG much better than a clinician, it is quite likely that our skills, which are becoming less and less demanded, will gradually disappear, with an advanced course gradually losing interest in this subject.
We will then become dependent on this technology, in ECG interpretation, as in a number of other fields, and in medicine, as elsewhere. We have a little time left to properly plan the path leading to this new easement.