Medicine and AI: Theory and Practice (2/2)

Artificial intelligence is already dethroning the medical practitioner in some very specific fields. What role will humans be allocated in an AI-predominant medical field?

Dr. Joris Galland breaks done the implications of artificial intelligence for medicine

About the author: Dr. Joris Galland is a specialist in internal medicine. After practicing at the Lariboisière Hospital (Paris, France) he joined the Bourg-en-Bresse Hospital (France). Passionate about new technologies, he offers to explain the issues at stake for the future of AI in medical research, practice, and policy.

Translated from the original French version.

In terms of technological progress, AI (rather considered on the "software" side of IT) was historically opposed to robotics. Nowadays, AI and robotics seem to converge more and more to create intelligent humanoid robots capable of performing tasks of increasing complexity.

In recent years, it must nevertheless be noted that AI has developed much faster than robotics. Does this mean that intellectual professions will disappear in favor of AIs, which are more efficient and would cost less to the employer? The answer is not that clear-cut. Let's take a concrete example: image recognition. Digital images can be counted in billions of billions, and, as we know, the more data an AI receives, the more it learns, and the more efficient and precise it is.

In the field of health, the specialties most impacted by AI are radiology, dermatology, and anatomical pathology. Indeed, these specialties are the source of a considerable quantity of labeled images for a precise diagnosis, i.e. what is called "clean" data that can be used immediately. It is sufficient to instruct the AI by indicating a binary result: cancer versus no cancer, melanoma versus no melanoma, infection versus no infection, etc.

The strength here lies in the fact that the diagnosis indicated by the machine is reinforced by what is known about the patient a posteriori: a subsequent biopsy having confirmed the malignant character of the suspect pulmonary nodule, a biomarker having confirmed the sensitivity of breast cancer to such and such a hormone therapy, etc. A human is therefore still very useful in the machine learning process.

The eye of the tiger

Current advances in imaging equipment make it possible to distinguish minute differences in tissue densities. AI is capable of differentiating thousands of grey densities present on an image, where the human eye - even that of an experienced radiologist - can only distinguish about ten1. Thanks to this precision of image analysis and to deep learning, AI today does as well as, or even better than, the radiologist.

This was the case in a study published in 2019 in the journal Nature2: a team of researchers developed an algorithm to predict the risk of lung cancer from low-dose CT images. When a previous image of the patient was available for comparison with the current image, the AI's performance was identical to that of the radiologist. But when no previous image was available, the AI performed better than the human eye, with a sensitivity of 94.4% and a reduction of 11% in false positives and 5% in false negatives compared to the radiologist. 

It takes years to train a radiologist, as opposed to a few days to train an AI in visual recognition: if the latter does better than the radiologist, isn't this profession doomed to disappear? "The death of radiologists is a matter of years: the machine will soon do their work much better than they do," says Laurent Alexandre.3

The practitioner's assistant

Thanks to a pixel-by-pixel analysis, AI is currently a very good image analyzer. In the future, it will probably be present in many other fields. However, we have not yet reached the stage where the AI can replace the physician. On the contrary, it could be a very valuable aid for the flesh and bone practitioner. One even speaks of an "artificial prosthesis for the biological neuron".

The IDx-DR software using AI for the detection of moderate to severe diabetic retinopathies was authorized by the FDA in 2018. It can be used by any primary care physician in the United States. All they have to do is enter the image of a patient's retina into the software's cloud, which provides the physician with one of two results:

  1. Diabetic retinopathy detected (to be referred to a specialist)
  2. Negative diabetic retinopathy

This authorization for screening practice, extended to general practitioners, is intended to respond to a real public health problem across the Atlantic: 50% of diabetic patients do not consult their ophthalmologist as they should.

In our primary care practices, the daily use of screening or diagnostic software using AI would allow us to save precious time. A "check-up" in the practice would improve the prevention of certain diseases, optimize their management and avoid unjustified specialist advice; a key to reducing the demand for care, but also to reducing the geographical divide in access to it. The only grey area: in the event of a medical error, whose fault is it? The physician who uses AI or the company that markets it?

AI in medicine: Three scenarios

In the years to come, the evolution of technologies but also of mentalities and patient acceptance will profoundly modify the use of AI in clinical practice. Three scenarios are possible:

A disruptive scenario: AI and robotics have evolved and surpass humans. “Strong" AI has become the norm. The machine is capable of "scanning" all pathologies in record time, makes the diagnosis, and offers personalized treatment based on the latest studies. The genetic code of all individuals is systematically "screened" from birth. As Kaï-Fu Lee predicts, the physician becomes the AI's nurse.3

Constant growth: AIs remain "weak", as technology has not allowed for strong AIs. Public acceptance of AIs is good. Patients have no problem with AIs assisting the physician in diagnosis or therapy. The physician continues to perform the clinical examination of the patient; the physician-patient relationship is preserved.

A decline: the need to obtain more and more big data to improve AI sets up a climate of mistrust in the population, which considers itself "spied on". The population is dissociating itself from new technologies. The growth of AI is in free fall. We maintain the physician-patient relationship as we and our ancestors have always known it.

Conclusion

It is impossible to know how this physician-AI relationship will evolve. Nevertheless, new technologies have already changed our daily lives. Social networks, Internet of Things, 4 and soon 5G... Society is increasingly digital and generates a profusion of big data. Everything is going faster and faster and is more and more precise.

With or without AI, medicine is becoming that of the 4Ps: predictive, preventive, personalized, and participative. What about the role of the physician? Kai-Fu Lee - the father of voice recognition and a figure of speech recognition - sketches out a hint of things to come: if AI helps the physician in administrative tasks or reduces the time spent on clinical reasoning, perhaps it could help to refocus on the "human" aspect of medicine?4 One aspect that will always escape AI and the machine is certain, according to Lee: "We are far from understanding the human heart and light-years away from reproducing it".

You can visit the first part of this article here.

References:
1. Hosny A, Parmar C, Quackenbush J, Schwartz LH, Aerts HJWL. Artificial intelligence in radiology. Nat Rev Cancer. 2018;18(8):500–10.
2. Ardila D, Kiraly AP, Bharadwaj S, Choi B, Reicher JJ, Peng L, et al. End-to-end lung cancer screening with three-dimensional deep learning on low-dose chest computed tomography. Nat Med. 2019;25(6):954–61.
3. La Guerre des intelligences : intelligence artificielle versus intelligence humaine, Laurent Alexandre, Paris, JC Lattès, 2017, 250 p.
4. I.A. La plus grande mutation de l’histoire, Kai-Fu Lee, Paris, Les Arènes, 2018