|Year : 2019 | Volume
| Issue : 2 | Page : 98-99
Artificial intelligence in medicine: The way forward
Department of Biology, Indian Institute of Science Education and Research, Pune, Maharashtra, India
|Date of Web Publication||25-Mar-2019|
Department of Biology, Indian Institute of Science Education and Research, Pune, Maharashtra
Source of Support: None, Conflict of Interest: None
|How to cite this article:|
Goel P. Artificial intelligence in medicine: The way forward. Med J DY Patil Vidyapeeth 2019;12:98-9
In an important essay, Physician and Pulitzer Prize winning author Siddhartha Mukherjee attempts to understand artificial intelligence (AI) in relation to its future in medicine. He compares the learning that a human diagnostician must undergo to the “learning” process of an algorithm attempting to predict that diagnosis. The key question is whether a machine's performance can improve as more and more examples are made available to it, that is, if a machine can truly be expected to “learn from experience” (AI researchers refer to this problem as supervised learning). It has now been demonstrated that algorithms such as (deep) neural networks are capable of impressive performance on such problems as detecting cancerous growth from dermatological photographs. A number of such computer programs are becoming available commercially, which makes us pause to ask: What role can we expect machine learning systems to play in the next generation of clinical practice?
Extrapolating from this (early) experience might lead an alarmist to conclude that doctors will eventually be replaced by machines. Most would argue more realistically, however, that AI will be used like any other technology in the past: Freeing the physician from the mundane to engage in the more sophisticated aspects of their craft. Viewed in this manner, AI becomes a tool of Intelligence Augmentation., In other words, machine intelligence will ultimately serve as one component in a complex decision-making process, centered on providing care.
There are important lessons to be learned as AI rapidly becomes a mainstream in medicine. AI systems are – at least, for the moment – largely “human-imitative.”, That is, they typically aim to reproduce (the diagnoses) which humans have achieved. Inasmuch as this is a considerable advance in terms of machines achieving learning, it also has the implication that they are only as good as the examples provided by physicians.
AI systems have several limitations at present. They are often out-maneuvered by “adversarial examples.” That is, an AI system that performs fine on thousands of examples may still happen to be outwitted by a stray example (possibly because of some unexpected feature of the example, “noise,” as it were). This suggests that a complete reliance on AI systems, without human audits or oversight mechanisms, may not be a foreseeable reality.
AI systems may be the only reasonably scalable healthcare mechanism for a large population, especially for noncommunicable chronic diseases. Curiously – as AI practitioners are well aware – humans can often learn from a few examples, whereas the current paradigm for training AI systems is to feed them a multitude of examples. In some ways, this represents an opportunity waiting to be taken advantage of. A large population implies that healthcare is continuously demanded by and provided to a great number of patients; in principle, this can aid in designing better algorithms (since all the requirements of supervised learning are satisfied: Large numbers of examples, including “control” as well as affected patients, and the corresponding physician diagnoses). In other words, a data dividend appears to be just around the corner: If the right resources could be marshaled into aligning current medical practice with data systems, scalable systems can indeed become possible in the very near future.
In the end, Mukherjee argues that there is another, deeper aspect of the medical practice that goes further than diagnostics or individual care. That of fundamental research: Investigating the very pathophysiology of disease itself, by working from clinical observations inwards to its biology. The new focus on machine learning may bring human communities closer across disciplinary boundaries: It is now not too hard to imagine a clinician reaching out to a data scientist – perhaps a faculty in a nearby University – to help make sense of blackbox AI, and to help integrate it into practice. The data scientist, in turn, is delighted for the chance to design computer simulations for medicine, in ways more powerful than ever before. This may be AI's greatest gift, after all.
| References|| |
The Lancet. Artificial intelligence in health care: Within touching distance. Lancet 2018;390:2739.
Waldrop MM. News feature: What are the limits of deep learning? Proc Natl Acad Sci U S A 2019;116:1074-7.