Tuesday, October 10, 2017

The Resurgence and Limitations of Artificial Intelligence in Medicine

I came of age in the biomedical informatics world in the late 1980s, which was near the end of the first era of artificial intelligence (AI). A good deal of work in what we called medical informatics at that time focused on developing “expert systems” that would aim to mimic, and perhaps someday replace, the cognition of physicians and others in healthcare.

But it was not to be, as excessive hype, stoked with misguided fears about losing out to Japan, led to the dreaded “AI winter.” Fortunately I had chosen to pursue research in information retrieval (search), which of course blossomed in the 1990s with the advent of the World Wide Web. The “decision support” aspect of AI did not go away, but rather was replaced with focused decision support that aimed to augment the cognition of physicians and not replace it.

In recent years, it seemed that the term AI had almost disappeared from the vernacular. My only use of it came in my teaching, where I consider it essential to learning to understand the history of the informatics field.

But now the term is seeing a resurgence in use [1]. Furthermore, modern AI systems take different approaches. Rather than trying to represent the world and create algorithms that operate on those representations, AI has reemerged due to the convergence of large amounts of real-world data, increases in storage and computational capabilities of hardware, and new computation methods, especially in machine learning.

This has given rise to a new generation of applications that again try to outperform human experts in medical diagnosis and treatment recommendations. Most of these successful applications employ machine learning, sometimes so-called “deep learning,” and include:
  • Diagnosing skin lesions – keratinocyte carcinomas vs. benign seborrheic keratoses and malignant melanomas vs. benign nevi [2]
  • Classifying metastatic breast cancer on pathology slide images [3]
  • Predicting longevity from CT imaging [4]
  • Predicting cardiovascular risk factors from retinal fundus photographs [5]
  • Detecting arrhythmias comparable to cardiologists [6]
Unfortunately, the hype is building back too, perhaps exemplified by the IBM Watson system [7]. I recently came across an interesting article by MIT Emeritus Professor Rodney Brooks that put a nice perspective on it and stimulated some of my own thinking [8].

From my perspective, the most interesting part of Brook's piece concerns “performance vs. competence.” He warns that we must not confuse performance on a single task, such as making the diagnosis from an image, with the larger task of competence, such as being a physician. As he states, “People hear that some robot or some AI system has performed some task. They then generalize from that performance to a competence that a person performing the same task could be expected to have. And they apply that generalization to the robot or AI system.”

I have no doubt that algorithmic accomplishments in the above medical examples will be used by physicians in the future, just as they now uses automated interpretation of EKGs and other tests that comers, in part, from earlier AI work. But I have a hard time believing that the practice of medicine will evolve to patients submitting pictures or blood samples to computers to obtain an automated diagnosis and treatment plan. It will be a long time before computers can replace the larger perspective that an experienced physician brings to a patient’s condition, to say nothing of the emotional and other support that goes along with the context of the diagnosis and its treatment. Indeed, the doctors of Star Trek are augmented by automated tools but in the end, still compassionate individuals who diagnose and treat patients.

Somewhat tongue in cheek, I won’t say that machines replacing physicians is impossible, since there is a quote in a different part of the article, attributed to Arthur C. Clarke, aimed at people like myself: “When a distinguished but elderly scientist states that something is possible, he is almost certainly right. When he states that something is impossible, he is very probably wrong.” As someone who does not consider himself quite yet to be elderly, but is has worked in the field for several decades, I want be careful to not say that something is “impossible.”

But on the other hand, while I am certain that we will see growing numbers of tools to improve the practice of medicine based on machine learning and other analysis of data, it is very difficult for me to see no continued role for the empathetic physician who puts the findings in context and supports in other ways the patient whose diagnosis and treatment are augmented by AI.


1. Stockert, J (2017). Artificial intelligence is coming to medicine — don’t be afraid. STAT, August 18, 2017. https://www.statnews.com/2017/08/18/artificial-intelligence-medicine/.
2. Esteva, A, Kuprel, B, et al. (2017). Dermatologist-level classification of skin cancer with deep neural networks. Nature. 542: 115-118.
3. Liu, Y, Gadepalli, K, et al. (2017). Detecting cancer metastases on gigapixel pathology images. arXiv.org: arXiv:1703.02442. https://arxiv.org/abs/1703.02442.
4. Oakden-Rayner, L, Carneiro, G, et al. (2017). Precision radiology: predicting longevity using feature engineering and deep learning methods in a radiomics framework. Scientific Reports. 7: 1648. https://www.nature.com/articles/s41598-017-01931-w.
5. Poplin, R, Varadarajan, AV, et al. (2017). Predicting Cardiovascular Risk Factors from Retinal Fundus Photographs using Deep Learning, Arxiv.org. https://arxiv.org/abs/1708.09843.
6. Rajpurkar, P, Hannun, AY, et al. (2017). Cardiologist-Level Arrhythmia Detection with Convolutional Neural Networks, Arxiv.org. https://arxiv.org/abs/1707.01836.
7. Ross, C and Swetlit, I (2017). IBM pitched its Watson supercomputer as a revolution in cancer care. It’s nowhere close. STAT, September 5, 2017. https://www.statnews.com/2017/09/05/watson-ibm-cancer/.
8. Brooks, R (2017). The Seven Deadly Sins of AI Predictions. MIT Technology Review, October 6, 2017. https://www.technologyreview.com/s/609048/the-seven-deadly-sins-of-ai-predictions/.

No comments:

Post a Comment