The term “artificial intelligence” (AI) was coined on August 31, 1955, when John McCarthy, Marvin L. Minsky, Nathaniel Rochester, and Claude E. Shannon submitted “A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence.” [¹, ²]. It was, however, Alan Turing who during a public lecture in London in 1947 mentioned computer intelligence, and in 1948 he introduced many of the central concepts of AI in a report entitled “Intelligent Machinery.” [³]. Moreover, Turing proposed in 1950 the test, originally called the imitation game, and later known as the

Turing test, as a way to confirm that the intelligent behavior of a machine was equivalent to that of a human. A human evaluator is asked to determine the nature of a partner (human or machine) based on a text-only conversation [¹–³]. After decades of slow progress since the

Turing test was proposed, AI has finally blossomed. Many new technologies and applications are available, and there is great enthusiasm about the promise of AI in health care. It holds the potential to improve patient and practitioner outcomes, reduce costs by preventing errors and unnecessary procedures, and provide population-wide health improvements. We have entered the fourth stage of the

Industrial Revolution that began in the eighteenth century, and its defining feature may well be the use of AI technologies (Fig. 1.1). The results of an annual competition known as the

ImageNet Large Scale Visual Recognition Challenge (ILSVRC) provide interesting insights into recent

developments in AI technology (Fig. 1.2). Over the years 2010–2016 there was a steady decrease in the error rates of the algorithms presented, and in 2017, 29 of the 38 competing teams had error rates lower than 5% (considered to be the human threshold). Thus in 10 years AI algorithms exceeded human performance in image recognition.

The four main stages of the

Industrial Revolution that began in the eighteenth century

Error-rate

history on ImageNet

There are many promising applications for AI in health care, addressing a variety of

aims and taking many different approaches (Table 1.1). For example, misdiagnoses constitute a huge, although poorly recognized, medical problem. A study published in 2014 estimated that diagnostic errors affect at least 5% of US adults (12 million people) per year [⁴]. More recently a systematic review and meta-analysis reported that the rate of diagnostic errors causing adverse events among hospitalized patients was 0.7% [⁵]. Furthermore, diagnostic error is the most important reason for malpractice litigation in the United States, accounting for 31% of malpractice lawsuits in 2017 [²]. The creation of AI

programs to identify and analyze diagnostic errors could be an important step in addressing this problem [⁶].

Some

ambitious expectations for AI in health care. Adapted from Topol E. Deep Medicine: How Artificial Intelligence Can Make Healthcare Human Again. Basic Books, New York 2019

Eric Topol has proposed that AI could help shift into “deep medicine,” by allowing the physicians to devote more time to crucial relationships with their patients—an aspect of medicine that cannot be replaced by any AI technology [²]. It is also interesting to consider whether AI might enrich the

doctor-patient relationship, enabling a shift from the present “

shallow medicine” into “

deep medicine,” based on deep empathy and connection [²]. Success in building such relationships is very much related to the amount of time doctors can spare for patients and the extent of the personal contact they have with their patients. The average time of a clinic visit in the United States for an established patient is 7 min and for a new patient 12 min. In many Asian countries, clinic visits last as little as 2 min per patient [²]. Making this situation even worse, part of this time must be devoted to completing electronic health records, further limiting personal contact. A study published in 2017 that asked patients to describe how they perceive their physician found that the most common negative responses were “rushed,” “busy,” and “hurried.” [⁷]. These

reactions are manifestations of “

shallow medicine.”

One of the arguments supporting the use of AI in medicine is that

human cognitive capacity to effectively manage information is often exceeded by the quantity of data generated. Each year the world produces zettabytes of data (roughly, enough to fill a trillion smartphones) [²]. Moreover, unlike humans, who have bad days and emotions, and who get tired, with subsequent decreases in performance and accuracy, AI works 24/7 without vacations or complaints [²].

AI-based technologies employing deep-learning (DL)

approaches

have proven effective in supporting decisions in many

medical specialties, including radiology, cardiology, oncology, dermatology, ophthalmology, and others. For example, AI/DL algorithms (also referred to as AI/DL models in the following text) have been shown to reduce waiting times, improve medication adherence, customize insulin dosages, and help interpret magnetic resonance images. The number of AI

life-science papers listed in PubMed increased from 596 in 2010 to 12,422 in 2019 [⁸]. The number of papers on the use of AI in the field of ophthalmology also has increased dramatically (Figs. 1.3 and 1.4).

The number of PubMed

articles on AI and the eye that were published between 1974 and 2020

The number of

articles relating to AI and eye diseases published in 2020

algorithms

have been used to detect diseases based on image analysis, with fundus photos and optical coherence tomography (OCT) scans analyzed for retinal diseases, chest radiographs assessed for lung diseases, and skin photos analyzed for skin disorders. Retinal photos have also been used to identify risk factors related to cardiovascular disorders, including blood pressure, smoking, and body mass index [⁹]. Using

DL models

trained on data from over 280,000 patients and validated on two independent data sets, Poplin et al. predicted cardiovascular

risk factors not previously thought to be present or quantifiable in retinal images, such as age (mean absolute error within 3.26 years), gender (area under the receiver operating characteristic curve = 0.97), smoking status (AUC = 0.71), systolic blood pressure (mean absolute error within 11.23 mmHg) and major adverse cardiac events (AUC = 0.70) (Fig. 1.5) [⁹].

Attention

maps for a single retinal fundus image. The top left image is a sample retinal image in color from the UK Biobank data set. The remaining images show the same retinal image in black-and-white. The soft attention heat map for each prediction is overlaid in green, indicating the areas of the heat map that the neural-network model is using to make the prediction for the image. Source: Poplin R, Varadarajan AV, Blumer K, Liu Y, McConnell MV, Corrado GS, Peng L, Webster DR. Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning. Nat Biomed Eng. 2018 Mar;2(3):158–164

The

COVID-19 pandemic has raised expectations for the use of AI in data analysis. So far it has been used in epidemic modeling, detection of misinformation, diagnostics, vaccine and drug development, triage and patient outcomes, and identification of regions of greatest need [¹⁰].