In 1950, British mathematician Alan Turing proposed a test to determine whether a computer is capable of thought. Now known as the Turing test, the evaluation involves a remote human interrogator who must distinguish between a computer and an actual person based on their replies to various questions. He predicted that, by the year 2000, a computer would be able to pass the Turing test after only five minutes of questioning.
While artificial intelligence (AI) has made huge strides in the past half-century, still no computer has come close to reaching this point. What has happened, however, is that AI has permeated our day-to-day lives in ways that perhaps even Turing could not have foreseen. Just a few of today’s real-world applications include speech recognition in virtual assistants like Siri or Google, recommendation engines that suggest new movies to watch on Netflix or products to buy on Amazon, and computer vision for self-driving cars or photo tagging on social media.
One area where AI-based technologies are advancing rapidly – but have not yet become an everyday reality – is medicine. Researchers are finding ways to incorporate AI into clinical settings as a tool to uncover relevant information from large amounts of data and assist in better decision-making. Essentially, the hope is that AI will overcome human blindspots and limitations to provide improved diagnostics, treatment, and overall health outcomes.
Humans vs. AI
Evidence that supports the use of AI in medicine has been concentrated in fields with a strong image-based or visual component, such as radiology, pathology, ophthalmology, and dermatology. Algorithms using a subtype of AI called deep learning can perform pattern recognition in medical scans, pathology slides, retinal images, and other pictures. Digitized inputs, such as an image or speech, go through multiple layers of “neurons” – known collectively as a deep neural network – that progressively detect physical features.
Multiple studies have demonstrated the ability of deep neural network algorithms to match or even outperform human physicians on a specific task. In 2017, an algorithm to classify skin lesions as malignant or benign achieved performance on par with 21 board-certified dermatologists when given biopsy-confirmed clinical images. Another system designed to detect malignant pulmonary nodules in chest X-rays did a better job when pitted against a group of radiologists in 2018. A 2017 study gave over 130,000 retinal fundus photographs to both algorithms and ophthalmologists in order to diagnose age-related macular degeneration. The accuracy of the algorithms ranged from 88% to 92%, which almost reached the level of the most expert ophthalmologists.
Despite the hype surrounding such contests of man versus machine, the vast majority of AI-based technologies have yet to prove their true clinical utility. In other words, they have not been evaluated in controlled clinical studies to evaluate their impact on healthcare decisions and patient outcomes.
Replacing human doctors?
Even for those algorithms approved by federal regulatory agencies like the U.S. Food and Drug Administration (FDA), the decisions were not based on randomized clinical trial evidence that the products actually improved care. One of the earliest AI-based technologies to complete the regulatory process in the U.S. is IDx-DR, the first autonomous diagnostic system approved by the FDA back in 2018. It analyzes images of a patient’s retinas for diabetes-related eye problems using an AI algorithm.
Interestingly, it does not require a clinician to interpret or review the images before giving the patient a test result. The motivation for such a system is to increase patient access to needed healthcare by allowing people to take the test in settings where an optometrist or ophthalmologist is not readily available.
However, the concept of AI-based systems replacing doctors or making decisions for them instead of simply augmenting their current role is still somewhat controversial. And more generally, AI is no magic bullet, and many ethical and logistical questions around the use of such technologies in medicine remain unanswered. If a patient suffers harm as a result of an AI-guided decision, who is responsible? How will the quality and efficacy of algorithms be evaluated for clinical use? What about inadvertent bias introduced into the algorithm by its human authors?
Avoiding the pitfalls
Many experts believe that AI will nevertheless have a major impact in certain fields of medicine, less as a replacement for human doctors and more as a way to make their jobs easier. For example, AI algorithms are predicted to save countless hours in radiology by performing tasks like automatic segmentation of structures within medical images. Pre-analysis of images by an AI tool may become routine clinical practice in the next decade, but it will likely still be followed by a comprehensive review and final decision from a human radiologist.
In addition, AI-based technologies remain vulnerable to many pitfalls. Training datasets fed to the AI strongly influence the resulting algorithm, so they need to be of high quality and very large in order to produce a reliable system. Also, any biases inherent to the dataset, such as only focusing on a certain demographic, can translate to inaccurate results when applied to a more general patient population.
The promise of AI to transform medicine is still a possibility, however, and many believe that it has the greatest potential for low-resource settings and regions where adequate medical infrastructure is lacking. The speed and efficiency of AI-based tools may serve as a key advantage in places where very few doctors are overwhelmed by a large number of patients who require care.