Emmet Feerick takes a look at the latest significant breakthrough in the application of machine learning to medicine.

 

Few professionals are more acquainted with the idea of “lifelong learning” than medical doctors. Every year, stacks upon stacks of new medical research are produced, and keeping on top of this exponentially growing pile is part of a doctor’s job description. Thankfully, this growth in knowledge is accompanied by a proliferation of new medical machinery. This enables medical professionals to best take advantage of this new information, improving the accuracy of their diagnoses, and the efficacy of their treatments.

In ophthalmology (the branch of medicine dealing with the eye), an imaging technique called optical coherence tomography (OCT) has become widespread since its introduction in the 1990s. It involves taking a 3D scan of the retina, which can then be viewed and manipulated by an ophthalmologist to help diagnose a wide range of eye conditions. In medicine, OCT is seen as the gold standard imaging technique for the initial assessment of many diseases of the eye, with 5.35 million OCT scans being performed in the US Medicare population in 2014 alone.

“In medicine, OCT is seen as the gold standard imaging technique for the initial assessment of many diseases of the eye”

Perversely, the widespread adoption of this technique has meant that there are now more scans available than there are doctors available to view them. This has created a backlog in many countries, where patients with potentially vision-threatening eye diseases wait weeks and months for a diagnosis and referral. It is this backlog that led researchers from Google’s Deep Mind to develop software which can read OCT scans and correctly diagnose more than 50 common eye conditions.

This software was trained using over 14,000 OCT scans from over 7,000 patients who were referred to the hospital with symptoms of macular pathology (diseases of the central part of the retina that is required for high-resolution, colour vision). As well as diagnosing eye conditions, the AI system was tasked with referring cases to the relevant specialists. Its performance on this task was compared against several experts, each of which had slightly different success rates in their referral (a worrying thought in itself).

“This software was trained using over 14,000 OCT scans from over 7,000 patients who were referred to the hospital with symptoms of macular pathology”

When all they had to go on was an OCT scan, the performance of the A.I was comparable to the two best-performing retina specialists in the experiment. The error rate of the A.I. was 5.5%, versus 6.7% and 6.8% for the two specialists. Additionally, it did not make a single clinically-serious wrong decision. When the human experts were allowed access to other images and summary notes of the patients, five came close to the performance level of the A.I, while three others still lagged significantly behind.

Expert reactions to the research, published in Nature Medicine this August, have been overwhelmingly positive. Many claim it to be an exemplary application of machine learning to medical diagnosis. Prof. Noel Sharkey, Emeritus Professor of Artificial Intelligence and Robotics, University of Sheffield, said “If you were looking for beneficial applications of deep learning, this has got to be one of them. It’s the type of task that these learning techniques are cut out for.”. However, he warned about the need for a well-crafted interface between machine and medical staff in order to prevent blind deference to the A.I.’s decisions.

“[The A.I.] did not make a single clinically-serious wrong decision”

Likewise, Prof. Duc Pham, Chance Professor of Engineering, University of Birmingham, emphasised the role of this technology as one of assistance in the decision-making process of medical staff, rather than of outright replacement. This, he says, is because “deep learning is inductive, i.e. it forms general rules and principles from specific training examples. Inductive systems cannot be guaranteed to produce 100% correct results, no matter how many training examples they used or how much training they received.” It (almost!) goes without saying that the concept of a second opinion is hardly a new one in medicine.  

Cautionary note notwithstanding, this technology shows much promise for the future treatment of eye diseases. It enables faster and more accurate diagnosis and referrals, which leads to safer and more prevention-focused care. The same trend of A.I.-assisted clinical decision making appears set to affect other branches of medicine too, from radiography to cardiology. We may now be living at the beginning of a time of exponential growth in the effectiveness of healthcare, where the merging of man and machine brings us all to a state of better health. Sláinte!