Google, IBM, and a number of other large companies are exploring the application of artificial intelligence to medicine. Recently at a national meeting, IBM Watson demonstrated that its nomogram for treatment for particular cancers had a 96.4% concordance with the course of care selected by the treating physicians.)
IBM Watson for Oncology was trained at Memorial Sloan Kettering and has therefore learned to emulate those physicians, much like a physician in training who is doing a fellowship at the same institution. When the same IBM Watson protocols were applied at different facilities, the rates of concordance were lower, indicating that not all hospitals follow the same practices.
Could we apply the same type of machine learning to help with treatment of ophthalmology patients? Computers can now read fundus photos to determine the presence and extent of diabetic retinopathy. Could we then implement a nomogram where the machine could also suggest an appropriate course of treatment?
The use of artificial intelligence is not new in our daily lives. Post offices and banks use machines to read the entire spectrum of human handwriting with a high level of accuracy. Using just about 100 lines of computer code, we can make a neural network that will read hand-written numerals with greater than 99% accuracy. Machines can even be programmed to master video games by trial and error, learn from mistakes, and repeat actions that had a positive outcome. A YouTube search turns up hundreds of videos of very young programmers teaching their computers to perfectly master Super Mario Bros—something that I can personally tell you is quite difficult.
One of the most promising uses of artificial intelligence in ophthalmology is in the realm of lens power calculations for cataract surgery. The math behind this is more complex than it appears, because there are many variables that must be accounted for simultaneously to determine the effective lens position of the optic. In preliminary testing, Ladas and colleagues have reported that 90% or more of patients achieve an accurate outcome (defined as hitting target +/- 0.5D), which is far better than the 60 to 80% achieved by most current formulas.
As our diagnostic evaluation moves toward more computerized devices, the value of the slitlamp examination may seem to decline. If your patient undergoes auto-refraction, higher-order aberration scanning, corneal topography, corneal tomography, anterior segment photos, optical coherence tomographic imaging, wide-field retinal scans, and more—it would seem that most diagnoses and treatment algorithms could be done without a physician seeing the patient.
But maybe there is something missing in the machine-based protocols—the human touch. When I see a patient with a complex issue, I realize that we need to do something more, to break from protocol, to personalize the treatment to the patient and the situation. When a patient comes into our county teaching hospital with neovascular glaucoma, proliferative diabetic retinopathy, and cataracts, we may elect to do a complex single-stage procedure instead of following traditional protocols. We may inject anti-VEGF compounds, take the patient to the operating room to remove the cataract, perform panretinal photocoagulation with the indirect ophthalmoscope laser, implant an appropriate lens, and insert of a glaucoma drainage device. It is what I would do for my own family. And that’s something a computer would have a hard time achieving.