Artificial Intelligence in Medicine
Artificial intelligence conjures mixed emotions including those emanating from the 1984 movie The Terminator wherein the fictional Artificial Intelligence (AI) named “Skynet” views humans as a threat and endeavors to eradicate us from planet Earth. Some modern-day renditions are more benevolent, including IBM Watson as popularized on the TV game show Jeopardy, and the more recent Elon Musk sponsored companies Open AI and Neuralink.
AI represents a double-edged sword in that it offers to solve many of humankind’s most difficult problems, while also creating new risks should bad actors and rogue states use the technology for malicious purpose. Readers might ask, “Well, how relevant is AI really and how close are we in fact to significant AI?” One can simply do a search on our National Institutes of Health website, PubMed, to see there are almost 74,000 published articles related to the search terms “artificial intelligence” and almost 12,000 published papers regarding the search terms “artificial intelligence medicine.” After reading some of these papers, readers would realize we are in the growth stage and there is no current general AI sentient being; however, “Narrow AI” applications are robust, effective, and expanding.1
Here we will focus on the opportunities that AI currently offers in health care.
Technological advancements over the years have provided tremendous efficiencies in helping keep us healthy, in discovering disease, and in providing medical treatment. Some of these technologies include the ability to detect heart attacks, strokes, infection, cancer, metabolic abnormalities, and others. AI may be considered, in some respects, a layering of these technologies which interact through “deep/machine learning” algorithms. For instance, AI concepts have been used for many years in computer-aided diagnosis (CAD) to assist radiologists in discovering abnormal breast tissue on mammograms that could be cancerous. Specialized software analyzes the digital mammogram through prescribed and validated algorithms to search for tissue architectural distortion. Software algorithms highlight regions of interest and present this information to the radiologist for additional review. These programs are essentially designed to say, “Hey, Doctor, what about this area? Is this concerning for potential cancer?”
In 2016, researchers used two different Deep Convolutional Neural Networks (DCNNs), AlexNet and GoogleNet, to analyze 1,007 chest radiographs for tuberculosis.2 The data sets were split into 68% training, 17% validation, and 15% test resulting in a best-performing classifier area under the curve (AUC) of 0.99.2 AUC is a general measure of diagnostic accuracy and 1.0 is considered perfect.2 Of 1,007 studies analyzed, there was disagreement between the two artificial neural networks between only 13 studies.2 These were interrogated by the cardiothoracic-trained radiologist and the radiologist augmented approach resulted in a sensitivity of 97.3% and specificity of 100%.2
Conclusion: deep learning with DCNNs can accurately diagnose tuberculosis. Much more interesting, however, is the ability for AI deep learning to decipher yet unknown pathologic associations.3 This was demonstrated recently wherein a deep learning supercomputer was able to analyze data from over 295,000 patients to create associations with those that had sustained a heart attack or not.3 The four resulting computer-trained algorithms were then challenged to predict the rate of heart attacks in another 82,000 patients whose heart attack history was already known.3 Amazingly all four of the AI-trained algorithms were superior to the American College of Cardiology and American Heart Association guidelines.3 The best of these trained algorithms was able to predict 7.2% more heart attacks correctly and reduce false alarms by 1.6%.3 This suggests that 355 more lives of those 82,000 patients could have been saved had that algorithm been used.3 Of note, over 30% of deaths in the industrialized world are due to heart disease.
The key to AI in the context of machine/deep learning is moving beyond fixed coded software sequences designed by computer programmers to software algorithms that can in fact learn.4 Human minds are routinely presented with enormous amounts of data that must be sorted, processed, and applied effectively. Our brains use neural networks supported by over 80 billion neurons, over 100 trillion neuronal connections, and over 100,000 miles of nerve fibers sending messages at over 250 miles per hour!5
Many different ways of measuring brain and computer processing power have been promoted. Two popular measures of computing power are floating operations per seconds (FLOPS) and traversed edges per second (TEPS). The human brain is appraised to process 1 exaflop. This is compared to one of the most powerful supercomputers in 2016, the “Sunway TaihuLight,” which was rated at 93 petaflops continuous (an exaflop is 1,000 times a petaflop, and researchers maintain supercomputers will exceed the human brain in exaflop calculations by the year 2020). The K Computer by Fujitsu was measured at 3.9 X101,3 TEPS in 2016 and the human brain is rated at 6.4 X101,4 TEPS. These two measurements suggest that the world’s fastest supercomputers are currently only a factor of 10 times slower than our brains and they are rapidly gathering speed! In fact, corporations and start-ups are ramping up to be a part of the estimated USD 16 billion AI market projected by 2022; growing at a 63% CAGR from 2016.6
Obviously, comparing our conscious to computers in these singular measures does not do justice to our ability to navigate the totality of life. A modern supercomputer’s ability, however, to process astonishing amounts of “big data” and learn from the experience, is where they leave us behind in the dust. This is precisely where they can help us discern subtle and otherwise occult plethora of information to come to accurate diagnoses and treatments.
Dr. Hancock is a board certified neuroradiologist with Desert Medical Imaging and can be reached at (760) 694.9559. He is also a member of Desert Doctors. For more information visit www.DesertMedicalImaging.com or www.DesertDoctors.org.
References: 1) AI in Medicine: Rise of The Machines, Paul Hsieh; Forbes. April 30, 2017; 2) Deep Learning at Chest Radiography: Automated Classification of Pulmonary Tuberculosis by Using Convolutional Neural Networks. Paras Lakhani, MD, Baskaran Sundaram, MD, Radiology Volume 284 Issue 2. August 2017; 3) Can machine-learning improve cardiovascular risk prediction using routine clinical data? Weng, S. PLOS-One, August 17, 2017; 4) Machine Learning for Medical Diagnosis: History, State of the Art, and Perspective. Kononeko, I., Artificial Intelligence in Medicine. 23 (2001) 89-109; 5) Numbers: The Nervous System, From 268-MPH Signals to Trillions of Synapses Ross, V., Discover March, 2011; 6) Artificial Intelligence Chipsets. Market and Markets. November 2016.