Few things seem more apocalyptic than sentient robots, controlled by a Skynet-like hive mind, taking over the planet. Some may consider this to be restricted to the domain of Hollywood and sci-fi movies, with close-to-zero odds of occurring in reality. People may deem it so far-fetched, given current technology, that no one needs to worry about it right now. Machine learning, however, is very real, as far as its uses are concerned. As a concept, it involves the ability of a “machine” (i.e. a computer program) to “learn” from given data and extrapolate to fresh data. As software, it involves a computer program that can take in an image, such as a medical scan, recognize the features of the image, and correlate these features, with a certain probability, to a specific condition or classification. Its applications, known as “algorithms”, are currently used in a variety of specialties. On one end of this “specialty spectrum” lies Google HQ. At Google, machine learning algorithms help researchers delve into “language, speech, translation, visual processing, ranking and prediction” in order to improve and refine the search engine. On the other end of the spectrum lie various hospitals and clinics worldwide, where machine learning algorithms assist physicians in creating or confirming diagnoses. With the current prevalence of big data in the medical field, machine learning algorithms can be used to condense or find patterns in the data. In addition, they can place individual patients’ conditions in the context of this data, which could allow physicians to draw faster, more precise, and personalized conclusions down the road.
To get to this stage, one first “trains” the machine learning algorithms, such as convoluted neural networks and deep learning, by introducing various inputs, such as brain MRIs, one at a time. Each of these images, upon being introduced, is manually associated to a specific condition, such as Multiple Sclerosis. The algorithm subsequently passes the image through various “layers”, filtering the original image down into an array of information that the computer can process, such as RGB values (representing color intensities). This is where the pattern detection takes place: the algorithm associates certain features in the images (based on the information in the arrays) to the given disease. Once enough images have been analyzed, the algorithm would ideally be able to predict whether a fresh image reflects characteristics of the disease, based on the presence of patterns and common features found in the training images.
Diagnostic radiologist Krishna Tirumala, who works at Baptist Medical Center, FL, recently informed me that he and his colleagues work with a program called Computer-Aided Diagnosis (CAD). This software, according to Dr. Tirumala, incorporates various types of machine learning algorithms to help radiologists review and interpret medical imaging studies and data. “However,” he explains, “there is one problem: CAD is sensitive, but not very specific.” That is, CAD flags a lot of features on an image (for example, an X-ray) as dangerous or indicative of disease. Surely, actual signs of disease may be among the flagged signs, but they would be swamped by the overwhelming number of false positives, which, to a trained diagnostic radiologist’s eye, may otherwise seem irrelevant or harmless. For this reason, the incorporation of machine learning into medicine seems more like it would fulfill an assistant’s role, rather than replace a specialist, at least for the time being.
Despite this, large advances are taking place that could render machine learning algorithms more specific while maintaining high sensitivity, thereby reducing the number of false positives and making the algorithms much more applicable clinically. For instance, its use in analyzing lengthy extensive medical journal articles or a vast amount of medical data could help physicians with their workload. Current medical journal articles contain dozens of images, which the average healthcare professional likely has little time to analyze in-depth. This is not a good thing: in many occasions, not having time to interpret all of the newer images and the data contained in these papers would prevent physicians from giving up-to-date diagnoses. However, machine learning can be tremendously useful here, especially when interfaced with big data.
According to Dr. Tirumala, if machine learning algorithms can tap into big data (the large amount of medical data that has been and is currently being produced), they could prove invaluable to the field in general. For example, IBM’s Watson is a computer that can use deep learning to mine patient data and research journal articles from the past and the present. It can then associate the patient data to relevant findings in these articles. As one could imagine, this can help physicians make quicker, more informed, and more personalized diagnoses for patients with newer diseases, or newly discovered variations on older diseases.
In the future, machine learning algorithms could be used in one of two ways. First, they can be used to reduce the amount of data that a physician has to peruse to diagnose patients accurately. Second, they can be used to supplement a physician’s knowledge with information taken from more recent research, which the physician may or may not have encountered. As new antibiotic-resistant strains of bacteria are discovered, or new genetic diseases and their underlying mutations are characterized, machine learning algorithms might become a necessity down the road, especially for idiopathic or previously unseen cases. That said, it is unlikely that a Terminator-style, sentient robot takeover would occur, unless machine learning algorithms somehow learn to train themselves. Whether these ongoing improvements in machine learning indicate the start of the intelligent robot takeover, or of more precise, personalized diagnoses, these advancements mark a turning point in the future of healthcare.