Artificial intelligence (AI) systems for interpreting complex medical datasets
AI research began with the 1950s invention of the computer. Early systems employed simple logic and “rules” to emulate human experts. What has changed in the last decade? First, data collection has exploded; most industries measure their processes to monitor inputs, outputs, costs, and quality. Electronic medical records store notes, lab data, imaging data, and billing/claims data. Genome‐scale measurements are inexpensive and critical to cancer, immunology, and other specialties. Personal mobile monitors provide information about motion, physiology, and environmental exposures. Second, computing power has continued to increase exponentially, doubling roughly every 18 months since the early 1960s. Recent advances come from massively parallel processors (originally developed for gaming systems) that can perform fast mathematical computations—today, the fastest computer can perform 100,000,000 billion (1016) arithmetic operations per second. The combination of data and computation creates a perfect storm for machine intelligence. Medical data are an attractive target, as the volume threatens to overwhelm individual practitioners.
A rapidly developing subdiscipline of AI is machine learning. “Deep Learning” models are based on an analogy to the human brain—networks of computational units (neurons) are densely connected (as by axons and dendrites) and are programmed to recognize and find similarities between objects. They are learning complex functions that map input to output, and the large number of “neurons” enable them to encode complex nonlinear function. The initial victories for deep learning have been in image analysis. Using huge numbers of images from the internet, these models learned to recognize objects (cats, cars, people…) with high accuracy. Recently, these systems have been combined with text‐generation software to analyze images and label them with captions2: “Two pieces of pizza on a stove top.” or “Two children playing near a firetruck.” Naturally, systems to assist radiologists and pathologists in the examination of image data are now appearing, and are expected to become part of routine clinical care—“Small worrisome nodule in right lower quadrant.” Before these systems are deployed, regulated, and reimbursed, there are issues to address of liability, billing, professional standards, communication, and validation.
Deep learning models are characterized by a single architectural principle (the analogy to neuronal systems), whereas other systems are more federated and employ multiple modules based on different technologies. The IBM Watson architecture for Jeopardy! was not a monolithic architecture, but a set of special‐purpose tools (with capabilities for text, images, search, inference) that together formed a system for playing the game. In general, there are two types of machine learning tasks: supervised learning and unsupervised learning. In supervised learning, the machine is given a set of data for each entity of interest, and a “label” that indicates how the object or event should be classified. For example, RNA expression datasets from breast tissue samples can be provided with labels indicating whether they are “cancer” or “benign.