Web Exclusive

What Machine Learning in Radiology Has in Common With Baseball's Electronic Strike Zone


Medicine and sports can be better when humans and machines work together.

The goal of health care IT is to provide the best information to an individual so they can make the best possible judgment. This goal exists in many industries, but technology's place in radiology can be illustrated by an example from an unlikely source: baseball.

Growing up in St. Louis—arguably the home of the best fan base in baseball—I quickly became an avid student and player of the sport. While the game may seem simple, every guy I grew up playing baseball with enjoyed dissecting its technicalities and all the moving parts that come together to form each and every play. Take the pitcher and the hitter, for example. The matchup would seem to be a mere matter of a pitcher throwing a ball in different ways and a hitter attempting to successfully make contact. In reality, there are a complex set of processes communicated nonverbally to every player in the field before any pitch is thrown.

Strategic data are relayed to each fielder through pitching and batting stances, the position and signals from the catcher before the ball is tossed, rhythm and body language from the batter, hand signals from the coaches, specific pitching strategies established in the locker room before the game starts, and injuries to a batter that could impact how a player might run, throw, or swing a bat. This analysis happens on every pitch within a matter of seconds.

From there, umpiring takes over an essential and critical role in administering highly subjective judgments following the rules of the game. For example, in microseconds, umpires must quickly discern the strike zone's structure for each batter on each pitch. This is arguably the most critical and consistent element of the game of baseball.

Despite years of training and experience, each umpire has their own interpretation of the strike zone. The increasing velocity of the thrown ball by the athletes of today combined with innovative ways to manipulate new pitches are putting stresses on umpires to demonstrate that they have consistently fair and equal judgements with every ball that crosses the plate. As players and technology continue to improve, umpire subjectivity and accuracy will be magnified for everyone to see or second-guess. The ability to implement these skills in a consistent manner, as shown by accurately calling balls and strikes, can affect a player's ability to execute in crucial game situations and, ultimately, directly impact their compensation, prestige, and ability to help win games.

Technology Creates a Level Playing Field
An electronic strike zone is one answer to minimize errors. In theory, the electronic format would provide precise measurement, but it is not independent of human variability. The electronic strike zone also has an element of subjectivity because it must be calculated for each pitch. If it were implemented, the umpire's role would change but not be replaced.

The electronic strike zone can be used for training, reducing errors, and speeding up the game. Much like players who review game footage coupled with the electronic strike zone to find pitching patterns, umpires can use it to improve their skills at calling balls and strikes because it allows new ways of analyzing myriad variables requiring snap decisions.

The same could be said of the use of machine intelligence in radiology and, particularly, in neuroradiology. The number of variables that a radiologist must incorporate into their reading analyses are similar in quantity to those that baseball umpires must analyze, but, instead of balls and strikes, radiologists' decisions impact a patient's health care.

As with the latest advancements in electronic measurement tools for baseball, current computer-aided imaging advances are not sufficiently developed to replace the judgment, creativity, intuition, and intelligence of a reading radiologist. Machine intelligence capabilities can significantly enhance the quality and timeliness of a radiologist's output—decreasing read times and providing a more rapid review of other not-so-obvious abnormalities including those found in detailed scanning data.

The Right Patient Data at Your Radiologist's Fingertips
Innovative tools that can assist the reading physician in producing a more accurate diagnosis are now emerging at a rapid pace. Imagine a world where a radiologist has the ability, with a push of a few buttons, to correlate and pattern-match patient data with other cohort patient groups—patients whose experiences might inform their diagnoses and suggest treatment alternatives available to their medical teams ... on demand, anywhere, and at any time.

What if these programs went a step farther and incorporated information sources, such as proteomic and genomic data, to enhance, validate, or dismiss subjective observations? Having immediate and accurate information that has only been available from cumbersome manual nonautomated systems is now emerging from new cloud-based data warehouses and high-performance computing systems.

Opportunities for machine learning in radiology have already begun to take shape, particularly in the oncology space. Companies such as HealthMyne and Enlitic are using machine intelligence to create algorithms that can automatically identify and analyze tumors and predict cancer prognoses. Enlitic claims to have developed a lung nodule detector that can reach positive predictive values 50% higher than those achievable by a radiologist, and HealthMyne's platform unites medical images and EHRs to provide quantitative imaging decision support to physicians.

Machine Intelligence in Radiology Enhances, Not Replaces
While advanced radiology programs can automatically delineate and quantify abnormalities, radiologists will still play a crucial role in dissecting the information to provide referring physicians with the best prognoses. Prepopulated, mathematically derived data obtained from imaging provide objective results and, hopefully, better patient outcomes. Rather than being a threat to radiologists, machine learning offers a net gain for those in radiology who use smart machines wisely.

Whether you're a physician or an umpire, machine intelligence provides quicker access to information needed to make better, faster decisions. Together, humans and machines have the potential to change the worlds of health care, baseball, and beyond. Smart people will succeed in the future by embracing the smart systems of today and tomorrow.

— John A. Kelley, Jr, is chairman and CEO of CereScan, a functional brain diagnostics company. Previously, he served as the chairman, president, and CEO of McData Corp. Prior to McData, he was executive vice president of networks at Qwest Communications.