Computers are now able to search through petabytes of data, detect patterns, and learn new functions without being explicitly programmed. What is “machine learning,” what are its many applications, and how might it transform healthcare?
The smartphone in your pocket is loaded with apps that are essentially personalized services, every interaction customized to your unique location, payment method, and widely expanding range of individual preferences. Summon a ride, order a meal, go shopping, scan the news, c0nnect with friends, stream entertainment—each transaction providing even more user data to these dynamic services, further optimizing your consumer experience.
From Uber to Amazon to Spotify, these services “learn” your tastes and habits to become progressively better at anticipating and acting on your needs—and encouraging your repeat business. That ability to “understand exactly what you mean and give you back exactly what you want” is the essence of a “perfect search engine” as Google co-founder Larry Page once described, and lies at the heart of the customer service revolution in the digital age.
Mobile apps are already so adept at matching meaning to want that they’ve significantly reduced mobile search, creating enormous disruption in how audiences connect with brands, and how brands in turn come to better understand their consumers. The functionality behind the fanfare is some of the most sophisticated computer science ever coded, evolving from the study of pattern recognition and computational learning in artificial intelligence.
The timeline of “machine learning” is fascinating, branching away from formal AI in the 1990s when expert systems and neural networks tried to emulate human decision-making but encountered growing pains. In contrast, machine learning shifted focus to practical problem solving using advanced statistical analysis and data retrieval techniques. So instead of struggling to somehow make computers think, scientists succeeded in helping them learn.
Machine learning systems essentially consist of three components: A model that makes predictions; parameters that are the factors used to make decisions; and a learner that adjusts the parameters and the model based on how its predictions fair in comparison to actual outcomes. In practice, a “teaching set” of data is provided as input, which is then analyzed based on the model’s parameters. Outputs expand the “teaching set,” teaching the system. Let’s take a look…
Netflix, Jeopardy, Cat Videos, and GO
Ever since “Colossus” was first created to crack German codes during WWII, digital computers and the software run on them have been complex tools designed to solve straightforward but extremely difficult challenges. Fast forward to 2006 and the Netflix Prize offered a million dollar award to the first computational contender to improve the accuracy of their “CineMatch” rating system by 10% using machine learning and data mining.
Sounds counter-intuitive, but the algorithm that optimized the Netflix system for recommending movies based on users’ personal preferences took more than three years to develop, and represented a significant milestone in machine learning. A seven-member team including two AT&T researchers dubbed “BellKor’s Pragmatic Chaos” beat the second place winner by only 24 minutes, and represented a significant improvement over prior entries.
In essence, the algorithm searched the Netflix database for users rating the same movie, for example “Star Wars,” and then determined which of these users also rated a second film, such as “Indiana Jones”. The statistical likelihood that users who liked “Star Wars” also liked “Indiana Jones” was calculated, the process then extrapolated across the ratings of numerous other films with the ultimate goal of recognizing patterns of correlation.
Vast troves of data are necessary for machine learning to work, in the Netflix case over 100 million ratings of nearly 18,000 movies from nearly half a million users! In 2011, IBM Watson applied similar techniques plus natural language processing to help the super computer beat two of the best human Jeopardy players: In this case, over 200 million pages worth of data was ranked based on what Watson had learned through prior interactions.
Unprecedented pattern recognition was accomplished for visual input about a year later by the “Google Brain” of X-Lab. Consisting of 16 thousand processors totaling over a billion connections, the neural network was exposed to 10 million randomly selected YouTube videos and “learned” how to recognize cats with 74.8 percent accuracy. What’s cool is that the brain was never instructed to identify a “cat,” instead teaching itself the concept.
Since then Facebook has applied machine learning to facial recognition with over 97% accuracy, while Google’s DeepMind has done what few thought possible—a computer beating the world’s top-ranked Go player. Here machine learning was combined with “tree search” techniques that evaluated millions of potential options to determine an optimal move, the program predicting success-failure scenarios based on learning from prior patterns.
Approaches & Applications
Machine learning is divided into categories, each consisting of diverse approaches resulting in a wide array of applications. In “supervised learning,” the computer is provided sample inputs and outputs, the goal to learn a general rule mapping them together; “unsupervised learning” has the computer on its own to discover hidden patterns; and in “reinforcement learning” the computer interacts with a dynamic environment, such as driving a car.
From decision tree learning to artificial neural networks, clustering to genetic algorithms, rule-based to learning classifier systems, machine learning applications have become widespread to the point of astonishing ubiquity. Virtually anywhere enormous amounts of data defy human analysis but lend themselves to subtle pattern recognition and advanced predictive modeling, machine learning can work its sophisticated computational magic.
Many game changing applications already include advances in security, such as identifying malware, money laundering, consumer fraud, and improving data protection; gamification and self-driving cars; speech, handwriting, textual analysis; and personalized marketing, in the form of search engine algorithms, dynamic content, and as we’ve seen preference-powered apps that make recommendations based on prior user behaviors and habits.
Arguably the most significant applications impact healthcare, de facto data-intensive and user experience focused. At every milestone along the patient journey from diagnostics to treatment through compliance, machine learning is already adding value and providing unique benefits, including medical imaging, oncology, diabetes, surgery, and for our purposes most significantly in pharma for drug discovery and other Big Data opportunities.
Machine learning is also the driving force behind many new organizations brimming with potential, including a startup that uses advanced data analytics to help payers and providers ensure patients are treated according to prescribed guidelines. Machine learning algorithms are applied to integrated claims, EHR, and remote monitoring data, with results fed back to the practicing physician to optimize treatment and ensure compliance.
But as we’ve seen from our recent post on blockchain, the gap between latest digital buzz word and paradigm changing benefit—especially in healthcare—is often considerable. Obstacles to machine learning for healthcare span the spectrum from privacy concerns to operational feasibility, exemplified by the Netflix Prize that, ironically enough, was never implemented or expanded due to both these concerns, respectively. Caveat learner!
School of Klick Health
Despite its many challenges, the incredible potential of machine learning for healthcare beckons pharma brands. Are you partnered with a digital health expert that not only understands the technology, but has already successfully applied it? Here at Klick Health, we’ve used machine learning to recognize hidden patterns in mountains of audience behavioral data, revealing entirely new and previously unengaged segments. Want to learn more? We can’t wait to learn together.