The following is a guest blog post by Gurjeet Singh, Executive Chairman and Co-founder of Ayasdi.
The combination of electronic medical records, financial data, clinical data, and advanced analytics promised to revolutionize healthcare.
It hasn’t happened.
The common excuse is that healthcare wasn’t really prepared for the enormity and complexity of the data challenge and that, over time, with the next EMR implementation, that healthcare will be positioned to reap the benefits. Unfortunately, the next generation of EMR, or the one after that, isn’t going to solve the problem.
They problem is on the analytics side.
Healthcare analytics are still driven by a question-first approach. The start of our analytics journey still begins with the question. The challenge is which question? The more data we have at our disposal, the more potential questions there are and the lower the likelihood that we will ask the one that generates new value for the patient, the provider, or the payer. Even when we are successful in asking the right question, we have engaged in a confirmatory process – we have confirmed something we already knew.
Some will suggest that predictive analytics solves the problem, but it too is hypothesis driven – just in a different way. With predictive analytics, the set of variables selected, the choice of algorithms are, in effect, guesses as to what will produce the best outcome.
Ultimately, both approaches are flawed.
We need a new approach that surfaces trends we humans haven’t even considered, and that delivers a host of meaningful insights to clinicians before they even ask any questions. We need technology solutions that combine the best qualities of human intelligence (artificial intelligence) with the best computing capabilities that exceed human ability (machine learning). When these technologies are operationalized systematically across an enterprise, it’s called Applied AI. Applied AI is here to replace healthcare analytics, and we all stand to benefit.
Five Keys to Applied AI
Applied AI has already begun driving care improvement, cost-reduction, and improved clinical and financial decision-making across the healthcare enterprise – and the entire healthcare continuum. Applied AI is not a concept, but a series of intelligent applications that target discrete healthcare problems from clinical variation to population health. These intelligent applications have a collection of capabilities that make them intelligent – of which all need to be present. Let’s look at those capabilities:
Discovery – Intelligent applications need to support both unsupervised and semi-supervised discovery. These capabilities are quite rare but serve as the foundation for our efforts to move past hypothesis driven inquiry. In practical terms, this means that an intelligent application considers all the data and all the possibilities within that data to detect the patterns, groups or anomalies that elude traditional approaches. Using their own systems of records, including EMRs, financial data, patient-generated data, and socio-economic data, healthcare organizations can automatically discover groups of patients that share unique combinations of characteristics. These groups can then be used to tailor and personalize diagnostics and care paths, for example. Alternatively, healthcare organizations may also discover unique patterns or outliers within their claims data to aid in member retention or preventing fraud or waste. This type of holistic discovery is unique to AI and improves prediction and makes operational insights possible.
Predictions – Intelligent applications must also be able to predict the future with high accuracy. Holistic discovery enables even better predictive models through the unbiased creation of groups or the identification of patterns. Superior prediction gives healthcare organizations foresight into the future needs, costs, disease burden, and risks of patients. For example, intelligent applications can determine the groups of patients projected to have the highest escalation of costs over time, as well as other outcomes such as the conditions likely to appear for each group, and an individual’s predicted change in utilization. Predictions can be made across multiple targets and are multi-faceted, considering all factors whether they’re health- or non-healthcare-related occurring outside of the healthcare system.
Justification – An intelligent solution must justify its predictions, discoveries, and actions in a transparent way so human operators feel confident to act upon its recommendations. For example, a healthcare app may reveal differentiating characteristics of patient risk trajectories, what factors make them high or low-risk, and descriptions of individual factors that lead to variation in cost and quality. Justification is key because without a thorough understanding of the “why” behind predictions, organizations are unable to adopt AI into day-to-day decision-making.
Action – An intelligent system that is not effectively operationalized will become less intelligent over time. Actionable information that guides and augments human decision-making is what makes AI a part of daily operations. For these systems to deliver optimal value they need humans in the loop providing feedback and governance. Whether it be a recommended care path or a detailed risk profile, intelligent applications allow organizations to collaborate on the best actions tailored for each patient population, or to physicians or organizations. Across the care continuum, within health systems and health plans, this allows them to better assess individuals and the best course of care, and more confidently prescribe care and programs for each individual.
Learning – Intelligent applications “learn” to improve predictions over time. As more and more data is analyzed, the technology learns from these complex data points to improve predictions over time. Whether it be claims, medical records, or socio-economic data, AI taps into these data points to generate more accurate, personalized predictions that continuously improve. Further, intelligent apps learn the impact of actions over time to support and continuously improve decision making.
Applied AI in action
A large hospital system decided it wanted to reduce clinical variation across its enterprise to improve outcomes for all patients. It implemented machine intelligence, including unsupervised machine learning techniques that run algorithms using the system’s own data—not benchmarks—to uncover actionable insights. The technology correlates and analyzes electronic medical record and financial data including treatments prescribed, procedures performed, drugs administered, length of stay, and costs per patient. The goal was to discover and refine clinical pathways that are optimized to drive higher quality of care and lower costs.
The machine intelligence solution identified a group of orthopedic surgeons who consistently had better outcomes among their knee replacement patients. These patients had shorter hospital stays and shorter time to ambulation than other total knee surgery replacements across the system. The solution also told clinicians why: these doctors prescribed a unique, not widely used medication at an earlier postsurgical time than their peers. The medication reduced patients’ pain so they could get out of bed and walk around sooner – improving their outcomes and reducing costs.
Clinicians hadn’t previously known to look for variation based on what medication was given post-operatively. But machine intelligence identified a pod of doctors with better outcomes that were statistically significant. By comparing very large numbers of data points, the solution quickly uncovered why. Now the hospital system has operationalized these best practices throughout their hospitals, lowering costs for knee replacement by more than 5 percent, and reducing pain for patients.
The last piece of the puzzle – AI applications
As healthcare organizations increasingly see the value of Applied AI, they may worry that more robust technology means greatly increased technical headcount to manage this strategy. But an important component of a successful Applied AI strategy is that it leverages the unique capabilities of both machines and humans. Hiring a dozen data scientists won’t make the most of the human intelligence within your organization. That’s because these new data scientists likely would not have the subject matter expertise needed to recognize and deploy the meaningful insights that surface. Meanwhile, the people who are the best suited to learn from the data, domain experts, usually do not have an interface to read data themselves. Subject matter experts typically only interact with data using rudimentary applications like PowerPoint or Excel.
So, the last key to a successful Applied AI strategy is to wrap the results of machine learning and artificial intelligence into business-facing applications. These applications can be customized for the types of insights they uncover, such as the optimal way to perform surgical procedures. It’s critical that the results of machine learning and machine intelligence actually make it to clinicians, instead of ending up siloed somewhere in the IT department. The successor technology to healthcare analytics must not only be more powerful and more precise, it must also be more user-friendly.
Healthcare analytics simply aren’t living up to their promise. We can wring our hands, we can wait, we can soldier on with insights that only marginally move the needle to improve outcomes and lower costs. Or we can combine artificial intelligence with powerful machine learning to turn enormous datasets into business insights that really matter. Then we can deliver those insights, via easy-to-use business applications, to the best clinician minds, to operationalize this machine intelligence approach across the enterprise. That’s Applied AI, and it’s a bright future.
About Gurjeet Singh
Gurjeet Singh is Ayasdi’s Executive Chairman and co-founder. As the Executive Chairman, he leads a technology movement that emphasizes the importance of extracting insight from data, not just storing and organizing it.
Gurjeet developed key mathematical and machine learning algorithms for Topological Data Analysis (TDA) and their applications during his tenure as graduate student in Stanford’s Mathematics Department where he was advised by Ayasdi co-founder Prof. Gunnar Carlsson.
Gurjeet is the author of numerous patents and has published in a variety of top mathematics and computer science journals. Before starting Ayasdi, he worked at Google and Texas Instruments. Gurjeet was named by Silicon Valley Business Journal as one of their 40 Under 40 in 2015.
Gurjeet holds a B.Tech. from Delhi University, and a Ph.D. in Computational Mathematics from Stanford University. He lives in Palo Alto with his wife and two children, and develops multi-legged robots in his spare time.