There is a lot of hype centered on machine learning in health care today. With new headlines and speculations hitting the news daily, health care leaders are asking if it’s worth all the buzz – but is that really the right question?
We’re no longer waiting to see where machine learning can be applied to improve care, but rather learning from the success stories and failures of early adopters and innovators. While there’s certainly plenty of runway for new applications, adopting a wait-and-see attitude now will leave you far behind the innovation curve in coming years. The question is no longer if you should adopt this technology but where and at what scale.
Still, despite headline triumphs, there remains some skepticism due to implementation challenges and real-world limitations of machine learning. Assessing the applications that have proven successful as well as those that haven’t added value as expected can be instructive in deciding which approach is right for your organization.
Ask the right questions of your data
Machine learning is best suited to solve well defined business problems with plenty of data to build on. Common reasons for failure or unintuitive results stem from poorly defined problems, lack of enough data, or lack of high-quality data.
The growing availability of clinical and claims databases has led to a surge in investment in analytics. This coupled with the availability of data on social determinants of health— such as socioeconomic status, education, living status, and social networks —presents an opportunity for providers to understand individual patients on a much deeper level, opening the door for precision medicine to become a reality.
Successful deployments of machine learning in other industries
To understand how this concept might work on the ground in health care, let’s first consider an application of machine learning that has successfully transformed another industry. Fraud detection in financial services has drastically improved over the last decade using machine learning techniques. Their models constantly ingest activity patterns, search for well-defined departures that may indicate fraud, and automatically escalate the issue, either by alerting a customer care representative or by directly notifying the cardholder.
This problem checks all of the boxes for technical success in machine learning; the problem is narrow and well-defined (it won't bleed unpredictably into other aspects of customer experience); there is sufficient data to sort the signal from the noise; and a financial services institution has trusted sources of both wide and deep information about each cardholder.
So, what lessons have they learned along the way? They have integrated their full range of data sources – including social determinants – into these systems to set up technical success. They have also given their machine learning systems a longitudinal view of their data, so that they can identify patterns that occur over time, rather than trying to build rules based on individual events. Additionally, to avoid some pitfalls in deploying machine learning, they actively recalibrate their system to keep up with fraud patterns as they change. Perhaps most importantly, these systems have been tuned to deliver in-the-moment insights – they send the right alert, at the right time, to the right person, and with all of the information they need to take immediate action.
Successfully employing machine learning in patient care
Key elements of this approach can be applied to many well-defined challenges in health care today to enable precise and personalized care, risk-stratify patient populations, understand variation in care patterns, and power dynamic care journey management and optimization. Linking and analyzing disparate and often siloed data sets provides access to a longitudinal view of patient care to expose patterns and variation, thus enabling automated triggers to drive the efficient prioritization of interventions. However, the most critical capability remains the delivery of insights on a real-time basis – more timely insights make a bigger impact.
It's important to think not just about the power of the insights your systems can provide, but also about how readily your team will accept and adopt them. In other words, pick a solution that solves the trust problem not just the business problem.
The holy grail in health care is not fancier technology and tools, it is physician and patient behavior change. Machine learning will truly come of age when it can systematically and reliably do one of two things – improve the decision-making of clinicians and patients or improve their efficiency in carrying out the actions that follow from those decisions. The speed at which advanced analytics becomes a reality for an organization will be heavily dependent on how quickly currently siloed clinical, claims, demographic, social and patient-reported data sets can be integrated.
Over a 20 to 30-year timeframe, clinicians should not underestimate the potential for advanced analytics and AI to dramatically change the very nature of their individual specialties and the concept of what it means to be a clinician. Just as the rise of interventional approaches (e.g., stents) have transformed both cardiac and neurosurgery by shifting procedures to interventional radiologists and cardiologists, we already see the potential to automate many aspects of radiology, pathology, and anesthesiology. It will be possible to dramatically reduce the time it takes to train a human being to be a ‘radiologist’ or ‘anesthesiologist’, because ‘the machine’ will be able to guide a lesser-trained individual to similar or superior outcomes. The extreme, of course, is where the machine entirely replaces the need for the individual.
The potential exists to provide patients with 24/7 guidance and monitoring, because code, unlike human beings, can be in multiple places at the same time and does not need to sleep. The concept of every individual having a physician in his or her pocket is no longer far-fetched. Indeed, there is already talk of small chips being injected just under the skin and providing a constant data feed of metabolic and activity data. Algorithms could then monitor and assess that data before alerting a human being, be that the patient or a physician.
Futuristic applications aside, machine learning and AI techniques can already be deployed today to reveal powerful insights. Hospitals and physicians should prepare by following the same true north that has guided them through the ages – do what is right for the patient. That means being open to implementing approaches that are proven to be effective. It means showing a far greater willingness to lower the barriers to patients obtaining and aggregating their own data. And it means accepting, as has been the case in professions from air traffic control through financial services, that human judgment is prone to error – thereby creating a role for systematic and machine-driven approaches.