The Role of Machine Learning in Making EHRs Worth It

Recently, a great op-ed published in The Wall Street Journal called “Turn Off the Computer and Listen to the Patient” brought a critical healthcare issue to the forefront of the national discussion. The physician authors, Caleb Gardner, MD and John Levinson, MD, describe the frustrations physicians experience with poor design, federal incentives, and the “one-size-fits-all rules for medical practice” implemented in today’s electronic medical records (EMRs).

From the start, the counter to any criticism of the EMR was that the collection of digital health data will finally make it possible to discover opportunities to improve the quality of care, prevent error, and steer resources to where they are needed most. This is, after all, the story of nearly every other industry post-digitization.

However, many organizations are learning the hard way that the business intelligence tools that were so successful in helping other industries learn from their quantified and reliable sales, inventory, and finance data can be limited in trying to make sense of healthcare’s unstructured, sparse, and often inaccurate clinical data.

Data warehouses and reporting tools — the foundation for understanding quantified and reliable sales, inventory, and finance data of other industries – are useful for required reporting of process measures for CMS, ACO, AQC, and who knows what mandates are next. However, it should be made clear that these multi-year, multi-million dollar investments are designed to address the concerns of fee-for-service care: what happened, to whom, and when. They will not begin to answer the questions most critical to value-based care: what is likely to happen, to whom, and what should be done about it.

Rapidly advancing analytic approaches are well suited for healthcare data and designed to answer the questions of value-based care. Unfortunately, journalists and vendors alike have done a terrible job in communicating the value, potential, and nature of these approaches.

Hidden beneath a veneer of buzzwords including artificial intelligence, big data, cognitive computing, data science, data mining, and machine learning is a set of methods that have proven capable of answering the “what’s next” questions of value-based care across clinical domains including cardiothoracic surgery, urology, orthopedic surgery, plastic surgery, otolaryngology, general surgery, transplant, trauma, and neurosurgery, cancer prediction and prognosis, and intensive care unit morbidity. Despite 20+ years of empirical evidence demonstrating superior predictive performance, these approaches have remained the nearly exclusive property of academics.

The rhetoric surrounding these methods is bimodal and not particularly helpful. Either big data will cure cancer in just a few years or clinicians proudly list the reasons they will not be replaced by virtual AI versions of themselves. Both are fun reads, but neither address the immediate opportunity to capitalize on the painstakingly entered data to deliver care more efficiently today.

More productive is a framing of machine learning as what it actually is — an emerging tool. Like all tools, machine learning has inherent pros and cons that should be considered.

In the pro column is the ability of these methods to consider many more data points than traditional risk score or rules-based approaches. Also important for medicine is the fact that machine learning-based approaches don’t require that data be well formatted or standardized in order to learn from it. Combined with natural language processing, machine learning can consider the free text impressions of clinicians or case managers in predicting which patient is most likely to benefit from attention sooner. Like clinical care, these approaches learn with new experience, allowing insights to evolve based on the ever-changing dynamics of care delivery.

To illustrate, the organization I work with was recently enlisted to identify members of a health plan most likely to dis-enroll after one year of membership. This is a particularly sensitive loss for organizations that take on the financial responsibility of delivering care, as considerable investments are made in Year 1 stabilizing and maintaining the health of the member.

Using software designed to employ these methods, we consumed 30 file types, from case management notes, to claims, to call center transcripts. Comparing all of the data of members that dis-enrolled after one year versus those that stayed in the plan, we learned the patterns that most highly correlate with disenrollment. Our partner uses these insights to proactively call members before they dis-enroll. As their call center employs strategies to reduce specific causes of dissatisfaction, members’ reasons for wanting to leave change. So, too do the patterns emerging from the software.

The result is greater member satisfaction, record low dis-enrollment rates, and a more proactive approach to addressing member concerns. It’s not the cure for cancer, but it is one of a growing number of questions that require addressing when the success of an organization is dependent on using resources efficiently.

The greatest limitation of machine learning to date has been inaccessibility. Like the mainframe before it, this new technology has remained the exclusive domain of experts. In most applications, each model is developed over the course of months using tools designed for data scientists. The results are delivered as recommendations, not HIPAA-compliant software ready to be plugged in when and where needed. Like the evolution of computing, all of that’s about to change.

Just hours after reading the Gardner and Levinson op-ed, I sat across from a primary care doc friend as she ended a long day of practice by charting out the last few patients. Her frustration was palpable as she fought her way through screen after screen of diabetes-related reporting requirements having “nothing to do with keeping [her] patients healthy.” Her thoughts on the benefits of using her organization’s industry-leading EMR were less measured than Drs. Gardner and Levinson: “I’d rather poke my eyes out.”

I agree fully with Drs. Gardner and Levinson. The answer isn’t abandoning electronic systems, but rather striking a balance between EMR usability and the valuable information that they provide. But I’ve been in healthcare long enough to know clinicians won’t be enjoying well-designed EMRs any time soon. In the meantime, it’s nice to know we don’t need to wait to begin generating returns from all their hard work.

Leonard D’Avolio, PhD is assistant professor at Harvard Medical School CEO and co-founder of Cyft of Cambridge, MA.

Categories: Uncategorized